diff --git a/OpenCL/inc_cipher_aes.cl b/OpenCL/inc_cipher_aes.cl
index 74b9c3700..f5280e179 100644
--- a/OpenCL/inc_cipher_aes.cl
+++ b/OpenCL/inc_cipher_aes.cl
@@ -692,7 +692,7 @@ __constant static u32a rcon[] =
 
 // 128 bit key
 
-void aes128_ExpandKey (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void aes128_ExpandKey (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   ks[0] = ukey[0];
   ks[1] = ukey[1];
@@ -718,7 +718,7 @@ void aes128_ExpandKey (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u
   }
 }
 
-void aes128_InvertKey (u32 *ks, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void aes128_InvertKey (u32 *ks, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   for (u32 i = 0, j = 40; i < j; i += 4, j -= 4)
   {
@@ -758,7 +758,7 @@ void aes128_InvertKey (u32 *ks, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TY
   }
 }
 
-void aes128_set_encrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void aes128_set_encrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   u32 ukey_s[4];
 
@@ -770,7 +770,7 @@ void aes128_set_encrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_
   aes128_ExpandKey (ks, ukey_s, s_te0, s_te1, s_te2, s_te3, s_te4);
 }
 
-void aes128_set_decrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void aes128_set_decrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 ukey_s[4];
 
@@ -784,7 +784,7 @@ void aes128_set_decrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_
   aes128_InvertKey (ks, s_te0, s_te1, s_te2, s_te3, s_te4, s_td0, s_td1, s_td2, s_td3, s_td4);
 }
 
-void aes128_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void aes128_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   const u32 in_s0 = swap32_S (in[0]);
   const u32 in_s1 = swap32_S (in[1]);
@@ -847,7 +847,7 @@ void aes128_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0
   out[3] = swap32_S (out[3]);
 }
 
-void aes128_decrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void aes128_decrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   const u32 in_s0 = swap32_S (in[0]);
   const u32 in_s1 = swap32_S (in[1]);
@@ -912,7 +912,7 @@ void aes128_decrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_td0
 
 // 256 bit key
 
-void aes256_ExpandKey (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void aes256_ExpandKey (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   ks[0] = ukey[0];
   ks[1] = ukey[1];
@@ -957,7 +957,7 @@ void aes256_ExpandKey (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u
   }
 }
 
-void aes256_InvertKey (u32 *ks, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void aes256_InvertKey (u32 *ks, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   for (u32 i = 0, j = 56; i < j; i += 4, j -= 4)
   {
@@ -997,7 +997,7 @@ void aes256_InvertKey (u32 *ks, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TY
   }
 }
 
-void aes256_set_encrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void aes256_set_encrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   u32 ukey_s[8];
 
@@ -1013,7 +1013,7 @@ void aes256_set_encrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_
   aes256_ExpandKey (ks, ukey_s, s_te0, s_te1, s_te2, s_te3, s_te4);
 }
 
-void aes256_set_decrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void aes256_set_decrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 ukey_s[8];
 
@@ -1031,7 +1031,7 @@ void aes256_set_decrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_
   aes256_InvertKey (ks, s_te0, s_te1, s_te2, s_te3, s_te4, s_td0, s_td1, s_td2, s_td3, s_td4);
 }
 
-void aes256_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void aes256_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   const u32 in_s0 = swap32_S (in[0]);
   const u32 in_s1 = swap32_S (in[1]);
@@ -1094,7 +1094,7 @@ void aes256_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0
   out[3] = swap32_S (out[3]);
 }
 
-void aes256_decrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void aes256_decrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   const u32 in_s0 = swap32_S (in[0]);
   const u32 in_s1 = swap32_S (in[1]);
@@ -1159,7 +1159,7 @@ void aes256_decrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_td0
 
 // wrapper to avoid swap32_S() confusion in the kernel code
 
-void AES128_set_encrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void AES128_set_encrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   u32 ukey_s[4];
 
@@ -1171,7 +1171,7 @@ void AES128_set_encrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_
   aes128_set_encrypt_key (ks, ukey_s, s_te0, s_te1, s_te2, s_te3, s_te4);
 }
 
-void AES128_set_decrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void AES128_set_decrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 ukey_s[4];
 
@@ -1183,7 +1183,7 @@ void AES128_set_decrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_
   aes128_set_decrypt_key (ks, ukey_s, s_te0, s_te1, s_te2, s_te3, s_te4, s_td0, s_td1, s_td2, s_td3, s_td4);
 }
 
-void AES128_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void AES128_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   u32 in_s[4];
 
@@ -1202,7 +1202,7 @@ void AES128_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0
   out[3] = swap32_S (out_s[3]);
 }
 
-void AES128_decrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void AES128_decrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 in_s[4];
 
@@ -1221,7 +1221,7 @@ void AES128_decrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_td0
   out[3] = swap32_S (out_s[3]);
 }
 
-void AES256_set_encrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void AES256_set_encrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   u32 ukey_s[8];
 
@@ -1237,7 +1237,7 @@ void AES256_set_encrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_
   aes256_set_encrypt_key (ks, ukey_s, s_te0, s_te1, s_te2, s_te3, s_te4);
 }
 
-void AES256_set_decrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void AES256_set_decrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 ukey_s[8];
 
@@ -1253,7 +1253,7 @@ void AES256_set_decrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, SHM_
   aes256_set_decrypt_key (ks, ukey_s, s_te0, s_te1, s_te2, s_te3, s_te4, s_td0, s_td1, s_td2, s_td3, s_td4);
 }
 
-void AES256_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void AES256_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   u32 in_s[4];
 
@@ -1272,7 +1272,7 @@ void AES256_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0
   out[3] = swap32_S (out_s[3]);
 }
 
-void AES256_decrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void AES256_decrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 in_s[4];
 
diff --git a/OpenCL/inc_cipher_serpent.cl b/OpenCL/inc_cipher_serpent.cl
index 783ad5738..20a077a38 100644
--- a/OpenCL/inc_cipher_serpent.cl
+++ b/OpenCL/inc_cipher_serpent.cl
@@ -403,7 +403,7 @@
 
 // 128 bit key
 
-void serpent128_set_key (u32 *ks, const u32 *ukey)
+DECLSPEC void serpent128_set_key (u32 *ks, const u32 *ukey)
 {
   #ifdef _unroll
   #pragma unroll
@@ -469,7 +469,7 @@ void serpent128_set_key (u32 *ks, const u32 *ukey)
   k_set(32,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get(32,e,f,g,h);
 }
 
-void serpent128_encrypt (const u32 *ks, const u32 *in, u32 *out)
+DECLSPEC void serpent128_encrypt (const u32 *ks, const u32 *in, u32 *out)
 {
   u32  a,b,c,d,e,f,g,h;
   u32  t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15,t16;
@@ -519,7 +519,7 @@ void serpent128_encrypt (const u32 *ks, const u32 *in, u32 *out)
   out[3] = d;
 }
 
-void serpent128_decrypt (const u32 *ks, const u32 *in, u32 *out)
+DECLSPEC void serpent128_decrypt (const u32 *ks, const u32 *in, u32 *out)
 {
   u32  a,b,c,d,e,f,g,h;
   u32  t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15,t16;
@@ -571,7 +571,7 @@ void serpent128_decrypt (const u32 *ks, const u32 *in, u32 *out)
 
 // 256 bit key
 
-void serpent256_set_key (u32 *ks, const u32 *ukey)
+DECLSPEC void serpent256_set_key (u32 *ks, const u32 *ukey)
 {
   #ifdef _unroll
   #pragma unroll
@@ -627,7 +627,7 @@ void serpent256_set_key (u32 *ks, const u32 *ukey)
   k_set(32,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get(32,e,f,g,h);
 }
 
-void serpent256_encrypt (const u32 *ks, const u32 *in, u32 *out)
+DECLSPEC void serpent256_encrypt (const u32 *ks, const u32 *in, u32 *out)
 {
   u32  a,b,c,d,e,f,g,h;
   u32  t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15,t16;
@@ -677,7 +677,7 @@ void serpent256_encrypt (const u32 *ks, const u32 *in, u32 *out)
   out[3] = d;
 }
 
-void serpent256_decrypt (const u32 *ks, const u32 *in, u32 *out)
+DECLSPEC void serpent256_decrypt (const u32 *ks, const u32 *in, u32 *out)
 {
   u32  a,b,c,d,e,f,g,h;
   u32  t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15,t16;
diff --git a/OpenCL/inc_cipher_twofish.cl b/OpenCL/inc_cipher_twofish.cl
index 8a326c347..ddbbe6869 100644
--- a/OpenCL/inc_cipher_twofish.cl
+++ b/OpenCL/inc_cipher_twofish.cl
@@ -266,7 +266,7 @@ __constant static u32a  m_tab[4][256] =
 #define q42(x,k) q (1, q (0, q (0, q (0, x) ^ extract_byte (k[3], 2)) ^ extract_byte (k[2], 2)) ^ extract_byte (k[1], 2)) ^ extract_byte (k[0], 2)
 #define q43(x,k) q (1, q (1, q (0, q (1, x) ^ extract_byte (k[3], 3)) ^ extract_byte (k[2], 3)) ^ extract_byte (k[1], 3)) ^ extract_byte (k[0], 3)
 
-u32 mds_rem (u32 p0, u32 p1)
+DECLSPEC u32 mds_rem (u32 p0, u32 p1)
 {
   #define G_MOD 0x14d
 
@@ -332,7 +332,7 @@ u32 mds_rem (u32 p0, u32 p1)
   data[1] = rotr32_S (data[1] ^ (t2 + 2 * t3 + lk[4 * (i) +  9]), 1); \
 }
 
-u32 h_fun128 (u32 *sk, u32 *lk, const u32 x, const u32 *key)
+DECLSPEC u32 h_fun128 (u32 *sk, u32 *lk, const u32 x, const u32 *key)
 {
   u32  b0, b1, b2, b3;
 
@@ -349,7 +349,7 @@ u32 h_fun128 (u32 *sk, u32 *lk, const u32 x, const u32 *key)
   return mds (0, b0) ^ mds (1, b1) ^ mds (2, b2) ^ mds (3, b3);
 }
 
-void twofish128_set_key (u32 *sk, u32 *lk, const u32 *ukey)
+DECLSPEC void twofish128_set_key (u32 *sk, u32 *lk, const u32 *ukey)
 {
   u32 me_key[2];
 
@@ -379,7 +379,7 @@ void twofish128_set_key (u32 *sk, u32 *lk, const u32 *ukey)
   }
 }
 
-void twofish128_encrypt (const u32 *sk, const u32 *lk, const u32 *in, u32 *out)
+DECLSPEC void twofish128_encrypt (const u32 *sk, const u32 *lk, const u32 *in, u32 *out)
 {
   u32 data[4];
 
@@ -403,7 +403,7 @@ void twofish128_encrypt (const u32 *sk, const u32 *lk, const u32 *in, u32 *out)
   out[3] = data[1] ^ lk[7];
 }
 
-void twofish128_decrypt (const u32 *sk, const u32 *lk, const u32 *in, u32 *out)
+DECLSPEC void twofish128_decrypt (const u32 *sk, const u32 *lk, const u32 *in, u32 *out)
 {
   u32 data[4];
 
@@ -465,7 +465,7 @@ void twofish128_decrypt (const u32 *sk, const u32 *lk, const u32 *in, u32 *out)
   data[1] = rotr32_S (data[1] ^ (t2 + 2 * t3 + lk[4 * (i) +  9]), 1); \
 }
 
-u32 h_fun256 (u32 *sk, u32 *lk, const u32 x, const u32 *key)
+DECLSPEC u32 h_fun256 (u32 *sk, u32 *lk, const u32 x, const u32 *key)
 {
   u32  b0, b1, b2, b3;
 
@@ -492,7 +492,7 @@ u32 h_fun256 (u32 *sk, u32 *lk, const u32 x, const u32 *key)
   return mds (0, b0) ^ mds (1, b1) ^ mds (2, b2) ^ mds (3, b3);
 }
 
-void twofish256_set_key (u32 *sk, u32 *lk, const u32 *ukey)
+DECLSPEC void twofish256_set_key (u32 *sk, u32 *lk, const u32 *ukey)
 {
   u32 me_key[4];
 
@@ -528,7 +528,7 @@ void twofish256_set_key (u32 *sk, u32 *lk, const u32 *ukey)
   }
 }
 
-void twofish256_encrypt (const u32 *sk, const u32 *lk, const u32 *in, u32 *out)
+DECLSPEC void twofish256_encrypt (const u32 *sk, const u32 *lk, const u32 *in, u32 *out)
 {
   u32 data[4];
 
@@ -552,7 +552,7 @@ void twofish256_encrypt (const u32 *sk, const u32 *lk, const u32 *in, u32 *out)
   out[3] = data[1] ^ lk[7];
 }
 
-void twofish256_decrypt (const u32 *sk, const u32 *lk, const u32 *in, u32 *out)
+DECLSPEC void twofish256_decrypt (const u32 *sk, const u32 *lk, const u32 *in, u32 *out)
 {
   u32 data[4];
 
diff --git a/OpenCL/inc_common.cl b/OpenCL/inc_common.cl
index 44d7bd8c5..28cb089b4 100644
--- a/OpenCL/inc_common.cl
+++ b/OpenCL/inc_common.cl
@@ -75,7 +75,7 @@ __constant static u32a c_append_helper[64][16] =
  * pure scalar functions
  */
 
-int ffz (const u32 v)
+DECLSPEC int ffz (const u32 v)
 {
   #ifdef _unroll
   #pragma unroll
@@ -90,7 +90,7 @@ int ffz (const u32 v)
   return -1;
 }
 
-int hash_comp (const u32 d1[4], __global const u32 *d2)
+DECLSPEC int hash_comp (const u32 d1[4], __global const u32 *d2)
 {
   if (d1[3] > d2[DGST_R3]) return ( 1);
   if (d1[3] < d2[DGST_R3]) return (-1);
@@ -104,7 +104,7 @@ int hash_comp (const u32 d1[4], __global const u32 *d2)
   return (0);
 }
 
-int find_hash (const u32 digest[4], const u32 digests_cnt, __global const digest_t *digests_buf)
+DECLSPEC int find_hash (const u32 digest[4], const u32 digests_cnt, __global const digest_t *digests_buf)
 {
   for (u32 l = 0, r = digests_cnt; r; r >>= 1)
   {
@@ -127,12 +127,12 @@ int find_hash (const u32 digest[4], const u32 digests_cnt, __global const digest
   return (-1);
 }
 
-u32 check_bitmap (__global const u32 *bitmap, const u32 bitmap_mask, const u32 bitmap_shift, const u32 digest)
+DECLSPEC u32 check_bitmap (__global const u32 *bitmap, const u32 bitmap_mask, const u32 bitmap_shift, const u32 digest)
 {
   return (bitmap[(digest >> bitmap_shift) & bitmap_mask] & (1 << (digest & 0x1f)));
 }
 
-u32 check (const u32 digest[4], __global const u32 *bitmap_s1_a, __global const u32 *bitmap_s1_b, __global const u32 *bitmap_s1_c, __global const u32 *bitmap_s1_d, __global const u32 *bitmap_s2_a, __global const u32 *bitmap_s2_b, __global const u32 *bitmap_s2_c, __global const u32 *bitmap_s2_d, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2)
+DECLSPEC u32 check (const u32 digest[4], __global const u32 *bitmap_s1_a, __global const u32 *bitmap_s1_b, __global const u32 *bitmap_s1_c, __global const u32 *bitmap_s1_d, __global const u32 *bitmap_s2_a, __global const u32 *bitmap_s2_b, __global const u32 *bitmap_s2_c, __global const u32 *bitmap_s2_d, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2)
 {
   if (check_bitmap (bitmap_s1_a, bitmap_mask, bitmap_shift1, digest[0]) == 0) return (0);
   if (check_bitmap (bitmap_s1_b, bitmap_mask, bitmap_shift1, digest[1]) == 0) return (0);
@@ -147,7 +147,7 @@ u32 check (const u32 digest[4], __global const u32 *bitmap_s1_a, __global const
   return (1);
 }
 
-void mark_hash (__global plain_t *plains_buf, __global u32 *d_result, const u32 salt_pos, const u32 digests_cnt, const u32 digest_pos, const u32 hash_pos, const u32 gid, const u32 il_pos)
+DECLSPEC void mark_hash (__global plain_t *plains_buf, __global u32 *d_result, const u32 salt_pos, const u32 digests_cnt, const u32 digest_pos, const u32 hash_pos, const u32 gid, const u32 il_pos)
 {
   const u32 idx = atomic_inc (d_result);
 
@@ -168,7 +168,7 @@ void mark_hash (__global plain_t *plains_buf, __global u32 *d_result, const u32
   plains_buf[idx].il_pos      = il_pos;
 }
 
-int count_char (const u32 *buf, const int elems, const u32 c)
+DECLSPEC int count_char (const u32 *buf, const int elems, const u32 c)
 {
   int r = 0;
 
@@ -185,7 +185,7 @@ int count_char (const u32 *buf, const int elems, const u32 c)
   return r;
 }
 
-float get_entropy (const u32 *buf, const int elems)
+DECLSPEC float get_entropy (const u32 *buf, const int elems)
 {
   const int length = elems * 4;
 
@@ -208,7 +208,7 @@ float get_entropy (const u32 *buf, const int elems)
   return entropy;
 }
 
-int is_valid_hex_8 (const u8 v)
+DECLSPEC int is_valid_hex_8 (const u8 v)
 {
   // direct lookup table is slower thanks to CMOV
 
@@ -218,7 +218,7 @@ int is_valid_hex_8 (const u8 v)
   return 0;
 }
 
-int is_valid_hex_32 (const u32 v)
+DECLSPEC int is_valid_hex_32 (const u32 v)
 {
   if (is_valid_hex_8 ((u8) (v >>  0)) == 0) return 0;
   if (is_valid_hex_8 ((u8) (v >>  8)) == 0) return 0;
@@ -232,7 +232,7 @@ int is_valid_hex_32 (const u32 v)
  * vector functions
  */
 
-void make_utf16be (const u32x in[4], u32x out1[4], u32x out2[4])
+DECLSPEC void make_utf16be (const u32x in[4], u32x out1[4], u32x out2[4])
 {
   #if defined IS_NV
 
@@ -270,7 +270,7 @@ void make_utf16be (const u32x in[4], u32x out1[4], u32x out2[4])
   #endif
 }
 
-void make_utf16beN (const u32x in[4], u32x out1[4], u32x out2[4])
+DECLSPEC void make_utf16beN (const u32x in[4], u32x out1[4], u32x out2[4])
 {
   #if defined IS_NV
 
@@ -308,7 +308,7 @@ void make_utf16beN (const u32x in[4], u32x out1[4], u32x out2[4])
   #endif
 }
 
-void make_utf16le (const u32x in[4], u32x out1[4], u32x out2[4])
+DECLSPEC void make_utf16le (const u32x in[4], u32x out1[4], u32x out2[4])
 {
   #if defined IS_NV
 
@@ -346,7 +346,7 @@ void make_utf16le (const u32x in[4], u32x out1[4], u32x out2[4])
   #endif
 }
 
-void make_utf16leN (const u32x in[4], u32x out1[4], u32x out2[4])
+DECLSPEC void make_utf16leN (const u32x in[4], u32x out1[4], u32x out2[4])
 {
   #if defined IS_NV
 
@@ -384,7 +384,7 @@ void make_utf16leN (const u32x in[4], u32x out1[4], u32x out2[4])
   #endif
 }
 
-void undo_utf16be (const u32x in1[4], const u32x in2[4], u32x out[4])
+DECLSPEC void undo_utf16be (const u32x in1[4], const u32x in2[4], u32x out[4])
 {
   #if defined IS_NV
 
@@ -414,7 +414,7 @@ void undo_utf16be (const u32x in1[4], const u32x in2[4], u32x out[4])
   #endif
 }
 
-void undo_utf16le (const u32x in1[4], const u32x in2[4], u32x out[4])
+DECLSPEC void undo_utf16le (const u32x in1[4], const u32x in2[4], u32x out[4])
 {
   #if defined IS_NV
 
@@ -444,7 +444,7 @@ void undo_utf16le (const u32x in1[4], const u32x in2[4], u32x out[4])
   #endif
 }
 
-void append_0x80_1x4 (u32x w0[4], const u32 offset)
+DECLSPEC void append_0x80_1x4 (u32x w0[4], const u32 offset)
 {
   w0[0] |= 0x80808080 & c_append_helper[offset][0];
   w0[1] |= 0x80808080 & c_append_helper[offset][1];
@@ -452,7 +452,7 @@ void append_0x80_1x4 (u32x w0[4], const u32 offset)
   w0[3] |= 0x80808080 & c_append_helper[offset][3];
 }
 
-void append_0x80_2x4 (u32x w0[4], u32x w1[4], const u32 offset)
+DECLSPEC void append_0x80_2x4 (u32x w0[4], u32x w1[4], const u32 offset)
 {
   w0[0] |= 0x80808080 & c_append_helper[offset][0];
   w0[1] |= 0x80808080 & c_append_helper[offset][1];
@@ -464,7 +464,7 @@ void append_0x80_2x4 (u32x w0[4], u32x w1[4], const u32 offset)
   w1[3] |= 0x80808080 & c_append_helper[offset][7];
 }
 
-void append_0x80_3x4 (u32x w0[4], u32x w1[4], u32x w2[4], const u32 offset)
+DECLSPEC void append_0x80_3x4 (u32x w0[4], u32x w1[4], u32x w2[4], const u32 offset)
 {
   w0[0] |= 0x80808080 & c_append_helper[offset][ 0];
   w0[1] |= 0x80808080 & c_append_helper[offset][ 1];
@@ -480,7 +480,7 @@ void append_0x80_3x4 (u32x w0[4], u32x w1[4], u32x w2[4], const u32 offset)
   w2[3] |= 0x80808080 & c_append_helper[offset][11];
 }
 
-void append_0x80_4x4 (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32 offset)
+DECLSPEC void append_0x80_4x4 (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32 offset)
 {
   w0[0] |= 0x80808080 & c_append_helper[offset][ 0];
   w0[1] |= 0x80808080 & c_append_helper[offset][ 1];
@@ -500,7 +500,7 @@ void append_0x80_4x4 (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32
   w3[3] |= 0x80808080 & c_append_helper[offset][15];
 }
 
-void append_0x80_8x4 (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const u32 offset)
+DECLSPEC void append_0x80_8x4 (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const u32 offset)
 {
   switch (offset)
   {
@@ -1018,7 +1018,7 @@ void append_0x80_8x4 (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4]
   }
 }
 
-void append_0x80_1x16 (u32x w[16], const u32 offset)
+DECLSPEC void append_0x80_1x16 (u32x w[16], const u32 offset)
 {
   switch (offset)
   {
@@ -1280,7 +1280,7 @@ void append_0x80_1x16 (u32x w[16], const u32 offset)
   }
 }
 
-void switch_buffer_by_offset_le (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_le (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32 offset)
 {
   const int offset_mod_4 = offset & 3;
 
@@ -1987,7 +1987,7 @@ void switch_buffer_by_offset_le (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4],
   #endif
 }
 
-void switch_buffer_by_offset_carry_le (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x c0[4], u32x c1[4], u32x c2[4], u32x c3[4], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_carry_le (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x c0[4], u32x c1[4], u32x c2[4], u32x c3[4], const u32 offset)
 {
   const int offset_mod_4 = offset & 3;
 
@@ -3309,7 +3309,7 @@ void switch_buffer_by_offset_carry_le (u32x w0[4], u32x w1[4], u32x w2[4], u32x
   #endif
 }
 
-void switch_buffer_by_offset_be (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_be (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32 offset)
 {
   #ifdef IS_AMD
   volatile const int offset_switch = offset / 4;
@@ -3980,7 +3980,7 @@ void switch_buffer_by_offset_be (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4],
   #endif
 }
 
-void switch_buffer_by_offset_carry_be (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x c0[4], u32x c1[4], u32x c2[4], u32x c3[4], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_carry_be (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x c0[4], u32x c1[4], u32x c2[4], u32x c3[4], const u32 offset)
 {
   #ifdef IS_AMD
   volatile const int offset_switch = offset / 4;
@@ -4920,7 +4920,7 @@ void switch_buffer_by_offset_carry_be (u32x w0[4], u32x w1[4], u32x w2[4], u32x
   #endif
 }
 
-void switch_buffer_by_offset_8x4_le (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_8x4_le (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const u32 offset)
 {
   const int offset_mod_4 = offset & 3;
 
@@ -6730,7 +6730,7 @@ void switch_buffer_by_offset_8x4_le (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3
   #endif
 }
 
-void switch_buffer_by_offset_8x4_be (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_8x4_be (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const u32 offset)
 {
   #ifdef IS_AMD
   volatile const int offset_switch = offset / 4;
@@ -9062,7 +9062,7 @@ void switch_buffer_by_offset_8x4_be (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3
   #endif
 }
 
-void switch_buffer_by_offset_8x4_carry_be (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u32x c0[4], u32x c1[4], u32x c2[4], u32x c3[4], u32x c4[4], u32x c5[4], u32x c6[4], u32x c7[4], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u32x c0[4], u32x c1[4], u32x c2[4], u32x c3[4], u32x c4[4], u32x c5[4], u32x c6[4], u32x c7[4], const u32 offset)
 {
   #ifdef IS_AMD
   volatile const int offset_switch = offset / 4;
@@ -12450,7 +12450,7 @@ void switch_buffer_by_offset_8x4_carry_be (u32x w0[4], u32x w1[4], u32x w2[4], u
   #endif
 }
 
-void switch_buffer_by_offset_1x64_le (u32x w[64], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_1x64_le (u32x w[64], const u32 offset)
 {
   const int offset_mod_4 = offset & 3;
 
@@ -21194,7 +21194,7 @@ void switch_buffer_by_offset_1x64_le (u32x w[64], const u32 offset)
   #endif
 }
 
-void switch_buffer_by_offset_1x64_be (u32x w[64], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_1x64_be (u32x w[64], const u32 offset)
 {
   #ifdef IS_AMD
   volatile const int offset_switch = offset / 4;
@@ -29930,7 +29930,7 @@ void switch_buffer_by_offset_1x64_be (u32x w[64], const u32 offset)
  * vector functions as scalar (for outer loop usage)
  */
 
-void truncate_block_4x4_le_S (u32 w0[4], const u32 len)
+DECLSPEC void truncate_block_4x4_le_S (u32 w0[4], const u32 len)
 {
   switch (len)
   {
@@ -30040,7 +30040,7 @@ void truncate_block_4x4_le_S (u32 w0[4], const u32 len)
   }
 }
 
-void truncate_block_4x4_be_S (u32 w0[4], const u32 len)
+DECLSPEC void truncate_block_4x4_be_S (u32 w0[4], const u32 len)
 {
   switch (len)
   {
@@ -30150,7 +30150,7 @@ void truncate_block_4x4_be_S (u32 w0[4], const u32 len)
   }
 }
 
-void truncate_block_16x4_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 len)
+DECLSPEC void truncate_block_16x4_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 len)
 {
   switch (len)
   {
@@ -30956,7 +30956,7 @@ void truncate_block_16x4_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const
   }
 }
 
-void truncate_block_16x4_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 len)
+DECLSPEC void truncate_block_16x4_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 len)
 {
   switch (len)
   {
@@ -31762,7 +31762,7 @@ void truncate_block_16x4_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const
   }
 }
 
-void append_0x01_2x4_S (u32 w0[4], u32 w1[4], const u32 offset)
+DECLSPEC void append_0x01_2x4_S (u32 w0[4], u32 w1[4], const u32 offset)
 {
   w0[0] |= 0x01010101 & c_append_helper[offset][0];
   w0[1] |= 0x01010101 & c_append_helper[offset][1];
@@ -31774,7 +31774,7 @@ void append_0x01_2x4_S (u32 w0[4], u32 w1[4], const u32 offset)
   w1[3] |= 0x01010101 & c_append_helper[offset][7];
 }
 
-void append_0x80_1x4_S (u32 w0[4], const u32 offset)
+DECLSPEC void append_0x80_1x4_S (u32 w0[4], const u32 offset)
 {
   w0[0] |= 0x80808080 & c_append_helper[offset][0];
   w0[1] |= 0x80808080 & c_append_helper[offset][1];
@@ -31782,7 +31782,7 @@ void append_0x80_1x4_S (u32 w0[4], const u32 offset)
   w0[3] |= 0x80808080 & c_append_helper[offset][3];
 }
 
-void append_0x80_2x4_S (u32 w0[4], u32 w1[4], const u32 offset)
+DECLSPEC void append_0x80_2x4_S (u32 w0[4], u32 w1[4], const u32 offset)
 {
   w0[0] |= 0x80808080 & c_append_helper[offset][0];
   w0[1] |= 0x80808080 & c_append_helper[offset][1];
@@ -31794,7 +31794,7 @@ void append_0x80_2x4_S (u32 w0[4], u32 w1[4], const u32 offset)
   w1[3] |= 0x80808080 & c_append_helper[offset][7];
 }
 
-void append_0x80_3x4_S (u32 w0[4], u32 w1[4], u32 w2[4], const u32 offset)
+DECLSPEC void append_0x80_3x4_S (u32 w0[4], u32 w1[4], u32 w2[4], const u32 offset)
 {
   w0[0] |= 0x80808080 & c_append_helper[offset][ 0];
   w0[1] |= 0x80808080 & c_append_helper[offset][ 1];
@@ -31810,7 +31810,7 @@ void append_0x80_3x4_S (u32 w0[4], u32 w1[4], u32 w2[4], const u32 offset)
   w2[3] |= 0x80808080 & c_append_helper[offset][11];
 }
 
-void append_0x80_4x4_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 offset)
+DECLSPEC void append_0x80_4x4_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 offset)
 {
   w0[0] |= 0x80808080 & c_append_helper[offset][ 0];
   w0[1] |= 0x80808080 & c_append_helper[offset][ 1];
@@ -31830,7 +31830,7 @@ void append_0x80_4x4_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 of
   w3[3] |= 0x80808080 & c_append_helper[offset][15];
 }
 
-void append_0x80_8x4_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const u32 offset)
+DECLSPEC void append_0x80_8x4_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const u32 offset)
 {
   switch (offset)
   {
@@ -32348,7 +32348,7 @@ void append_0x80_8x4_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u
   }
 }
 
-void make_utf16be_S (const u32 in[4], u32 out1[4], u32 out2[4])
+DECLSPEC void make_utf16be_S (const u32 in[4], u32 out1[4], u32 out2[4])
 {
   #if defined IS_NV
 
@@ -32386,7 +32386,7 @@ void make_utf16be_S (const u32 in[4], u32 out1[4], u32 out2[4])
   #endif
 }
 
-void make_utf16le_S (const u32 in[4], u32 out1[4], u32 out2[4])
+DECLSPEC void make_utf16le_S (const u32 in[4], u32 out1[4], u32 out2[4])
 {
   #if defined IS_NV
 
@@ -32424,7 +32424,7 @@ void make_utf16le_S (const u32 in[4], u32 out1[4], u32 out2[4])
   #endif
 }
 
-void undo_utf16be_S (const u32 in1[4], const u32 in2[4], u32 out[4])
+DECLSPEC void undo_utf16be_S (const u32 in1[4], const u32 in2[4], u32 out[4])
 {
   #if defined IS_NV
 
@@ -32454,7 +32454,7 @@ void undo_utf16be_S (const u32 in1[4], const u32 in2[4], u32 out[4])
   #endif
 }
 
-void undo_utf16le_S (const u32 in1[4], const u32 in2[4], u32 out[4])
+DECLSPEC void undo_utf16le_S (const u32 in1[4], const u32 in2[4], u32 out[4])
 {
   #if defined IS_NV
 
@@ -32484,7 +32484,7 @@ void undo_utf16le_S (const u32 in1[4], const u32 in2[4], u32 out[4])
   #endif
 }
 
-void switch_buffer_by_offset_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 offset)
 {
   const int offset_mod_4 = offset & 3;
 
@@ -33190,7 +33190,7 @@ void switch_buffer_by_offset_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], c
   #endif
 }
 
-void switch_buffer_by_offset_carry_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 c0[4], u32 c1[4], u32 c2[4], u32 c3[4], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_carry_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 c0[4], u32 c1[4], u32 c2[4], u32 c3[4], const u32 offset)
 {
   const int offset_mod_4 = offset & 3;
 
@@ -34512,7 +34512,7 @@ void switch_buffer_by_offset_carry_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3
   #endif
 }
 
-void switch_buffer_by_offset_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 offset)
 {
   #ifdef IS_AMD
   volatile const int offset_switch = offset / 4;
@@ -35180,7 +35180,7 @@ void switch_buffer_by_offset_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], c
   #endif
 }
 
-void switch_buffer_by_offset_carry_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 c0[4], u32 c1[4], u32 c2[4], u32 c3[4], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_carry_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 c0[4], u32 c1[4], u32 c2[4], u32 c3[4], const u32 offset)
 {
   #ifdef IS_AMD
   volatile const int offset_switch = offset / 4;
@@ -36120,7 +36120,7 @@ void switch_buffer_by_offset_carry_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3
   #endif
 }
 
-void switch_buffer_by_offset_8x4_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_8x4_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const u32 offset)
 {
   const int offset_mod_4 = offset & 3;
 
@@ -37930,7 +37930,7 @@ void switch_buffer_by_offset_8x4_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4
   #endif
 }
 
-void switch_buffer_by_offset_8x4_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const u32 offset)
 {
   #ifdef IS_AMD
   volatile const int offset_switch = offset / 4;
@@ -40262,7 +40262,7 @@ void switch_buffer_by_offset_8x4_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4
   #endif
 }
 
-void switch_buffer_by_offset_8x4_carry_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], u32 c0[4], u32 c1[4], u32 c2[4], u32 c3[4], u32 c4[4], u32 c5[4], u32 c6[4], u32 c7[4], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], u32 c0[4], u32 c1[4], u32 c2[4], u32 c3[4], u32 c4[4], u32 c5[4], u32 c6[4], u32 c7[4], const u32 offset)
 {
   #ifdef IS_AMD
   volatile const int offset_switch = offset / 4;
@@ -43650,7 +43650,7 @@ void switch_buffer_by_offset_8x4_carry_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u3
   #endif
 }
 
-void switch_buffer_by_offset_1x64_le_S (u32 w[64], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_1x64_le_S (u32 w[64], const u32 offset)
 {
   const int offset_mod_4 = offset & 3;
 
@@ -52394,7 +52394,7 @@ void switch_buffer_by_offset_1x64_le_S (u32 w[64], const u32 offset)
   #endif
 }
 
-void switch_buffer_by_offset_1x64_be_S (u32 w[64], const u32 offset)
+DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 w[64], const u32 offset)
 {
   #ifdef IS_AMD
   volatile const int offset_switch = offset / 4;
@@ -61190,7 +61190,7 @@ void switch_buffer_by_offset_1x64_be_S (u32 w[64], const u32 offset)
   PACKSV4 (s6, v6, e);                                              \
   PACKSV4 (s7, v7, e);
 
-void switch_buffer_by_offset_le_VV (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32x offset)
+DECLSPEC void switch_buffer_by_offset_le_VV (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32x offset)
 {
   #if VECT_SIZE == 1
 
@@ -61250,7 +61250,7 @@ void switch_buffer_by_offset_le_VV (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[
   #endif
 }
 
-void switch_buffer_by_offset_8x4_le_VV (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const u32x offset)
+DECLSPEC void switch_buffer_by_offset_8x4_le_VV (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const u32x offset)
 {
   #if VECT_SIZE == 1
 
@@ -61430,7 +61430,7 @@ void switch_buffer_by_offset_8x4_le_VV (u32x w0[4], u32x w1[4], u32x w2[4], u32x
   #endif
 }
 
-void append_0x01_2x4_VV (u32x w0[4], u32x w1[4], const u32x offset)
+DECLSPEC void append_0x01_2x4_VV (u32x w0[4], u32x w1[4], const u32x offset)
 {
   #if VECT_SIZE == 1
 
@@ -61488,7 +61488,7 @@ void append_0x01_2x4_VV (u32x w0[4], u32x w1[4], const u32x offset)
   #endif
 }
 
-void append_0x80_2x4_VV (u32x w0[4], u32x w1[4], const u32x offset)
+DECLSPEC void append_0x80_2x4_VV (u32x w0[4], u32x w1[4], const u32x offset)
 {
   #if VECT_SIZE == 1
 
@@ -61546,7 +61546,7 @@ void append_0x80_2x4_VV (u32x w0[4], u32x w1[4], const u32x offset)
   #endif
 }
 
-void append_0x80_4x4_VV (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32x offset)
+DECLSPEC void append_0x80_4x4_VV (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32x offset)
 {
   #if VECT_SIZE == 1
 
@@ -61606,7 +61606,7 @@ void append_0x80_4x4_VV (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u
   #endif
 }
 
-static void gpu_decompress_entry (__global pw_idx_t *pws_idx, __global u32 *pws_comp, pw_t *pw, const u64 gid)
+DECLSPEC void gpu_decompress_entry (__global pw_idx_t *pws_idx, __global u32 *pws_comp, pw_t *pw, const u64 gid)
 {
   const u32 off = pws_idx[gid].off;
   const u32 cnt = pws_idx[gid].cnt;
diff --git a/OpenCL/inc_hash_md4.cl b/OpenCL/inc_hash_md4.cl
index eb4b7dc00..ec8994704 100644
--- a/OpenCL/inc_hash_md4.cl
+++ b/OpenCL/inc_hash_md4.cl
@@ -17,7 +17,7 @@ typedef struct md4_ctx
 
 } md4_ctx_t;
 
-void md4_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
+DECLSPEC void md4_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
 {
   u32 a = digest[0];
   u32 b = digest[1];
@@ -81,7 +81,7 @@ void md4_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32
   digest[3] += d;
 }
 
-void md4_init (md4_ctx_t *ctx)
+DECLSPEC void md4_init (md4_ctx_t *ctx)
 {
   ctx->h[0] = MD4M_A;
   ctx->h[1] = MD4M_B;
@@ -108,7 +108,7 @@ void md4_init (md4_ctx_t *ctx)
   ctx->len = 0;
 }
 
-void md4_update_64 (md4_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void md4_update_64 (md4_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -186,7 +186,7 @@ void md4_update_64 (md4_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4],
   }
 }
 
-void md4_update (md4_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md4_update (md4_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -238,7 +238,7 @@ void md4_update (md4_ctx_t *ctx, const u32 *w, const int len)
   md4_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void md4_update_swap (md4_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md4_update_swap (md4_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -324,7 +324,7 @@ void md4_update_swap (md4_ctx_t *ctx, const u32 *w, const int len)
   md4_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void md4_update_utf16le (md4_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md4_update_utf16le (md4_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -366,7 +366,7 @@ void md4_update_utf16le (md4_ctx_t *ctx, const u32 *w, const int len)
   md4_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void md4_update_utf16le_swap (md4_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md4_update_utf16le_swap (md4_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -442,7 +442,7 @@ void md4_update_utf16le_swap (md4_ctx_t *ctx, const u32 *w, const int len)
   md4_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void md4_update_global (md4_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md4_update_global (md4_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -494,7 +494,7 @@ void md4_update_global (md4_ctx_t *ctx, const __global u32 *w, const int len)
   md4_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void md4_update_global_swap (md4_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md4_update_global_swap (md4_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -580,7 +580,7 @@ void md4_update_global_swap (md4_ctx_t *ctx, const __global u32 *w, const int le
   md4_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void md4_update_global_utf16le (md4_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md4_update_global_utf16le (md4_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -622,7 +622,7 @@ void md4_update_global_utf16le (md4_ctx_t *ctx, const __global u32 *w, const int
   md4_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void md4_update_global_utf16le_swap (md4_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md4_update_global_utf16le_swap (md4_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -698,7 +698,7 @@ void md4_update_global_utf16le_swap (md4_ctx_t *ctx, const __global u32 *w, cons
   md4_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void md4_final (md4_ctx_t *ctx)
+DECLSPEC void md4_final (md4_ctx_t *ctx)
 {
   const int pos = ctx->len & 63;
 
@@ -741,7 +741,7 @@ typedef struct md4_hmac_ctx
 
 } md4_hmac_ctx_t;
 
-void md4_hmac_init_64 (md4_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4])
+DECLSPEC void md4_hmac_init_64 (md4_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4])
 {
   u32 t0[4];
   u32 t1[4];
@@ -795,7 +795,7 @@ void md4_hmac_init_64 (md4_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], co
   md4_update_64 (&ctx->opad, t0, t1, t2, t3, 64);
 }
 
-void md4_hmac_init (md4_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md4_hmac_init (md4_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -852,7 +852,7 @@ void md4_hmac_init (md4_hmac_ctx_t *ctx, const u32 *w, const int len)
   md4_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void md4_hmac_init_swap (md4_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md4_hmac_init_swap (md4_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -909,7 +909,7 @@ void md4_hmac_init_swap (md4_hmac_ctx_t *ctx, const u32 *w, const int len)
   md4_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void md4_hmac_init_global (md4_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void md4_hmac_init_global (md4_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -966,7 +966,7 @@ void md4_hmac_init_global (md4_hmac_ctx_t *ctx, __global const u32 *w, const int
   md4_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void md4_hmac_init_global_swap (md4_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void md4_hmac_init_global_swap (md4_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1023,52 +1023,52 @@ void md4_hmac_init_global_swap (md4_hmac_ctx_t *ctx, __global const u32 *w, cons
   md4_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void md4_hmac_update_64 (md4_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void md4_hmac_update_64 (md4_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   md4_update_64 (&ctx->ipad, w0, w1, w2, w3, len);
 }
 
-void md4_hmac_update (md4_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md4_hmac_update (md4_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   md4_update (&ctx->ipad, w, len);
 }
 
-void md4_hmac_update_swap (md4_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md4_hmac_update_swap (md4_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   md4_update_swap (&ctx->ipad, w, len);
 }
 
-void md4_hmac_update_utf16le (md4_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md4_hmac_update_utf16le (md4_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   md4_update_utf16le (&ctx->ipad, w, len);
 }
 
-void md4_hmac_update_utf16le_swap (md4_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md4_hmac_update_utf16le_swap (md4_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   md4_update_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void md4_hmac_update_global (md4_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md4_hmac_update_global (md4_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   md4_update_global (&ctx->ipad, w, len);
 }
 
-void md4_hmac_update_global_swap (md4_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md4_hmac_update_global_swap (md4_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   md4_update_global_swap (&ctx->ipad, w, len);
 }
 
-void md4_hmac_update_global_utf16le (md4_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md4_hmac_update_global_utf16le (md4_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   md4_update_global_utf16le (&ctx->ipad, w, len);
 }
 
-void md4_hmac_update_global_utf16le_swap (md4_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md4_hmac_update_global_utf16le_swap (md4_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   md4_update_global_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void md4_hmac_final (md4_hmac_ctx_t *ctx)
+DECLSPEC void md4_hmac_final (md4_hmac_ctx_t *ctx)
 {
   md4_final (&ctx->ipad);
 
@@ -1114,7 +1114,7 @@ typedef struct md4_ctx_vector
 
 } md4_ctx_vector_t;
 
-void md4_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
+DECLSPEC void md4_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
 {
   u32x a = digest[0];
   u32x b = digest[1];
@@ -1178,7 +1178,7 @@ void md4_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4],
   digest[3] += d;
 }
 
-void md4_init_vector (md4_ctx_vector_t *ctx)
+DECLSPEC void md4_init_vector (md4_ctx_vector_t *ctx)
 {
   ctx->h[0] = MD4M_A;
   ctx->h[1] = MD4M_B;
@@ -1205,7 +1205,7 @@ void md4_init_vector (md4_ctx_vector_t *ctx)
   ctx->len = 0;
 }
 
-void md4_init_vector_from_scalar (md4_ctx_vector_t *ctx, md4_ctx_t *ctx0)
+DECLSPEC void md4_init_vector_from_scalar (md4_ctx_vector_t *ctx, md4_ctx_t *ctx0)
 {
   ctx->h[0] = ctx0->h[0];
   ctx->h[1] = ctx0->h[1];
@@ -1232,7 +1232,7 @@ void md4_init_vector_from_scalar (md4_ctx_vector_t *ctx, md4_ctx_t *ctx0)
   ctx->len = ctx0->len;
 }
 
-void md4_update_vector_64 (md4_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
+DECLSPEC void md4_update_vector_64 (md4_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -1310,7 +1310,7 @@ void md4_update_vector_64 (md4_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w
   }
 }
 
-void md4_update_vector (md4_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void md4_update_vector (md4_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1362,7 +1362,7 @@ void md4_update_vector (md4_ctx_vector_t *ctx, const u32x *w, const int len)
   md4_update_vector_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void md4_update_vector_swap (md4_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void md4_update_vector_swap (md4_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1448,7 +1448,7 @@ void md4_update_vector_swap (md4_ctx_vector_t *ctx, const u32x *w, const int len
   md4_update_vector_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void md4_update_vector_utf16le (md4_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void md4_update_vector_utf16le (md4_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1490,7 +1490,7 @@ void md4_update_vector_utf16le (md4_ctx_vector_t *ctx, const u32x *w, const int
   md4_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void md4_update_vector_utf16le_swap (md4_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void md4_update_vector_utf16le_swap (md4_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1566,7 +1566,7 @@ void md4_update_vector_utf16le_swap (md4_ctx_vector_t *ctx, const u32x *w, const
   md4_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void md4_final_vector (md4_ctx_vector_t *ctx)
+DECLSPEC void md4_final_vector (md4_ctx_vector_t *ctx)
 {
   const int pos = ctx->len & 63;
 
@@ -1609,7 +1609,7 @@ typedef struct md4_hmac_ctx_vector
 
 } md4_hmac_ctx_vector_t;
 
-void md4_hmac_init_vector_64 (md4_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4])
+DECLSPEC void md4_hmac_init_vector_64 (md4_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4])
 {
   u32x t0[4];
   u32x t1[4];
@@ -1663,7 +1663,7 @@ void md4_hmac_init_vector_64 (md4_hmac_ctx_vector_t *ctx, const u32x w0[4], cons
   md4_update_vector_64 (&ctx->opad, t0, t1, t2, t3, 64);
 }
 
-void md4_hmac_init_vector (md4_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void md4_hmac_init_vector (md4_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1720,17 +1720,17 @@ void md4_hmac_init_vector (md4_hmac_ctx_vector_t *ctx, const u32x *w, const int
   md4_hmac_init_vector_64 (ctx, w0, w1, w2, w3);
 }
 
-void md4_hmac_update_vector_64 (md4_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
+DECLSPEC void md4_hmac_update_vector_64 (md4_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
 {
   md4_update_vector_64 (&ctx->ipad, w0, w1, w2, w3, len);
 }
 
-void md4_hmac_update_vector (md4_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void md4_hmac_update_vector (md4_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   md4_update_vector (&ctx->ipad, w, len);
 }
 
-void md4_hmac_final_vector (md4_hmac_ctx_vector_t *ctx)
+DECLSPEC void md4_hmac_final_vector (md4_hmac_ctx_vector_t *ctx)
 {
   md4_final_vector (&ctx->ipad);
 
diff --git a/OpenCL/inc_hash_md5.cl b/OpenCL/inc_hash_md5.cl
index 37e7c7edc..48a090760 100644
--- a/OpenCL/inc_hash_md5.cl
+++ b/OpenCL/inc_hash_md5.cl
@@ -17,7 +17,7 @@ typedef struct md5_ctx
 
 } md5_ctx_t;
 
-void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
+DECLSPEC void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
 {
   u32 a = digest[0];
   u32 b = digest[1];
@@ -117,7 +117,7 @@ void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32
   digest[3] += d;
 }
 
-void md5_init (md5_ctx_t *ctx)
+DECLSPEC void md5_init (md5_ctx_t *ctx)
 {
   ctx->h[0] = MD5M_A;
   ctx->h[1] = MD5M_B;
@@ -144,7 +144,7 @@ void md5_init (md5_ctx_t *ctx)
   ctx->len = 0;
 }
 
-void md5_update_64 (md5_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void md5_update_64 (md5_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -222,7 +222,7 @@ void md5_update_64 (md5_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4],
   }
 }
 
-void md5_update (md5_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md5_update (md5_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -274,7 +274,7 @@ void md5_update (md5_ctx_t *ctx, const u32 *w, const int len)
   md5_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void md5_update_swap (md5_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md5_update_swap (md5_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -360,7 +360,7 @@ void md5_update_swap (md5_ctx_t *ctx, const u32 *w, const int len)
   md5_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void md5_update_utf16le (md5_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md5_update_utf16le (md5_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -402,7 +402,7 @@ void md5_update_utf16le (md5_ctx_t *ctx, const u32 *w, const int len)
   md5_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void md5_update_utf16le_swap (md5_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md5_update_utf16le_swap (md5_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -478,7 +478,7 @@ void md5_update_utf16le_swap (md5_ctx_t *ctx, const u32 *w, const int len)
   md5_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void md5_update_global (md5_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md5_update_global (md5_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -530,7 +530,7 @@ void md5_update_global (md5_ctx_t *ctx, const __global u32 *w, const int len)
   md5_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void md5_update_global_swap (md5_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md5_update_global_swap (md5_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -616,7 +616,7 @@ void md5_update_global_swap (md5_ctx_t *ctx, const __global u32 *w, const int le
   md5_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void md5_update_global_utf16le (md5_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md5_update_global_utf16le (md5_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -658,7 +658,7 @@ void md5_update_global_utf16le (md5_ctx_t *ctx, const __global u32 *w, const int
   md5_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void md5_update_global_utf16le_swap (md5_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md5_update_global_utf16le_swap (md5_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -734,7 +734,7 @@ void md5_update_global_utf16le_swap (md5_ctx_t *ctx, const __global u32 *w, cons
   md5_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void md5_final (md5_ctx_t *ctx)
+DECLSPEC void md5_final (md5_ctx_t *ctx)
 {
   const int pos = ctx->len & 63;
 
@@ -777,7 +777,7 @@ typedef struct md5_hmac_ctx
 
 } md5_hmac_ctx_t;
 
-void md5_hmac_init_64 (md5_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4])
+DECLSPEC void md5_hmac_init_64 (md5_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4])
 {
   u32 t0[4];
   u32 t1[4];
@@ -831,7 +831,7 @@ void md5_hmac_init_64 (md5_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], co
   md5_update_64 (&ctx->opad, t0, t1, t2, t3, 64);
 }
 
-void md5_hmac_init (md5_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md5_hmac_init (md5_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -888,7 +888,7 @@ void md5_hmac_init (md5_hmac_ctx_t *ctx, const u32 *w, const int len)
   md5_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void md5_hmac_init_swap (md5_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md5_hmac_init_swap (md5_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -945,7 +945,7 @@ void md5_hmac_init_swap (md5_hmac_ctx_t *ctx, const u32 *w, const int len)
   md5_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void md5_hmac_init_global (md5_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void md5_hmac_init_global (md5_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1002,7 +1002,7 @@ void md5_hmac_init_global (md5_hmac_ctx_t *ctx, __global const u32 *w, const int
   md5_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void md5_hmac_init_global_swap (md5_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void md5_hmac_init_global_swap (md5_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1059,52 +1059,52 @@ void md5_hmac_init_global_swap (md5_hmac_ctx_t *ctx, __global const u32 *w, cons
   md5_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void md5_hmac_update_64 (md5_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void md5_hmac_update_64 (md5_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   md5_update_64 (&ctx->ipad, w0, w1, w2, w3, len);
 }
 
-void md5_hmac_update (md5_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md5_hmac_update (md5_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   md5_update (&ctx->ipad, w, len);
 }
 
-void md5_hmac_update_swap (md5_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md5_hmac_update_swap (md5_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   md5_update_swap (&ctx->ipad, w, len);
 }
 
-void md5_hmac_update_utf16le (md5_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md5_hmac_update_utf16le (md5_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   md5_update_utf16le (&ctx->ipad, w, len);
 }
 
-void md5_hmac_update_utf16le_swap (md5_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void md5_hmac_update_utf16le_swap (md5_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   md5_update_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void md5_hmac_update_global (md5_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md5_hmac_update_global (md5_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   md5_update_global (&ctx->ipad, w, len);
 }
 
-void md5_hmac_update_global_swap (md5_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md5_hmac_update_global_swap (md5_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   md5_update_global_swap (&ctx->ipad, w, len);
 }
 
-void md5_hmac_update_global_utf16le (md5_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md5_hmac_update_global_utf16le (md5_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   md5_update_global_utf16le (&ctx->ipad, w, len);
 }
 
-void md5_hmac_update_global_utf16le_swap (md5_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void md5_hmac_update_global_utf16le_swap (md5_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   md5_update_global_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void md5_hmac_final (md5_hmac_ctx_t *ctx)
+DECLSPEC void md5_hmac_final (md5_hmac_ctx_t *ctx)
 {
   md5_final (&ctx->ipad);
 
@@ -1150,7 +1150,7 @@ typedef struct md5_ctx_vector
 
 } md5_ctx_vector_t;
 
-void md5_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
+DECLSPEC void md5_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
 {
   u32x a = digest[0];
   u32x b = digest[1];
@@ -1250,7 +1250,7 @@ void md5_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4],
   digest[3] += d;
 }
 
-void md5_init_vector (md5_ctx_vector_t *ctx)
+DECLSPEC void md5_init_vector (md5_ctx_vector_t *ctx)
 {
   ctx->h[0] = MD5M_A;
   ctx->h[1] = MD5M_B;
@@ -1277,7 +1277,7 @@ void md5_init_vector (md5_ctx_vector_t *ctx)
   ctx->len = 0;
 }
 
-void md5_init_vector_from_scalar (md5_ctx_vector_t *ctx, md5_ctx_t *ctx0)
+DECLSPEC void md5_init_vector_from_scalar (md5_ctx_vector_t *ctx, md5_ctx_t *ctx0)
 {
   ctx->h[0] = ctx0->h[0];
   ctx->h[1] = ctx0->h[1];
@@ -1304,7 +1304,7 @@ void md5_init_vector_from_scalar (md5_ctx_vector_t *ctx, md5_ctx_t *ctx0)
   ctx->len = ctx0->len;
 }
 
-void md5_update_vector_64 (md5_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
+DECLSPEC void md5_update_vector_64 (md5_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -1382,7 +1382,7 @@ void md5_update_vector_64 (md5_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w
   }
 }
 
-void md5_update_vector (md5_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void md5_update_vector (md5_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1434,7 +1434,7 @@ void md5_update_vector (md5_ctx_vector_t *ctx, const u32x *w, const int len)
   md5_update_vector_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void md5_update_vector_swap (md5_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void md5_update_vector_swap (md5_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1520,7 +1520,7 @@ void md5_update_vector_swap (md5_ctx_vector_t *ctx, const u32x *w, const int len
   md5_update_vector_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void md5_update_vector_utf16le (md5_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void md5_update_vector_utf16le (md5_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1562,7 +1562,7 @@ void md5_update_vector_utf16le (md5_ctx_vector_t *ctx, const u32x *w, const int
   md5_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void md5_update_vector_utf16le_swap (md5_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void md5_update_vector_utf16le_swap (md5_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1638,7 +1638,7 @@ void md5_update_vector_utf16le_swap (md5_ctx_vector_t *ctx, const u32x *w, const
   md5_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void md5_final_vector (md5_ctx_vector_t *ctx)
+DECLSPEC void md5_final_vector (md5_ctx_vector_t *ctx)
 {
   const int pos = ctx->len & 63;
 
@@ -1681,7 +1681,7 @@ typedef struct md5_hmac_ctx_vector
 
 } md5_hmac_ctx_vector_t;
 
-void md5_hmac_init_vector_64 (md5_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4])
+DECLSPEC void md5_hmac_init_vector_64 (md5_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4])
 {
   u32x t0[4];
   u32x t1[4];
@@ -1735,7 +1735,7 @@ void md5_hmac_init_vector_64 (md5_hmac_ctx_vector_t *ctx, const u32x w0[4], cons
   md5_update_vector_64 (&ctx->opad, t0, t1, t2, t3, 64);
 }
 
-void md5_hmac_init_vector (md5_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void md5_hmac_init_vector (md5_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1792,17 +1792,17 @@ void md5_hmac_init_vector (md5_hmac_ctx_vector_t *ctx, const u32x *w, const int
   md5_hmac_init_vector_64 (ctx, w0, w1, w2, w3);
 }
 
-void md5_hmac_update_vector_64 (md5_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
+DECLSPEC void md5_hmac_update_vector_64 (md5_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
 {
   md5_update_vector_64 (&ctx->ipad, w0, w1, w2, w3, len);
 }
 
-void md5_hmac_update_vector (md5_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void md5_hmac_update_vector (md5_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   md5_update_vector (&ctx->ipad, w, len);
 }
 
-void md5_hmac_final_vector (md5_hmac_ctx_vector_t *ctx)
+DECLSPEC void md5_hmac_final_vector (md5_hmac_ctx_vector_t *ctx)
 {
   md5_final_vector (&ctx->ipad);
 
diff --git a/OpenCL/inc_hash_ripemd160.cl b/OpenCL/inc_hash_ripemd160.cl
index bf5d2ec42..64b72603e 100644
--- a/OpenCL/inc_hash_ripemd160.cl
+++ b/OpenCL/inc_hash_ripemd160.cl
@@ -17,7 +17,7 @@ typedef struct ripemd160_ctx
 
 } ripemd160_ctx_t;
 
-void ripemd160_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5])
+DECLSPEC void ripemd160_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5])
 {
   u32 a1 = digest[0];
   u32 b1 = digest[1];
@@ -214,7 +214,7 @@ void ripemd160_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], con
   digest[4] = e;
 }
 
-void ripemd160_init (ripemd160_ctx_t *ctx)
+DECLSPEC void ripemd160_init (ripemd160_ctx_t *ctx)
 {
   ctx->h[0] = RIPEMD160M_A;
   ctx->h[1] = RIPEMD160M_B;
@@ -242,7 +242,7 @@ void ripemd160_init (ripemd160_ctx_t *ctx)
   ctx->len = 0;
 }
 
-void ripemd160_update_64 (ripemd160_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void ripemd160_update_64 (ripemd160_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -320,7 +320,7 @@ void ripemd160_update_64 (ripemd160_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4],
   }
 }
 
-void ripemd160_update (ripemd160_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void ripemd160_update (ripemd160_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -372,7 +372,7 @@ void ripemd160_update (ripemd160_ctx_t *ctx, const u32 *w, const int len)
   ripemd160_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void ripemd160_update_swap (ripemd160_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void ripemd160_update_swap (ripemd160_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -458,7 +458,7 @@ void ripemd160_update_swap (ripemd160_ctx_t *ctx, const u32 *w, const int len)
   ripemd160_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void ripemd160_update_utf16le (ripemd160_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void ripemd160_update_utf16le (ripemd160_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -500,7 +500,7 @@ void ripemd160_update_utf16le (ripemd160_ctx_t *ctx, const u32 *w, const int len
   ripemd160_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void ripemd160_update_utf16le_swap (ripemd160_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void ripemd160_update_utf16le_swap (ripemd160_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -576,7 +576,7 @@ void ripemd160_update_utf16le_swap (ripemd160_ctx_t *ctx, const u32 *w, const in
   ripemd160_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void ripemd160_update_global (ripemd160_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void ripemd160_update_global (ripemd160_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -628,7 +628,7 @@ void ripemd160_update_global (ripemd160_ctx_t *ctx, const __global u32 *w, const
   ripemd160_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void ripemd160_update_global_swap (ripemd160_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void ripemd160_update_global_swap (ripemd160_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -714,7 +714,7 @@ void ripemd160_update_global_swap (ripemd160_ctx_t *ctx, const __global u32 *w,
   ripemd160_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void ripemd160_update_global_utf16le (ripemd160_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void ripemd160_update_global_utf16le (ripemd160_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -756,7 +756,7 @@ void ripemd160_update_global_utf16le (ripemd160_ctx_t *ctx, const __global u32 *
   ripemd160_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void ripemd160_update_global_utf16le_swap (ripemd160_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void ripemd160_update_global_utf16le_swap (ripemd160_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -832,7 +832,7 @@ void ripemd160_update_global_utf16le_swap (ripemd160_ctx_t *ctx, const __global
   ripemd160_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void ripemd160_final (ripemd160_ctx_t *ctx)
+DECLSPEC void ripemd160_final (ripemd160_ctx_t *ctx)
 {
   const int pos = ctx->len & 63;
 
@@ -875,7 +875,7 @@ typedef struct ripemd160_hmac_ctx
 
 } ripemd160_hmac_ctx_t;
 
-void ripemd160_hmac_init_64 (ripemd160_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4])
+DECLSPEC void ripemd160_hmac_init_64 (ripemd160_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4])
 {
   u32 t0[4];
   u32 t1[4];
@@ -929,7 +929,7 @@ void ripemd160_hmac_init_64 (ripemd160_hmac_ctx_t *ctx, const u32 w0[4], const u
   ripemd160_update_64 (&ctx->opad, t0, t1, t2, t3, 64);
 }
 
-void ripemd160_hmac_init (ripemd160_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void ripemd160_hmac_init (ripemd160_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -986,7 +986,7 @@ void ripemd160_hmac_init (ripemd160_hmac_ctx_t *ctx, const u32 *w, const int len
   ripemd160_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void ripemd160_hmac_init_swap (ripemd160_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void ripemd160_hmac_init_swap (ripemd160_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1043,7 +1043,7 @@ void ripemd160_hmac_init_swap (ripemd160_hmac_ctx_t *ctx, const u32 *w, const in
   ripemd160_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void ripemd160_hmac_init_global (ripemd160_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void ripemd160_hmac_init_global (ripemd160_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1100,7 +1100,7 @@ void ripemd160_hmac_init_global (ripemd160_hmac_ctx_t *ctx, __global const u32 *
   ripemd160_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void ripemd160_hmac_init_global_swap (ripemd160_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void ripemd160_hmac_init_global_swap (ripemd160_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1157,52 +1157,52 @@ void ripemd160_hmac_init_global_swap (ripemd160_hmac_ctx_t *ctx, __global const
   ripemd160_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void ripemd160_hmac_update_64 (ripemd160_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void ripemd160_hmac_update_64 (ripemd160_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   ripemd160_update_64 (&ctx->ipad, w0, w1, w2, w3, len);
 }
 
-void ripemd160_hmac_update (ripemd160_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void ripemd160_hmac_update (ripemd160_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   ripemd160_update (&ctx->ipad, w, len);
 }
 
-void ripemd160_hmac_update_swap (ripemd160_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void ripemd160_hmac_update_swap (ripemd160_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   ripemd160_update_swap (&ctx->ipad, w, len);
 }
 
-void ripemd160_hmac_update_utf16le (ripemd160_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void ripemd160_hmac_update_utf16le (ripemd160_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   ripemd160_update_utf16le (&ctx->ipad, w, len);
 }
 
-void ripemd160_hmac_update_utf16le_swap (ripemd160_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void ripemd160_hmac_update_utf16le_swap (ripemd160_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   ripemd160_update_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void ripemd160_hmac_update_global (ripemd160_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void ripemd160_hmac_update_global (ripemd160_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   ripemd160_update_global (&ctx->ipad, w, len);
 }
 
-void ripemd160_hmac_update_global_swap (ripemd160_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void ripemd160_hmac_update_global_swap (ripemd160_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   ripemd160_update_global_swap (&ctx->ipad, w, len);
 }
 
-void ripemd160_hmac_update_global_utf16le (ripemd160_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void ripemd160_hmac_update_global_utf16le (ripemd160_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   ripemd160_update_global_utf16le (&ctx->ipad, w, len);
 }
 
-void ripemd160_hmac_update_global_utf16le_swap (ripemd160_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void ripemd160_hmac_update_global_utf16le_swap (ripemd160_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   ripemd160_update_global_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void ripemd160_hmac_final (ripemd160_hmac_ctx_t *ctx)
+DECLSPEC void ripemd160_hmac_final (ripemd160_hmac_ctx_t *ctx)
 {
   ripemd160_final (&ctx->ipad);
 
@@ -1248,7 +1248,7 @@ typedef struct ripemd160_ctx_vector
 
 } ripemd160_ctx_vector_t;
 
-void ripemd160_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[5])
+DECLSPEC void ripemd160_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[5])
 {
   u32x a1 = digest[0];
   u32x b1 = digest[1];
@@ -1445,7 +1445,7 @@ void ripemd160_transform_vector (const u32x w0[4], const u32x w1[4], const u32x
   digest[4] = e;
 }
 
-void ripemd160_init_vector (ripemd160_ctx_vector_t *ctx)
+DECLSPEC void ripemd160_init_vector (ripemd160_ctx_vector_t *ctx)
 {
   ctx->h[0] = RIPEMD160M_A;
   ctx->h[1] = RIPEMD160M_B;
@@ -1473,7 +1473,7 @@ void ripemd160_init_vector (ripemd160_ctx_vector_t *ctx)
   ctx->len = 0;
 }
 
-void ripemd160_init_vector_from_scalar (ripemd160_ctx_vector_t *ctx, ripemd160_ctx_t *ctx0)
+DECLSPEC void ripemd160_init_vector_from_scalar (ripemd160_ctx_vector_t *ctx, ripemd160_ctx_t *ctx0)
 {
   ctx->h[0] = ctx0->h[0];
   ctx->h[1] = ctx0->h[1];
@@ -1501,7 +1501,7 @@ void ripemd160_init_vector_from_scalar (ripemd160_ctx_vector_t *ctx, ripemd160_c
   ctx->len = ctx0->len;
 }
 
-void ripemd160_update_vector_64 (ripemd160_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
+DECLSPEC void ripemd160_update_vector_64 (ripemd160_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -1579,7 +1579,7 @@ void ripemd160_update_vector_64 (ripemd160_ctx_vector_t *ctx, u32x w0[4], u32x w
   }
 }
 
-void ripemd160_update_vector (ripemd160_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void ripemd160_update_vector (ripemd160_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1631,7 +1631,7 @@ void ripemd160_update_vector (ripemd160_ctx_vector_t *ctx, const u32x *w, const
   ripemd160_update_vector_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void ripemd160_update_vector_swap (ripemd160_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void ripemd160_update_vector_swap (ripemd160_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1717,7 +1717,7 @@ void ripemd160_update_vector_swap (ripemd160_ctx_vector_t *ctx, const u32x *w, c
   ripemd160_update_vector_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void ripemd160_update_vector_utf16le (ripemd160_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void ripemd160_update_vector_utf16le (ripemd160_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1759,7 +1759,7 @@ void ripemd160_update_vector_utf16le (ripemd160_ctx_vector_t *ctx, const u32x *w
   ripemd160_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void ripemd160_update_vector_utf16le_swap (ripemd160_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void ripemd160_update_vector_utf16le_swap (ripemd160_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1835,7 +1835,7 @@ void ripemd160_update_vector_utf16le_swap (ripemd160_ctx_vector_t *ctx, const u3
   ripemd160_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void ripemd160_final_vector (ripemd160_ctx_vector_t *ctx)
+DECLSPEC void ripemd160_final_vector (ripemd160_ctx_vector_t *ctx)
 {
   const int pos = ctx->len & 63;
 
@@ -1878,7 +1878,7 @@ typedef struct ripemd160_hmac_ctx_vector
 
 } ripemd160_hmac_ctx_vector_t;
 
-void ripemd160_hmac_init_vector_64 (ripemd160_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4])
+DECLSPEC void ripemd160_hmac_init_vector_64 (ripemd160_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4])
 {
   u32x t0[4];
   u32x t1[4];
@@ -1932,7 +1932,7 @@ void ripemd160_hmac_init_vector_64 (ripemd160_hmac_ctx_vector_t *ctx, const u32x
   ripemd160_update_vector_64 (&ctx->opad, t0, t1, t2, t3, 64);
 }
 
-void ripemd160_hmac_init_vector (ripemd160_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void ripemd160_hmac_init_vector (ripemd160_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1989,17 +1989,17 @@ void ripemd160_hmac_init_vector (ripemd160_hmac_ctx_vector_t *ctx, const u32x *w
   ripemd160_hmac_init_vector_64 (ctx, w0, w1, w2, w3);
 }
 
-void ripemd160_hmac_update_vector_64 (ripemd160_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
+DECLSPEC void ripemd160_hmac_update_vector_64 (ripemd160_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
 {
   ripemd160_update_vector_64 (&ctx->ipad, w0, w1, w2, w3, len);
 }
 
-void ripemd160_hmac_update_vector (ripemd160_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void ripemd160_hmac_update_vector (ripemd160_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   ripemd160_update_vector (&ctx->ipad, w, len);
 }
 
-void ripemd160_hmac_final_vector (ripemd160_hmac_ctx_vector_t *ctx)
+DECLSPEC void ripemd160_hmac_final_vector (ripemd160_hmac_ctx_vector_t *ctx)
 {
   ripemd160_final_vector (&ctx->ipad);
 
diff --git a/OpenCL/inc_hash_sha1.cl b/OpenCL/inc_hash_sha1.cl
index 9eea99fce..873d713b3 100644
--- a/OpenCL/inc_hash_sha1.cl
+++ b/OpenCL/inc_hash_sha1.cl
@@ -17,7 +17,7 @@ typedef struct sha1_ctx
 
 } sha1_ctx_t;
 
-void sha1_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5])
+DECLSPEC void sha1_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5])
 {
   u32 a = digest[0];
   u32 b = digest[1];
@@ -146,7 +146,7 @@ void sha1_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u3
   digest[4] += e;
 }
 
-void sha1_init (sha1_ctx_t *ctx)
+DECLSPEC void sha1_init (sha1_ctx_t *ctx)
 {
   ctx->h[0] = SHA1M_A;
   ctx->h[1] = SHA1M_B;
@@ -174,7 +174,7 @@ void sha1_init (sha1_ctx_t *ctx)
   ctx->len = 0;
 }
 
-void sha1_update_64 (sha1_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void sha1_update_64 (sha1_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -212,7 +212,7 @@ void sha1_update_64 (sha1_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4]
     u32 c2[4] = { 0 };
     u32 c3[4] = { 0 };
 
-    switch_buffer_by_offset_carry_be_S (w0, w1, w2, w3, c0, c1, c2, c3, pos);
+    //switch_buffer_by_offset_carry_be_S (w0, w1, w2, w3, c0, c1, c2, c3, pos);
 
     ctx->w0[0] |= w0[0];
     ctx->w0[1] |= w0[1];
@@ -252,7 +252,7 @@ void sha1_update_64 (sha1_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4]
   }
 }
 
-void sha1_update (sha1_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha1_update (sha1_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -304,7 +304,7 @@ void sha1_update (sha1_ctx_t *ctx, const u32 *w, const int len)
   sha1_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha1_update_swap (sha1_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha1_update_swap (sha1_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -390,7 +390,7 @@ void sha1_update_swap (sha1_ctx_t *ctx, const u32 *w, const int len)
   sha1_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha1_update_utf16le (sha1_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha1_update_utf16le (sha1_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -432,7 +432,7 @@ void sha1_update_utf16le (sha1_ctx_t *ctx, const u32 *w, const int len)
   sha1_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha1_update_utf16le_swap (sha1_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha1_update_utf16le_swap (sha1_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -508,7 +508,7 @@ void sha1_update_utf16le_swap (sha1_ctx_t *ctx, const u32 *w, const int len)
   sha1_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha1_update_utf16be (sha1_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha1_update_utf16be (sha1_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -550,7 +550,7 @@ void sha1_update_utf16be (sha1_ctx_t *ctx, const u32 *w, const int len)
   sha1_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha1_update_utf16be_swap (sha1_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha1_update_utf16be_swap (sha1_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -626,7 +626,7 @@ void sha1_update_utf16be_swap (sha1_ctx_t *ctx, const u32 *w, const int len)
   sha1_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha1_update_global (sha1_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha1_update_global (sha1_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -678,7 +678,7 @@ void sha1_update_global (sha1_ctx_t *ctx, const __global u32 *w, const int len)
   sha1_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha1_update_global_swap (sha1_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha1_update_global_swap (sha1_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -764,7 +764,7 @@ void sha1_update_global_swap (sha1_ctx_t *ctx, const __global u32 *w, const int
   sha1_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha1_update_global_utf16le (sha1_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha1_update_global_utf16le (sha1_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -806,7 +806,7 @@ void sha1_update_global_utf16le (sha1_ctx_t *ctx, const __global u32 *w, const i
   sha1_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha1_update_global_utf16le_swap (sha1_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha1_update_global_utf16le_swap (sha1_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -882,7 +882,7 @@ void sha1_update_global_utf16le_swap (sha1_ctx_t *ctx, const __global u32 *w, co
   sha1_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha1_update_global_utf16be (sha1_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha1_update_global_utf16be (sha1_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -924,7 +924,7 @@ void sha1_update_global_utf16be (sha1_ctx_t *ctx, const __global u32 *w, const i
   sha1_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha1_update_global_utf16be_swap (sha1_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha1_update_global_utf16be_swap (sha1_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1000,7 +1000,7 @@ void sha1_update_global_utf16be_swap (sha1_ctx_t *ctx, const __global u32 *w, co
   sha1_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha1_final (sha1_ctx_t *ctx)
+DECLSPEC void sha1_final (sha1_ctx_t *ctx)
 {
   const int pos = ctx->len & 63;
 
@@ -1043,7 +1043,7 @@ typedef struct sha1_hmac_ctx
 
 } sha1_hmac_ctx_t;
 
-void sha1_hmac_init_64 (sha1_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4])
+DECLSPEC void sha1_hmac_init_64 (sha1_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4])
 {
   u32 t0[4];
   u32 t1[4];
@@ -1097,7 +1097,7 @@ void sha1_hmac_init_64 (sha1_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4],
   sha1_update_64 (&ctx->opad, t0, t1, t2, t3, 64);
 }
 
-void sha1_hmac_init (sha1_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha1_hmac_init (sha1_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1154,7 +1154,7 @@ void sha1_hmac_init (sha1_hmac_ctx_t *ctx, const u32 *w, const int len)
   sha1_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void sha1_hmac_init_swap (sha1_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha1_hmac_init_swap (sha1_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1211,7 +1211,7 @@ void sha1_hmac_init_swap (sha1_hmac_ctx_t *ctx, const u32 *w, const int len)
   sha1_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void sha1_hmac_init_global (sha1_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void sha1_hmac_init_global (sha1_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1268,7 +1268,7 @@ void sha1_hmac_init_global (sha1_hmac_ctx_t *ctx, __global const u32 *w, const i
   sha1_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void sha1_hmac_init_global_swap (sha1_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void sha1_hmac_init_global_swap (sha1_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1325,52 +1325,52 @@ void sha1_hmac_init_global_swap (sha1_hmac_ctx_t *ctx, __global const u32 *w, co
   sha1_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void sha1_hmac_update_64 (sha1_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void sha1_hmac_update_64 (sha1_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   sha1_update_64 (&ctx->ipad, w0, w1, w2, w3, len);
 }
 
-void sha1_hmac_update (sha1_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha1_hmac_update (sha1_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha1_update (&ctx->ipad, w, len);
 }
 
-void sha1_hmac_update_swap (sha1_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha1_hmac_update_swap (sha1_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha1_update_swap (&ctx->ipad, w, len);
 }
 
-void sha1_hmac_update_utf16le (sha1_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha1_hmac_update_utf16le (sha1_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha1_update_utf16le (&ctx->ipad, w, len);
 }
 
-void sha1_hmac_update_utf16le_swap (sha1_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha1_hmac_update_utf16le_swap (sha1_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha1_update_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void sha1_hmac_update_global (sha1_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha1_hmac_update_global (sha1_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha1_update_global (&ctx->ipad, w, len);
 }
 
-void sha1_hmac_update_global_swap (sha1_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha1_hmac_update_global_swap (sha1_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha1_update_global_swap (&ctx->ipad, w, len);
 }
 
-void sha1_hmac_update_global_utf16le (sha1_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha1_hmac_update_global_utf16le (sha1_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha1_update_global_utf16le (&ctx->ipad, w, len);
 }
 
-void sha1_hmac_update_global_utf16le_swap (sha1_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha1_hmac_update_global_utf16le_swap (sha1_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha1_update_global_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void sha1_hmac_final (sha1_hmac_ctx_t *ctx)
+DECLSPEC void sha1_hmac_final (sha1_hmac_ctx_t *ctx)
 {
   sha1_final (&ctx->ipad);
 
@@ -1416,7 +1416,7 @@ typedef struct sha1_ctx_vector
 
 } sha1_ctx_vector_t;
 
-void sha1_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[5])
+DECLSPEC void sha1_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[5])
 {
   u32x a = digest[0];
   u32x b = digest[1];
@@ -1545,7 +1545,7 @@ void sha1_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4]
   digest[4] += e;
 }
 
-void sha1_init_vector (sha1_ctx_vector_t *ctx)
+DECLSPEC void sha1_init_vector (sha1_ctx_vector_t *ctx)
 {
   ctx->h[0] = SHA1M_A;
   ctx->h[1] = SHA1M_B;
@@ -1573,7 +1573,7 @@ void sha1_init_vector (sha1_ctx_vector_t *ctx)
   ctx->len = 0;
 }
 
-void sha1_init_vector_from_scalar (sha1_ctx_vector_t *ctx, sha1_ctx_t *ctx0)
+DECLSPEC void sha1_init_vector_from_scalar (sha1_ctx_vector_t *ctx, sha1_ctx_t *ctx0)
 {
   ctx->h[0] = ctx0->h[0];
   ctx->h[1] = ctx0->h[1];
@@ -1601,7 +1601,7 @@ void sha1_init_vector_from_scalar (sha1_ctx_vector_t *ctx, sha1_ctx_t *ctx0)
   ctx->len = ctx0->len;
 }
 
-void sha1_update_vector_64 (sha1_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
+DECLSPEC void sha1_update_vector_64 (sha1_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -1613,7 +1613,7 @@ void sha1_update_vector_64 (sha1_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x
 
   if ((pos + len) < 64)
   {
-    switch_buffer_by_offset_be (w0, w1, w2, w3, pos);
+    //switch_buffer_by_offset_be (w0, w1, w2, w3, pos);
 
     ctx->w0[0] |= w0[0];
     ctx->w0[1] |= w0[1];
@@ -1639,7 +1639,7 @@ void sha1_update_vector_64 (sha1_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x
     u32x c2[4] = { 0 };
     u32x c3[4] = { 0 };
 
-    switch_buffer_by_offset_carry_be (w0, w1, w2, w3, c0, c1, c2, c3, pos);
+    //switch_buffer_by_offset_carry_be (w0, w1, w2, w3, c0, c1, c2, c3, pos);
 
     ctx->w0[0] |= w0[0];
     ctx->w0[1] |= w0[1];
@@ -1679,7 +1679,7 @@ void sha1_update_vector_64 (sha1_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x
   }
 }
 
-void sha1_update_vector (sha1_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha1_update_vector (sha1_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1731,7 +1731,7 @@ void sha1_update_vector (sha1_ctx_vector_t *ctx, const u32x *w, const int len)
   sha1_update_vector_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha1_update_vector_swap (sha1_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha1_update_vector_swap (sha1_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1817,7 +1817,7 @@ void sha1_update_vector_swap (sha1_ctx_vector_t *ctx, const u32x *w, const int l
   sha1_update_vector_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha1_update_vector_utf16le (sha1_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha1_update_vector_utf16le (sha1_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1859,7 +1859,7 @@ void sha1_update_vector_utf16le (sha1_ctx_vector_t *ctx, const u32x *w, const in
   sha1_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha1_update_vector_utf16le_swap (sha1_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha1_update_vector_utf16le_swap (sha1_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1935,7 +1935,7 @@ void sha1_update_vector_utf16le_swap (sha1_ctx_vector_t *ctx, const u32x *w, con
   sha1_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha1_update_vector_utf16leN (sha1_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha1_update_vector_utf16leN (sha1_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1977,7 +1977,7 @@ void sha1_update_vector_utf16leN (sha1_ctx_vector_t *ctx, const u32x *w, const i
   sha1_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha1_update_vector_utf16beN (sha1_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha1_update_vector_utf16beN (sha1_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2019,7 +2019,7 @@ void sha1_update_vector_utf16beN (sha1_ctx_vector_t *ctx, const u32x *w, const i
   sha1_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha1_final_vector (sha1_ctx_vector_t *ctx)
+DECLSPEC void sha1_final_vector (sha1_ctx_vector_t *ctx)
 {
   const int pos = ctx->len & 63;
 
@@ -2062,7 +2062,7 @@ typedef struct sha1_hmac_ctx_vector
 
 } sha1_hmac_ctx_vector_t;
 
-void sha1_hmac_init_vector_64 (sha1_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4])
+DECLSPEC void sha1_hmac_init_vector_64 (sha1_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4])
 {
   u32x t0[4];
   u32x t1[4];
@@ -2116,7 +2116,7 @@ void sha1_hmac_init_vector_64 (sha1_hmac_ctx_vector_t *ctx, const u32x w0[4], co
   sha1_update_vector_64 (&ctx->opad, t0, t1, t2, t3, 64);
 }
 
-void sha1_hmac_init_vector (sha1_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha1_hmac_init_vector (sha1_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2173,17 +2173,17 @@ void sha1_hmac_init_vector (sha1_hmac_ctx_vector_t *ctx, const u32x *w, const in
   sha1_hmac_init_vector_64 (ctx, w0, w1, w2, w3);
 }
 
-void sha1_hmac_update_vector_64 (sha1_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
+DECLSPEC void sha1_hmac_update_vector_64 (sha1_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
 {
   sha1_update_vector_64 (&ctx->ipad, w0, w1, w2, w3, len);
 }
 
-void sha1_hmac_update_vector (sha1_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha1_hmac_update_vector (sha1_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   sha1_update_vector (&ctx->ipad, w, len);
 }
 
-void sha1_hmac_final_vector (sha1_hmac_ctx_vector_t *ctx)
+DECLSPEC void sha1_hmac_final_vector (sha1_hmac_ctx_vector_t *ctx)
 {
   sha1_final_vector (&ctx->ipad);
 
diff --git a/OpenCL/inc_hash_sha224.cl b/OpenCL/inc_hash_sha224.cl
index 51436aa28..dd3a81e6e 100644
--- a/OpenCL/inc_hash_sha224.cl
+++ b/OpenCL/inc_hash_sha224.cl
@@ -37,7 +37,7 @@ typedef struct sha224_ctx
 
 } sha224_ctx_t;
 
-void sha224_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[8])
+DECLSPEC void sha224_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[8])
 {
   u32 a = digest[0];
   u32 b = digest[1];
@@ -128,7 +128,7 @@ void sha224_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const
   digest[7] += h;
 }
 
-void sha224_init (sha224_ctx_t *ctx)
+DECLSPEC void sha224_init (sha224_ctx_t *ctx)
 {
   ctx->h[0] = SHA224M_A;
   ctx->h[1] = SHA224M_B;
@@ -159,7 +159,7 @@ void sha224_init (sha224_ctx_t *ctx)
   ctx->len = 0;
 }
 
-void sha224_update_64 (sha224_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void sha224_update_64 (sha224_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -237,7 +237,7 @@ void sha224_update_64 (sha224_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w
   }
 }
 
-void sha224_update (sha224_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha224_update (sha224_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -289,7 +289,7 @@ void sha224_update (sha224_ctx_t *ctx, const u32 *w, const int len)
   sha224_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha224_update_swap (sha224_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha224_update_swap (sha224_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -375,7 +375,7 @@ void sha224_update_swap (sha224_ctx_t *ctx, const u32 *w, const int len)
   sha224_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha224_update_utf16le (sha224_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha224_update_utf16le (sha224_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -417,7 +417,7 @@ void sha224_update_utf16le (sha224_ctx_t *ctx, const u32 *w, const int len)
   sha224_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha224_update_utf16le_swap (sha224_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha224_update_utf16le_swap (sha224_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -493,7 +493,7 @@ void sha224_update_utf16le_swap (sha224_ctx_t *ctx, const u32 *w, const int len)
   sha224_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha224_update_global (sha224_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha224_update_global (sha224_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -545,7 +545,7 @@ void sha224_update_global (sha224_ctx_t *ctx, const __global u32 *w, const int l
   sha224_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha224_update_global_swap (sha224_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha224_update_global_swap (sha224_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -631,7 +631,7 @@ void sha224_update_global_swap (sha224_ctx_t *ctx, const __global u32 *w, const
   sha224_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha224_update_global_utf16le (sha224_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha224_update_global_utf16le (sha224_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -673,7 +673,7 @@ void sha224_update_global_utf16le (sha224_ctx_t *ctx, const __global u32 *w, con
   sha224_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha224_update_global_utf16le_swap (sha224_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha224_update_global_utf16le_swap (sha224_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -749,7 +749,7 @@ void sha224_update_global_utf16le_swap (sha224_ctx_t *ctx, const __global u32 *w
   sha224_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha224_final (sha224_ctx_t *ctx)
+DECLSPEC void sha224_final (sha224_ctx_t *ctx)
 {
   const int pos = ctx->len & 63;
 
@@ -792,7 +792,7 @@ typedef struct sha224_hmac_ctx
 
 } sha224_hmac_ctx_t;
 
-void sha224_hmac_init_64 (sha224_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4])
+DECLSPEC void sha224_hmac_init_64 (sha224_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4])
 {
   u32 t0[4];
   u32 t1[4];
@@ -846,7 +846,7 @@ void sha224_hmac_init_64 (sha224_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[
   sha224_update_64 (&ctx->opad, t0, t1, t2, t3, 64);
 }
 
-void sha224_hmac_init (sha224_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha224_hmac_init (sha224_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -903,7 +903,7 @@ void sha224_hmac_init (sha224_hmac_ctx_t *ctx, const u32 *w, const int len)
   sha224_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void sha224_hmac_init_swap (sha224_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha224_hmac_init_swap (sha224_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -960,7 +960,7 @@ void sha224_hmac_init_swap (sha224_hmac_ctx_t *ctx, const u32 *w, const int len)
   sha224_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void sha224_hmac_init_global (sha224_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void sha224_hmac_init_global (sha224_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1017,7 +1017,7 @@ void sha224_hmac_init_global (sha224_hmac_ctx_t *ctx, __global const u32 *w, con
   sha224_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void sha224_hmac_init_global_swap (sha224_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void sha224_hmac_init_global_swap (sha224_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1074,52 +1074,52 @@ void sha224_hmac_init_global_swap (sha224_hmac_ctx_t *ctx, __global const u32 *w
   sha224_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void sha224_hmac_update_64 (sha224_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void sha224_hmac_update_64 (sha224_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   sha224_update_64 (&ctx->ipad, w0, w1, w2, w3, len);
 }
 
-void sha224_hmac_update (sha224_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha224_hmac_update (sha224_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha224_update (&ctx->ipad, w, len);
 }
 
-void sha224_hmac_update_swap (sha224_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha224_hmac_update_swap (sha224_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha224_update_swap (&ctx->ipad, w, len);
 }
 
-void sha224_hmac_update_utf16le (sha224_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha224_hmac_update_utf16le (sha224_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha224_update_utf16le (&ctx->ipad, w, len);
 }
 
-void sha224_hmac_update_utf16le_swap (sha224_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha224_hmac_update_utf16le_swap (sha224_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha224_update_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void sha224_hmac_update_global (sha224_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha224_hmac_update_global (sha224_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha224_update_global (&ctx->ipad, w, len);
 }
 
-void sha224_hmac_update_global_swap (sha224_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha224_hmac_update_global_swap (sha224_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha224_update_global_swap (&ctx->ipad, w, len);
 }
 
-void sha224_hmac_update_global_utf16le (sha224_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha224_hmac_update_global_utf16le (sha224_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha224_update_global_utf16le (&ctx->ipad, w, len);
 }
 
-void sha224_hmac_update_global_utf16le_swap (sha224_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha224_hmac_update_global_utf16le_swap (sha224_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha224_update_global_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void sha224_hmac_final (sha224_hmac_ctx_t *ctx)
+DECLSPEC void sha224_hmac_final (sha224_hmac_ctx_t *ctx)
 {
   sha224_final (&ctx->ipad);
 
@@ -1165,7 +1165,7 @@ typedef struct sha224_ctx_vector
 
 } sha224_ctx_vector_t;
 
-void sha224_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[8])
+DECLSPEC void sha224_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[8])
 {
   u32x a = digest[0];
   u32x b = digest[1];
@@ -1256,7 +1256,7 @@ void sha224_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[
   digest[7] += h;
 }
 
-void sha224_init_vector (sha224_ctx_vector_t *ctx)
+DECLSPEC void sha224_init_vector (sha224_ctx_vector_t *ctx)
 {
   ctx->h[0] = SHA224M_A;
   ctx->h[1] = SHA224M_B;
@@ -1287,7 +1287,7 @@ void sha224_init_vector (sha224_ctx_vector_t *ctx)
   ctx->len = 0;
 }
 
-void sha224_init_vector_from_scalar (sha224_ctx_vector_t *ctx, sha224_ctx_t *ctx0)
+DECLSPEC void sha224_init_vector_from_scalar (sha224_ctx_vector_t *ctx, sha224_ctx_t *ctx0)
 {
   ctx->h[0] = ctx0->h[0];
   ctx->h[1] = ctx0->h[1];
@@ -1318,7 +1318,7 @@ void sha224_init_vector_from_scalar (sha224_ctx_vector_t *ctx, sha224_ctx_t *ctx
   ctx->len = ctx0->len;
 }
 
-void sha224_update_vector_64 (sha224_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
+DECLSPEC void sha224_update_vector_64 (sha224_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -1396,7 +1396,7 @@ void sha224_update_vector_64 (sha224_ctx_vector_t *ctx, u32x w0[4], u32x w1[4],
   }
 }
 
-void sha224_update_vector (sha224_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha224_update_vector (sha224_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1448,7 +1448,7 @@ void sha224_update_vector (sha224_ctx_vector_t *ctx, const u32x *w, const int le
   sha224_update_vector_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha224_update_vector_swap (sha224_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha224_update_vector_swap (sha224_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1534,7 +1534,7 @@ void sha224_update_vector_swap (sha224_ctx_vector_t *ctx, const u32x *w, const i
   sha224_update_vector_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha224_update_vector_utf16le (sha224_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha224_update_vector_utf16le (sha224_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1576,7 +1576,7 @@ void sha224_update_vector_utf16le (sha224_ctx_vector_t *ctx, const u32x *w, cons
   sha224_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha224_update_vector_utf16le_swap (sha224_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha224_update_vector_utf16le_swap (sha224_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1652,7 +1652,7 @@ void sha224_update_vector_utf16le_swap (sha224_ctx_vector_t *ctx, const u32x *w,
   sha224_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha224_update_vector_utf16beN (sha224_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha224_update_vector_utf16beN (sha224_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1694,7 +1694,7 @@ void sha224_update_vector_utf16beN (sha224_ctx_vector_t *ctx, const u32x *w, con
   sha224_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha224_final_vector (sha224_ctx_vector_t *ctx)
+DECLSPEC void sha224_final_vector (sha224_ctx_vector_t *ctx)
 {
   const int pos = ctx->len & 63;
 
@@ -1737,7 +1737,7 @@ typedef struct sha224_hmac_ctx_vector
 
 } sha224_hmac_ctx_vector_t;
 
-void sha224_hmac_init_vector_64 (sha224_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4])
+DECLSPEC void sha224_hmac_init_vector_64 (sha224_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4])
 {
   u32x t0[4];
   u32x t1[4];
@@ -1791,7 +1791,7 @@ void sha224_hmac_init_vector_64 (sha224_hmac_ctx_vector_t *ctx, const u32x w0[4]
   sha224_update_vector_64 (&ctx->opad, t0, t1, t2, t3, 64);
 }
 
-void sha224_hmac_init_vector (sha224_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha224_hmac_init_vector (sha224_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1848,17 +1848,17 @@ void sha224_hmac_init_vector (sha224_hmac_ctx_vector_t *ctx, const u32x *w, cons
   sha224_hmac_init_vector_64 (ctx, w0, w1, w2, w3);
 }
 
-void sha224_hmac_update_vector_64 (sha224_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
+DECLSPEC void sha224_hmac_update_vector_64 (sha224_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
 {
   sha224_update_vector_64 (&ctx->ipad, w0, w1, w2, w3, len);
 }
 
-void sha224_hmac_update_vector (sha224_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha224_hmac_update_vector (sha224_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   sha224_update_vector (&ctx->ipad, w, len);
 }
 
-void sha224_hmac_final_vector (sha224_hmac_ctx_vector_t *ctx)
+DECLSPEC void sha224_hmac_final_vector (sha224_hmac_ctx_vector_t *ctx)
 {
   sha224_final_vector (&ctx->ipad);
 
diff --git a/OpenCL/inc_hash_sha256.cl b/OpenCL/inc_hash_sha256.cl
index 0ae1e5b75..66387626b 100644
--- a/OpenCL/inc_hash_sha256.cl
+++ b/OpenCL/inc_hash_sha256.cl
@@ -37,7 +37,7 @@ typedef struct sha256_ctx
 
 } sha256_ctx_t;
 
-void sha256_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[8])
+DECLSPEC void sha256_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[8])
 {
   u32 a = digest[0];
   u32 b = digest[1];
@@ -128,7 +128,7 @@ void sha256_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const
   digest[7] += h;
 }
 
-void sha256_init (sha256_ctx_t *ctx)
+DECLSPEC void sha256_init (sha256_ctx_t *ctx)
 {
   ctx->h[0] = SHA256M_A;
   ctx->h[1] = SHA256M_B;
@@ -159,7 +159,7 @@ void sha256_init (sha256_ctx_t *ctx)
   ctx->len = 0;
 }
 
-void sha256_update_64 (sha256_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void sha256_update_64 (sha256_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -237,7 +237,7 @@ void sha256_update_64 (sha256_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w
   }
 }
 
-void sha256_update (sha256_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha256_update (sha256_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -289,7 +289,7 @@ void sha256_update (sha256_ctx_t *ctx, const u32 *w, const int len)
   sha256_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha256_update_swap (sha256_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha256_update_swap (sha256_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -375,7 +375,7 @@ void sha256_update_swap (sha256_ctx_t *ctx, const u32 *w, const int len)
   sha256_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha256_update_utf16le (sha256_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha256_update_utf16le (sha256_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -417,7 +417,7 @@ void sha256_update_utf16le (sha256_ctx_t *ctx, const u32 *w, const int len)
   sha256_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha256_update_utf16le_swap (sha256_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha256_update_utf16le_swap (sha256_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -493,7 +493,7 @@ void sha256_update_utf16le_swap (sha256_ctx_t *ctx, const u32 *w, const int len)
   sha256_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha256_update_global (sha256_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha256_update_global (sha256_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -545,7 +545,7 @@ void sha256_update_global (sha256_ctx_t *ctx, const __global u32 *w, const int l
   sha256_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha256_update_global_swap (sha256_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha256_update_global_swap (sha256_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -631,7 +631,7 @@ void sha256_update_global_swap (sha256_ctx_t *ctx, const __global u32 *w, const
   sha256_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha256_update_global_utf16le (sha256_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha256_update_global_utf16le (sha256_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -673,7 +673,7 @@ void sha256_update_global_utf16le (sha256_ctx_t *ctx, const __global u32 *w, con
   sha256_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha256_update_global_utf16le_swap (sha256_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha256_update_global_utf16le_swap (sha256_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -749,7 +749,7 @@ void sha256_update_global_utf16le_swap (sha256_ctx_t *ctx, const __global u32 *w
   sha256_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha256_final (sha256_ctx_t *ctx)
+DECLSPEC void sha256_final (sha256_ctx_t *ctx)
 {
   const int pos = ctx->len & 63;
 
@@ -792,7 +792,7 @@ typedef struct sha256_hmac_ctx
 
 } sha256_hmac_ctx_t;
 
-void sha256_hmac_init_64 (sha256_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4])
+DECLSPEC void sha256_hmac_init_64 (sha256_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4])
 {
   u32 t0[4];
   u32 t1[4];
@@ -846,7 +846,7 @@ void sha256_hmac_init_64 (sha256_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[
   sha256_update_64 (&ctx->opad, t0, t1, t2, t3, 64);
 }
 
-void sha256_hmac_init (sha256_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha256_hmac_init (sha256_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -903,7 +903,7 @@ void sha256_hmac_init (sha256_hmac_ctx_t *ctx, const u32 *w, const int len)
   sha256_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void sha256_hmac_init_swap (sha256_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha256_hmac_init_swap (sha256_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -960,7 +960,7 @@ void sha256_hmac_init_swap (sha256_hmac_ctx_t *ctx, const u32 *w, const int len)
   sha256_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void sha256_hmac_init_global (sha256_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void sha256_hmac_init_global (sha256_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1017,7 +1017,7 @@ void sha256_hmac_init_global (sha256_hmac_ctx_t *ctx, __global const u32 *w, con
   sha256_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void sha256_hmac_init_global_swap (sha256_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void sha256_hmac_init_global_swap (sha256_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1074,52 +1074,52 @@ void sha256_hmac_init_global_swap (sha256_hmac_ctx_t *ctx, __global const u32 *w
   sha256_hmac_init_64 (ctx, w0, w1, w2, w3);
 }
 
-void sha256_hmac_update_64 (sha256_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void sha256_hmac_update_64 (sha256_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   sha256_update_64 (&ctx->ipad, w0, w1, w2, w3, len);
 }
 
-void sha256_hmac_update (sha256_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha256_hmac_update (sha256_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha256_update (&ctx->ipad, w, len);
 }
 
-void sha256_hmac_update_swap (sha256_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha256_hmac_update_swap (sha256_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha256_update_swap (&ctx->ipad, w, len);
 }
 
-void sha256_hmac_update_utf16le (sha256_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha256_hmac_update_utf16le (sha256_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha256_update_utf16le (&ctx->ipad, w, len);
 }
 
-void sha256_hmac_update_utf16le_swap (sha256_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha256_hmac_update_utf16le_swap (sha256_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha256_update_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void sha256_hmac_update_global (sha256_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha256_hmac_update_global (sha256_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha256_update_global (&ctx->ipad, w, len);
 }
 
-void sha256_hmac_update_global_swap (sha256_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha256_hmac_update_global_swap (sha256_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha256_update_global_swap (&ctx->ipad, w, len);
 }
 
-void sha256_hmac_update_global_utf16le (sha256_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha256_hmac_update_global_utf16le (sha256_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha256_update_global_utf16le (&ctx->ipad, w, len);
 }
 
-void sha256_hmac_update_global_utf16le_swap (sha256_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha256_hmac_update_global_utf16le_swap (sha256_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha256_update_global_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void sha256_hmac_final (sha256_hmac_ctx_t *ctx)
+DECLSPEC void sha256_hmac_final (sha256_hmac_ctx_t *ctx)
 {
   sha256_final (&ctx->ipad);
 
@@ -1165,7 +1165,7 @@ typedef struct sha256_ctx_vector
 
 } sha256_ctx_vector_t;
 
-void sha256_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[8])
+DECLSPEC void sha256_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[8])
 {
   u32x a = digest[0];
   u32x b = digest[1];
@@ -1256,7 +1256,7 @@ void sha256_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[
   digest[7] += h;
 }
 
-void sha256_init_vector (sha256_ctx_vector_t *ctx)
+DECLSPEC void sha256_init_vector (sha256_ctx_vector_t *ctx)
 {
   ctx->h[0] = SHA256M_A;
   ctx->h[1] = SHA256M_B;
@@ -1287,7 +1287,7 @@ void sha256_init_vector (sha256_ctx_vector_t *ctx)
   ctx->len = 0;
 }
 
-void sha256_init_vector_from_scalar (sha256_ctx_vector_t *ctx, sha256_ctx_t *ctx0)
+DECLSPEC void sha256_init_vector_from_scalar (sha256_ctx_vector_t *ctx, sha256_ctx_t *ctx0)
 {
   ctx->h[0] = ctx0->h[0];
   ctx->h[1] = ctx0->h[1];
@@ -1318,7 +1318,7 @@ void sha256_init_vector_from_scalar (sha256_ctx_vector_t *ctx, sha256_ctx_t *ctx
   ctx->len = ctx0->len;
 }
 
-void sha256_update_vector_64 (sha256_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
+DECLSPEC void sha256_update_vector_64 (sha256_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -1396,7 +1396,7 @@ void sha256_update_vector_64 (sha256_ctx_vector_t *ctx, u32x w0[4], u32x w1[4],
   }
 }
 
-void sha256_update_vector (sha256_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha256_update_vector (sha256_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1448,7 +1448,7 @@ void sha256_update_vector (sha256_ctx_vector_t *ctx, const u32x *w, const int le
   sha256_update_vector_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha256_update_vector_swap (sha256_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha256_update_vector_swap (sha256_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1534,7 +1534,7 @@ void sha256_update_vector_swap (sha256_ctx_vector_t *ctx, const u32x *w, const i
   sha256_update_vector_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void sha256_update_vector_utf16le (sha256_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha256_update_vector_utf16le (sha256_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1576,7 +1576,7 @@ void sha256_update_vector_utf16le (sha256_ctx_vector_t *ctx, const u32x *w, cons
   sha256_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha256_update_vector_utf16le_swap (sha256_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha256_update_vector_utf16le_swap (sha256_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1652,7 +1652,7 @@ void sha256_update_vector_utf16le_swap (sha256_ctx_vector_t *ctx, const u32x *w,
   sha256_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha256_update_vector_utf16beN (sha256_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha256_update_vector_utf16beN (sha256_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1694,7 +1694,7 @@ void sha256_update_vector_utf16beN (sha256_ctx_vector_t *ctx, const u32x *w, con
   sha256_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void sha256_final_vector (sha256_ctx_vector_t *ctx)
+DECLSPEC void sha256_final_vector (sha256_ctx_vector_t *ctx)
 {
   const int pos = ctx->len & 63;
 
@@ -1737,7 +1737,7 @@ typedef struct sha256_hmac_ctx_vector
 
 } sha256_hmac_ctx_vector_t;
 
-void sha256_hmac_init_vector_64 (sha256_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4])
+DECLSPEC void sha256_hmac_init_vector_64 (sha256_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4])
 {
   u32x t0[4];
   u32x t1[4];
@@ -1791,7 +1791,7 @@ void sha256_hmac_init_vector_64 (sha256_hmac_ctx_vector_t *ctx, const u32x w0[4]
   sha256_update_vector_64 (&ctx->opad, t0, t1, t2, t3, 64);
 }
 
-void sha256_hmac_init_vector (sha256_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha256_hmac_init_vector (sha256_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -1848,17 +1848,17 @@ void sha256_hmac_init_vector (sha256_hmac_ctx_vector_t *ctx, const u32x *w, cons
   sha256_hmac_init_vector_64 (ctx, w0, w1, w2, w3);
 }
 
-void sha256_hmac_update_vector_64 (sha256_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
+DECLSPEC void sha256_hmac_update_vector_64 (sha256_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
 {
   sha256_update_vector_64 (&ctx->ipad, w0, w1, w2, w3, len);
 }
 
-void sha256_hmac_update_vector (sha256_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha256_hmac_update_vector (sha256_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   sha256_update_vector (&ctx->ipad, w, len);
 }
 
-void sha256_hmac_final_vector (sha256_hmac_ctx_vector_t *ctx)
+DECLSPEC void sha256_hmac_final_vector (sha256_hmac_ctx_vector_t *ctx)
 {
   sha256_final_vector (&ctx->ipad);
 
diff --git a/OpenCL/inc_hash_sha384.cl b/OpenCL/inc_hash_sha384.cl
index c5471c011..bc63dfce3 100644
--- a/OpenCL/inc_hash_sha384.cl
+++ b/OpenCL/inc_hash_sha384.cl
@@ -45,7 +45,7 @@ typedef struct sha384_ctx
 
 } sha384_ctx_t;
 
-void sha384_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], const u32 w4[4], const u32 w5[4], const u32 w6[4], const u32 w7[4], u64 digest[8])
+DECLSPEC void sha384_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], const u32 w4[4], const u32 w5[4], const u32 w6[4], const u32 w7[4], u64 digest[8])
 {
   u64 a = digest[0];
   u64 b = digest[1];
@@ -136,7 +136,7 @@ void sha384_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const
   digest[7] += h;
 }
 
-void sha384_init (sha384_ctx_t *ctx)
+DECLSPEC void sha384_init (sha384_ctx_t *ctx)
 {
   ctx->h[0] = SHA384M_A;
   ctx->h[1] = SHA384M_B;
@@ -183,7 +183,7 @@ void sha384_init (sha384_ctx_t *ctx)
   ctx->len = 0;
 }
 
-void sha384_update_128 (sha384_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const int len)
+DECLSPEC void sha384_update_128 (sha384_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 127;
@@ -313,7 +313,7 @@ void sha384_update_128 (sha384_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32
   }
 }
 
-void sha384_update (sha384_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha384_update (sha384_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -401,7 +401,7 @@ void sha384_update (sha384_ctx_t *ctx, const u32 *w, const int len)
   sha384_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, len - pos1);
 }
 
-void sha384_update_swap (sha384_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha384_update_swap (sha384_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -555,7 +555,7 @@ void sha384_update_swap (sha384_ctx_t *ctx, const u32 *w, const int len)
   sha384_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, len - pos1);
 }
 
-void sha384_update_utf16le (sha384_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha384_update_utf16le (sha384_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -621,7 +621,7 @@ void sha384_update_utf16le (sha384_ctx_t *ctx, const u32 *w, const int len)
   sha384_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, (len - pos1) * 2);
 }
 
-void sha384_update_utf16le_swap (sha384_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha384_update_utf16le_swap (sha384_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -753,7 +753,7 @@ void sha384_update_utf16le_swap (sha384_ctx_t *ctx, const u32 *w, const int len)
   sha384_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, (len - pos1) * 2);
 }
 
-void sha384_update_global (sha384_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha384_update_global (sha384_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -841,7 +841,7 @@ void sha384_update_global (sha384_ctx_t *ctx, const __global u32 *w, const int l
   sha384_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, len - pos1);
 }
 
-void sha384_update_global_swap (sha384_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha384_update_global_swap (sha384_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -995,7 +995,7 @@ void sha384_update_global_swap (sha384_ctx_t *ctx, const __global u32 *w, const
   sha384_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, len - pos1);
 }
 
-void sha384_update_global_utf16le (sha384_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha384_update_global_utf16le (sha384_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1061,7 +1061,7 @@ void sha384_update_global_utf16le (sha384_ctx_t *ctx, const __global u32 *w, con
   sha384_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, (len - pos1) * 2);
 }
 
-void sha384_update_global_utf16le_swap (sha384_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha384_update_global_utf16le_swap (sha384_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1193,7 +1193,7 @@ void sha384_update_global_utf16le_swap (sha384_ctx_t *ctx, const __global u32 *w
   sha384_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, (len - pos1) * 2);
 }
 
-void sha384_final (sha384_ctx_t *ctx)
+DECLSPEC void sha384_final (sha384_ctx_t *ctx)
 {
   const int pos = ctx->len & 127;
 
@@ -1252,7 +1252,7 @@ typedef struct sha384_hmac_ctx
 
 } sha384_hmac_ctx_t;
 
-void sha384_hmac_init_128 (sha384_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], const u32 w4[4], const u32 w5[4], const u32 w6[4], const u32 w7[4])
+DECLSPEC void sha384_hmac_init_128 (sha384_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], const u32 w4[4], const u32 w5[4], const u32 w6[4], const u32 w7[4])
 {
   u32 t0[4];
   u32 t1[4];
@@ -1342,7 +1342,7 @@ void sha384_hmac_init_128 (sha384_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1
   sha384_update_128 (&ctx->opad, t0, t1, t2, t3, t4, t5, t6, t7, 128);
 }
 
-void sha384_hmac_init (sha384_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha384_hmac_init (sha384_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1435,7 +1435,7 @@ void sha384_hmac_init (sha384_hmac_ctx_t *ctx, const u32 *w, const int len)
   sha384_hmac_init_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7);
 }
 
-void sha384_hmac_init_swap (sha384_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha384_hmac_init_swap (sha384_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1528,7 +1528,7 @@ void sha384_hmac_init_swap (sha384_hmac_ctx_t *ctx, const u32 *w, const int len)
   sha384_hmac_init_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7);
 }
 
-void sha384_hmac_init_global (sha384_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void sha384_hmac_init_global (sha384_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1621,7 +1621,7 @@ void sha384_hmac_init_global (sha384_hmac_ctx_t *ctx, __global const u32 *w, con
   sha384_hmac_init_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7);
 }
 
-void sha384_hmac_init_global_swap (sha384_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void sha384_hmac_init_global_swap (sha384_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1714,52 +1714,52 @@ void sha384_hmac_init_global_swap (sha384_hmac_ctx_t *ctx, __global const u32 *w
   sha384_hmac_init_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7);
 }
 
-void sha384_hmac_update_128 (sha384_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const int len)
+DECLSPEC void sha384_hmac_update_128 (sha384_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const int len)
 {
   sha384_update_128 (&ctx->ipad, w0, w1, w2, w3, w4, w5, w6, w7, len);
 }
 
-void sha384_hmac_update (sha384_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha384_hmac_update (sha384_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha384_update (&ctx->ipad, w, len);
 }
 
-void sha384_hmac_update_swap (sha384_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha384_hmac_update_swap (sha384_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha384_update_swap (&ctx->ipad, w, len);
 }
 
-void sha384_hmac_update_utf16le (sha384_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha384_hmac_update_utf16le (sha384_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha384_update_utf16le (&ctx->ipad, w, len);
 }
 
-void sha384_hmac_update_utf16le_swap (sha384_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha384_hmac_update_utf16le_swap (sha384_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha384_update_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void sha384_hmac_update_global (sha384_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha384_hmac_update_global (sha384_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha384_update_global (&ctx->ipad, w, len);
 }
 
-void sha384_hmac_update_global_swap (sha384_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha384_hmac_update_global_swap (sha384_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha384_update_global_swap (&ctx->ipad, w, len);
 }
 
-void sha384_hmac_update_global_utf16le (sha384_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha384_hmac_update_global_utf16le (sha384_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha384_update_global_utf16le (&ctx->ipad, w, len);
 }
 
-void sha384_hmac_update_global_utf16le_swap (sha384_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha384_hmac_update_global_utf16le_swap (sha384_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha384_update_global_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void sha384_hmac_final (sha384_hmac_ctx_t *ctx)
+DECLSPEC void sha384_hmac_final (sha384_hmac_ctx_t *ctx)
 {
   sha384_final (&ctx->ipad);
 
@@ -1829,7 +1829,7 @@ typedef struct sha384_ctx_vector
 
 } sha384_ctx_vector_t;
 
-void sha384_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x w4[4], const u32x w5[4], const u32x w6[4], const u32x w7[4], u64x digest[8])
+DECLSPEC void sha384_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x w4[4], const u32x w5[4], const u32x w6[4], const u32x w7[4], u64x digest[8])
 {
   u64x a = digest[0];
   u64x b = digest[1];
@@ -1920,7 +1920,7 @@ void sha384_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[
   digest[7] += h;
 }
 
-void sha384_init_vector (sha384_ctx_vector_t *ctx)
+DECLSPEC void sha384_init_vector (sha384_ctx_vector_t *ctx)
 {
   ctx->h[0] = SHA384M_A;
   ctx->h[1] = SHA384M_B;
@@ -1967,7 +1967,7 @@ void sha384_init_vector (sha384_ctx_vector_t *ctx)
   ctx->len = 0;
 }
 
-void sha384_init_vector_from_scalar (sha384_ctx_vector_t *ctx, sha384_ctx_t *ctx0)
+DECLSPEC void sha384_init_vector_from_scalar (sha384_ctx_vector_t *ctx, sha384_ctx_t *ctx0)
 {
   ctx->h[0] = ctx0->h[0];
   ctx->h[1] = ctx0->h[1];
@@ -2014,7 +2014,7 @@ void sha384_init_vector_from_scalar (sha384_ctx_vector_t *ctx, sha384_ctx_t *ctx
   ctx->len = ctx0->len;
 }
 
-void sha384_update_vector_128 (sha384_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const int len)
+DECLSPEC void sha384_update_vector_128 (sha384_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 127;
@@ -2144,7 +2144,7 @@ void sha384_update_vector_128 (sha384_ctx_vector_t *ctx, u32x w0[4], u32x w1[4],
   }
 }
 
-void sha384_update_vector (sha384_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha384_update_vector (sha384_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2232,7 +2232,7 @@ void sha384_update_vector (sha384_ctx_vector_t *ctx, const u32x *w, const int le
   sha384_update_vector_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, len - pos1);
 }
 
-void sha384_update_vector_swap (sha384_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha384_update_vector_swap (sha384_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2386,7 +2386,7 @@ void sha384_update_vector_swap (sha384_ctx_vector_t *ctx, const u32x *w, const i
   sha384_update_vector_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, len - pos1);
 }
 
-void sha384_update_vector_utf16le (sha384_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha384_update_vector_utf16le (sha384_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2452,7 +2452,7 @@ void sha384_update_vector_utf16le (sha384_ctx_vector_t *ctx, const u32x *w, cons
   sha384_update_vector_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, (len - pos1) * 2);
 }
 
-void sha384_update_vector_utf16le_swap (sha384_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha384_update_vector_utf16le_swap (sha384_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2584,7 +2584,7 @@ void sha384_update_vector_utf16le_swap (sha384_ctx_vector_t *ctx, const u32x *w,
   sha384_update_vector_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, (len - pos1) * 2);
 }
 
-void sha384_update_vector_utf16beN (sha384_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha384_update_vector_utf16beN (sha384_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2650,7 +2650,7 @@ void sha384_update_vector_utf16beN (sha384_ctx_vector_t *ctx, const u32x *w, con
   sha384_update_vector_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, (len - pos1) * 2);
 }
 
-void sha384_final_vector (sha384_ctx_vector_t *ctx)
+DECLSPEC void sha384_final_vector (sha384_ctx_vector_t *ctx)
 {
   const int pos = ctx->len & 127;
 
@@ -2709,7 +2709,7 @@ typedef struct sha384_hmac_ctx_vector
 
 } sha384_hmac_ctx_vector_t;
 
-void sha384_hmac_init_vector_128 (sha384_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x w4[4], const u32x w5[4], const u32x w6[4], const u32x w7[4])
+DECLSPEC void sha384_hmac_init_vector_128 (sha384_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x w4[4], const u32x w5[4], const u32x w6[4], const u32x w7[4])
 {
   u32x t0[4];
   u32x t1[4];
@@ -2799,7 +2799,7 @@ void sha384_hmac_init_vector_128 (sha384_hmac_ctx_vector_t *ctx, const u32x w0[4
   sha384_update_vector_128 (&ctx->opad, t0, t1, t2, t3, t4, t5, t6, t7, 128);
 }
 
-void sha384_hmac_init_vector (sha384_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha384_hmac_init_vector (sha384_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2892,17 +2892,17 @@ void sha384_hmac_init_vector (sha384_hmac_ctx_vector_t *ctx, const u32x *w, cons
   sha384_hmac_init_vector_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7);
 }
 
-void sha384_hmac_update_vector_128 (sha384_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const int len)
+DECLSPEC void sha384_hmac_update_vector_128 (sha384_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const int len)
 {
   sha384_update_vector_128 (&ctx->ipad, w0, w1, w2, w3, w4, w5, w6, w7, len);
 }
 
-void sha384_hmac_update_vector (sha384_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha384_hmac_update_vector (sha384_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   sha384_update_vector (&ctx->ipad, w, len);
 }
 
-void sha384_hmac_final_vector (sha384_hmac_ctx_vector_t *ctx)
+DECLSPEC void sha384_hmac_final_vector (sha384_hmac_ctx_vector_t *ctx)
 {
   sha384_final_vector (&ctx->ipad);
 
diff --git a/OpenCL/inc_hash_sha512.cl b/OpenCL/inc_hash_sha512.cl
index d8f7d9ca8..9fd37c639 100644
--- a/OpenCL/inc_hash_sha512.cl
+++ b/OpenCL/inc_hash_sha512.cl
@@ -45,7 +45,7 @@ typedef struct sha512_ctx
 
 } sha512_ctx_t;
 
-void sha512_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], const u32 w4[4], const u32 w5[4], const u32 w6[4], const u32 w7[4], u64 digest[8])
+DECLSPEC void sha512_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], const u32 w4[4], const u32 w5[4], const u32 w6[4], const u32 w7[4], u64 digest[8])
 {
   u64 a = digest[0];
   u64 b = digest[1];
@@ -136,7 +136,7 @@ void sha512_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const
   digest[7] += h;
 }
 
-void sha512_init (sha512_ctx_t *ctx)
+DECLSPEC void sha512_init (sha512_ctx_t *ctx)
 {
   ctx->h[0] = SHA512M_A;
   ctx->h[1] = SHA512M_B;
@@ -183,7 +183,7 @@ void sha512_init (sha512_ctx_t *ctx)
   ctx->len = 0;
 }
 
-void sha512_update_128 (sha512_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const int len)
+DECLSPEC void sha512_update_128 (sha512_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 127;
@@ -313,7 +313,7 @@ void sha512_update_128 (sha512_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32
   }
 }
 
-void sha512_update (sha512_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha512_update (sha512_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -401,7 +401,7 @@ void sha512_update (sha512_ctx_t *ctx, const u32 *w, const int len)
   sha512_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, len - pos1);
 }
 
-void sha512_update_swap (sha512_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha512_update_swap (sha512_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -555,7 +555,7 @@ void sha512_update_swap (sha512_ctx_t *ctx, const u32 *w, const int len)
   sha512_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, len - pos1);
 }
 
-void sha512_update_utf16le (sha512_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha512_update_utf16le (sha512_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -621,7 +621,7 @@ void sha512_update_utf16le (sha512_ctx_t *ctx, const u32 *w, const int len)
   sha512_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, (len - pos1) * 2);
 }
 
-void sha512_update_utf16le_swap (sha512_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha512_update_utf16le_swap (sha512_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -753,7 +753,7 @@ void sha512_update_utf16le_swap (sha512_ctx_t *ctx, const u32 *w, const int len)
   sha512_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, (len - pos1) * 2);
 }
 
-void sha512_update_global (sha512_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha512_update_global (sha512_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -841,7 +841,7 @@ void sha512_update_global (sha512_ctx_t *ctx, const __global u32 *w, const int l
   sha512_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, len - pos1);
 }
 
-void sha512_update_global_swap (sha512_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha512_update_global_swap (sha512_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -995,7 +995,7 @@ void sha512_update_global_swap (sha512_ctx_t *ctx, const __global u32 *w, const
   sha512_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, len - pos1);
 }
 
-void sha512_update_global_utf16le (sha512_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha512_update_global_utf16le (sha512_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1061,7 +1061,7 @@ void sha512_update_global_utf16le (sha512_ctx_t *ctx, const __global u32 *w, con
   sha512_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, (len - pos1) * 2);
 }
 
-void sha512_update_global_utf16le_swap (sha512_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha512_update_global_utf16le_swap (sha512_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1193,7 +1193,7 @@ void sha512_update_global_utf16le_swap (sha512_ctx_t *ctx, const __global u32 *w
   sha512_update_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, (len - pos1) * 2);
 }
 
-void sha512_final (sha512_ctx_t *ctx)
+DECLSPEC void sha512_final (sha512_ctx_t *ctx)
 {
   const int pos = ctx->len & 127;
 
@@ -1252,7 +1252,7 @@ typedef struct sha512_hmac_ctx
 
 } sha512_hmac_ctx_t;
 
-void sha512_hmac_init_128 (sha512_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], const u32 w4[4], const u32 w5[4], const u32 w6[4], const u32 w7[4])
+DECLSPEC void sha512_hmac_init_128 (sha512_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], const u32 w4[4], const u32 w5[4], const u32 w6[4], const u32 w7[4])
 {
   u32 t0[4];
   u32 t1[4];
@@ -1342,7 +1342,7 @@ void sha512_hmac_init_128 (sha512_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1
   sha512_update_128 (&ctx->opad, t0, t1, t2, t3, t4, t5, t6, t7, 128);
 }
 
-void sha512_hmac_init (sha512_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha512_hmac_init (sha512_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1435,7 +1435,7 @@ void sha512_hmac_init (sha512_hmac_ctx_t *ctx, const u32 *w, const int len)
   sha512_hmac_init_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7);
 }
 
-void sha512_hmac_init_swap (sha512_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha512_hmac_init_swap (sha512_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1528,7 +1528,7 @@ void sha512_hmac_init_swap (sha512_hmac_ctx_t *ctx, const u32 *w, const int len)
   sha512_hmac_init_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7);
 }
 
-void sha512_hmac_init_global (sha512_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void sha512_hmac_init_global (sha512_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1621,7 +1621,7 @@ void sha512_hmac_init_global (sha512_hmac_ctx_t *ctx, __global const u32 *w, con
   sha512_hmac_init_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7);
 }
 
-void sha512_hmac_init_global_swap (sha512_hmac_ctx_t *ctx, __global const u32 *w, const int len)
+DECLSPEC void sha512_hmac_init_global_swap (sha512_hmac_ctx_t *ctx, __global const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1714,52 +1714,52 @@ void sha512_hmac_init_global_swap (sha512_hmac_ctx_t *ctx, __global const u32 *w
   sha512_hmac_init_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7);
 }
 
-void sha512_hmac_update_128 (sha512_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const int len)
+DECLSPEC void sha512_hmac_update_128 (sha512_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const int len)
 {
   sha512_update_128 (&ctx->ipad, w0, w1, w2, w3, w4, w5, w6, w7, len);
 }
 
-void sha512_hmac_update (sha512_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha512_hmac_update (sha512_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha512_update (&ctx->ipad, w, len);
 }
 
-void sha512_hmac_update_swap (sha512_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha512_hmac_update_swap (sha512_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha512_update_swap (&ctx->ipad, w, len);
 }
 
-void sha512_hmac_update_utf16le (sha512_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha512_hmac_update_utf16le (sha512_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha512_update_utf16le (&ctx->ipad, w, len);
 }
 
-void sha512_hmac_update_utf16le_swap (sha512_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void sha512_hmac_update_utf16le_swap (sha512_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   sha512_update_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void sha512_hmac_update_global (sha512_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha512_hmac_update_global (sha512_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha512_update_global (&ctx->ipad, w, len);
 }
 
-void sha512_hmac_update_global_swap (sha512_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha512_hmac_update_global_swap (sha512_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha512_update_global_swap (&ctx->ipad, w, len);
 }
 
-void sha512_hmac_update_global_utf16le (sha512_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha512_hmac_update_global_utf16le (sha512_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha512_update_global_utf16le (&ctx->ipad, w, len);
 }
 
-void sha512_hmac_update_global_utf16le_swap (sha512_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void sha512_hmac_update_global_utf16le_swap (sha512_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   sha512_update_global_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void sha512_hmac_final (sha512_hmac_ctx_t *ctx)
+DECLSPEC void sha512_hmac_final (sha512_hmac_ctx_t *ctx)
 {
   sha512_final (&ctx->ipad);
 
@@ -1829,7 +1829,7 @@ typedef struct sha512_ctx_vector
 
 } sha512_ctx_vector_t;
 
-void sha512_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x w4[4], const u32x w5[4], const u32x w6[4], const u32x w7[4], u64x digest[8])
+DECLSPEC void sha512_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x w4[4], const u32x w5[4], const u32x w6[4], const u32x w7[4], u64x digest[8])
 {
   u64x a = digest[0];
   u64x b = digest[1];
@@ -1920,7 +1920,7 @@ void sha512_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[
   digest[7] += h;
 }
 
-void sha512_init_vector (sha512_ctx_vector_t *ctx)
+DECLSPEC void sha512_init_vector (sha512_ctx_vector_t *ctx)
 {
   ctx->h[0] = SHA512M_A;
   ctx->h[1] = SHA512M_B;
@@ -1967,7 +1967,7 @@ void sha512_init_vector (sha512_ctx_vector_t *ctx)
   ctx->len = 0;
 }
 
-void sha512_init_vector_from_scalar (sha512_ctx_vector_t *ctx, sha512_ctx_t *ctx0)
+DECLSPEC void sha512_init_vector_from_scalar (sha512_ctx_vector_t *ctx, sha512_ctx_t *ctx0)
 {
   ctx->h[0] = ctx0->h[0];
   ctx->h[1] = ctx0->h[1];
@@ -2014,7 +2014,7 @@ void sha512_init_vector_from_scalar (sha512_ctx_vector_t *ctx, sha512_ctx_t *ctx
   ctx->len = ctx0->len;
 }
 
-void sha512_update_vector_128 (sha512_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const int len)
+DECLSPEC void sha512_update_vector_128 (sha512_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 127;
@@ -2144,7 +2144,7 @@ void sha512_update_vector_128 (sha512_ctx_vector_t *ctx, u32x w0[4], u32x w1[4],
   }
 }
 
-void sha512_update_vector (sha512_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha512_update_vector (sha512_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2232,7 +2232,7 @@ void sha512_update_vector (sha512_ctx_vector_t *ctx, const u32x *w, const int le
   sha512_update_vector_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, len - pos1);
 }
 
-void sha512_update_vector_swap (sha512_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha512_update_vector_swap (sha512_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2386,7 +2386,7 @@ void sha512_update_vector_swap (sha512_ctx_vector_t *ctx, const u32x *w, const i
   sha512_update_vector_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, len - pos1);
 }
 
-void sha512_update_vector_utf16le (sha512_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha512_update_vector_utf16le (sha512_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2452,7 +2452,7 @@ void sha512_update_vector_utf16le (sha512_ctx_vector_t *ctx, const u32x *w, cons
   sha512_update_vector_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, (len - pos1) * 2);
 }
 
-void sha512_update_vector_utf16le_swap (sha512_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha512_update_vector_utf16le_swap (sha512_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2584,7 +2584,7 @@ void sha512_update_vector_utf16le_swap (sha512_ctx_vector_t *ctx, const u32x *w,
   sha512_update_vector_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, (len - pos1) * 2);
 }
 
-void sha512_update_vector_utf16beN (sha512_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha512_update_vector_utf16beN (sha512_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2650,7 +2650,7 @@ void sha512_update_vector_utf16beN (sha512_ctx_vector_t *ctx, const u32x *w, con
   sha512_update_vector_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, (len - pos1) * 2);
 }
 
-void sha512_final_vector (sha512_ctx_vector_t *ctx)
+DECLSPEC void sha512_final_vector (sha512_ctx_vector_t *ctx)
 {
   const int pos = ctx->len & 127;
 
@@ -2709,7 +2709,7 @@ typedef struct sha512_hmac_ctx_vector
 
 } sha512_hmac_ctx_vector_t;
 
-void sha512_hmac_init_vector_128 (sha512_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x w4[4], const u32x w5[4], const u32x w6[4], const u32x w7[4])
+DECLSPEC void sha512_hmac_init_vector_128 (sha512_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x w4[4], const u32x w5[4], const u32x w6[4], const u32x w7[4])
 {
   u32x t0[4];
   u32x t1[4];
@@ -2799,7 +2799,7 @@ void sha512_hmac_init_vector_128 (sha512_hmac_ctx_vector_t *ctx, const u32x w0[4
   sha512_update_vector_128 (&ctx->opad, t0, t1, t2, t3, t4, t5, t6, t7, 128);
 }
 
-void sha512_hmac_init_vector (sha512_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha512_hmac_init_vector (sha512_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2892,17 +2892,17 @@ void sha512_hmac_init_vector (sha512_hmac_ctx_vector_t *ctx, const u32x *w, cons
   sha512_hmac_init_vector_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7);
 }
 
-void sha512_hmac_update_vector_128 (sha512_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const int len)
+DECLSPEC void sha512_hmac_update_vector_128 (sha512_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const int len)
 {
   sha512_update_vector_128 (&ctx->ipad, w0, w1, w2, w3, w4, w5, w6, w7, len);
 }
 
-void sha512_hmac_update_vector (sha512_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void sha512_hmac_update_vector (sha512_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   sha512_update_vector (&ctx->ipad, w, len);
 }
 
-void sha512_hmac_final_vector (sha512_hmac_ctx_vector_t *ctx)
+DECLSPEC void sha512_hmac_final_vector (sha512_hmac_ctx_vector_t *ctx)
 {
   sha512_final_vector (&ctx->ipad);
 
diff --git a/OpenCL/inc_hash_whirlpool.cl b/OpenCL/inc_hash_whirlpool.cl
index 66db07440..def248be5 100644
--- a/OpenCL/inc_hash_whirlpool.cl
+++ b/OpenCL/inc_hash_whirlpool.cl
@@ -1130,7 +1130,7 @@ typedef struct whirlpool_ctx
 
 } whirlpool_ctx_t;
 
-void whirlpool_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void whirlpool_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   u32 Kh[8];
   u32 Kl[8];
@@ -1300,7 +1300,7 @@ void whirlpool_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], con
   digest[15] ^= statel[7] ^ w3[3];
 }
 
-void whirlpool_init (whirlpool_ctx_t *ctx, SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void whirlpool_init (whirlpool_ctx_t *ctx, SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   ctx->h[ 0] = 0;
   ctx->h[ 1] = 0;
@@ -1342,7 +1342,7 @@ void whirlpool_init (whirlpool_ctx_t *ctx, SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u
   ctx->s_Cl = s_Cl;
 }
 
-void whirlpool_update_64 (whirlpool_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void whirlpool_update_64 (whirlpool_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -1420,7 +1420,7 @@ void whirlpool_update_64 (whirlpool_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4],
   }
 }
 
-void whirlpool_update (whirlpool_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void whirlpool_update (whirlpool_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1472,7 +1472,7 @@ void whirlpool_update (whirlpool_ctx_t *ctx, const u32 *w, const int len)
   whirlpool_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void whirlpool_update_swap (whirlpool_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void whirlpool_update_swap (whirlpool_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1558,7 +1558,7 @@ void whirlpool_update_swap (whirlpool_ctx_t *ctx, const u32 *w, const int len)
   whirlpool_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void whirlpool_update_utf16le (whirlpool_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void whirlpool_update_utf16le (whirlpool_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1600,7 +1600,7 @@ void whirlpool_update_utf16le (whirlpool_ctx_t *ctx, const u32 *w, const int len
   whirlpool_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void whirlpool_update_utf16le_swap (whirlpool_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void whirlpool_update_utf16le_swap (whirlpool_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1676,7 +1676,7 @@ void whirlpool_update_utf16le_swap (whirlpool_ctx_t *ctx, const u32 *w, const in
   whirlpool_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void whirlpool_update_global (whirlpool_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void whirlpool_update_global (whirlpool_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1728,7 +1728,7 @@ void whirlpool_update_global (whirlpool_ctx_t *ctx, const __global u32 *w, const
   whirlpool_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void whirlpool_update_global_swap (whirlpool_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void whirlpool_update_global_swap (whirlpool_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1814,7 +1814,7 @@ void whirlpool_update_global_swap (whirlpool_ctx_t *ctx, const __global u32 *w,
   whirlpool_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void whirlpool_update_global_utf16le (whirlpool_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void whirlpool_update_global_utf16le (whirlpool_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1856,7 +1856,7 @@ void whirlpool_update_global_utf16le (whirlpool_ctx_t *ctx, const __global u32 *
   whirlpool_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void whirlpool_update_global_utf16le_swap (whirlpool_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void whirlpool_update_global_utf16le_swap (whirlpool_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1932,7 +1932,7 @@ void whirlpool_update_global_utf16le_swap (whirlpool_ctx_t *ctx, const __global
   whirlpool_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void whirlpool_final (whirlpool_ctx_t *ctx)
+DECLSPEC void whirlpool_final (whirlpool_ctx_t *ctx)
 {
   const int pos = ctx->len & 63;
 
@@ -1975,7 +1975,7 @@ typedef struct whirlpool_hmac_ctx
 
 } whirlpool_hmac_ctx_t;
 
-void whirlpool_hmac_init_64 (whirlpool_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void whirlpool_hmac_init_64 (whirlpool_hmac_ctx_t *ctx, const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   u32 t0[4];
   u32 t1[4];
@@ -2029,7 +2029,7 @@ void whirlpool_hmac_init_64 (whirlpool_hmac_ctx_t *ctx, const u32 w0[4], const u
   whirlpool_update_64 (&ctx->opad, t0, t1, t2, t3, 64);
 }
 
-void whirlpool_hmac_init (whirlpool_hmac_ctx_t *ctx, const u32 *w, const int len, SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void whirlpool_hmac_init (whirlpool_hmac_ctx_t *ctx, const u32 *w, const int len, SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   u32 w0[4];
   u32 w1[4];
@@ -2086,7 +2086,7 @@ void whirlpool_hmac_init (whirlpool_hmac_ctx_t *ctx, const u32 *w, const int len
   whirlpool_hmac_init_64 (ctx, w0, w1, w2, w3, s_Ch, s_Cl);
 }
 
-void whirlpool_hmac_init_swap (whirlpool_hmac_ctx_t *ctx, const u32 *w, const int len, SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void whirlpool_hmac_init_swap (whirlpool_hmac_ctx_t *ctx, const u32 *w, const int len, SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   u32 w0[4];
   u32 w1[4];
@@ -2143,7 +2143,7 @@ void whirlpool_hmac_init_swap (whirlpool_hmac_ctx_t *ctx, const u32 *w, const in
   whirlpool_hmac_init_64 (ctx, w0, w1, w2, w3, s_Ch, s_Cl);
 }
 
-void whirlpool_hmac_init_global (whirlpool_hmac_ctx_t *ctx, __global const u32 *w, const int len, SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void whirlpool_hmac_init_global (whirlpool_hmac_ctx_t *ctx, __global const u32 *w, const int len, SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   u32 w0[4];
   u32 w1[4];
@@ -2200,7 +2200,7 @@ void whirlpool_hmac_init_global (whirlpool_hmac_ctx_t *ctx, __global const u32 *
   whirlpool_hmac_init_64 (ctx, w0, w1, w2, w3, s_Ch, s_Cl);
 }
 
-void whirlpool_hmac_init_global_swap (whirlpool_hmac_ctx_t *ctx, __global const u32 *w, const int len, SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void whirlpool_hmac_init_global_swap (whirlpool_hmac_ctx_t *ctx, __global const u32 *w, const int len, SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   u32 w0[4];
   u32 w1[4];
@@ -2257,52 +2257,52 @@ void whirlpool_hmac_init_global_swap (whirlpool_hmac_ctx_t *ctx, __global const
   whirlpool_hmac_init_64 (ctx, w0, w1, w2, w3, s_Ch, s_Cl);
 }
 
-void whirlpool_hmac_update_64 (whirlpool_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void whirlpool_hmac_update_64 (whirlpool_hmac_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   whirlpool_update_64 (&ctx->ipad, w0, w1, w2, w3, len);
 }
 
-void whirlpool_hmac_update (whirlpool_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void whirlpool_hmac_update (whirlpool_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   whirlpool_update (&ctx->ipad, w, len);
 }
 
-void whirlpool_hmac_update_swap (whirlpool_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void whirlpool_hmac_update_swap (whirlpool_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   whirlpool_update_swap (&ctx->ipad, w, len);
 }
 
-void whirlpool_hmac_update_utf16le (whirlpool_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void whirlpool_hmac_update_utf16le (whirlpool_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   whirlpool_update_utf16le (&ctx->ipad, w, len);
 }
 
-void whirlpool_hmac_update_utf16le_swap (whirlpool_hmac_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void whirlpool_hmac_update_utf16le_swap (whirlpool_hmac_ctx_t *ctx, const u32 *w, const int len)
 {
   whirlpool_update_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void whirlpool_hmac_update_global (whirlpool_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void whirlpool_hmac_update_global (whirlpool_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   whirlpool_update_global (&ctx->ipad, w, len);
 }
 
-void whirlpool_hmac_update_global_swap (whirlpool_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void whirlpool_hmac_update_global_swap (whirlpool_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   whirlpool_update_global_swap (&ctx->ipad, w, len);
 }
 
-void whirlpool_hmac_update_global_utf16le (whirlpool_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void whirlpool_hmac_update_global_utf16le (whirlpool_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   whirlpool_update_global_utf16le (&ctx->ipad, w, len);
 }
 
-void whirlpool_hmac_update_global_utf16le_swap (whirlpool_hmac_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void whirlpool_hmac_update_global_utf16le_swap (whirlpool_hmac_ctx_t *ctx, const __global u32 *w, const int len)
 {
   whirlpool_update_global_utf16le_swap (&ctx->ipad, w, len);
 }
 
-void whirlpool_hmac_final (whirlpool_hmac_ctx_t *ctx)
+DECLSPEC void whirlpool_hmac_final (whirlpool_hmac_ctx_t *ctx)
 {
   whirlpool_final (&ctx->ipad);
 
@@ -2351,7 +2351,7 @@ typedef struct whirlpool_ctx_vector
 
 } whirlpool_ctx_vector_t;
 
-void whirlpool_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void whirlpool_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   u32x Kh[8];
   u32x Kl[8];
@@ -2521,7 +2521,7 @@ void whirlpool_transform_vector (const u32x w0[4], const u32x w1[4], const u32x
   digest[15] ^= statel[7] ^ w3[3];
 }
 
-void whirlpool_init_vector (whirlpool_ctx_vector_t *ctx, SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void whirlpool_init_vector (whirlpool_ctx_vector_t *ctx, SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   ctx->h[ 0] = 0;
   ctx->h[ 1] = 0;
@@ -2563,7 +2563,7 @@ void whirlpool_init_vector (whirlpool_ctx_vector_t *ctx, SHM_TYPE u32 (*s_Ch)[25
   ctx->s_Cl = s_Cl;
 }
 
-void whirlpool_init_vector_from_scalar (whirlpool_ctx_vector_t *ctx, whirlpool_ctx_t *ctx0)
+DECLSPEC void whirlpool_init_vector_from_scalar (whirlpool_ctx_vector_t *ctx, whirlpool_ctx_t *ctx0)
 {
   ctx->h[ 0] = ctx0->h[ 0];
   ctx->h[ 1] = ctx0->h[ 1];
@@ -2605,7 +2605,7 @@ void whirlpool_init_vector_from_scalar (whirlpool_ctx_vector_t *ctx, whirlpool_c
   ctx->s_Cl = ctx0->s_Cl;
 }
 
-void whirlpool_update_vector_64 (whirlpool_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
+DECLSPEC void whirlpool_update_vector_64 (whirlpool_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -2683,7 +2683,7 @@ void whirlpool_update_vector_64 (whirlpool_ctx_vector_t *ctx, u32x w0[4], u32x w
   }
 }
 
-void whirlpool_update_vector (whirlpool_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void whirlpool_update_vector (whirlpool_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2735,7 +2735,7 @@ void whirlpool_update_vector (whirlpool_ctx_vector_t *ctx, const u32x *w, const
   whirlpool_update_vector_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void whirlpool_update_vector_swap (whirlpool_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void whirlpool_update_vector_swap (whirlpool_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2821,7 +2821,7 @@ void whirlpool_update_vector_swap (whirlpool_ctx_vector_t *ctx, const u32x *w, c
   whirlpool_update_vector_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
-void whirlpool_update_vector_utf16le (whirlpool_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void whirlpool_update_vector_utf16le (whirlpool_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2863,7 +2863,7 @@ void whirlpool_update_vector_utf16le (whirlpool_ctx_vector_t *ctx, const u32x *w
   whirlpool_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void whirlpool_update_vector_utf16le_swap (whirlpool_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void whirlpool_update_vector_utf16le_swap (whirlpool_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -2939,7 +2939,7 @@ void whirlpool_update_vector_utf16le_swap (whirlpool_ctx_vector_t *ctx, const u3
   whirlpool_update_vector_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
-void whirlpool_final_vector (whirlpool_ctx_vector_t *ctx)
+DECLSPEC void whirlpool_final_vector (whirlpool_ctx_vector_t *ctx)
 {
   const int pos = ctx->len & 63;
 
@@ -2982,7 +2982,7 @@ typedef struct whirlpool_hmac_ctx_vector
 
 } whirlpool_hmac_ctx_vector_t;
 
-void whirlpool_hmac_init_vector_64 (whirlpool_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void whirlpool_hmac_init_vector_64 (whirlpool_hmac_ctx_vector_t *ctx, const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   u32x t0[4];
   u32x t1[4];
@@ -3036,7 +3036,7 @@ void whirlpool_hmac_init_vector_64 (whirlpool_hmac_ctx_vector_t *ctx, const u32x
   whirlpool_update_vector_64 (&ctx->opad, t0, t1, t2, t3, 64);
 }
 
-void whirlpool_hmac_init_vector (whirlpool_hmac_ctx_vector_t *ctx, const u32x *w, const int len, SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void whirlpool_hmac_init_vector (whirlpool_hmac_ctx_vector_t *ctx, const u32x *w, const int len, SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   u32x w0[4];
   u32x w1[4];
@@ -3093,17 +3093,17 @@ void whirlpool_hmac_init_vector (whirlpool_hmac_ctx_vector_t *ctx, const u32x *w
   whirlpool_hmac_init_vector_64 (ctx, w0, w1, w2, w3, s_Ch, s_Cl);
 }
 
-void whirlpool_hmac_update_vector_64 (whirlpool_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
+DECLSPEC void whirlpool_hmac_update_vector_64 (whirlpool_hmac_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
 {
   whirlpool_update_vector_64 (&ctx->ipad, w0, w1, w2, w3, len);
 }
 
-void whirlpool_hmac_update_vector (whirlpool_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void whirlpool_hmac_update_vector (whirlpool_hmac_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   whirlpool_update_vector (&ctx->ipad, w, len);
 }
 
-void whirlpool_hmac_final_vector (whirlpool_hmac_ctx_vector_t *ctx)
+DECLSPEC void whirlpool_hmac_final_vector (whirlpool_hmac_ctx_vector_t *ctx)
 {
   whirlpool_final_vector (&ctx->ipad);
 
diff --git a/OpenCL/inc_luks_aes.cl b/OpenCL/inc_luks_aes.cl
index a26fc3969..d6192ab06 100644
--- a/OpenCL/inc_luks_aes.cl
+++ b/OpenCL/inc_luks_aes.cl
@@ -1,7 +1,7 @@
 
 // cbc-essiv
 
-void aes128_decrypt_cbc (const u32 *ks1, const u32 *in, u32 *out, u32 *essiv, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void aes128_decrypt_cbc (const u32 *ks1, const u32 *in, u32 *out, u32 *essiv, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   aes128_decrypt (ks1, in, out, s_td0, s_td1, s_td2, s_td3, s_td4);
 
@@ -16,7 +16,7 @@ void aes128_decrypt_cbc (const u32 *ks1, const u32 *in, u32 *out, u32 *essiv, SH
   essiv[3] = in[3];
 }
 
-void aes256_decrypt_cbc (const u32 *ks1, const u32 *in, u32 *out, u32 *essiv, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void aes256_decrypt_cbc (const u32 *ks1, const u32 *in, u32 *out, u32 *essiv, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   aes256_decrypt (ks1, in, out, s_td0, s_td1, s_td2, s_td3, s_td4);
 
@@ -31,7 +31,7 @@ void aes256_decrypt_cbc (const u32 *ks1, const u32 *in, u32 *out, u32 *essiv, SH
   essiv[3] = in[3];
 }
 
-void luks_decrypt_sector_aes_cbc_essiv128 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv128 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -65,7 +65,7 @@ void luks_decrypt_sector_aes_cbc_essiv128 (__global const u32 *in, u32 *out, con
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv128_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv128_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -102,7 +102,7 @@ void luks_decrypt_sector_aes_cbc_essiv128_mk_sha1 (__global const u32 *in, u32 *
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv128_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv128_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -162,7 +162,7 @@ void luks_decrypt_sector_aes_cbc_essiv128_mk_sha1_final (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv128_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv128_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -199,7 +199,7 @@ void luks_decrypt_sector_aes_cbc_essiv128_mk_sha256 (__global const u32 *in, u32
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv128_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv128_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -259,7 +259,7 @@ void luks_decrypt_sector_aes_cbc_essiv128_mk_sha256_final (__global const u32 *i
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv128_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv128_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -296,7 +296,7 @@ void luks_decrypt_sector_aes_cbc_essiv128_mk_sha512 (__global const u32 *in, u32
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv128_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv128_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -356,7 +356,7 @@ void luks_decrypt_sector_aes_cbc_essiv128_mk_sha512_final (__global const u32 *i
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv128_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv128_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -393,7 +393,7 @@ void luks_decrypt_sector_aes_cbc_essiv128_mk_ripemd160 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv128_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv128_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -453,7 +453,7 @@ void luks_decrypt_sector_aes_cbc_essiv128_mk_ripemd160_final (__global const u32
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv256 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv256 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -487,7 +487,7 @@ void luks_decrypt_sector_aes_cbc_essiv256 (__global const u32 *in, u32 *out, con
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -524,7 +524,7 @@ void luks_decrypt_sector_aes_cbc_essiv256_mk_sha1 (__global const u32 *in, u32 *
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -584,7 +584,7 @@ void luks_decrypt_sector_aes_cbc_essiv256_mk_sha1_final (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -621,7 +621,7 @@ void luks_decrypt_sector_aes_cbc_essiv256_mk_sha256 (__global const u32 *in, u32
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -681,7 +681,7 @@ void luks_decrypt_sector_aes_cbc_essiv256_mk_sha256_final (__global const u32 *i
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -718,7 +718,7 @@ void luks_decrypt_sector_aes_cbc_essiv256_mk_sha512 (__global const u32 *in, u32
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -778,7 +778,7 @@ void luks_decrypt_sector_aes_cbc_essiv256_mk_sha512_final (__global const u32 *i
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -815,7 +815,7 @@ void luks_decrypt_sector_aes_cbc_essiv256_mk_ripemd160 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_aes_cbc_essiv256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_essiv256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -877,7 +877,7 @@ void luks_decrypt_sector_aes_cbc_essiv256_mk_ripemd160_final (__global const u32
 
 // cbc-plain
 
-void luks_decrypt_sector_aes_cbc_plain128 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain128 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -907,7 +907,7 @@ void luks_decrypt_sector_aes_cbc_plain128 (__global const u32 *in, u32 *out, con
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain128_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain128_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -940,7 +940,7 @@ void luks_decrypt_sector_aes_cbc_plain128_mk_sha1 (__global const u32 *in, u32 *
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain128_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain128_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -996,7 +996,7 @@ void luks_decrypt_sector_aes_cbc_plain128_mk_sha1_final (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain128_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain128_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1029,7 +1029,7 @@ void luks_decrypt_sector_aes_cbc_plain128_mk_sha256 (__global const u32 *in, u32
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain128_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain128_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1085,7 +1085,7 @@ void luks_decrypt_sector_aes_cbc_plain128_mk_sha256_final (__global const u32 *i
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain128_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain128_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1118,7 +1118,7 @@ void luks_decrypt_sector_aes_cbc_plain128_mk_sha512 (__global const u32 *in, u32
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain128_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain128_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1174,7 +1174,7 @@ void luks_decrypt_sector_aes_cbc_plain128_mk_sha512_final (__global const u32 *i
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain128_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain128_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1207,7 +1207,7 @@ void luks_decrypt_sector_aes_cbc_plain128_mk_ripemd160 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain128_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain128_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1263,7 +1263,7 @@ void luks_decrypt_sector_aes_cbc_plain128_mk_ripemd160_final (__global const u32
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain256 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain256 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1293,7 +1293,7 @@ void luks_decrypt_sector_aes_cbc_plain256 (__global const u32 *in, u32 *out, con
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1326,7 +1326,7 @@ void luks_decrypt_sector_aes_cbc_plain256_mk_sha1 (__global const u32 *in, u32 *
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1382,7 +1382,7 @@ void luks_decrypt_sector_aes_cbc_plain256_mk_sha1_final (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1415,7 +1415,7 @@ void luks_decrypt_sector_aes_cbc_plain256_mk_sha256 (__global const u32 *in, u32
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1471,7 +1471,7 @@ void luks_decrypt_sector_aes_cbc_plain256_mk_sha256_final (__global const u32 *i
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1504,7 +1504,7 @@ void luks_decrypt_sector_aes_cbc_plain256_mk_sha512 (__global const u32 *in, u32
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1560,7 +1560,7 @@ void luks_decrypt_sector_aes_cbc_plain256_mk_sha512_final (__global const u32 *i
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1593,7 +1593,7 @@ void luks_decrypt_sector_aes_cbc_plain256_mk_ripemd160 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_aes_cbc_plain256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_cbc_plain256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1651,7 +1651,7 @@ void luks_decrypt_sector_aes_cbc_plain256_mk_ripemd160_final (__global const u32
 
 // xts-plain
 
-void aes128_decrypt_xts (const u32 *ks1, const u32 *in, u32 *out, u32 *T, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void aes128_decrypt_xts (const u32 *ks1, const u32 *in, u32 *out, u32 *T, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   out[0] = in[0];
   out[1] = in[1];
@@ -1673,7 +1673,7 @@ void aes128_decrypt_xts (const u32 *ks1, const u32 *in, u32 *out, u32 *T, SHM_TY
   xts_mul2 (T, T);
 }
 
-void aes256_decrypt_xts (const u32 *ks1, const u32 *in, u32 *out, u32 *T, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void aes256_decrypt_xts (const u32 *ks1, const u32 *in, u32 *out, u32 *T, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   out[0] = in[0];
   out[1] = in[1];
@@ -1695,7 +1695,7 @@ void aes256_decrypt_xts (const u32 *ks1, const u32 *in, u32 *out, u32 *T, SHM_TY
   xts_mul2 (T, T);
 }
 
-void luks_decrypt_sector_aes_xts_plain256 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain256 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1729,7 +1729,7 @@ void luks_decrypt_sector_aes_xts_plain256 (__global const u32 *in, u32 *out, con
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1766,7 +1766,7 @@ void luks_decrypt_sector_aes_xts_plain256_mk_sha1 (__global const u32 *in, u32 *
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1826,7 +1826,7 @@ void luks_decrypt_sector_aes_xts_plain256_mk_sha1_final (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1863,7 +1863,7 @@ void luks_decrypt_sector_aes_xts_plain256_mk_sha256 (__global const u32 *in, u32
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1923,7 +1923,7 @@ void luks_decrypt_sector_aes_xts_plain256_mk_sha256_final (__global const u32 *i
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1960,7 +1960,7 @@ void luks_decrypt_sector_aes_xts_plain256_mk_sha512 (__global const u32 *in, u32
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2020,7 +2020,7 @@ void luks_decrypt_sector_aes_xts_plain256_mk_sha512_final (__global const u32 *i
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2057,7 +2057,7 @@ void luks_decrypt_sector_aes_xts_plain256_mk_ripemd160 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2117,7 +2117,7 @@ void luks_decrypt_sector_aes_xts_plain256_mk_ripemd160_final (__global const u32
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain512 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain512 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2151,7 +2151,7 @@ void luks_decrypt_sector_aes_xts_plain512 (__global const u32 *in, u32 *out, con
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain512_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain512_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2188,7 +2188,7 @@ void luks_decrypt_sector_aes_xts_plain512_mk_sha1 (__global const u32 *in, u32 *
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain512_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain512_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2248,7 +2248,7 @@ void luks_decrypt_sector_aes_xts_plain512_mk_sha1_final (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain512_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain512_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2285,7 +2285,7 @@ void luks_decrypt_sector_aes_xts_plain512_mk_sha256 (__global const u32 *in, u32
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain512_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain512_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2345,7 +2345,7 @@ void luks_decrypt_sector_aes_xts_plain512_mk_sha256_final (__global const u32 *i
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain512_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain512_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2382,7 +2382,7 @@ void luks_decrypt_sector_aes_xts_plain512_mk_sha512 (__global const u32 *in, u32
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain512_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain512_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2442,7 +2442,7 @@ void luks_decrypt_sector_aes_xts_plain512_mk_sha512_final (__global const u32 *i
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain512_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain512_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2479,7 +2479,7 @@ void luks_decrypt_sector_aes_xts_plain512_mk_ripemd160 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_aes_xts_plain512_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_decrypt_sector_aes_xts_plain512_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, const u32 sector, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2541,7 +2541,7 @@ void luks_decrypt_sector_aes_xts_plain512_mk_ripemd160_final (__global const u32
 
 // luks helper
 
-void luks_af_sha1_then_aes_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_af_sha1_then_aes_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   const u32 key_size    = luks_bufs->key_size;
   const u32 cipher_mode = luks_bufs->cipher_mode;
@@ -2932,7 +2932,7 @@ void luks_af_sha1_then_aes_decrypt (__global const luks_t *luks_bufs, __global l
   }
 }
 
-void luks_af_sha256_then_aes_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_af_sha256_then_aes_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   const u32 key_size    = luks_bufs->key_size;
   const u32 cipher_mode = luks_bufs->cipher_mode;
@@ -3291,7 +3291,7 @@ void luks_af_sha256_then_aes_decrypt (__global const luks_t *luks_bufs, __global
   }
 }
 
-void luks_af_sha512_then_aes_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_af_sha512_then_aes_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   const u32 key_size    = luks_bufs->key_size;
   const u32 cipher_mode = luks_bufs->cipher_mode;
@@ -3669,7 +3669,7 @@ void luks_af_sha512_then_aes_decrypt (__global const luks_t *luks_bufs, __global
   }
 }
 
-void luks_af_ripemd160_then_aes_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void luks_af_ripemd160_then_aes_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   const u32 key_size    = luks_bufs->key_size;
   const u32 cipher_mode = luks_bufs->cipher_mode;
diff --git a/OpenCL/inc_luks_af.cl b/OpenCL/inc_luks_af.cl
index 9946c209c..84f781bcd 100644
--- a/OpenCL/inc_luks_af.cl
+++ b/OpenCL/inc_luks_af.cl
@@ -45,7 +45,7 @@ __constant static u64a AF_k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void AF_sha1_transform_S (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5])
+DECLSPEC void AF_sha1_transform_S (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5])
 {
   u32 A = digest[0];
   u32 B = digest[1];
@@ -173,7 +173,7 @@ void AF_sha1_transform_S (const u32 w0[4], const u32 w1[4], const u32 w2[4], con
   digest[4] += E;
 }
 
-void AF_sha256_transform_S (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[8])
+DECLSPEC void AF_sha256_transform_S (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[8])
 {
   u32 a = digest[0];
   u32 b = digest[1];
@@ -261,7 +261,7 @@ void AF_sha256_transform_S (const u32 w0[4], const u32 w1[4], const u32 w2[4], c
   digest[7] += h;
 }
 
-void AF_sha512_transform_S (const u64 w0[4], const u64 w1[4], const u64 w2[4], const u64 w3[4], u64 digest[8])
+DECLSPEC void AF_sha512_transform_S (const u64 w0[4], const u64 w1[4], const u64 w2[4], const u64 w3[4], u64 digest[8])
 {
   u64 a = digest[0];
   u64 b = digest[1];
@@ -349,7 +349,7 @@ void AF_sha512_transform_S (const u64 w0[4], const u64 w1[4], const u64 w2[4], c
   digest[7] += h;
 }
 
-void AF_ripemd160_transform_S (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5])
+DECLSPEC void AF_ripemd160_transform_S (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5])
 {
   u32 w0_t = w0[0];
   u32 w1_t = w0[1];
@@ -565,7 +565,7 @@ void AF_ripemd160_transform_S (const u32 w0[4], const u32 w1[4], const u32 w2[4]
 
 // diffuse functions
 
-void AF_sha1_diffuse16 (u32 *out)
+DECLSPEC void AF_sha1_diffuse16 (u32 *out)
 {
   u32 w0[4];
   u32 w1[4];
@@ -607,7 +607,7 @@ void AF_sha1_diffuse16 (u32 *out)
   out[3] = swap32_S (digest[3]);
 }
 
-void AF_sha1_diffuse32 (u32 *out)
+DECLSPEC void AF_sha1_diffuse32 (u32 *out)
 {
   u32 w0[4];
   u32 w1[4];
@@ -681,7 +681,7 @@ void AF_sha1_diffuse32 (u32 *out)
   out[7] = swap32_S (digest[2]);
 }
 
-void AF_sha1_diffuse64 (u32 *out)
+DECLSPEC void AF_sha1_diffuse64 (u32 *out)
 {
   u32 w0[4];
   u32 w1[4];
@@ -819,7 +819,7 @@ void AF_sha1_diffuse64 (u32 *out)
   out[15] = swap32_S (digest[0]);
 }
 
-void AF_sha256_diffuse16 (u32 *out)
+DECLSPEC void AF_sha256_diffuse16 (u32 *out)
 {
   u32 w0[4];
   u32 w1[4];
@@ -864,7 +864,7 @@ void AF_sha256_diffuse16 (u32 *out)
   out[3] = swap32_S (digest[3]);
 }
 
-void AF_sha256_diffuse32 (u32 *out)
+DECLSPEC void AF_sha256_diffuse32 (u32 *out)
 {
   u32 w0[4];
   u32 w1[4];
@@ -913,7 +913,7 @@ void AF_sha256_diffuse32 (u32 *out)
   out[7] = swap32_S (digest[7]);
 }
 
-void AF_sha256_diffuse64 (u32 *out)
+DECLSPEC void AF_sha256_diffuse64 (u32 *out)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1001,7 +1001,7 @@ void AF_sha256_diffuse64 (u32 *out)
   out[15] = swap32_S (digest[7]);
 }
 
-void AF_sha512_diffuse16 (u32 *out)
+DECLSPEC void AF_sha512_diffuse16 (u32 *out)
 {
   u64 w0[4];
   u64 w1[4];
@@ -1046,7 +1046,7 @@ void AF_sha512_diffuse16 (u32 *out)
   out[3] = swap32_S (l32_from_64_S (digest[1]));
 }
 
-void AF_sha512_diffuse32 (u32 *out)
+DECLSPEC void AF_sha512_diffuse32 (u32 *out)
 {
   u64 w0[4];
   u64 w1[4];
@@ -1095,7 +1095,7 @@ void AF_sha512_diffuse32 (u32 *out)
   out[7] = swap32_S (l32_from_64_S (digest[3]));
 }
 
-void AF_sha512_diffuse64 (u32 *out)
+DECLSPEC void AF_sha512_diffuse64 (u32 *out)
 {
   u64 w0[4];
   u64 w1[4];
@@ -1152,7 +1152,7 @@ void AF_sha512_diffuse64 (u32 *out)
   out[15] = swap32_S (l32_from_64_S (digest[7]));
 }
 
-void AF_ripemd160_diffuse16 (u32 *out)
+DECLSPEC void AF_ripemd160_diffuse16 (u32 *out)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1194,7 +1194,7 @@ void AF_ripemd160_diffuse16 (u32 *out)
   out[3] = digest[3];
 }
 
-void AF_ripemd160_diffuse32 (u32 *out)
+DECLSPEC void AF_ripemd160_diffuse32 (u32 *out)
 {
   u32 w0[4];
   u32 w1[4];
@@ -1268,7 +1268,7 @@ void AF_ripemd160_diffuse32 (u32 *out)
   out[7] = digest[2];
 }
 
-void AF_ripemd160_diffuse64 (u32 *out)
+DECLSPEC void AF_ripemd160_diffuse64 (u32 *out)
 {
   u32 w0[4];
   u32 w1[4];
diff --git a/OpenCL/inc_luks_essiv.cl b/OpenCL/inc_luks_essiv.cl
index ecf3e1da4..00391cc54 100644
--- a/OpenCL/inc_luks_essiv.cl
+++ b/OpenCL/inc_luks_essiv.cl
@@ -19,7 +19,7 @@ __constant static u32a ESSIV_k_sha256[64] =
 };
 
 // basically a normal sha256_transform() but with a different name to avoid collisions with function nameing
-void ESSIV_sha256_transform_S (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[8])
+DECLSPEC void ESSIV_sha256_transform_S (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[8])
 {
   u32 a = digest[0];
   u32 b = digest[1];
@@ -107,7 +107,7 @@ void ESSIV_sha256_transform_S (const u32 w0[4], const u32 w1[4], const u32 w2[4]
   digest[7] += h;
 }
 
-void ESSIV_sha256_init128 (u32 *key, u32 *essivhash)
+DECLSPEC void ESSIV_sha256_init128 (u32 *key, u32 *essivhash)
 {
   essivhash[0] = SHA256M_A;
   essivhash[1] = SHA256M_B;
@@ -152,7 +152,7 @@ void ESSIV_sha256_init128 (u32 *key, u32 *essivhash)
   essivhash[7] = swap32_S (essivhash[7]);
 }
 
-void ESSIV_sha256_init256 (u32 *key, u32 *essivhash)
+DECLSPEC void ESSIV_sha256_init256 (u32 *key, u32 *essivhash)
 {
   essivhash[0] = SHA256M_A;
   essivhash[1] = SHA256M_B;
diff --git a/OpenCL/inc_luks_serpent.cl b/OpenCL/inc_luks_serpent.cl
index 771792818..fc4e1f26c 100644
--- a/OpenCL/inc_luks_serpent.cl
+++ b/OpenCL/inc_luks_serpent.cl
@@ -1,7 +1,7 @@
 
 // cbc-essiv
 
-void serpent128_decrypt_cbc (const u32 *ks1, const u32 *in, u32 *out, u32 *essiv)
+DECLSPEC void serpent128_decrypt_cbc (const u32 *ks1, const u32 *in, u32 *out, u32 *essiv)
 {
   serpent128_decrypt (ks1, in, out);
 
@@ -16,7 +16,7 @@ void serpent128_decrypt_cbc (const u32 *ks1, const u32 *in, u32 *out, u32 *essiv
   essiv[3] = in[3];
 }
 
-void serpent256_decrypt_cbc (const u32 *ks1, const u32 *in, u32 *out, u32 *essiv)
+DECLSPEC void serpent256_decrypt_cbc (const u32 *ks1, const u32 *in, u32 *out, u32 *essiv)
 {
   serpent256_decrypt (ks1, in, out);
 
@@ -31,7 +31,7 @@ void serpent256_decrypt_cbc (const u32 *ks1, const u32 *in, u32 *out, u32 *essiv
   essiv[3] = in[3];
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv128 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv128 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -65,7 +65,7 @@ void luks_decrypt_sector_serpent_cbc_essiv128 (__global const u32 *in, u32 *out,
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -102,7 +102,7 @@ void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha1 (__global const u32 *in, u
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -162,7 +162,7 @@ void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha1_final (__global const u32
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -199,7 +199,7 @@ void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha256 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -259,7 +259,7 @@ void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha256_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -296,7 +296,7 @@ void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha512 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -356,7 +356,7 @@ void luks_decrypt_sector_serpent_cbc_essiv128_mk_sha512_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv128_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv128_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -393,7 +393,7 @@ void luks_decrypt_sector_serpent_cbc_essiv128_mk_ripemd160 (__global const u32 *
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv128_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv128_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -453,7 +453,7 @@ void luks_decrypt_sector_serpent_cbc_essiv128_mk_ripemd160_final (__global const
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv256 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv256 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -487,7 +487,7 @@ void luks_decrypt_sector_serpent_cbc_essiv256 (__global const u32 *in, u32 *out,
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -524,7 +524,7 @@ void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha1 (__global const u32 *in, u
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -584,7 +584,7 @@ void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha1_final (__global const u32
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -621,7 +621,7 @@ void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha256 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -681,7 +681,7 @@ void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha256_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -718,7 +718,7 @@ void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha512 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -778,7 +778,7 @@ void luks_decrypt_sector_serpent_cbc_essiv256_mk_sha512_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -815,7 +815,7 @@ void luks_decrypt_sector_serpent_cbc_essiv256_mk_ripemd160 (__global const u32 *
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_essiv256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_essiv256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -877,7 +877,7 @@ void luks_decrypt_sector_serpent_cbc_essiv256_mk_ripemd160_final (__global const
 
 // cbc-plain
 
-void luks_decrypt_sector_serpent_cbc_plain128 (__global const u32 *in, u32 *out, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain128 (__global const u32 *in, u32 *out, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -907,7 +907,7 @@ void luks_decrypt_sector_serpent_cbc_plain128 (__global const u32 *in, u32 *out,
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain128_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain128_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -940,7 +940,7 @@ void luks_decrypt_sector_serpent_cbc_plain128_mk_sha1 (__global const u32 *in, u
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain128_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain128_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -996,7 +996,7 @@ void luks_decrypt_sector_serpent_cbc_plain128_mk_sha1_final (__global const u32
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain128_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain128_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1029,7 +1029,7 @@ void luks_decrypt_sector_serpent_cbc_plain128_mk_sha256 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain128_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain128_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1085,7 +1085,7 @@ void luks_decrypt_sector_serpent_cbc_plain128_mk_sha256_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain128_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain128_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1118,7 +1118,7 @@ void luks_decrypt_sector_serpent_cbc_plain128_mk_sha512 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain128_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain128_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1174,7 +1174,7 @@ void luks_decrypt_sector_serpent_cbc_plain128_mk_sha512_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain128_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain128_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1207,7 +1207,7 @@ void luks_decrypt_sector_serpent_cbc_plain128_mk_ripemd160 (__global const u32 *
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain128_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain128_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1263,7 +1263,7 @@ void luks_decrypt_sector_serpent_cbc_plain128_mk_ripemd160_final (__global const
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain256 (__global const u32 *in, u32 *out, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain256 (__global const u32 *in, u32 *out, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1293,7 +1293,7 @@ void luks_decrypt_sector_serpent_cbc_plain256 (__global const u32 *in, u32 *out,
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1326,7 +1326,7 @@ void luks_decrypt_sector_serpent_cbc_plain256_mk_sha1 (__global const u32 *in, u
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1382,7 +1382,7 @@ void luks_decrypt_sector_serpent_cbc_plain256_mk_sha1_final (__global const u32
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1415,7 +1415,7 @@ void luks_decrypt_sector_serpent_cbc_plain256_mk_sha256 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1471,7 +1471,7 @@ void luks_decrypt_sector_serpent_cbc_plain256_mk_sha256_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1504,7 +1504,7 @@ void luks_decrypt_sector_serpent_cbc_plain256_mk_sha512 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1560,7 +1560,7 @@ void luks_decrypt_sector_serpent_cbc_plain256_mk_sha512_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1593,7 +1593,7 @@ void luks_decrypt_sector_serpent_cbc_plain256_mk_ripemd160 (__global const u32 *
   }
 }
 
-void luks_decrypt_sector_serpent_cbc_plain256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_cbc_plain256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1651,7 +1651,7 @@ void luks_decrypt_sector_serpent_cbc_plain256_mk_ripemd160_final (__global const
 
 // xts-plain
 
-void serpent128_decrypt_xts (const u32 *ks1, const u32 *in, u32 *out, u32 *T)
+DECLSPEC void serpent128_decrypt_xts (const u32 *ks1, const u32 *in, u32 *out, u32 *T)
 {
   out[0] = in[0];
   out[1] = in[1];
@@ -1673,7 +1673,7 @@ void serpent128_decrypt_xts (const u32 *ks1, const u32 *in, u32 *out, u32 *T)
   xts_mul2 (T, T);
 }
 
-void serpent256_decrypt_xts (const u32 *ks1, const u32 *in, u32 *out, u32 *T)
+DECLSPEC void serpent256_decrypt_xts (const u32 *ks1, const u32 *in, u32 *out, u32 *T)
 {
   out[0] = in[0];
   out[1] = in[1];
@@ -1695,7 +1695,7 @@ void serpent256_decrypt_xts (const u32 *ks1, const u32 *in, u32 *out, u32 *T)
   xts_mul2 (T, T);
 }
 
-void luks_decrypt_sector_serpent_xts_plain256 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain256 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1729,7 +1729,7 @@ void luks_decrypt_sector_serpent_xts_plain256 (__global const u32 *in, u32 *out,
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1766,7 +1766,7 @@ void luks_decrypt_sector_serpent_xts_plain256_mk_sha1 (__global const u32 *in, u
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1826,7 +1826,7 @@ void luks_decrypt_sector_serpent_xts_plain256_mk_sha1_final (__global const u32
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1863,7 +1863,7 @@ void luks_decrypt_sector_serpent_xts_plain256_mk_sha256 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1923,7 +1923,7 @@ void luks_decrypt_sector_serpent_xts_plain256_mk_sha256_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1960,7 +1960,7 @@ void luks_decrypt_sector_serpent_xts_plain256_mk_sha512 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2020,7 +2020,7 @@ void luks_decrypt_sector_serpent_xts_plain256_mk_sha512_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2057,7 +2057,7 @@ void luks_decrypt_sector_serpent_xts_plain256_mk_ripemd160 (__global const u32 *
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2117,7 +2117,7 @@ void luks_decrypt_sector_serpent_xts_plain256_mk_ripemd160_final (__global const
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain512 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain512 (__global const u32 *in, u32 *out, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2151,7 +2151,7 @@ void luks_decrypt_sector_serpent_xts_plain512 (__global const u32 *in, u32 *out,
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain512_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain512_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2188,7 +2188,7 @@ void luks_decrypt_sector_serpent_xts_plain512_mk_sha1 (__global const u32 *in, u
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain512_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain512_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2248,7 +2248,7 @@ void luks_decrypt_sector_serpent_xts_plain512_mk_sha1_final (__global const u32
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain512_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain512_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2285,7 +2285,7 @@ void luks_decrypt_sector_serpent_xts_plain512_mk_sha256 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain512_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain512_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2345,7 +2345,7 @@ void luks_decrypt_sector_serpent_xts_plain512_mk_sha256_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain512_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain512_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2382,7 +2382,7 @@ void luks_decrypt_sector_serpent_xts_plain512_mk_sha512 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain512_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain512_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2442,7 +2442,7 @@ void luks_decrypt_sector_serpent_xts_plain512_mk_sha512_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain512_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain512_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2479,7 +2479,7 @@ void luks_decrypt_sector_serpent_xts_plain512_mk_ripemd160 (__global const u32 *
   }
 }
 
-void luks_decrypt_sector_serpent_xts_plain512_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
+DECLSPEC void luks_decrypt_sector_serpent_xts_plain512_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *ks1, const u32 *ks2, volatile const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2541,7 +2541,7 @@ void luks_decrypt_sector_serpent_xts_plain512_mk_ripemd160_final (__global const
 
 // luks helper
 
-void luks_af_sha1_then_serpent_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
+DECLSPEC void luks_af_sha1_then_serpent_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
 {
   const u32 key_size    = luks_bufs->key_size;
   const u32 cipher_mode = luks_bufs->cipher_mode;
@@ -2900,7 +2900,7 @@ void luks_af_sha1_then_serpent_decrypt (__global const luks_t *luks_bufs, __glob
   }
 }
 
-void luks_af_sha256_then_serpent_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
+DECLSPEC void luks_af_sha256_then_serpent_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
 {
   const u32 key_size    = luks_bufs->key_size;
   const u32 cipher_mode = luks_bufs->cipher_mode;
@@ -3259,7 +3259,7 @@ void luks_af_sha256_then_serpent_decrypt (__global const luks_t *luks_bufs, __gl
   }
 }
 
-void luks_af_sha512_then_serpent_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
+DECLSPEC void luks_af_sha512_then_serpent_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
 {
   const u32 key_size    = luks_bufs->key_size;
   const u32 cipher_mode = luks_bufs->cipher_mode;
@@ -3637,7 +3637,7 @@ void luks_af_sha512_then_serpent_decrypt (__global const luks_t *luks_bufs, __gl
   }
 }
 
-void luks_af_ripemd160_then_serpent_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
+DECLSPEC void luks_af_ripemd160_then_serpent_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
 {
   const u32 key_size    = luks_bufs->key_size;
   const u32 cipher_mode = luks_bufs->cipher_mode;
diff --git a/OpenCL/inc_luks_twofish.cl b/OpenCL/inc_luks_twofish.cl
index 51a01a0df..e8349c87e 100644
--- a/OpenCL/inc_luks_twofish.cl
+++ b/OpenCL/inc_luks_twofish.cl
@@ -1,7 +1,7 @@
 
 // cbc-essiv
 
-void twofish128_decrypt_cbc (const u32 *sk1, const u32 *lk1, const u32 *in, u32 *out, u32 *essiv)
+DECLSPEC void twofish128_decrypt_cbc (const u32 *sk1, const u32 *lk1, const u32 *in, u32 *out, u32 *essiv)
 {
   twofish128_decrypt (sk1, lk1, in, out);
 
@@ -16,7 +16,7 @@ void twofish128_decrypt_cbc (const u32 *sk1, const u32 *lk1, const u32 *in, u32
   essiv[3] = in[3];
 }
 
-void twofish256_decrypt_cbc (const u32 *sk1, const u32 *lk1, const u32 *in, u32 *out, u32 *essiv)
+DECLSPEC void twofish256_decrypt_cbc (const u32 *sk1, const u32 *lk1, const u32 *in, u32 *out, u32 *essiv)
 {
   twofish256_decrypt (sk1, lk1, in, out);
 
@@ -31,7 +31,7 @@ void twofish256_decrypt_cbc (const u32 *sk1, const u32 *lk1, const u32 *in, u32
   essiv[3] = in[3];
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv128 (__global const u32 *in, u32 *out, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv128 (__global const u32 *in, u32 *out, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -65,7 +65,7 @@ void luks_decrypt_sector_twofish_cbc_essiv128 (__global const u32 *in, u32 *out,
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -102,7 +102,7 @@ void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha1 (__global const u32 *in, u
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -162,7 +162,7 @@ void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha1_final (__global const u32
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -199,7 +199,7 @@ void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha256 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -259,7 +259,7 @@ void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha256_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -296,7 +296,7 @@ void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha512 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -356,7 +356,7 @@ void luks_decrypt_sector_twofish_cbc_essiv128_mk_sha512_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv128_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv128_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -393,7 +393,7 @@ void luks_decrypt_sector_twofish_cbc_essiv128_mk_ripemd160 (__global const u32 *
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv128_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv128_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -453,7 +453,7 @@ void luks_decrypt_sector_twofish_cbc_essiv128_mk_ripemd160_final (__global const
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv256 (__global const u32 *in, u32 *out, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv256 (__global const u32 *in, u32 *out, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -487,7 +487,7 @@ void luks_decrypt_sector_twofish_cbc_essiv256 (__global const u32 *in, u32 *out,
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -524,7 +524,7 @@ void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha1 (__global const u32 *in, u
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -584,7 +584,7 @@ void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha1_final (__global const u32
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -621,7 +621,7 @@ void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha256 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -681,7 +681,7 @@ void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha256_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -718,7 +718,7 @@ void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha512 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -778,7 +778,7 @@ void luks_decrypt_sector_twofish_cbc_essiv256_mk_sha512_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -815,7 +815,7 @@ void luks_decrypt_sector_twofish_cbc_essiv256_mk_ripemd160 (__global const u32 *
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_essiv256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_essiv256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -877,7 +877,7 @@ void luks_decrypt_sector_twofish_cbc_essiv256_mk_ripemd160_final (__global const
 
 // cbc-plain
 
-void luks_decrypt_sector_twofish_cbc_plain128 (__global const u32 *in, u32 *out, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain128 (__global const u32 *in, u32 *out, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -907,7 +907,7 @@ void luks_decrypt_sector_twofish_cbc_plain128 (__global const u32 *in, u32 *out,
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain128_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain128_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -940,7 +940,7 @@ void luks_decrypt_sector_twofish_cbc_plain128_mk_sha1 (__global const u32 *in, u
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain128_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain128_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -996,7 +996,7 @@ void luks_decrypt_sector_twofish_cbc_plain128_mk_sha1_final (__global const u32
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain128_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain128_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1029,7 +1029,7 @@ void luks_decrypt_sector_twofish_cbc_plain128_mk_sha256 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain128_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain128_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1085,7 +1085,7 @@ void luks_decrypt_sector_twofish_cbc_plain128_mk_sha256_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain128_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain128_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1118,7 +1118,7 @@ void luks_decrypt_sector_twofish_cbc_plain128_mk_sha512 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain128_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain128_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1174,7 +1174,7 @@ void luks_decrypt_sector_twofish_cbc_plain128_mk_sha512_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain128_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain128_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1207,7 +1207,7 @@ void luks_decrypt_sector_twofish_cbc_plain128_mk_ripemd160 (__global const u32 *
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain128_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain128_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1263,7 +1263,7 @@ void luks_decrypt_sector_twofish_cbc_plain128_mk_ripemd160_final (__global const
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain256 (__global const u32 *in, u32 *out, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain256 (__global const u32 *in, u32 *out, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1293,7 +1293,7 @@ void luks_decrypt_sector_twofish_cbc_plain256 (__global const u32 *in, u32 *out,
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1326,7 +1326,7 @@ void luks_decrypt_sector_twofish_cbc_plain256_mk_sha1 (__global const u32 *in, u
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1382,7 +1382,7 @@ void luks_decrypt_sector_twofish_cbc_plain256_mk_sha1_final (__global const u32
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1415,7 +1415,7 @@ void luks_decrypt_sector_twofish_cbc_plain256_mk_sha256 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1471,7 +1471,7 @@ void luks_decrypt_sector_twofish_cbc_plain256_mk_sha256_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1504,7 +1504,7 @@ void luks_decrypt_sector_twofish_cbc_plain256_mk_sha512 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1560,7 +1560,7 @@ void luks_decrypt_sector_twofish_cbc_plain256_mk_sha512_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1593,7 +1593,7 @@ void luks_decrypt_sector_twofish_cbc_plain256_mk_ripemd160 (__global const u32 *
   }
 }
 
-void luks_decrypt_sector_twofish_cbc_plain256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_cbc_plain256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1651,7 +1651,7 @@ void luks_decrypt_sector_twofish_cbc_plain256_mk_ripemd160_final (__global const
 
 // xts-plain
 
-void twofish128_decrypt_xts (const u32 *sk1, const u32 *lk1, const u32 *in, u32 *out, u32 *T)
+DECLSPEC void twofish128_decrypt_xts (const u32 *sk1, const u32 *lk1, const u32 *in, u32 *out, u32 *T)
 {
   out[0] = in[0];
   out[1] = in[1];
@@ -1673,7 +1673,7 @@ void twofish128_decrypt_xts (const u32 *sk1, const u32 *lk1, const u32 *in, u32
   xts_mul2 (T, T);
 }
 
-void twofish256_decrypt_xts (const u32 *sk1, const u32 *lk1, const u32 *in, u32 *out, u32 *T)
+DECLSPEC void twofish256_decrypt_xts (const u32 *sk1, const u32 *lk1, const u32 *in, u32 *out, u32 *T)
 {
   out[0] = in[0];
   out[1] = in[1];
@@ -1695,7 +1695,7 @@ void twofish256_decrypt_xts (const u32 *sk1, const u32 *lk1, const u32 *in, u32
   xts_mul2 (T, T);
 }
 
-void luks_decrypt_sector_twofish_xts_plain256 (__global const u32 *in, u32 *out, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain256 (__global const u32 *in, u32 *out, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1729,7 +1729,7 @@ void luks_decrypt_sector_twofish_xts_plain256 (__global const u32 *in, u32 *out,
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain256_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1766,7 +1766,7 @@ void luks_decrypt_sector_twofish_xts_plain256_mk_sha1 (__global const u32 *in, u
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain256_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1826,7 +1826,7 @@ void luks_decrypt_sector_twofish_xts_plain256_mk_sha1_final (__global const u32
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain256_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1863,7 +1863,7 @@ void luks_decrypt_sector_twofish_xts_plain256_mk_sha256 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain256_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1923,7 +1923,7 @@ void luks_decrypt_sector_twofish_xts_plain256_mk_sha256_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain256_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -1960,7 +1960,7 @@ void luks_decrypt_sector_twofish_xts_plain256_mk_sha512 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain256_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2020,7 +2020,7 @@ void luks_decrypt_sector_twofish_xts_plain256_mk_sha512_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain256_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2057,7 +2057,7 @@ void luks_decrypt_sector_twofish_xts_plain256_mk_ripemd160 (__global const u32 *
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain256_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2117,7 +2117,7 @@ void luks_decrypt_sector_twofish_xts_plain256_mk_ripemd160_final (__global const
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain512 (__global const u32 *in, u32 *out, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain512 (__global const u32 *in, u32 *out, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2151,7 +2151,7 @@ void luks_decrypt_sector_twofish_xts_plain512 (__global const u32 *in, u32 *out,
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain512_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain512_mk_sha1 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2188,7 +2188,7 @@ void luks_decrypt_sector_twofish_xts_plain512_mk_sha1 (__global const u32 *in, u
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain512_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain512_mk_sha1_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2248,7 +2248,7 @@ void luks_decrypt_sector_twofish_xts_plain512_mk_sha1_final (__global const u32
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain512_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain512_mk_sha256 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2285,7 +2285,7 @@ void luks_decrypt_sector_twofish_xts_plain512_mk_sha256 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain512_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain512_mk_sha256_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2345,7 +2345,7 @@ void luks_decrypt_sector_twofish_xts_plain512_mk_sha256_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain512_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain512_mk_sha512 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2382,7 +2382,7 @@ void luks_decrypt_sector_twofish_xts_plain512_mk_sha512 (__global const u32 *in,
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain512_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain512_mk_sha512_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2442,7 +2442,7 @@ void luks_decrypt_sector_twofish_xts_plain512_mk_sha512_final (__global const u3
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain512_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain512_mk_ripemd160 (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2479,7 +2479,7 @@ void luks_decrypt_sector_twofish_xts_plain512_mk_ripemd160 (__global const u32 *
   }
 }
 
-void luks_decrypt_sector_twofish_xts_plain512_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
+DECLSPEC void luks_decrypt_sector_twofish_xts_plain512_mk_ripemd160_final (__global const u32 *in, u32 *mk, const u32 *sk1, const u32 *lk1, const u32 *sk2, const u32 *lk2, const u32 sector)
 {
   u32 S[4] = { sector, 0, 0, 0 };
 
@@ -2541,7 +2541,7 @@ void luks_decrypt_sector_twofish_xts_plain512_mk_ripemd160_final (__global const
 
 // luks helper
 
-void luks_af_sha1_then_twofish_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
+DECLSPEC void luks_af_sha1_then_twofish_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
 {
   const u32 key_size    = luks_bufs->key_size;
   const u32 cipher_mode = luks_bufs->cipher_mode;
@@ -2900,7 +2900,7 @@ void luks_af_sha1_then_twofish_decrypt (__global const luks_t *luks_bufs, __glob
   }
 }
 
-void luks_af_sha256_then_twofish_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
+DECLSPEC void luks_af_sha256_then_twofish_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
 {
   const u32 key_size    = luks_bufs->key_size;
   const u32 cipher_mode = luks_bufs->cipher_mode;
@@ -3259,7 +3259,7 @@ void luks_af_sha256_then_twofish_decrypt (__global const luks_t *luks_bufs, __gl
   }
 }
 
-void luks_af_sha512_then_twofish_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
+DECLSPEC void luks_af_sha512_then_twofish_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
 {
   const u32 key_size    = luks_bufs->key_size;
   const u32 cipher_mode = luks_bufs->cipher_mode;
@@ -3637,7 +3637,7 @@ void luks_af_sha512_then_twofish_decrypt (__global const luks_t *luks_bufs, __gl
   }
 }
 
-void luks_af_ripemd160_then_twofish_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
+DECLSPEC void luks_af_ripemd160_then_twofish_decrypt (__global const luks_t *luks_bufs, __global luks_tmp_t *tmps, u32 *pt_buf)
 {
   const u32 key_size    = luks_bufs->key_size;
   const u32 cipher_mode = luks_bufs->cipher_mode;
diff --git a/OpenCL/inc_luks_xts.cl b/OpenCL/inc_luks_xts.cl
index 1af200181..c2a36937e 100644
--- a/OpenCL/inc_luks_xts.cl
+++ b/OpenCL/inc_luks_xts.cl
@@ -1,4 +1,4 @@
-void xts_mul2 (u32 *in, u32 *out)
+DECLSPEC void xts_mul2 (u32 *in, u32 *out)
 {
   const u32 c = in[3] >> 31;
 
diff --git a/OpenCL/inc_rp.cl b/OpenCL/inc_rp.cl
index 25ab5b62a..78ea98fdd 100644
--- a/OpenCL/inc_rp.cl
+++ b/OpenCL/inc_rp.cl
@@ -5,7 +5,7 @@
 
 #define MAYBE_UNUSED
 
-u32 generate_cmask (const u32 value)
+DECLSPEC u32 generate_cmask (const u32 value)
 {
   const u32 rmask =  ((value & 0x40404040u) >> 1u)
                   & ~((value & 0x80808080u) >> 2u);
@@ -16,7 +16,7 @@ u32 generate_cmask (const u32 value)
   return rmask & ~hmask & lmask;
 }
 
-void append_four_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, const int off_dst)
+DECLSPEC void append_four_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, const int off_dst)
 {
   const int sd  = off_src / 4;
   const int sm  = off_src & 3;
@@ -39,7 +39,7 @@ void append_four_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, cons
   buf_dst[dd + 1] |= t1;
 }
 
-void append_three_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, const int off_dst)
+DECLSPEC void append_three_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, const int off_dst)
 {
   const int sd  = off_src / 4;
   const int sm  = off_src & 3;
@@ -62,7 +62,7 @@ void append_three_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, con
   buf_dst[dd + 1] |= t1;
 }
 
-void append_two_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, const int off_dst)
+DECLSPEC void append_two_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, const int off_dst)
 {
   const int sd  = off_src / 4;
   const int sm  = off_src & 3;
@@ -85,7 +85,7 @@ void append_two_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, const
   buf_dst[dd + 1] |= t1;
 }
 
-void append_one_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, const int off_dst)
+DECLSPEC void append_one_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, const int off_dst)
 {
   const int sd  = off_src / 4;
   const int sm  = off_src & 3;
@@ -104,7 +104,7 @@ void append_one_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, const
   buf_dst[dd] |= t;
 }
 
-void append_block (const u32 *buf_src, const int off_src, u32 *buf_dst, const int off_dst, const int len)
+DECLSPEC void append_block (const u32 *buf_src, const int off_src, u32 *buf_dst, const int off_dst, const int len)
 {
   int i;
 
@@ -123,7 +123,7 @@ void append_block (const u32 *buf_src, const int off_src, u32 *buf_dst, const in
   }
 }
 
-void exchange_byte (u32 *buf, const int off_src, const int off_dst)
+DECLSPEC void exchange_byte (u32 *buf, const int off_src, const int off_dst)
 {
   u8 *ptr = (u8 *) buf;
 
@@ -162,7 +162,7 @@ void exchange_byte (u32 *buf, const int off_src, const int off_dst)
   */
 }
 
-int mangle_lrest (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
+DECLSPEC int mangle_lrest (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
 {
   for (int i = 0, idx = 0; i < len; i += 4, idx += 1)
   {
@@ -174,7 +174,7 @@ int mangle_lrest (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf,
   return (len);
 }
 
-int mangle_lrest_ufirst (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
+DECLSPEC int mangle_lrest_ufirst (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
 {
   for (int i = 0, idx = 0; i < len; i += 4, idx += 1)
   {
@@ -190,7 +190,7 @@ int mangle_lrest_ufirst (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32
   return (len);
 }
 
-int mangle_urest (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
+DECLSPEC int mangle_urest (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
 {
   for (int i = 0, idx = 0; i < len; i += 4, idx += 1)
   {
@@ -202,7 +202,7 @@ int mangle_urest (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf,
   return (len);
 }
 
-int mangle_urest_lfirst (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
+DECLSPEC int mangle_urest_lfirst (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
 {
   for (int i = 0, idx = 0; i < len; i += 4, idx += 1)
   {
@@ -218,7 +218,7 @@ int mangle_urest_lfirst (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32
   return (len);
 }
 
-int mangle_trest (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
+DECLSPEC int mangle_trest (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
 {
   for (int i = 0, idx = 0; i < len; i += 4, idx += 1)
   {
@@ -230,7 +230,7 @@ int mangle_trest (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf,
   return (len);
 }
 
-int mangle_toggle_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
+DECLSPEC int mangle_toggle_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
 {
   if (p0 >= len) return (len);
 
@@ -246,7 +246,7 @@ int mangle_toggle_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *b
   return (len);
 }
 
-int mangle_reverse (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
+DECLSPEC int mangle_reverse (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
 {
   for (int l = 0; l < len / 2; l++)
   {
@@ -258,7 +258,7 @@ int mangle_reverse (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf
   return (len);
 }
 
-int mangle_dupeword (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
+DECLSPEC int mangle_dupeword (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
 {
   const int out_len = len * 2;
 
@@ -269,7 +269,7 @@ int mangle_dupeword (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *bu
   return (out_len);
 }
 
-int mangle_dupeword_times (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_dupeword_times (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   const int out_len = (len * p0) + len;
 
@@ -282,7 +282,7 @@ int mangle_dupeword_times (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u
   return (out_len);
 }
 
-int mangle_reflect (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
+DECLSPEC int mangle_reflect (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
 {
   const int out_len = len * 2;
 
@@ -300,7 +300,7 @@ int mangle_reflect (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf
   return out_len;
 }
 
-int mangle_append (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_append (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   const int out_len = len + 1;
 
@@ -311,7 +311,7 @@ int mangle_append (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf,
   return (out_len);
 }
 
-int mangle_prepend (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_prepend (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   const int out_len = len + 1;
 
@@ -327,7 +327,7 @@ int mangle_prepend (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf,
   return (out_len);
 }
 
-int mangle_rotate_left (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
+DECLSPEC int mangle_rotate_left (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
 {
   for (int l = 0, r = len - 1; r > l; r--)
   {
@@ -337,7 +337,7 @@ int mangle_rotate_left (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32
   return (len);
 }
 
-int mangle_rotate_right (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
+DECLSPEC int mangle_rotate_right (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
 {
   for (int l = 0, r = len - 1; l < r; l++)
   {
@@ -347,7 +347,7 @@ int mangle_rotate_right (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32
   return (len);
 }
 
-int mangle_delete_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_delete_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   if (p0 >= len) return (len);
 
@@ -361,19 +361,19 @@ int mangle_delete_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *bu
   return (len - 1);
 }
 
-int mangle_delete_first (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_delete_first (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   return mangle_delete_at (0, p1, buf, len);
 }
 
-int mangle_delete_last (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_delete_last (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   if (len == 0) return 0;
 
   return mangle_delete_at (len - 1, p1, buf, len);
 }
 
-int mangle_extract (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_extract (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   if (p0 >= len) return (len);
 
@@ -392,7 +392,7 @@ int mangle_extract (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf,
   return (p1);
 }
 
-int mangle_omit (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_omit (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   if (p0 >= len) return (len);
 
@@ -411,7 +411,7 @@ int mangle_omit (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, co
   return (len - p1);
 }
 
-int mangle_insert (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_insert (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   if (p0 >= len + 1) return (len);
 
@@ -429,7 +429,7 @@ int mangle_insert (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf,
   return (out_len);
 }
 
-int mangle_overstrike (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_overstrike (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   if (p0 >= len) return (len);
 
@@ -438,7 +438,7 @@ int mangle_overstrike (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *b
   return (len);
 }
 
-int mangle_truncate_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_truncate_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   if (p0 >= len) return (len);
 
@@ -450,7 +450,7 @@ int mangle_truncate_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *
   return (p0);
 }
 
-int mangle_replace (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_replace (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   for (int pos = 0; pos < len; pos++)
   {
@@ -462,7 +462,7 @@ int mangle_replace (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf,
   return (len);
 }
 
-int mangle_purgechar (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_purgechar (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   int out_len = 0;
 
@@ -483,7 +483,7 @@ int mangle_purgechar (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *bu
   return (out_len);
 }
 
-int mangle_dupechar_first (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_dupechar_first (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   const int out_len = len + p0;
 
@@ -499,7 +499,7 @@ int mangle_dupechar_first (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u
   return (out_len);
 }
 
-int mangle_dupechar_last (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_dupechar_last (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   const int out_len = len + p0;
 
@@ -516,7 +516,7 @@ int mangle_dupechar_last (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8
   return (out_len);
 }
 
-int mangle_dupechar_all (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_dupechar_all (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   const int out_len = len + len;
 
@@ -534,7 +534,7 @@ int mangle_dupechar_all (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8
   return (out_len);
 }
 
-int mangle_switch_first (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
+DECLSPEC int mangle_switch_first (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
 {
   if (len < 2) return (len);
 
@@ -543,7 +543,7 @@ int mangle_switch_first (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32
   return (len);
 }
 
-int mangle_switch_last (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
+DECLSPEC int mangle_switch_last (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
 {
   if (len < 2) return (len);
 
@@ -552,7 +552,7 @@ int mangle_switch_last (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32
   return (len);
 }
 
-int mangle_switch_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
+DECLSPEC int mangle_switch_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
 {
   if (p0 >= len) return (len);
   if (p1 >= len) return (len);
@@ -562,7 +562,7 @@ int mangle_switch_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *b
   return (len);
 }
 
-int mangle_chr_shiftl (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_chr_shiftl (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   if (p0 >= len) return (len);
 
@@ -571,7 +571,7 @@ int mangle_chr_shiftl (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *b
   return (len);
 }
 
-int mangle_chr_shiftr (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_chr_shiftr (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   if (p0 >= len) return (len);
 
@@ -580,7 +580,7 @@ int mangle_chr_shiftr (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *b
   return (len);
 }
 
-int mangle_chr_incr (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_chr_incr (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   if (p0 >= len) return (len);
 
@@ -589,7 +589,7 @@ int mangle_chr_incr (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf
   return (len);
 }
 
-int mangle_chr_decr (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_chr_decr (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   if (p0 >= len) return (len);
 
@@ -598,7 +598,7 @@ int mangle_chr_decr (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf
   return (len);
 }
 
-int mangle_replace_np1 (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_replace_np1 (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   if ((p0 + 1) >= len) return (len);
 
@@ -607,7 +607,7 @@ int mangle_replace_np1 (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *
   return (len);
 }
 
-int mangle_replace_nm1 (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_replace_nm1 (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   if (p0 == 0) return (len);
 
@@ -618,7 +618,7 @@ int mangle_replace_nm1 (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *
   return (len);
 }
 
-int mangle_dupeblock_first (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_dupeblock_first (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   if (p0 >= len) return (len);
 
@@ -636,7 +636,7 @@ int mangle_dupeblock_first (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1,
   return (out_len);
 }
 
-int mangle_dupeblock_last (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
+DECLSPEC int mangle_dupeblock_last (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len)
 {
   if (p0 >= len) return (len);
 
@@ -654,7 +654,7 @@ int mangle_dupeblock_last (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u
   return (out_len);
 }
 
-int mangle_title_sep (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
+DECLSPEC int mangle_title_sep (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len)
 {
   if ((len + 4) >= RP_PASSWORD_SIZE) return (len); // cheap way to not need to check for overflow of i + 1
 
@@ -679,7 +679,7 @@ int mangle_title_sep (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *b
   return (len);
 }
 
-int apply_rule (const u32 name, MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int in_len)
+DECLSPEC int apply_rule (const u32 name, MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int in_len)
 {
   int out_len = in_len;
 
@@ -730,7 +730,7 @@ int apply_rule (const u32 name, MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8
   return out_len;
 }
 
-int apply_rules (__constant const u32 *cmds, u32 *buf, const int in_len)
+DECLSPEC int apply_rules (__constant const u32 *cmds, u32 *buf, const int in_len)
 {
   int out_len = in_len;
 
diff --git a/OpenCL/inc_rp_optimized.cl b/OpenCL/inc_rp_optimized.cl
index b4062fac1..abda80575 100644
--- a/OpenCL/inc_rp_optimized.cl
+++ b/OpenCL/inc_rp_optimized.cl
@@ -5,7 +5,7 @@
 
 #define MAYBE_UNUSED
 
-u32 generate_cmask (const u32 value)
+DECLSPEC u32 generate_cmask (const u32 value)
 {
   const u32 rmask =  ((value & 0x40404040u) >> 1u)
                   & ~((value & 0x80808080u) >> 2u);
@@ -16,7 +16,7 @@ u32 generate_cmask (const u32 value)
   return rmask & ~hmask & lmask;
 }
 
-void truncate_right (u32 buf0[4], u32 buf1[4], const u32 offset)
+DECLSPEC void truncate_right (u32 buf0[4], u32 buf1[4], const u32 offset)
 {
   const u32 tmp = (1u << ((offset & 3u) * 8u)) - 1u;
 
@@ -75,7 +75,7 @@ void truncate_right (u32 buf0[4], u32 buf1[4], const u32 offset)
   }
 }
 
-void truncate_left (u32 buf0[4], u32 buf1[4], const u32 offset)
+DECLSPEC void truncate_left (u32 buf0[4], u32 buf1[4], const u32 offset)
 {
   const u32 tmp = ~((1u << ((offset & 3u) * 8u)) - 1u);
 
@@ -134,7 +134,7 @@ void truncate_left (u32 buf0[4], u32 buf1[4], const u32 offset)
   }
 }
 
-void lshift_block (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4])
+DECLSPEC void lshift_block (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4])
 {
   out0[0] = amd_bytealign_S (in0[1], in0[0], 1);
   out0[1] = amd_bytealign_S (in0[2], in0[1], 1);
@@ -146,7 +146,7 @@ void lshift_block (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4])
   out1[3] = amd_bytealign_S (     0, in1[3], 1);
 }
 
-void rshift_block (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4])
+DECLSPEC void rshift_block (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4])
 {
   out1[3] = amd_bytealign_S (in1[3], in1[2], 3);
   out1[2] = amd_bytealign_S (in1[2], in1[1], 3);
@@ -158,7 +158,7 @@ void rshift_block (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4])
   out0[0] = amd_bytealign_S (in0[0],      0, 3);
 }
 
-void lshift_block_N (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4], const u32 num)
+DECLSPEC void lshift_block_N (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4], const u32 num)
 {
   switch (num)
   {
@@ -453,7 +453,7 @@ void lshift_block_N (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4
   }
 }
 
-void rshift_block_N (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4], const u32 num)
+DECLSPEC void rshift_block_N (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4], const u32 num)
 {
   switch (num)
   {
@@ -748,7 +748,7 @@ void rshift_block_N (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4
   }
 }
 
-void append_block1 (const u32 offset, u32 buf0[4], u32 buf1[4], const u32 src_r0)
+DECLSPEC void append_block1 (const u32 offset, u32 buf0[4], u32 buf1[4], const u32 src_r0)
 {
   // this version works with 1 byte append only
   const u32 value = src_r0 & 0xff;
@@ -768,7 +768,7 @@ void append_block1 (const u32 offset, u32 buf0[4], u32 buf1[4], const u32 src_r0
   buf1[3] |= tmp & c_append_helper[offset][7];
 }
 
-void append_block8 (const u32 offset, u32 buf0[4], u32 buf1[4], const u32 src_l0[4], const u32 src_l1[4], const u32 src_r0[4], const u32 src_r1[4])
+DECLSPEC void append_block8 (const u32 offset, u32 buf0[4], u32 buf1[4], const u32 src_l0[4], const u32 src_l1[4], const u32 src_r0[4], const u32 src_r1[4])
 {
   u32 s0 = 0;
   u32 s1 = 0;
@@ -1023,7 +1023,7 @@ void append_block8 (const u32 offset, u32 buf0[4], u32 buf1[4], const u32 src_l0
   buf1[3] = src_l1[3] | s7;
 }
 
-void reverse_block (u32 in0[4], u32 in1[4], u32 out0[4], u32 out1[4], const u32 len)
+DECLSPEC void reverse_block (u32 in0[4], u32 in1[4], u32 out0[4], u32 out1[4], const u32 len)
 {
   rshift_block_N (in0, in1, out0, out1, 32 - len);
 
@@ -1049,7 +1049,7 @@ void reverse_block (u32 in0[4], u32 in1[4], u32 out0[4], u32 out1[4], const u32
   out1[3] = swap32_S (tib41[3]);
 }
 
-void exchange_byte (u32 *buf, const int off_src, const int off_dst)
+DECLSPEC void exchange_byte (u32 *buf, const int off_src, const int off_dst)
 {
   u8 *ptr = (u8 *) buf;
 
@@ -1059,7 +1059,7 @@ void exchange_byte (u32 *buf, const int off_src, const int off_dst)
   ptr[off_dst] = tmp;
 }
 
-u32 rule_op_mangle_lrest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_lrest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   u32 t;
 
@@ -1075,7 +1075,7 @@ u32 rule_op_mangle_lrest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1,
   return in_len;
 }
 
-u32 rule_op_mangle_urest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_urest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   u32 t;
 
@@ -1091,7 +1091,7 @@ u32 rule_op_mangle_urest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1,
   return in_len;
 }
 
-u32 rule_op_mangle_lrest_ufirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_lrest_ufirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   u32 t;
 
@@ -1109,7 +1109,7 @@ u32 rule_op_mangle_lrest_ufirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u
   return in_len;
 }
 
-u32 rule_op_mangle_urest_lfirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_urest_lfirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   u32 t;
 
@@ -1127,7 +1127,7 @@ u32 rule_op_mangle_urest_lfirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u
   return in_len;
 }
 
-u32 rule_op_mangle_trest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_trest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   u32 t;
 
@@ -1143,7 +1143,7 @@ u32 rule_op_mangle_trest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1,
   return in_len;
 }
 
-u32 rule_op_mangle_toggle_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_toggle_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -1176,14 +1176,14 @@ u32 rule_op_mangle_toggle_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32
   return (in_len);
 }
 
-u32 rule_op_mangle_reverse (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_reverse (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   reverse_block (buf0, buf1, buf0, buf1, in_len);
 
   return in_len;
 }
 
-u32 rule_op_mangle_dupeword (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_dupeword (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if ((in_len + in_len) >= 32) return in_len;
 
@@ -1196,7 +1196,7 @@ u32 rule_op_mangle_dupeword (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p
   return out_len;
 }
 
-u32 rule_op_mangle_dupeword_times (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_dupeword_times (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (((in_len * p0) + in_len) >= 32) return in_len;
 
@@ -1224,7 +1224,7 @@ u32 rule_op_mangle_dupeword_times (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const
   return out_len;
 }
 
-u32 rule_op_mangle_reflect (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_reflect (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if ((in_len + in_len) >= 32) return in_len;
 
@@ -1242,7 +1242,7 @@ u32 rule_op_mangle_reflect (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1
   return out_len;
 }
 
-u32 rule_op_mangle_append (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_append (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if ((in_len + 1) >= 32) return in_len;
 
@@ -1255,7 +1255,7 @@ u32 rule_op_mangle_append (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1,
   return out_len;
 }
 
-u32 rule_op_mangle_prepend (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_prepend (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if ((in_len + 1) >= 32) return in_len;
 
@@ -1270,7 +1270,7 @@ u32 rule_op_mangle_prepend (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1
   return out_len;
 }
 
-u32 rule_op_mangle_rotate_left (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_rotate_left (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (in_len == 0) return in_len;
 
@@ -1285,7 +1285,7 @@ u32 rule_op_mangle_rotate_left (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u3
   return in_len;
 }
 
-u32 rule_op_mangle_rotate_right (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_rotate_right (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (in_len == 0) return in_len;
 
@@ -1315,7 +1315,7 @@ u32 rule_op_mangle_rotate_right (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u
   return in_len;
 }
 
-u32 rule_op_mangle_delete_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_delete_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (in_len == 0) return in_len;
 
@@ -1326,7 +1326,7 @@ u32 rule_op_mangle_delete_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u
   return in_len1;
 }
 
-u32 rule_op_mangle_delete_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_delete_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (in_len == 0) return in_len;
 
@@ -1346,7 +1346,7 @@ u32 rule_op_mangle_delete_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u3
   return in_len1;
 }
 
-u32 rule_op_mangle_delete_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_delete_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -1427,7 +1427,7 @@ u32 rule_op_mangle_delete_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32
   return out_len;
 }
 
-u32 rule_op_mangle_extract (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_extract (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -1442,7 +1442,7 @@ u32 rule_op_mangle_extract (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1
   return out_len;
 }
 
-u32 rule_op_mangle_omit (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_omit (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -1534,7 +1534,7 @@ u32 rule_op_mangle_omit (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, M
   return out_len;
 }
 
-u32 rule_op_mangle_insert (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_insert (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (p0 > in_len) return in_len;
 
@@ -1612,7 +1612,7 @@ u32 rule_op_mangle_insert (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1,
   return out_len;
 }
 
-u32 rule_op_mangle_overstrike (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_overstrike (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -1647,7 +1647,7 @@ u32 rule_op_mangle_overstrike (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32
   return in_len;
 }
 
-u32 rule_op_mangle_truncate_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_truncate_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -1656,7 +1656,7 @@ u32 rule_op_mangle_truncate_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u3
   return p0;
 }
 
-u32 search_on_register (const u32 in, const u32 p0)
+DECLSPEC u32 search_on_register (const u32 in, const u32 p0)
 {
   u32 r = 0;
 
@@ -1668,7 +1668,7 @@ u32 search_on_register (const u32 in, const u32 p0)
   return r;
 }
 
-u32 replace_on_register (const u32 in, const u32 r, const u32 p1)
+DECLSPEC u32 replace_on_register (const u32 in, const u32 r, const u32 p1)
 {
   u32 out = in;
 
@@ -1680,7 +1680,7 @@ u32 replace_on_register (const u32 in, const u32 r, const u32 p1)
   return out;
 }
 
-u32 rule_op_mangle_replace (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_replace (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   const u32 r0 = search_on_register (buf0[0], p0);
   const u32 r1 = search_on_register (buf0[1], p0);
@@ -1707,7 +1707,7 @@ u32 rule_op_mangle_replace (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1
   return in_len;
 }
 
-u32 rule_op_mangle_purgechar (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_purgechar (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   const u32 r0 = search_on_register (buf0[0], p0);
   const u32 r1 = search_on_register (buf0[1], p0);
@@ -1761,7 +1761,7 @@ u32 rule_op_mangle_purgechar (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32
   return out_len;
 }
 
-u32 rule_op_mangle_dupechar_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_dupechar_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if ( in_len       ==  0) return in_len;
   if ((in_len + p0) >= 32) return in_len;
@@ -1796,7 +1796,7 @@ u32 rule_op_mangle_dupechar_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const
   return out_len;
 }
 
-u32 rule_op_mangle_dupechar_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_dupechar_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if ( in_len       ==  0) return in_len;
   if ((in_len + p0) >= 32) return in_len;
@@ -1830,7 +1830,7 @@ u32 rule_op_mangle_dupechar_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const
   return out_len;
 }
 
-u32 rule_op_mangle_dupechar_all (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_dupechar_all (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if ( in_len           ==  0) return in_len;
   if ((in_len + in_len) >= 32) return in_len;
@@ -1863,7 +1863,7 @@ u32 rule_op_mangle_dupechar_all (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u
   return out_len;
 }
 
-u32 rule_op_mangle_switch_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_switch_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (in_len < 2) return in_len;
 
@@ -1872,7 +1872,7 @@ u32 rule_op_mangle_switch_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u
   return in_len;
 }
 
-u32 rule_op_mangle_switch_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_switch_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (in_len < 2) return in_len;
 
@@ -1901,7 +1901,7 @@ u32 rule_op_mangle_switch_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u3
   return in_len;
 }
 
-u32 rule_op_mangle_switch_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_switch_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
   if (p1 >= in_len) return in_len;
@@ -1931,7 +1931,7 @@ u32 rule_op_mangle_switch_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32
   return in_len;
 }
 
-u32 rule_op_mangle_chr_shiftl (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_chr_shiftl (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -1965,7 +1965,7 @@ u32 rule_op_mangle_chr_shiftl (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32
   return in_len;
 }
 
-u32 rule_op_mangle_chr_shiftr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_chr_shiftr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -1999,7 +1999,7 @@ u32 rule_op_mangle_chr_shiftr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32
   return in_len;
 }
 
-u32 rule_op_mangle_chr_incr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_chr_incr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -2035,7 +2035,7 @@ u32 rule_op_mangle_chr_incr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p
   return in_len;
 }
 
-u32 rule_op_mangle_chr_decr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_chr_decr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -2071,7 +2071,7 @@ u32 rule_op_mangle_chr_decr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p
   return in_len;
 }
 
-u32 rule_op_mangle_replace_np1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_replace_np1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if ((p0 + 1) >= in_len) return in_len;
 
@@ -2111,7 +2111,7 @@ u32 rule_op_mangle_replace_np1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u3
   return in_len;
 }
 
-u32 rule_op_mangle_replace_nm1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_replace_nm1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (p0 == 0) return in_len;
 
@@ -2153,7 +2153,7 @@ u32 rule_op_mangle_replace_nm1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u3
   return in_len;
 }
 
-u32 rule_op_mangle_dupeblock_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_dupeblock_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (p0 > in_len) return in_len;
 
@@ -2191,7 +2191,7 @@ u32 rule_op_mangle_dupeblock_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED cons
   return out_len;
 }
 
-u32 rule_op_mangle_dupeblock_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_dupeblock_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (p0 > in_len) return in_len;
 
@@ -2220,7 +2220,7 @@ u32 rule_op_mangle_dupeblock_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const
   return out_len;
 }
 
-u32 toggle_on_register (const u32 in, const u32 r)
+DECLSPEC u32 toggle_on_register (const u32 in, const u32 r)
 {
   u32 out = in;
 
@@ -2234,7 +2234,7 @@ u32 toggle_on_register (const u32 in, const u32 r)
   return out;
 }
 
-u32 rule_op_mangle_title_sep (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
+DECLSPEC u32 rule_op_mangle_title_sep (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 buf0[4], MAYBE_UNUSED u32 buf1[4], const u32 in_len)
 {
   if (in_len == 0) return in_len;
 
@@ -2274,7 +2274,7 @@ u32 rule_op_mangle_title_sep (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32
   return in_len;
 }
 
-u32 apply_rule (const u32 name, const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
+DECLSPEC u32 apply_rule (const u32 name, const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
 {
   u32 out_len = in_len;
 
@@ -2326,7 +2326,7 @@ u32 apply_rule (const u32 name, const u32 p0, const u32 p1, u32 buf0[4], u32 buf
   return out_len;
 }
 
-u32 apply_rules (__constant const u32 *cmds, u32 buf0[4], u32 buf1[4], const u32 len)
+DECLSPEC u32 apply_rules (__constant const u32 *cmds, u32 buf0[4], u32 buf1[4], const u32 len)
 {
   u32 out_len = len;
 
@@ -2344,7 +2344,7 @@ u32 apply_rules (__constant const u32 *cmds, u32 buf0[4], u32 buf1[4], const u32
   return out_len;
 }
 
-u32x apply_rules_vect (const u32 pw_buf0[4], const u32 pw_buf1[4], const u32 pw_len, __constant const kernel_rule_t *rules_buf, const u32 il_pos, u32x buf0[4], u32x buf1[4])
+DECLSPEC u32x apply_rules_vect (const u32 pw_buf0[4], const u32 pw_buf1[4], const u32 pw_len, __constant const kernel_rule_t *rules_buf, const u32 il_pos, u32x buf0[4], u32x buf1[4])
 {
   #if VECT_SIZE == 1
 
diff --git a/OpenCL/inc_simd.cl b/OpenCL/inc_simd.cl
index 5aa920d09..2b3c02274 100644
--- a/OpenCL/inc_simd.cl
+++ b/OpenCL/inc_simd.cl
@@ -1048,7 +1048,7 @@
 
 // attack-mode 0
 
-u32x ix_create_bft (__global const bf_t *bfs_buf, const u32 il_pos)
+DECLSPEC u32x ix_create_bft (__global const bf_t *bfs_buf, const u32 il_pos)
 {
   #if   VECT_SIZE == 1
   const u32x ix = (u32x) (bfs_buf[il_pos + 0].i);
@@ -1067,7 +1067,7 @@ u32x ix_create_bft (__global const bf_t *bfs_buf, const u32 il_pos)
 
 // attack-mode 1
 
-u32x pwlenx_create_combt (__global const pw_t *combs_buf, const u32 il_pos)
+DECLSPEC u32x pwlenx_create_combt (__global const pw_t *combs_buf, const u32 il_pos)
 {
   #if   VECT_SIZE == 1
   const u32x pw_lenx = (u32x) (combs_buf[il_pos + 0].pw_len);
@@ -1084,7 +1084,7 @@ u32x pwlenx_create_combt (__global const pw_t *combs_buf, const u32 il_pos)
   return pw_lenx;
 }
 
-u32x ix_create_combt (__global const pw_t *combs_buf, const u32 il_pos, const int idx)
+DECLSPEC u32x ix_create_combt (__global const pw_t *combs_buf, const u32 il_pos, const int idx)
 {
   #if   VECT_SIZE == 1
   const u32x ix = (u32x) (combs_buf[il_pos + 0].i[idx]);
diff --git a/OpenCL/inc_truecrypt_crc32.cl b/OpenCL/inc_truecrypt_crc32.cl
index b8e464b74..7849e1fe7 100644
--- a/OpenCL/inc_truecrypt_crc32.cl
+++ b/OpenCL/inc_truecrypt_crc32.cl
@@ -66,7 +66,7 @@ __constant static u32a crc32tab[0x100] =
   0xb40bbe37, 0xc30c8ea1, 0x5a05df1b, 0x2d02ef8d
 };
 
-u32 round_crc32 (u32 a, const u32 v)
+DECLSPEC u32 round_crc32 (u32 a, const u32 v)
 {
   const u32 k = (a ^ v) & 0xff;
 
@@ -79,7 +79,7 @@ u32 round_crc32 (u32 a, const u32 v)
   return a;
 }
 
-u32 round_crc32_4 (const u32 w, const u32 iv)
+DECLSPEC u32 round_crc32_4 (const u32 w, const u32 iv)
 {
   u32 a = iv;
 
diff --git a/OpenCL/inc_truecrypt_xts.cl b/OpenCL/inc_truecrypt_xts.cl
index 6bf6684d2..bc6f319d0 100644
--- a/OpenCL/inc_truecrypt_xts.cl
+++ b/OpenCL/inc_truecrypt_xts.cl
@@ -1,4 +1,4 @@
-void xts_mul2 (u32 *in, u32 *out)
+DECLSPEC void xts_mul2 (u32 *in, u32 *out)
 {
   const u32 c = in[3] >> 31;
 
@@ -10,7 +10,7 @@ void xts_mul2 (u32 *in, u32 *out)
   out[0] ^= c * 0x87;
 }
 
-void aes256_decrypt_xts_first (const u32 *ukey1, const u32 *ukey2, const u32 *in, u32 *out, u32 *S, u32 *T, u32 *ks, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void aes256_decrypt_xts_first (const u32 *ukey1, const u32 *ukey2, const u32 *in, u32 *out, u32 *S, u32 *T, u32 *ks, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   out[0] = in[0];
   out[1] = in[1];
@@ -34,7 +34,7 @@ void aes256_decrypt_xts_first (const u32 *ukey1, const u32 *ukey2, const u32 *in
   out[3] ^= T[3];
 }
 
-void aes256_decrypt_xts_next (const u32 *in, u32 *out, u32 *T, u32 *ks, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC void aes256_decrypt_xts_next (const u32 *in, u32 *out, u32 *T, u32 *ks, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   out[0] = in[0];
   out[1] = in[1];
@@ -56,7 +56,7 @@ void aes256_decrypt_xts_next (const u32 *in, u32 *out, u32 *T, u32 *ks, SHM_TYPE
   out[3] ^= T[3];
 }
 
-void serpent256_decrypt_xts_first (const u32 *ukey1, const u32 *ukey2, const u32 *in, u32 *out, u32 *S, u32 *T, u32 *ks)
+DECLSPEC void serpent256_decrypt_xts_first (const u32 *ukey1, const u32 *ukey2, const u32 *in, u32 *out, u32 *S, u32 *T, u32 *ks)
 {
   out[0] = in[0];
   out[1] = in[1];
@@ -80,7 +80,7 @@ void serpent256_decrypt_xts_first (const u32 *ukey1, const u32 *ukey2, const u32
   out[3] ^= T[3];
 }
 
-void serpent256_decrypt_xts_next (const u32 *in, u32 *out, u32 *T, u32 *ks)
+DECLSPEC void serpent256_decrypt_xts_next (const u32 *in, u32 *out, u32 *T, u32 *ks)
 {
   out[0] = in[0];
   out[1] = in[1];
@@ -102,7 +102,7 @@ void serpent256_decrypt_xts_next (const u32 *in, u32 *out, u32 *T, u32 *ks)
   out[3] ^= T[3];
 }
 
-void twofish256_decrypt_xts_first (const u32 *ukey1, const u32 *ukey2, const u32 *in, u32 *out, u32 *S, u32 *T, u32 *sk, u32 *lk)
+DECLSPEC void twofish256_decrypt_xts_first (const u32 *ukey1, const u32 *ukey2, const u32 *in, u32 *out, u32 *S, u32 *T, u32 *sk, u32 *lk)
 {
   out[0] = in[0];
   out[1] = in[1];
@@ -126,7 +126,7 @@ void twofish256_decrypt_xts_first (const u32 *ukey1, const u32 *ukey2, const u32
   out[3] ^= T[3];
 }
 
-void twofish256_decrypt_xts_next (const u32 *in, u32 *out, u32 *T, u32 *sk, u32 *lk)
+DECLSPEC void twofish256_decrypt_xts_next (const u32 *in, u32 *out, u32 *T, u32 *sk, u32 *lk)
 {
   out[0] = in[0];
   out[1] = in[1];
@@ -150,7 +150,7 @@ void twofish256_decrypt_xts_next (const u32 *in, u32 *out, u32 *T, u32 *sk, u32
 
 // 512 bit
 
-int verify_header_aes (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC int verify_header_aes (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 ks_aes[60];
 
@@ -206,7 +206,7 @@ int verify_header_aes (__global const tc_t *esalt_bufs, const u32 *ukey1, const
   return 1;
 }
 
-int verify_header_serpent (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2)
+DECLSPEC int verify_header_serpent (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2)
 {
   u32 ks_serpent[140];
 
@@ -262,7 +262,7 @@ int verify_header_serpent (__global const tc_t *esalt_bufs, const u32 *ukey1, co
   return 1;
 }
 
-int verify_header_twofish (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2)
+DECLSPEC int verify_header_twofish (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2)
 {
   u32 sk_twofish[4];
   u32 lk_twofish[40];
@@ -321,7 +321,7 @@ int verify_header_twofish (__global const tc_t *esalt_bufs, const u32 *ukey1, co
 
 // 1024 bit
 
-int verify_header_aes_twofish (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2, const u32 *ukey3, const u32 *ukey4, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC int verify_header_aes_twofish (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2, const u32 *ukey3, const u32 *ukey4, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 ks_aes[60];
 
@@ -384,7 +384,7 @@ int verify_header_aes_twofish (__global const tc_t *esalt_bufs, const u32 *ukey1
   return 1;
 }
 
-int verify_header_serpent_aes (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2, const u32 *ukey3, const u32 *ukey4, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC int verify_header_serpent_aes (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2, const u32 *ukey3, const u32 *ukey4, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 ks_serpent[140];
   u32 ks_aes[60];
@@ -445,7 +445,7 @@ int verify_header_serpent_aes (__global const tc_t *esalt_bufs, const u32 *ukey1
   return 1;
 }
 
-int verify_header_twofish_serpent (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2, const u32 *ukey3, const u32 *ukey4)
+DECLSPEC int verify_header_twofish_serpent (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2, const u32 *ukey3, const u32 *ukey4)
 {
   u32 sk_twofish[4];
   u32 lk_twofish[40];
@@ -510,7 +510,7 @@ int verify_header_twofish_serpent (__global const tc_t *esalt_bufs, const u32 *u
 
 // 1536 bit
 
-int verify_header_aes_twofish_serpent (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2, const u32 *ukey3, const u32 *ukey4, const u32 *ukey5, const u32 *ukey6, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC int verify_header_aes_twofish_serpent (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2, const u32 *ukey3, const u32 *ukey4, const u32 *ukey5, const u32 *ukey6, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 ks_aes[60];
 
@@ -579,7 +579,7 @@ int verify_header_aes_twofish_serpent (__global const tc_t *esalt_bufs, const u3
   return 1;
 }
 
-int verify_header_serpent_twofish_aes (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2, const u32 *ukey3, const u32 *ukey4, const u32 *ukey5, const u32 *ukey6, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
+DECLSPEC int verify_header_serpent_twofish_aes (__global const tc_t *esalt_bufs, const u32 *ukey1, const u32 *ukey2, const u32 *ukey3, const u32 *ukey4, const u32 *ukey5, const u32 *ukey6, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
   u32 ks_serpent[140];
 
diff --git a/OpenCL/inc_types.cl b/OpenCL/inc_types.cl
index 248dfe74f..1fd549451 100644
--- a/OpenCL/inc_types.cl
+++ b/OpenCL/inc_types.cl
@@ -33,14 +33,14 @@ typedef VTYPE(uint,   VECT_SIZE) u32x;
 typedef VTYPE(ulong,  VECT_SIZE) u64x;
 #endif
 
-u32 l32_from_64_S (u64 a)
+DECLSPEC u32 l32_from_64_S (u64 a)
 {
   const u32 r = (u32) (a);
 
   return r;
 }
 
-u32 h32_from_64_S (u64 a)
+DECLSPEC u32 h32_from_64_S (u64 a)
 {
   a >>= 32;
 
@@ -49,12 +49,12 @@ u32 h32_from_64_S (u64 a)
   return r;
 }
 
-u64 hl32_to_64_S (const u32 a, const u32 b)
+DECLSPEC u64 hl32_to_64_S (const u32 a, const u32 b)
 {
   return as_ulong ((uint2) (b, a));
 }
 
-u32x l32_from_64 (u64x a)
+DECLSPEC u32x l32_from_64 (u64x a)
 {
   u32x r;
 
@@ -93,7 +93,7 @@ u32x l32_from_64 (u64x a)
   return r;
 }
 
-u32x h32_from_64 (u64x a)
+DECLSPEC u32x h32_from_64 (u64x a)
 {
   a >>= 32;
 
@@ -134,7 +134,7 @@ u32x h32_from_64 (u64x a)
   return r;
 }
 
-u64x hl32_to_64 (const u32x a, const u32x b)
+DECLSPEC u64x hl32_to_64 (const u32x a, const u32x b)
 {
   u64x r;
 
@@ -176,7 +176,7 @@ u64x hl32_to_64 (const u32x a, const u32x b)
 #ifdef IS_AMD
 
 #if AMD_GCN >= 3
-u32 swap32_S (const u32 v)
+DECLSPEC u32 swap32_S (const u32 v)
 {
   u32 r;
 
@@ -185,7 +185,7 @@ u32 swap32_S (const u32 v)
   return r;
 }
 
-u64 swap64_S (const u64 v)
+DECLSPEC u64 swap64_S (const u64 v)
 {
   const u32 v0 = h32_from_64_S (v);
   const u32 v1 = l32_from_64_S (v);
@@ -201,28 +201,28 @@ u64 swap64_S (const u64 v)
   return r;
 }
 #else
-u32 swap32_S (const u32 v)
+DECLSPEC u32 swap32_S (const u32 v)
 {
   return as_uint (as_uchar4 (v).s3210);
 }
 
-u64 swap64_S (const u64 v)
+DECLSPEC u64 swap64_S (const u64 v)
 {
   return (as_ulong (as_uchar8 (v).s76543210));
 }
 #endif
 
-u32 rotr32_S (const u32 a, const u32 n)
+DECLSPEC u32 rotr32_S (const u32 a, const u32 n)
 {
   return rotate (a, (32 - n));
 }
 
-u32 rotl32_S (const u32 a, const u32 n)
+DECLSPEC u32 rotl32_S (const u32 a, const u32 n)
 {
   return rotate (a, n);
 }
 
-u64 rotr64_S (const u64 a, const u32 n)
+DECLSPEC u64 rotr64_S (const u64 a, const u32 n)
 {
   const u32 a0 = h32_from_64_S (a);
   const u32 a1 = l32_from_64_S (a);
@@ -235,18 +235,18 @@ u64 rotr64_S (const u64 a, const u32 n)
   return r;
 }
 
-u64 rotl64_S (const u64 a, const u32 n)
+DECLSPEC u64 rotl64_S (const u64 a, const u32 n)
 {
   return rotr64_S (a, 64 - n);
 }
 
 #if AMD_GCN >= 3
-u32x swap32 (const u32x v)
+DECLSPEC u32x swap32 (const u32x v)
 {
   return bitselect (rotate (v, 24u), rotate (v, 8u), 0x00ff00ffu);
 }
 
-u64x swap64 (const u64x v)
+DECLSPEC u64x swap64 (const u64x v)
 {
   const u32x a0 = h32_from_64 (v);
   const u32x a1 = l32_from_64 (v);
@@ -308,12 +308,12 @@ u64x swap64 (const u64x v)
   return r;
 }
 #else
-u32x swap32 (const u32x v)
+DECLSPEC u32x swap32 (const u32x v)
 {
   return bitselect (rotate (v, 24u), rotate (v, 8u), 0x00ff00ffu);
 }
 
-u64x swap64 (const u64x v)
+DECLSPEC u64x swap64 (const u64x v)
 {
   return bitselect (bitselect (rotate (v, 24ul),
                                rotate (v,  8ul), 0x000000ff000000fful),
@@ -323,17 +323,17 @@ u64x swap64 (const u64x v)
 }
 #endif
 
-u32x rotr32 (const u32x a, const u32 n)
+DECLSPEC u32x rotr32 (const u32x a, const u32 n)
 {
   return rotate (a, (32 - n));
 }
 
-u32x rotl32 (const u32x a, const u32 n)
+DECLSPEC u32x rotl32 (const u32x a, const u32 n)
 {
   return rotate (a, n);
 }
 
-u64x rotr64 (const u64x a, const u32 n)
+DECLSPEC u64x rotr64 (const u64x a, const u32 n)
 {
   const u32x a0 = h32_from_64 (a);
   const u32x a1 = l32_from_64 (a);
@@ -346,28 +346,28 @@ u64x rotr64 (const u64x a, const u32 n)
   return r;
 }
 
-u64x rotl64 (const u64x a, const u32 n)
+DECLSPEC u64x rotl64 (const u64x a, const u32 n)
 {
   return rotr64 (a, 64 - n);
 }
 
-u32x __bfe (const u32x a, const u32x b, const u32x c)
+DECLSPEC u32x __bfe (const u32x a, const u32x b, const u32x c)
 {
   return amd_bfe (a, b, c);
 }
 
-u32 __bfe_S (const u32 a, const u32 b, const u32 c)
+DECLSPEC u32 __bfe_S (const u32 a, const u32 b, const u32 c)
 {
   return amd_bfe (a, b, c);
 }
 
-u32 amd_bytealign_S (const u32 a, const u32 b, const u32 c)
+DECLSPEC u32 amd_bytealign_S (const u32 a, const u32 b, const u32 c)
 {
   return amd_bytealign (a, b, c);
 }
 
 #if AMD_GCN >= 3
-u32x __byte_perm (const u32x a, const u32x b, const u32x c)
+DECLSPEC u32x __byte_perm (const u32x a, const u32x b, const u32x c)
 {
   u32x r;
 
@@ -420,7 +420,7 @@ u32x __byte_perm (const u32x a, const u32x b, const u32x c)
   return r;
 }
 
-u32 __byte_perm_S (const u32 a, const u32 b, const u32 c)
+DECLSPEC u32 __byte_perm_S (const u32 a, const u32 b, const u32 c)
 {
   u32 r;
 
@@ -431,7 +431,7 @@ u32 __byte_perm_S (const u32 a, const u32 b, const u32 c)
 #endif
 
 #if AMD_GCN >= 5
-u32x __add3 (const u32x a, const u32x b, const u32x c)
+DECLSPEC u32x __add3 (const u32x a, const u32x b, const u32x c)
 {
   u32x r;
 
@@ -484,7 +484,7 @@ u32x __add3 (const u32x a, const u32x b, const u32x c)
   return r;
 }
 
-u32 __add3_S (const u32 a, const u32 b, const u32 c)
+DECLSPEC u32 __add3_S (const u32 a, const u32 b, const u32 c)
 {
   u32 r;
 
@@ -493,12 +493,12 @@ u32 __add3_S (const u32 a, const u32 b, const u32 c)
   return r;
 }
 #else
-u32x __add3 (const u32x a, const u32x b, const u32x c)
+DECLSPEC u32x __add3 (const u32x a, const u32x b, const u32x c)
 {
   return a + b + c;
 }
 
-u32 __add3_S (const u32 a, const u32 b, const u32 c)
+DECLSPEC u32 __add3_S (const u32 a, const u32 b, const u32 c)
 {
   return a + b + c;
 }
@@ -507,7 +507,7 @@ u32 __add3_S (const u32 a, const u32 b, const u32 c)
 #endif
 
 #ifdef IS_NV
-u32 swap32_S (const u32 v)
+DECLSPEC u32 swap32_S (const u32 v)
 {
   u32 r;
 
@@ -516,7 +516,7 @@ u32 swap32_S (const u32 v)
   return r;
 }
 
-u64 swap64_S (const u64 v)
+DECLSPEC u64 swap64_S (const u64 v)
 {
   u32 il;
   u32 ir;
@@ -536,27 +536,27 @@ u64 swap64_S (const u64 v)
   return r;
 }
 
-u32 rotr32_S (const u32 a, const u32 n)
+DECLSPEC u32 rotr32_S (const u32 a, const u32 n)
 {
   return rotate (a, (32 - n));
 }
 
-u32 rotl32_S (const u32 a, const u32 n)
+DECLSPEC u32 rotl32_S (const u32 a, const u32 n)
 {
   return rotate (a, n);
 }
 
-u64 rotr64_S (const u64 a, const u32 n)
+DECLSPEC u64 rotr64_S (const u64 a, const u32 n)
 {
   return rotate (a, (u64) (64 - n));
 }
 
-u64 rotl64_S (const u64 a, const u32 n)
+DECLSPEC u64 rotl64_S (const u64 a, const u32 n)
 {
   return rotate (a, (u64) n);
 }
 
-u32x swap32 (const u32x v)
+DECLSPEC u32x swap32 (const u32x v)
 {
   u32x r;
 
@@ -595,7 +595,7 @@ u32x swap32 (const u32x v)
   return r;
 }
 
-u64x swap64 (const u64x v)
+DECLSPEC u64x swap64 (const u64x v)
 {
   u32x il;
   u32x ir;
@@ -721,27 +721,27 @@ u64x swap64 (const u64x v)
   return r;
 }
 
-u32x rotr32 (const u32x a, const u32 n)
+DECLSPEC u32x rotr32 (const u32x a, const u32 n)
 {
   return rotate (a, (32 - n));
 }
 
-u32x rotl32 (const u32x a, const u32 n)
+DECLSPEC u32x rotl32 (const u32x a, const u32 n)
 {
   return rotate (a, n);
 }
 
-u64x rotr64 (const u64x a, const u32 n)
+DECLSPEC u64x rotr64 (const u64x a, const u32 n)
 {
   return rotate (a, (u64x) (64 - n));
 }
 
-u64x rotl64 (const u64x a, const u32 n)
+DECLSPEC u64x rotl64 (const u64x a, const u32 n)
 {
   return rotate (a, (u64x) n);
 }
 
-u32x __byte_perm (const u32x a, const u32x b, const u32x c)
+DECLSPEC u32x __byte_perm (const u32x a, const u32x b, const u32x c)
 {
   u32x r;
 
@@ -780,7 +780,7 @@ u32x __byte_perm (const u32x a, const u32x b, const u32x c)
   return r;
 }
 
-u32 __byte_perm_S (const u32 a, const u32 b, const u32 c)
+DECLSPEC u32 __byte_perm_S (const u32 a, const u32 b, const u32 c)
 {
   u32 r;
 
@@ -789,7 +789,7 @@ u32 __byte_perm_S (const u32 a, const u32 b, const u32 c)
   return r;
 }
 
-u32x __bfe (const u32x a, const u32x b, const u32x c)
+DECLSPEC u32x __bfe (const u32x a, const u32x b, const u32x c)
 {
   u32x r;
 
@@ -828,7 +828,7 @@ u32x __bfe (const u32x a, const u32x b, const u32x c)
   return r;
 }
 
-u32 __bfe_S (const u32 a, const u32 b, const u32 c)
+DECLSPEC u32 __bfe_S (const u32 a, const u32 b, const u32 c)
 {
   u32 r;
 
@@ -837,7 +837,7 @@ u32 __bfe_S (const u32 a, const u32 b, const u32 c)
   return r;
 }
 
-u32x amd_bytealign (const u32x a, const u32x b, const u32x c)
+DECLSPEC u32x amd_bytealign (const u32x a, const u32x b, const u32x c)
 {
   u32x r;
 
@@ -884,7 +884,7 @@ u32x amd_bytealign (const u32x a, const u32x b, const u32x c)
   return r;
 }
 
-u32 amd_bytealign_S (const u32 a, const u32 b, const u32 c)
+DECLSPEC u32 amd_bytealign_S (const u32 a, const u32 b, const u32 c)
 {
   u32 r;
 
@@ -901,12 +901,12 @@ u32 amd_bytealign_S (const u32 a, const u32 b, const u32 c)
   return r;
 }
 
-u32x __add3 (const u32x a, const u32x b, const u32x c)
+DECLSPEC u32x __add3 (const u32x a, const u32x b, const u32x c)
 {
   return a + b + c;
 }
 
-u32 __add3_S (const u32 a, const u32 b, const u32 c)
+DECLSPEC u32 __add3_S (const u32 a, const u32 b, const u32 c)
 {
   return a + b + c;
 }
@@ -914,37 +914,37 @@ u32 __add3_S (const u32 a, const u32 b, const u32 c)
 #endif
 
 #ifdef IS_GENERIC
-u32 swap32_S (const u32 v)
+DECLSPEC u32 swap32_S (const u32 v)
 {
   return (as_uint (as_uchar4 (v).s3210));
 }
 
-u64 swap64_S (const u64 v)
+DECLSPEC u64 swap64_S (const u64 v)
 {
   return (as_ulong (as_uchar8 (v).s76543210));
 }
 
-u32 rotr32_S (const u32 a, const u32 n)
+DECLSPEC u32 rotr32_S (const u32 a, const u32 n)
 {
   return rotate (a, (32 - n));
 }
 
-u32 rotl32_S (const u32 a, const u32 n)
+DECLSPEC u32 rotl32_S (const u32 a, const u32 n)
 {
   return rotate (a, n);
 }
 
-u64 rotr64_S (const u64 a, const u32 n)
+DECLSPEC u64 rotr64_S (const u64 a, const u32 n)
 {
   return rotate (a, (u64) (64 - n));
 }
 
-u64 rotl64_S (const u64 a, const u32 n)
+DECLSPEC u64 rotl64_S (const u64 a, const u32 n)
 {
   return rotate (a, (u64) n);
 }
 
-u32x swap32 (const u32x v)
+DECLSPEC u32x swap32 (const u32x v)
 {
   return ((v >> 24) & 0x000000ff)
        | ((v >>  8) & 0x0000ff00)
@@ -952,7 +952,7 @@ u32x swap32 (const u32x v)
        | ((v << 24) & 0xff000000);
 }
 
-u64x swap64 (const u64x v)
+DECLSPEC u64x swap64 (const u64x v)
 {
   return ((v >> 56) & 0x00000000000000ff)
        | ((v >> 40) & 0x000000000000ff00)
@@ -964,27 +964,27 @@ u64x swap64 (const u64x v)
        | ((v << 56) & 0xff00000000000000);
 }
 
-u32x rotr32 (const u32x a, const u32 n)
+DECLSPEC u32x rotr32 (const u32x a, const u32 n)
 {
   return rotate (a, (32 - n));
 }
 
-u32x rotl32 (const u32x a, const u32 n)
+DECLSPEC u32x rotl32 (const u32x a, const u32 n)
 {
   return rotate (a, n);
 }
 
-u64x rotr64 (const u64x a, const u32 n)
+DECLSPEC u64x rotr64 (const u64x a, const u32 n)
 {
   return rotate (a, (u64x) (64 - n));
 }
 
-u64x rotl64 (const u64x a, const u32 n)
+DECLSPEC u64x rotl64 (const u64x a, const u32 n)
 {
   return rotate (a, (u64x) n);
 }
 
-u32x __bfe (const u32x a, const u32x b, const u32x c)
+DECLSPEC u32x __bfe (const u32x a, const u32x b, const u32x c)
 {
   #define BIT(x)      ((u32x) (1u) << (x))
   #define BIT_MASK(x) (BIT (x) - 1)
@@ -997,7 +997,7 @@ u32x __bfe (const u32x a, const u32x b, const u32x c)
   #undef BFE
 }
 
-u32 __bfe_S (const u32 a, const u32 b, const u32 c)
+DECLSPEC u32 __bfe_S (const u32 a, const u32 b, const u32 c)
 {
   #define BIT(x)      (1u << (x))
   #define BIT_MASK(x) (BIT (x) - 1)
@@ -1010,7 +1010,7 @@ u32 __bfe_S (const u32 a, const u32 b, const u32 c)
   #undef BFE
 }
 
-u32x amd_bytealign (const u32x a, const u32x b, const u32 c)
+DECLSPEC u32x amd_bytealign (const u32x a, const u32x b, const u32 c)
 {
   #if VECT_SIZE == 1
   const u64x tmp = ((((u64x) (a)) << 32) | ((u64x) (b))) >> ((c & 3) * 8);
@@ -1043,19 +1043,19 @@ u32x amd_bytealign (const u32x a, const u32x b, const u32 c)
   #endif
 }
 
-u32 amd_bytealign_S (const u32 a, const u32 b, const u32 c)
+DECLSPEC u32 amd_bytealign_S (const u32 a, const u32 b, const u32 c)
 {
   const u64 tmp = ((((u64) a) << 32) | ((u64) b)) >> ((c & 3) * 8);
 
   return (u32) (tmp);
 }
 
-u32x __add3 (const u32x a, const u32x b, const u32x c)
+DECLSPEC u32x __add3 (const u32x a, const u32x b, const u32x c)
 {
   return a + b + c;
 }
 
-u32 __add3_S (const u32 a, const u32 b, const u32 c)
+DECLSPEC u32 __add3_S (const u32 a, const u32 b, const u32 c)
 {
   return a + b + c;
 }
diff --git a/OpenCL/inc_vendor.cl b/OpenCL/inc_vendor.cl
index b49f7f181..04ac1319a 100644
--- a/OpenCL/inc_vendor.cl
+++ b/OpenCL/inc_vendor.cl
@@ -77,6 +77,12 @@
 #define IS_GENERIC
 #endif
 
+/**
+ * function declarations can have a large influence depending on the opencl runtime
+ */
+
+#define DECLSPEC static inline
+
 /**
  * AMD specific
  */
diff --git a/OpenCL/m00000_a3-optimized.cl b/OpenCL/m00000_a3-optimized.cl
index 55e5846ef..0c6b86574 100644
--- a/OpenCL/m00000_a3-optimized.cl
+++ b/OpenCL/m00000_a3-optimized.cl
@@ -30,7 +30,7 @@
   a -= t;                               \
 }
 
-void m00000m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00000m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -202,7 +202,7 @@ void m00000m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m00000s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00000s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00010_a3-optimized.cl b/OpenCL/m00010_a3-optimized.cl
index c13988e7c..f48c3cb4d 100644
--- a/OpenCL/m00010_a3-optimized.cl
+++ b/OpenCL/m00010_a3-optimized.cl
@@ -29,7 +29,7 @@
   a -= t;                               \
 }
 
-void m00010m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00010m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -250,7 +250,7 @@ void m00010m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m00010s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00010s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00020_a3-optimized.cl b/OpenCL/m00020_a3-optimized.cl
index 94cbb6c61..0b1f7e6bf 100644
--- a/OpenCL/m00020_a3-optimized.cl
+++ b/OpenCL/m00020_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m00020m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00020m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -187,7 +187,7 @@ void m00020m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m00020s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00020s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00030_a3-optimized.cl b/OpenCL/m00030_a3-optimized.cl
index a2100fb22..a4d0a7c09 100644
--- a/OpenCL/m00030_a3-optimized.cl
+++ b/OpenCL/m00030_a3-optimized.cl
@@ -29,7 +29,7 @@
   a -= t;                               \
 }
 
-void m00030m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00030m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -250,7 +250,7 @@ void m00030m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m00030s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00030s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00040_a3-optimized.cl b/OpenCL/m00040_a3-optimized.cl
index 2ad6c824b..f16ec4e03 100644
--- a/OpenCL/m00040_a3-optimized.cl
+++ b/OpenCL/m00040_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m00040m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00040m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -187,7 +187,7 @@ void m00040m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m00040s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00040s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00050_a0-optimized.cl b/OpenCL/m00050_a0-optimized.cl
index ce2c346fd..639129235 100644
--- a/OpenCL/m00050_a0-optimized.cl
+++ b/OpenCL/m00050_a0-optimized.cl
@@ -15,7 +15,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md5.cl"
 
-void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
+DECLSPEC void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -66,7 +66,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
+DECLSPEC void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m00050_a1-optimized.cl b/OpenCL/m00050_a1-optimized.cl
index 6e4b1e497..8093dfbb2 100644
--- a/OpenCL/m00050_a1-optimized.cl
+++ b/OpenCL/m00050_a1-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md5.cl"
 
-void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
+DECLSPEC void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -64,7 +64,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
+DECLSPEC void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m00050_a3-optimized.cl b/OpenCL/m00050_a3-optimized.cl
index 21eeb689e..8f07ebe6a 100644
--- a/OpenCL/m00050_a3-optimized.cl
+++ b/OpenCL/m00050_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md5.cl"
 
-void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
+DECLSPEC void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -64,7 +64,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
+DECLSPEC void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -98,7 +98,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, digest);
 }
 
-void m00050m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00050m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -203,7 +203,7 @@ void m00050m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m00050s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00050s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00060_a0-optimized.cl b/OpenCL/m00060_a0-optimized.cl
index 2220870dd..115f400a7 100644
--- a/OpenCL/m00060_a0-optimized.cl
+++ b/OpenCL/m00060_a0-optimized.cl
@@ -15,7 +15,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md5.cl"
 
-void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
+DECLSPEC void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -66,7 +66,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
+DECLSPEC void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m00060_a1-optimized.cl b/OpenCL/m00060_a1-optimized.cl
index 99bcf9cb8..04f1b5055 100644
--- a/OpenCL/m00060_a1-optimized.cl
+++ b/OpenCL/m00060_a1-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md5.cl"
 
-void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
+DECLSPEC void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -64,7 +64,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
+DECLSPEC void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m00060_a3-optimized.cl b/OpenCL/m00060_a3-optimized.cl
index ecd95babf..890bb40c9 100644
--- a/OpenCL/m00060_a3-optimized.cl
+++ b/OpenCL/m00060_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md5.cl"
 
-void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
+DECLSPEC void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -64,7 +64,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
+DECLSPEC void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -98,7 +98,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, digest);
 }
 
-void m00060m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esal_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00060m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esal_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -203,7 +203,7 @@ void m00060m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m00060s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00060s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00100_a3-optimized.cl b/OpenCL/m00100_a3-optimized.cl
index a22e98b13..714dd37eb 100644
--- a/OpenCL/m00100_a3-optimized.cl
+++ b/OpenCL/m00100_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m00100m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00100m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -256,7 +256,7 @@ void m00100m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m00100s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00100s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00110_a3-optimized.cl b/OpenCL/m00110_a3-optimized.cl
index ee5bd4fac..bdb7849c0 100644
--- a/OpenCL/m00110_a3-optimized.cl
+++ b/OpenCL/m00110_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m00110m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00110m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -304,7 +304,7 @@ void m00110m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m00110s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00110s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00120_a3-optimized.cl b/OpenCL/m00120_a3-optimized.cl
index ead815e29..8421487e0 100644
--- a/OpenCL/m00120_a3-optimized.cl
+++ b/OpenCL/m00120_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m00120m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00120m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -231,7 +231,7 @@ void m00120m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m00120s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00120s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00130_a3-optimized.cl b/OpenCL/m00130_a3-optimized.cl
index 3cc025e82..9974c5a70 100644
--- a/OpenCL/m00130_a3-optimized.cl
+++ b/OpenCL/m00130_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m00130m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00130m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -304,7 +304,7 @@ void m00130m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m00130s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00130s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00140_a3-optimized.cl b/OpenCL/m00140_a3-optimized.cl
index 8a11a2dd2..682d0f0dc 100644
--- a/OpenCL/m00140_a3-optimized.cl
+++ b/OpenCL/m00140_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m00140m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00140m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -231,7 +231,7 @@ void m00140m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m00140s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00140s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00150_a0-optimized.cl b/OpenCL/m00150_a0-optimized.cl
index c20bf6b33..25713a640 100644
--- a/OpenCL/m00150_a0-optimized.cl
+++ b/OpenCL/m00150_a0-optimized.cl
@@ -15,7 +15,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha1.cl"
 
-void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
+DECLSPEC void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -68,7 +68,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m00150_a1-optimized.cl b/OpenCL/m00150_a1-optimized.cl
index 3c90dde30..2d1e23aa7 100644
--- a/OpenCL/m00150_a1-optimized.cl
+++ b/OpenCL/m00150_a1-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha1.cl"
 
-void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
+DECLSPEC void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -66,7 +66,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m00150_a3-optimized.cl b/OpenCL/m00150_a3-optimized.cl
index 3123f191f..db3cf3d6e 100644
--- a/OpenCL/m00150_a3-optimized.cl
+++ b/OpenCL/m00150_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha1.cl"
 
-void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
+DECLSPEC void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -66,7 +66,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -102,7 +102,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, digest);
 }
 
-void m00150m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00150m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -207,7 +207,7 @@ void m00150m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m00150s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00150s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00160_a0-optimized.cl b/OpenCL/m00160_a0-optimized.cl
index 984e318d8..37c551cfb 100644
--- a/OpenCL/m00160_a0-optimized.cl
+++ b/OpenCL/m00160_a0-optimized.cl
@@ -15,7 +15,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha1.cl"
 
-void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
+DECLSPEC void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -68,7 +68,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m00160_a1-optimized.cl b/OpenCL/m00160_a1-optimized.cl
index 29a3c6b2b..0b4948252 100644
--- a/OpenCL/m00160_a1-optimized.cl
+++ b/OpenCL/m00160_a1-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha1.cl"
 
-void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
+DECLSPEC void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -66,7 +66,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m00160_a3-optimized.cl b/OpenCL/m00160_a3-optimized.cl
index 0df9688f5..613879ba8 100644
--- a/OpenCL/m00160_a3-optimized.cl
+++ b/OpenCL/m00160_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha1.cl"
 
-void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
+DECLSPEC void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -66,7 +66,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -102,7 +102,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, digest);
 }
 
-void m00160m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esal_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00160m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esal_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -207,7 +207,7 @@ void m00160m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m00160s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00160s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00200_a3-optimized.cl b/OpenCL/m00200_a3-optimized.cl
index e59af20c4..fadaab549 100644
--- a/OpenCL/m00200_a3-optimized.cl
+++ b/OpenCL/m00200_a3-optimized.cl
@@ -89,7 +89,7 @@
   }                                                               \
 }
 
-void m00200m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00200m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -235,7 +235,7 @@ void m00200m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m00200s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00200s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00300_a3-optimized.cl b/OpenCL/m00300_a3-optimized.cl
index 5f22c7f8c..21b041d6e 100644
--- a/OpenCL/m00300_a3-optimized.cl
+++ b/OpenCL/m00300_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m00300m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00300m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -381,7 +381,7 @@ void m00300m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m00300s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00300s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m00500-optimized.cl b/OpenCL/m00500-optimized.cl
index bf7da0e03..f124e7ae8 100644
--- a/OpenCL/m00500-optimized.cl
+++ b/OpenCL/m00500-optimized.cl
@@ -17,7 +17,7 @@
 
 #define md5crypt_magic 0x00243124u
 
-void memcat16 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[4])
+DECLSPEC void memcat16 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[4])
 {
   u32 tmp0;
   u32 tmp1;
@@ -130,7 +130,7 @@ void memcat16 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const
   }
 }
 
-void memcat16_x80 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[4])
+DECLSPEC void memcat16_x80 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[4])
 {
   u32 tmp0;
   u32 tmp1;
@@ -245,7 +245,7 @@ void memcat16_x80 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], c
   }
 }
 
-void memcat8 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[2])
+DECLSPEC void memcat8 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[2])
 {
   u32 tmp0;
   u32 tmp1;
@@ -334,7 +334,7 @@ void memcat8 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const
   }
 }
 
-void append_sign (u32 block0[4], u32 block1[4], const u32 block_len)
+DECLSPEC void append_sign (u32 block0[4], u32 block1[4], const u32 block_len)
 {
   switch (block_len)
   {
@@ -416,7 +416,7 @@ void append_sign (u32 block0[4], u32 block1[4], const u32 block_len)
   }
 }
 
-void append_1st (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 block_len, const u32 append)
+DECLSPEC void append_1st (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 block_len, const u32 append)
 {
   switch (block_len)
   {
diff --git a/OpenCL/m00600_a0-optimized.cl b/OpenCL/m00600_a0-optimized.cl
index 7b4fdfa1e..e05b67b88 100644
--- a/OpenCL/m00600_a0-optimized.cl
+++ b/OpenCL/m00600_a0-optimized.cl
@@ -41,7 +41,7 @@
     BLAKE2B_G (r,7,v[ 3],v[ 4],v[ 9],v[14]); \
 } while(0)
 
-void blake2b_transform (u64x h[8], u64x t[2], u64x f[2], u64x m[16], u64x v[16], const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x out_len, const u8 isFinal)
+DECLSPEC void blake2b_transform (u64x h[8], u64x t[2], u64x f[2], u64x m[16], u64x v[16], const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x out_len, const u8 isFinal)
 {
   if (isFinal)
     f[0] = -1;
diff --git a/OpenCL/m00600_a1-optimized.cl b/OpenCL/m00600_a1-optimized.cl
index a40b10613..717a6d76b 100644
--- a/OpenCL/m00600_a1-optimized.cl
+++ b/OpenCL/m00600_a1-optimized.cl
@@ -41,7 +41,7 @@
     BLAKE2B_G (r,7,v[ 3],v[ 4],v[ 9],v[14]); \
 } while(0)
 
-void blake2b_transform (u64x h[8], u64x t[2], u64x f[2], u64x m[16], u64x v[16], const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x out_len, const u8 isFinal)
+DECLSPEC void blake2b_transform (u64x h[8], u64x t[2], u64x f[2], u64x m[16], u64x v[16], const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x out_len, const u8 isFinal)
 {
   if (isFinal)
     f[0] = -1;
diff --git a/OpenCL/m00600_a3-optimized.cl b/OpenCL/m00600_a3-optimized.cl
index 398e054d8..1e574d2e2 100644
--- a/OpenCL/m00600_a3-optimized.cl
+++ b/OpenCL/m00600_a3-optimized.cl
@@ -39,7 +39,7 @@
     BLAKE2B_G (r,7,v[ 3],v[ 4],v[ 9],v[14]); \
 } while(0)
 
-void blake2b_transform (u64x h[8], u64x t[2], u64x f[2], u64x m[16], u64x v[16], const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x out_len, const u8 isFinal)
+DECLSPEC void blake2b_transform (u64x h[8], u64x t[2], u64x f[2], u64x m[16], u64x v[16], const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x out_len, const u8 isFinal)
 {
   if (isFinal)
     f[0] = -1;
diff --git a/OpenCL/m00900_a3-optimized.cl b/OpenCL/m00900_a3-optimized.cl
index 57c42a4f1..482d7bd99 100644
--- a/OpenCL/m00900_a3-optimized.cl
+++ b/OpenCL/m00900_a3-optimized.cl
@@ -27,7 +27,7 @@
   a -= t;                               \
 }
 
-void m00900m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00900m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -163,7 +163,7 @@ void m00900m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m00900s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m00900s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01000_a3-optimized.cl b/OpenCL/m01000_a3-optimized.cl
index a1ad18235..152155644 100644
--- a/OpenCL/m01000_a3-optimized.cl
+++ b/OpenCL/m01000_a3-optimized.cl
@@ -27,7 +27,7 @@
   a -= t;                               \
 }
 
-void m01000m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01000m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -163,7 +163,7 @@ void m01000m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m01000s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01000s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01100_a3-optimized.cl b/OpenCL/m01100_a3-optimized.cl
index 7da75ddcc..711984934 100644
--- a/OpenCL/m01100_a3-optimized.cl
+++ b/OpenCL/m01100_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m01100m (__local salt_t *s_salt_buf, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01100m (__local salt_t *s_salt_buf, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -247,7 +247,7 @@ void m01100m (__local salt_t *s_salt_buf, u32 w[16], const u32 pw_len, __global
   }
 }
 
-void m01100s (__local salt_t *s_salt_buf, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01100s (__local salt_t *s_salt_buf, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01300_a3-optimized.cl b/OpenCL/m01300_a3-optimized.cl
index 6c976f731..c1a66a969 100644
--- a/OpenCL/m01300_a3-optimized.cl
+++ b/OpenCL/m01300_a3-optimized.cl
@@ -25,7 +25,7 @@
   g = 0;                                        \
 }
 
-void m01300m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01300m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -144,7 +144,7 @@ void m01300m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m01300s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01300s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01400_a3-optimized.cl b/OpenCL/m01400_a3-optimized.cl
index 30a7cee5f..7b478734d 100644
--- a/OpenCL/m01400_a3-optimized.cl
+++ b/OpenCL/m01400_a3-optimized.cl
@@ -26,7 +26,7 @@
   h = 0;                                        \
 }
 
-void m01400m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01400m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -145,7 +145,7 @@ void m01400m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m01400s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01400s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01410_a3-optimized.cl b/OpenCL/m01410_a3-optimized.cl
index 4ce8a04f4..ed3905ed1 100644
--- a/OpenCL/m01410_a3-optimized.cl
+++ b/OpenCL/m01410_a3-optimized.cl
@@ -26,7 +26,7 @@
   h = 0;                                        \
 }
 
-void m01410m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01410m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -196,7 +196,7 @@ void m01410m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m01410s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01410s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01420_a3-optimized.cl b/OpenCL/m01420_a3-optimized.cl
index b69cfcf01..31e5170d4 100644
--- a/OpenCL/m01420_a3-optimized.cl
+++ b/OpenCL/m01420_a3-optimized.cl
@@ -26,7 +26,7 @@
   h = 0;                                        \
 }
 
-void m01420m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01420m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -220,7 +220,7 @@ void m01420m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m01420s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01420s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01430_a3-optimized.cl b/OpenCL/m01430_a3-optimized.cl
index 46d76575b..5eed1dbf4 100644
--- a/OpenCL/m01430_a3-optimized.cl
+++ b/OpenCL/m01430_a3-optimized.cl
@@ -26,7 +26,7 @@
   h = 0;                                        \
 }
 
-void m01430m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01430m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -196,7 +196,7 @@ void m01430m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m01430s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01430s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01440_a3-optimized.cl b/OpenCL/m01440_a3-optimized.cl
index d61a7243a..c698b1e47 100644
--- a/OpenCL/m01440_a3-optimized.cl
+++ b/OpenCL/m01440_a3-optimized.cl
@@ -26,7 +26,7 @@
   h = 0;                                        \
 }
 
-void m01440m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01440m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -220,7 +220,7 @@ void m01440m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m01440s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01440s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01450_a0-optimized.cl b/OpenCL/m01450_a0-optimized.cl
index 910057ffc..1d3eee7d5 100644
--- a/OpenCL/m01450_a0-optimized.cl
+++ b/OpenCL/m01450_a0-optimized.cl
@@ -15,7 +15,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha256.cl"
 
-void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8])
+DECLSPEC void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -74,7 +74,7 @@ void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[
   sha256_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha256_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m01450_a1-optimized.cl b/OpenCL/m01450_a1-optimized.cl
index e149172e4..4726d9c76 100644
--- a/OpenCL/m01450_a1-optimized.cl
+++ b/OpenCL/m01450_a1-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha256.cl"
 
-void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8])
+DECLSPEC void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -72,7 +72,7 @@ void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[
   sha256_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha256_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m01450_a3-optimized.cl b/OpenCL/m01450_a3-optimized.cl
index ed4b796c2..fc5e05c20 100644
--- a/OpenCL/m01450_a3-optimized.cl
+++ b/OpenCL/m01450_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha256.cl"
 
-void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8])
+DECLSPEC void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -72,7 +72,7 @@ void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[
   sha256_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha256_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -114,7 +114,7 @@ void hmac_sha256_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[
   sha256_transform_vector (w0, w1, w2, w3, digest);
 }
 
-void m01450m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01450m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -219,7 +219,7 @@ void m01450m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m01450s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01450s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01460_a0-optimized.cl b/OpenCL/m01460_a0-optimized.cl
index a643c7535..c5a6d5f81 100644
--- a/OpenCL/m01460_a0-optimized.cl
+++ b/OpenCL/m01460_a0-optimized.cl
@@ -15,7 +15,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha256.cl"
 
-void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8])
+DECLSPEC void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -74,7 +74,7 @@ void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[
   sha256_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha256_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m01460_a1-optimized.cl b/OpenCL/m01460_a1-optimized.cl
index 1890f8133..a535d92a9 100644
--- a/OpenCL/m01460_a1-optimized.cl
+++ b/OpenCL/m01460_a1-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha256.cl"
 
-void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8])
+DECLSPEC void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -72,7 +72,7 @@ void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[
   sha256_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha256_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m01460_a3-optimized.cl b/OpenCL/m01460_a3-optimized.cl
index 629cc04b0..88ee77016 100644
--- a/OpenCL/m01460_a3-optimized.cl
+++ b/OpenCL/m01460_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha256.cl"
 
-void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8])
+DECLSPEC void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -72,7 +72,7 @@ void hmac_sha256_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[
   sha256_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha256_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -114,7 +114,7 @@ void hmac_sha256_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[
   sha256_transform_vector (w0, w1, w2, w3, digest);
 }
 
-void m01460m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01460m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -219,7 +219,7 @@ void m01460m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m01460s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01460s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01500_a0.cl b/OpenCL/m01500_a0.cl
index 42c5496cf..a1389667f 100644
--- a/OpenCL/m01500_a0.cl
+++ b/OpenCL/m01500_a0.cl
@@ -342,7 +342,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
@@ -411,7 +411,7 @@ void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (
   }
 }
 
-void _des_crypt_encrypt (u32 iv[2], u32 mask, u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32 mask, u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   const u32 E1 = (mask >> 2) & 0x3f0;
   const u32 E0 = mask & 0x3f;
diff --git a/OpenCL/m01500_a1.cl b/OpenCL/m01500_a1.cl
index 177628442..fb59dc74d 100644
--- a/OpenCL/m01500_a1.cl
+++ b/OpenCL/m01500_a1.cl
@@ -340,7 +340,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
@@ -409,7 +409,7 @@ void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (
   }
 }
 
-void _des_crypt_encrypt (u32 iv[2], u32 mask, u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32 mask, u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   const u32 E1 = (mask >> 2) & 0x3f0;
   const u32 E0 = mask & 0x3f;
diff --git a/OpenCL/m01500_a3.cl b/OpenCL/m01500_a3.cl
index 4a89ad578..3a812046d 100644
--- a/OpenCL/m01500_a3.cl
+++ b/OpenCL/m01500_a3.cl
@@ -57,7 +57,7 @@
 
 #define LUT(a,b,c,d,e) u32 a; asm volatile ("lop3.b32 %0, %1, %2, %3, "#e";" : "=r"(a): "r"(b), "r"(c), "r"(d));
 
-void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(xAA55AA5500550055, a1, a4, a6, 0xC1)
   LUT(xA55AA55AF0F5F0F5, a3, a6, xAA55AA5500550055, 0x9E)
@@ -91,7 +91,7 @@ void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(xEEEEEEEE99999999, a1, a2, a6, 0x97)
   LUT(xFFFFEEEE66666666, a5, a6, xEEEEEEEE99999999, 0x67)
@@ -124,7 +124,7 @@ void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(xA50FA50FA50FA50F, a1, a3, a4, 0xC9)
   LUT(xF0F00F0FF0F0F0F0, a3, a5, a6, 0x4B)
@@ -158,7 +158,7 @@ void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(x55F055F055F055F0, a1, a3, a4, 0x72)
   LUT(xA500F5F0A500F5F0, a3, a5, x55F055F055F055F0, 0xAD)
@@ -185,7 +185,7 @@ void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(xA0A0A0A0FFFFFFFF, a1, a3, a6, 0xAB)
   LUT(xFFFF00005555FFFF, a1, a5, a6, 0xB9)
@@ -219,7 +219,7 @@ void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(x5050F5F55050F5F5, a1, a3, a5, 0xB2)
   LUT(x6363C6C66363C6C6, a1, a2, x5050F5F55050F5F5, 0x66)
@@ -252,7 +252,7 @@ void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(x88AA88AA88AA88AA, a1, a2, a4, 0x0B)
   LUT(xAAAAFF00AAAAFF00, a1, a4, a5, 0x27)
@@ -285,7 +285,7 @@ void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(xEEEE3333EEEE3333, a1, a2, a5, 0x9D)
   LUT(xBBBBBBBBBBBBBBBB, a1, a1, a2, 0x83)
@@ -343,7 +343,7 @@ void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
  * The effort has been sponsored by Rapid7: http://www.rapid7.com
  */
 
-void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x55005500, x5A0F5A0F, x3333FFFF, x66666666, x22226666, x2D2D6969,
         x25202160;
@@ -418,7 +418,7 @@ void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x33CC33CC;
     u32 x55550000, x00AA00FF, x33BB33FF;
@@ -489,7 +489,7 @@ void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x44444444, x0F0FF0F0, x4F4FF4F4, x00FFFF00, x00AAAA00, x4FE55EF4;
     u32 x3C3CC3C3, x3C3C0000, x7373F4F4, x0C840A00;
@@ -560,7 +560,7 @@ void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out3 ^= x21;
 }
 
-void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x5A5A5A5A, x0F0FF0F0;
     u32 x33FF33FF, x33FFCC00, x0C0030F0, x0C0CC0C0, x0CF3C03F, x5EFBDA7F,
@@ -614,7 +614,7 @@ void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x77777777, x77770000, x22225555, x11116666, x1F1F6F6F;
     u32 x70700000, x43433333, x00430033, x55557777, x55167744, x5A19784B;
@@ -687,7 +687,7 @@ void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out2 ^= x11;
 }
 
-void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x33CC33CC;
     u32 x3333FFFF, x11115555, x22DD6699, x22DD9966, x00220099;
@@ -760,7 +760,7 @@ void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out1 ^= x01;
 }
 
-void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x0FF00FF0, x3CC33CC3, x00003CC3, x0F000F00, x5A555A55, x00001841;
     u32 x00000F00, x33333C33, x7B777E77, x0FF0F00F, x74878E78;
@@ -831,7 +831,7 @@ void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out2 ^= x11;
 }
 
-void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x0C0C0C0C, x0000F0F0, x00FFF00F, x00555005, x00515001;
     u32 x33000330, x77555775, x30303030, x3030CFCF, x30104745, x30555745;
@@ -924,7 +924,7 @@ void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
  * The effort has been sponsored by Rapid7: http://www.rapid7.com
  */
 
-void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x55005500, x5A0F5A0F, x3333FFFF, x66666666, x22226666, x2D2D6969,
         x25202160;
@@ -999,7 +999,7 @@ void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x33CC33CC;
     u32 x55550000, x00AA00FF, x33BB33FF;
@@ -1070,7 +1070,7 @@ void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x44444444, x0F0FF0F0, x4F4FF4F4, x00FFFF00, x00AAAA00, x4FE55EF4;
     u32 x3C3CC3C3, x3C3C0000, x7373F4F4, x0C840A00;
@@ -1141,7 +1141,7 @@ void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out3 ^= x21;
 }
 
-void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x5A5A5A5A, x0F0FF0F0;
     u32 x33FF33FF, x33FFCC00, x0C0030F0, x0C0CC0C0, x0CF3C03F, x5EFBDA7F,
@@ -1195,7 +1195,7 @@ void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x77777777, x77770000, x22225555, x11116666, x1F1F6F6F;
     u32 x70700000, x43433333, x00430033, x55557777, x55167744, x5A19784B;
@@ -1268,7 +1268,7 @@ void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out2 ^= x11;
 }
 
-void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x33CC33CC;
     u32 x3333FFFF, x11115555, x22DD6699, x22DD9966, x00220099;
@@ -1341,7 +1341,7 @@ void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out1 ^= x01;
 }
 
-void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x0FF00FF0, x3CC33CC3, x00003CC3, x0F000F00, x5A555A55, x00001841;
     u32 x00000F00, x33333C33, x7B777E77, x0FF0F00F, x74878E78;
@@ -1412,7 +1412,7 @@ void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out2 ^= x11;
 }
 
-void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x0C0C0C0C, x0000F0F0, x00FFF00F, x00555005, x00515001;
     u32 x33000330, x77555775, x30303030, x3030CFCF, x30104745, x30555745;
@@ -1536,7 +1536,7 @@ void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
 
 #ifdef DESCRYPT_SALT
 
-void DESCrypt (const u32 SALT, const u32 K00, const u32 K01, const u32 K02, const u32 K03, const u32 K04, const u32 K05, const u32 K06, const u32 K07, const u32 K08, const u32 K09, const u32 K10, const u32 K11, const u32 K12, const u32 K13, const u32 K14, const u32 K15, const u32 K16, const u32 K17, const u32 K18, const u32 K19, const u32 K20, const u32 K21, const u32 K22, const u32 K23, const u32 K24, const u32 K25, const u32 K26, const u32 K27, const u32 K28, const u32 K29, const u32 K30, const u32 K31, const u32 K32, const u32 K33, const u32 K34, const u32 K35, const u32 K36, const u32 K37, const u32 K38, const u32 K39, const u32 K40, const u32 K41, const u32 K42, const u32 K43, const u32 K44, const u32 K45, const u32 K46, const u32 K47, const u32 K48, const u32 K49, const u32 K50, const u32 K51, const u32 K52, const u32 K53, const u32 K54, const u32 K55, u32 *D00, u32 *D01, u32 *D02, u32 *D03, u32 *D04, u32 *D05, u32 *D06, u32 *D07, u32 *D08, u32 *D09, u32 *D10, u32 *D11, u32 *D12, u32 *D13, u32 *D14, u32 *D15, u32 *D16, u32 *D17, u32 *D18, u32 *D19, u32 *D20, u32 *D21, u32 *D22, u32 *D23, u32 *D24, u32 *D25, u32 *D26, u32 *D27, u32 *D28, u32 *D29, u32 *D30, u32 *D31, u32 *D32, u32 *D33, u32 *D34, u32 *D35, u32 *D36, u32 *D37, u32 *D38, u32 *D39, u32 *D40, u32 *D41, u32 *D42, u32 *D43, u32 *D44, u32 *D45, u32 *D46, u32 *D47, u32 *D48, u32 *D49, u32 *D50, u32 *D51, u32 *D52, u32 *D53, u32 *D54, u32 *D55, u32 *D56, u32 *D57, u32 *D58, u32 *D59, u32 *D60, u32 *D61, u32 *D62, u32 *D63)
+DECLSPEC void DESCrypt (const u32 SALT, const u32 K00, const u32 K01, const u32 K02, const u32 K03, const u32 K04, const u32 K05, const u32 K06, const u32 K07, const u32 K08, const u32 K09, const u32 K10, const u32 K11, const u32 K12, const u32 K13, const u32 K14, const u32 K15, const u32 K16, const u32 K17, const u32 K18, const u32 K19, const u32 K20, const u32 K21, const u32 K22, const u32 K23, const u32 K24, const u32 K25, const u32 K26, const u32 K27, const u32 K28, const u32 K29, const u32 K30, const u32 K31, const u32 K32, const u32 K33, const u32 K34, const u32 K35, const u32 K36, const u32 K37, const u32 K38, const u32 K39, const u32 K40, const u32 K41, const u32 K42, const u32 K43, const u32 K44, const u32 K45, const u32 K46, const u32 K47, const u32 K48, const u32 K49, const u32 K50, const u32 K51, const u32 K52, const u32 K53, const u32 K54, const u32 K55, u32 *D00, u32 *D01, u32 *D02, u32 *D03, u32 *D04, u32 *D05, u32 *D06, u32 *D07, u32 *D08, u32 *D09, u32 *D10, u32 *D11, u32 *D12, u32 *D13, u32 *D14, u32 *D15, u32 *D16, u32 *D17, u32 *D18, u32 *D19, u32 *D20, u32 *D21, u32 *D22, u32 *D23, u32 *D24, u32 *D25, u32 *D26, u32 *D27, u32 *D28, u32 *D29, u32 *D30, u32 *D31, u32 *D32, u32 *D33, u32 *D34, u32 *D35, u32 *D36, u32 *D37, u32 *D38, u32 *D39, u32 *D40, u32 *D41, u32 *D42, u32 *D43, u32 *D44, u32 *D45, u32 *D46, u32 *D47, u32 *D48, u32 *D49, u32 *D50, u32 *D51, u32 *D52, u32 *D53, u32 *D54, u32 *D55, u32 *D56, u32 *D57, u32 *D58, u32 *D59, u32 *D60, u32 *D61, u32 *D62, u32 *D63)
 {
   sXXX_DECL u32 s001 = (0x001 & DESCRYPT_SALT) ? 1 : 0;
   sXXX_DECL u32 s002 = (0x002 & DESCRYPT_SALT) ? 1 : 0;
@@ -1664,7 +1664,7 @@ void DESCrypt (const u32 SALT, const u32 K00, const u32 K01, const u32 K02, cons
 
 #else
 
-void DESCrypt (const u32 SALT, const u32 K00, const u32 K01, const u32 K02, const u32 K03, const u32 K04, const u32 K05, const u32 K06, const u32 K07, const u32 K08, const u32 K09, const u32 K10, const u32 K11, const u32 K12, const u32 K13, const u32 K14, const u32 K15, const u32 K16, const u32 K17, const u32 K18, const u32 K19, const u32 K20, const u32 K21, const u32 K22, const u32 K23, const u32 K24, const u32 K25, const u32 K26, const u32 K27, const u32 K28, const u32 K29, const u32 K30, const u32 K31, const u32 K32, const u32 K33, const u32 K34, const u32 K35, const u32 K36, const u32 K37, const u32 K38, const u32 K39, const u32 K40, const u32 K41, const u32 K42, const u32 K43, const u32 K44, const u32 K45, const u32 K46, const u32 K47, const u32 K48, const u32 K49, const u32 K50, const u32 K51, const u32 K52, const u32 K53, const u32 K54, const u32 K55, u32 *D00, u32 *D01, u32 *D02, u32 *D03, u32 *D04, u32 *D05, u32 *D06, u32 *D07, u32 *D08, u32 *D09, u32 *D10, u32 *D11, u32 *D12, u32 *D13, u32 *D14, u32 *D15, u32 *D16, u32 *D17, u32 *D18, u32 *D19, u32 *D20, u32 *D21, u32 *D22, u32 *D23, u32 *D24, u32 *D25, u32 *D26, u32 *D27, u32 *D28, u32 *D29, u32 *D30, u32 *D31, u32 *D32, u32 *D33, u32 *D34, u32 *D35, u32 *D36, u32 *D37, u32 *D38, u32 *D39, u32 *D40, u32 *D41, u32 *D42, u32 *D43, u32 *D44, u32 *D45, u32 *D46, u32 *D47, u32 *D48, u32 *D49, u32 *D50, u32 *D51, u32 *D52, u32 *D53, u32 *D54, u32 *D55, u32 *D56, u32 *D57, u32 *D58, u32 *D59, u32 *D60, u32 *D61, u32 *D62, u32 *D63)
+DECLSPEC void DESCrypt (const u32 SALT, const u32 K00, const u32 K01, const u32 K02, const u32 K03, const u32 K04, const u32 K05, const u32 K06, const u32 K07, const u32 K08, const u32 K09, const u32 K10, const u32 K11, const u32 K12, const u32 K13, const u32 K14, const u32 K15, const u32 K16, const u32 K17, const u32 K18, const u32 K19, const u32 K20, const u32 K21, const u32 K22, const u32 K23, const u32 K24, const u32 K25, const u32 K26, const u32 K27, const u32 K28, const u32 K29, const u32 K30, const u32 K31, const u32 K32, const u32 K33, const u32 K34, const u32 K35, const u32 K36, const u32 K37, const u32 K38, const u32 K39, const u32 K40, const u32 K41, const u32 K42, const u32 K43, const u32 K44, const u32 K45, const u32 K46, const u32 K47, const u32 K48, const u32 K49, const u32 K50, const u32 K51, const u32 K52, const u32 K53, const u32 K54, const u32 K55, u32 *D00, u32 *D01, u32 *D02, u32 *D03, u32 *D04, u32 *D05, u32 *D06, u32 *D07, u32 *D08, u32 *D09, u32 *D10, u32 *D11, u32 *D12, u32 *D13, u32 *D14, u32 *D15, u32 *D16, u32 *D17, u32 *D18, u32 *D19, u32 *D20, u32 *D21, u32 *D22, u32 *D23, u32 *D24, u32 *D25, u32 *D26, u32 *D27, u32 *D28, u32 *D29, u32 *D30, u32 *D31, u32 *D32, u32 *D33, u32 *D34, u32 *D35, u32 *D36, u32 *D37, u32 *D38, u32 *D39, u32 *D40, u32 *D41, u32 *D42, u32 *D43, u32 *D44, u32 *D45, u32 *D46, u32 *D47, u32 *D48, u32 *D49, u32 *D50, u32 *D51, u32 *D52, u32 *D53, u32 *D54, u32 *D55, u32 *D56, u32 *D57, u32 *D58, u32 *D59, u32 *D60, u32 *D61, u32 *D62, u32 *D63)
 {
   sXXX_DECL u32 s001 = (0x001 & SALT) ? 0xffffffff : 0;
   sXXX_DECL u32 s002 = (0x002 & SALT) ? 0xffffffff : 0;
@@ -1792,7 +1792,7 @@ void DESCrypt (const u32 SALT, const u32 K00, const u32 K01, const u32 K02, cons
 
 #endif
 
-void transpose32c (u32 data[32])
+DECLSPEC void transpose32c (u32 data[32])
 {
   #define swap(x,y,j,m)               \
      t  = ((x) ^ ((y) >> (j))) & (m); \
@@ -1883,7 +1883,7 @@ void transpose32c (u32 data[32])
   swap (data[30], data[31],  1, 0x55555555);
 }
 
-void m01500m (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bs_word_t *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01500m (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bs_word_t *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * base
@@ -2255,7 +2255,7 @@ void m01500m (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __glo
   }
 }
 
-void m01500s (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bs_word_t *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01500s (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bs_word_t *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * base
diff --git a/OpenCL/m01600-optimized.cl b/OpenCL/m01600-optimized.cl
index 82dd70e9c..b1cede218 100644
--- a/OpenCL/m01600-optimized.cl
+++ b/OpenCL/m01600-optimized.cl
@@ -16,7 +16,7 @@
 #define md5apr1_magic0 0x72706124u
 #define md5apr1_magic1 0x00002431u
 
-void memcat16 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[4])
+DECLSPEC void memcat16 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[4])
 {
   u32 tmp0;
   u32 tmp1;
@@ -129,7 +129,7 @@ void memcat16 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const
   }
 }
 
-void memcat16_x80 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[4])
+DECLSPEC void memcat16_x80 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[4])
 {
   u32 tmp0;
   u32 tmp1;
@@ -244,7 +244,7 @@ void memcat16_x80 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], c
   }
 }
 
-void memcat8 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[2])
+DECLSPEC void memcat8 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[2])
 {
   u32 tmp0;
   u32 tmp1;
@@ -333,7 +333,7 @@ void memcat8 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const
   }
 }
 
-void append_sign (u32 block0[4], u32 block1[4], const u32 block_len)
+DECLSPEC void append_sign (u32 block0[4], u32 block1[4], const u32 block_len)
 {
   switch (block_len)
   {
@@ -431,7 +431,7 @@ void append_sign (u32 block0[4], u32 block1[4], const u32 block_len)
   }
 }
 
-void append_1st (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 block_len, const u32 append)
+DECLSPEC void append_1st (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 block_len, const u32 append)
 {
   switch (block_len)
   {
diff --git a/OpenCL/m01700_a0-optimized.cl b/OpenCL/m01700_a0-optimized.cl
index b891f704c..f30493d5e 100644
--- a/OpenCL/m01700_a0-optimized.cl
+++ b/OpenCL/m01700_a0-optimized.cl
@@ -38,7 +38,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
diff --git a/OpenCL/m01700_a1-optimized.cl b/OpenCL/m01700_a1-optimized.cl
index e55f1f9a3..2de5899ca 100644
--- a/OpenCL/m01700_a1-optimized.cl
+++ b/OpenCL/m01700_a1-optimized.cl
@@ -36,7 +36,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
diff --git a/OpenCL/m01700_a3-optimized.cl b/OpenCL/m01700_a3-optimized.cl
index e06cb28c8..ec2130db6 100644
--- a/OpenCL/m01700_a3-optimized.cl
+++ b/OpenCL/m01700_a3-optimized.cl
@@ -36,7 +36,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
@@ -135,7 +135,7 @@ void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], con
   digest[7] = h;
 }
 
-void m01700m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01700m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -200,7 +200,7 @@ void m01700m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m01700s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01700s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01710_a0-optimized.cl b/OpenCL/m01710_a0-optimized.cl
index 821abb685..03c6a6bc3 100644
--- a/OpenCL/m01710_a0-optimized.cl
+++ b/OpenCL/m01710_a0-optimized.cl
@@ -38,7 +38,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
diff --git a/OpenCL/m01710_a1-optimized.cl b/OpenCL/m01710_a1-optimized.cl
index 72fb2a125..11f4d9967 100644
--- a/OpenCL/m01710_a1-optimized.cl
+++ b/OpenCL/m01710_a1-optimized.cl
@@ -36,7 +36,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
diff --git a/OpenCL/m01710_a3-optimized.cl b/OpenCL/m01710_a3-optimized.cl
index 0a061a901..17f5dfe1c 100644
--- a/OpenCL/m01710_a3-optimized.cl
+++ b/OpenCL/m01710_a3-optimized.cl
@@ -36,7 +36,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
@@ -135,7 +135,7 @@ void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], con
   digest[7] = h;
 }
 
-void m01710m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01710m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -251,7 +251,7 @@ void m01710m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m01710s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01710s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01720_a0-optimized.cl b/OpenCL/m01720_a0-optimized.cl
index 47c07083a..7ca358ad3 100644
--- a/OpenCL/m01720_a0-optimized.cl
+++ b/OpenCL/m01720_a0-optimized.cl
@@ -38,7 +38,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
diff --git a/OpenCL/m01720_a1-optimized.cl b/OpenCL/m01720_a1-optimized.cl
index 72c2715a2..d41550bd5 100644
--- a/OpenCL/m01720_a1-optimized.cl
+++ b/OpenCL/m01720_a1-optimized.cl
@@ -36,7 +36,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
diff --git a/OpenCL/m01720_a3-optimized.cl b/OpenCL/m01720_a3-optimized.cl
index 4604afb06..00c5275a0 100644
--- a/OpenCL/m01720_a3-optimized.cl
+++ b/OpenCL/m01720_a3-optimized.cl
@@ -36,7 +36,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
@@ -135,7 +135,7 @@ void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], con
   digest[7] = h;
 }
 
-void m01720m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01720m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -253,7 +253,7 @@ void m01720m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m01720s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01720s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01730_a0-optimized.cl b/OpenCL/m01730_a0-optimized.cl
index efb24779a..f41a5fdf6 100644
--- a/OpenCL/m01730_a0-optimized.cl
+++ b/OpenCL/m01730_a0-optimized.cl
@@ -38,7 +38,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
diff --git a/OpenCL/m01730_a1-optimized.cl b/OpenCL/m01730_a1-optimized.cl
index dbb3141df..98e1cc267 100644
--- a/OpenCL/m01730_a1-optimized.cl
+++ b/OpenCL/m01730_a1-optimized.cl
@@ -36,7 +36,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
diff --git a/OpenCL/m01730_a3-optimized.cl b/OpenCL/m01730_a3-optimized.cl
index 26d2e4e43..b359cfe81 100644
--- a/OpenCL/m01730_a3-optimized.cl
+++ b/OpenCL/m01730_a3-optimized.cl
@@ -36,7 +36,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
@@ -135,7 +135,7 @@ void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], con
   digest[7] = h;
 }
 
-void m01730m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01730m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -251,7 +251,7 @@ void m01730m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m01730s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01730s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01740_a0-optimized.cl b/OpenCL/m01740_a0-optimized.cl
index a8a8eced1..4318e527a 100644
--- a/OpenCL/m01740_a0-optimized.cl
+++ b/OpenCL/m01740_a0-optimized.cl
@@ -38,7 +38,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
diff --git a/OpenCL/m01740_a1-optimized.cl b/OpenCL/m01740_a1-optimized.cl
index 310af40ae..4cb0cabe8 100644
--- a/OpenCL/m01740_a1-optimized.cl
+++ b/OpenCL/m01740_a1-optimized.cl
@@ -36,7 +36,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
diff --git a/OpenCL/m01740_a3-optimized.cl b/OpenCL/m01740_a3-optimized.cl
index dc2ac9796..19e561703 100644
--- a/OpenCL/m01740_a3-optimized.cl
+++ b/OpenCL/m01740_a3-optimized.cl
@@ -36,7 +36,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
@@ -135,7 +135,7 @@ void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], con
   digest[7] = h;
 }
 
-void m01740m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01740m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -253,7 +253,7 @@ void m01740m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m01740s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01740s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01750_a0-optimized.cl b/OpenCL/m01750_a0-optimized.cl
index 6d2c9358a..2ec9c91d4 100644
--- a/OpenCL/m01750_a0-optimized.cl
+++ b/OpenCL/m01750_a0-optimized.cl
@@ -15,7 +15,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha512.cl"
 
-void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], const u64x w2[4], const u64x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], const u64x w2[4], const u64x w3[4], u64x digest[8])
 {
   u32x t0[4];
   u32x t1[4];
@@ -62,7 +62,7 @@ void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], cons
   sha512_transform_vector (t0, t1, t2, t3, t4, t5, t6, t7, digest);
 }
 
-void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8])
+DECLSPEC void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8])
 {
   u64x w0_t[4];
   u64x w1_t[4];
@@ -126,7 +126,7 @@ void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[
   sha512_transform_transport_vector (w0_t, w1_t, w2_t, w3_t, opad);
 }
 
-void hmac_sha512_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   u64x w0_t[4];
   u64x w1_t[4];
diff --git a/OpenCL/m01750_a1-optimized.cl b/OpenCL/m01750_a1-optimized.cl
index 4978ce856..7be127de6 100644
--- a/OpenCL/m01750_a1-optimized.cl
+++ b/OpenCL/m01750_a1-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha512.cl"
 
-void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], const u64x w2[4], const u64x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], const u64x w2[4], const u64x w3[4], u64x digest[8])
 {
   u32x t0[4];
   u32x t1[4];
@@ -60,7 +60,7 @@ void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], cons
   sha512_transform_vector (t0, t1, t2, t3, t4, t5, t6, t7, digest);
 }
 
-void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8])
+DECLSPEC void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8])
 {
   u64x w0_t[4];
   u64x w1_t[4];
@@ -124,7 +124,7 @@ void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[
   sha512_transform_transport_vector (w0_t, w1_t, w2_t, w3_t, opad);
 }
 
-void hmac_sha512_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   u64x w0_t[4];
   u64x w1_t[4];
diff --git a/OpenCL/m01750_a3-optimized.cl b/OpenCL/m01750_a3-optimized.cl
index 78e895323..4f8ec688a 100644
--- a/OpenCL/m01750_a3-optimized.cl
+++ b/OpenCL/m01750_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha512.cl"
 
-void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], const u64x w2[4], const u64x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], const u64x w2[4], const u64x w3[4], u64x digest[8])
 {
   u32x t0[4];
   u32x t1[4];
@@ -60,7 +60,7 @@ void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], cons
   sha512_transform_vector (t0, t1, t2, t3, t4, t5, t6, t7, digest);
 }
 
-void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8])
+DECLSPEC void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8])
 {
   u64x w0_t[4];
   u64x w1_t[4];
@@ -124,7 +124,7 @@ void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[
   sha512_transform_transport_vector (w0_t, w1_t, w2_t, w3_t, opad);
 }
 
-void hmac_sha512_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   u64x w0_t[4];
   u64x w1_t[4];
@@ -188,7 +188,7 @@ void hmac_sha512_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[
   sha512_transform_transport_vector (w0_t, w1_t, w2_t, w3_t, digest);
 }
 
-void m01750m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01750m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -298,7 +298,7 @@ void m01750m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m01750s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01750s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01760_a0-optimized.cl b/OpenCL/m01760_a0-optimized.cl
index 24cc848f9..00fadf2b9 100644
--- a/OpenCL/m01760_a0-optimized.cl
+++ b/OpenCL/m01760_a0-optimized.cl
@@ -15,7 +15,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha512.cl"
 
-void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], const u64x w2[4], const u64x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], const u64x w2[4], const u64x w3[4], u64x digest[8])
 {
   u32x t0[4];
   u32x t1[4];
@@ -62,7 +62,7 @@ void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], cons
   sha512_transform_vector (t0, t1, t2, t3, t4, t5, t6, t7, digest);
 }
 
-void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8])
+DECLSPEC void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8])
 {
   u64x w0_t[4];
   u64x w1_t[4];
@@ -126,7 +126,7 @@ void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[
   sha512_transform_transport_vector (w0_t, w1_t, w2_t, w3_t, opad);
 }
 
-void hmac_sha512_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   u64x w0_t[4];
   u64x w1_t[4];
diff --git a/OpenCL/m01760_a1-optimized.cl b/OpenCL/m01760_a1-optimized.cl
index f8e69dee5..5b8815cf6 100644
--- a/OpenCL/m01760_a1-optimized.cl
+++ b/OpenCL/m01760_a1-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha512.cl"
 
-void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], const u64x w2[4], const u64x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], const u64x w2[4], const u64x w3[4], u64x digest[8])
 {
   u32x t0[4];
   u32x t1[4];
@@ -60,7 +60,7 @@ void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], cons
   sha512_transform_vector (t0, t1, t2, t3, t4, t5, t6, t7, digest);
 }
 
-void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8])
+DECLSPEC void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8])
 {
   u64x w0_t[4];
   u64x w1_t[4];
@@ -124,7 +124,7 @@ void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[
   sha512_transform_transport_vector (w0_t, w1_t, w2_t, w3_t, opad);
 }
 
-void hmac_sha512_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   u64x w0_t[4];
   u64x w1_t[4];
diff --git a/OpenCL/m01760_a3-optimized.cl b/OpenCL/m01760_a3-optimized.cl
index 550350c79..c15d7e351 100644
--- a/OpenCL/m01760_a3-optimized.cl
+++ b/OpenCL/m01760_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha512.cl"
 
-void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], const u64x w2[4], const u64x w3[4], u64x digest[8])
+DECLSPEC void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], const u64x w2[4], const u64x w3[4], u64x digest[8])
 {
   u32x t0[4];
   u32x t1[4];
@@ -60,7 +60,7 @@ void sha512_transform_transport_vector (const u64x w0[4], const u64x w1[4], cons
   sha512_transform_vector (t0, t1, t2, t3, t4, t5, t6, t7, digest);
 }
 
-void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8])
+DECLSPEC void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8])
 {
   u64x w0_t[4];
   u64x w1_t[4];
@@ -124,7 +124,7 @@ void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[
   sha512_transform_transport_vector (w0_t, w1_t, w2_t, w3_t, opad);
 }
 
-void hmac_sha512_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   u64x w0_t[4];
   u64x w1_t[4];
@@ -188,7 +188,7 @@ void hmac_sha512_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[
   sha512_transform_transport_vector (w0_t, w1_t, w2_t, w3_t, digest);
 }
 
-void m01760m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esal_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01760m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esal_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -298,7 +298,7 @@ void m01760m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m01760s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m01760s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m01800-optimized.cl b/OpenCL/m01800-optimized.cl
index 943d73c82..b4a2d856d 100644
--- a/OpenCL/m01800-optimized.cl
+++ b/OpenCL/m01800-optimized.cl
@@ -24,7 +24,7 @@ typedef struct
 
 } orig_sha512_ctx_t;
 
-void sha512_transform_transport (const u64 *w, u64 *digest)
+DECLSPEC void sha512_transform_transport (const u64 *w, u64 *digest)
 {
   u32 t0[4];
   u32 t1[4];
@@ -71,7 +71,7 @@ void sha512_transform_transport (const u64 *w, u64 *digest)
   sha512_transform (t0, t1, t2, t3, t4, t5, t6, t7, digest);
 }
 
-void orig_sha512_init (orig_sha512_ctx_t *sha512_ctx)
+DECLSPEC void orig_sha512_init (orig_sha512_ctx_t *sha512_ctx)
 {
   sha512_ctx->state[0] = SHA512M_A;
   sha512_ctx->state[1] = SHA512M_B;
@@ -85,7 +85,7 @@ void orig_sha512_init (orig_sha512_ctx_t *sha512_ctx)
   sha512_ctx->len = 0;
 }
 
-void orig_sha512_update (orig_sha512_ctx_t *sha512_ctx, const u64 *buf, int len)
+DECLSPEC void orig_sha512_update (orig_sha512_ctx_t *sha512_ctx, const u64 *buf, int len)
 {
   int pos = sha512_ctx->len & 0x7f;
 
@@ -118,7 +118,7 @@ void orig_sha512_update (orig_sha512_ctx_t *sha512_ctx, const u64 *buf, int len)
   }
 }
 
-void orig_sha512_final (orig_sha512_ctx_t *sha512_ctx)
+DECLSPEC void orig_sha512_final (orig_sha512_ctx_t *sha512_ctx)
 {
   int pos = sha512_ctx->len & 0x7f;
 
diff --git a/OpenCL/m02000_a0.cl b/OpenCL/m02000_a0.cl
index 1cb42cdec..5e5d7b439 100644
--- a/OpenCL/m02000_a0.cl
+++ b/OpenCL/m02000_a0.cl
@@ -3,32 +3,8 @@
  * License.....: MIT
  */
 
-typedef uint  u32;
-typedef ulong u64;
-
-typedef struct pw
-{
-  u32 i[64];
-
-  u32 pw_len;
-
-} pw_t;
-
-static u32 l32_from_64_S (u64 a)
-{
-  const u32 r = (u32) (a);
-
-  return r;
-}
-
-static u32 h32_from_64_S (u64 a)
-{
-  a >>= 32;
-
-  const u32 r = (u32) (a);
-
-  return r;
-}
+#include "inc_vendor.cl"
+#include "inc_types.cl"
 
 __kernel void gpu_memset (__global uint4 *buf, const u32 value, const u64 gid_max)
 {
diff --git a/OpenCL/m02000_a1.cl b/OpenCL/m02000_a1.cl
index 1cb42cdec..5e5d7b439 100644
--- a/OpenCL/m02000_a1.cl
+++ b/OpenCL/m02000_a1.cl
@@ -3,32 +3,8 @@
  * License.....: MIT
  */
 
-typedef uint  u32;
-typedef ulong u64;
-
-typedef struct pw
-{
-  u32 i[64];
-
-  u32 pw_len;
-
-} pw_t;
-
-static u32 l32_from_64_S (u64 a)
-{
-  const u32 r = (u32) (a);
-
-  return r;
-}
-
-static u32 h32_from_64_S (u64 a)
-{
-  a >>= 32;
-
-  const u32 r = (u32) (a);
-
-  return r;
-}
+#include "inc_vendor.cl"
+#include "inc_types.cl"
 
 __kernel void gpu_memset (__global uint4 *buf, const u32 value, const u64 gid_max)
 {
diff --git a/OpenCL/m02000_a3.cl b/OpenCL/m02000_a3.cl
index 1cb42cdec..5e5d7b439 100644
--- a/OpenCL/m02000_a3.cl
+++ b/OpenCL/m02000_a3.cl
@@ -3,32 +3,8 @@
  * License.....: MIT
  */
 
-typedef uint  u32;
-typedef ulong u64;
-
-typedef struct pw
-{
-  u32 i[64];
-
-  u32 pw_len;
-
-} pw_t;
-
-static u32 l32_from_64_S (u64 a)
-{
-  const u32 r = (u32) (a);
-
-  return r;
-}
-
-static u32 h32_from_64_S (u64 a)
-{
-  a >>= 32;
-
-  const u32 r = (u32) (a);
-
-  return r;
-}
+#include "inc_vendor.cl"
+#include "inc_types.cl"
 
 __kernel void gpu_memset (__global uint4 *buf, const u32 value, const u64 gid_max)
 {
diff --git a/OpenCL/m02100.cl b/OpenCL/m02100.cl
index 66420eb38..f7738b04b 100644
--- a/OpenCL/m02100.cl
+++ b/OpenCL/m02100.cl
@@ -17,7 +17,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m02400_a3-optimized.cl b/OpenCL/m02400_a3-optimized.cl
index 0a411e3d3..efd8edb2b 100644
--- a/OpenCL/m02400_a3-optimized.cl
+++ b/OpenCL/m02400_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m02400m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m02400m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -227,7 +227,7 @@ void m02400m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m02400s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m02400s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m02410_a3-optimized.cl b/OpenCL/m02410_a3-optimized.cl
index e91f4d3fa..89b35cc51 100644
--- a/OpenCL/m02410_a3-optimized.cl
+++ b/OpenCL/m02410_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m02410m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m02410m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -276,7 +276,7 @@ void m02410m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m02410s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m02410s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m02500.cl b/OpenCL/m02500.cl
index 74ac366a2..77a06d8fe 100644
--- a/OpenCL/m02500.cl
+++ b/OpenCL/m02500.cl
@@ -19,7 +19,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void make_kn (u32 *k)
+DECLSPEC void make_kn (u32 *k)
 {
   u32 kl[4];
   u32 kr[4];
@@ -49,7 +49,7 @@ void make_kn (u32 *k)
   k[3] ^= c * 0x87000000;
 }
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m02501.cl b/OpenCL/m02501.cl
index b61713868..20250476b 100644
--- a/OpenCL/m02501.cl
+++ b/OpenCL/m02501.cl
@@ -19,12 +19,12 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-u8 hex_convert (const u8 c)
+DECLSPEC u8 hex_convert (const u8 c)
 {
   return (c & 15) + (c >> 6) * 9;
 }
 
-u8 hex_to_u8 (const u8 hex[2])
+DECLSPEC u8 hex_to_u8 (const u8 hex[2])
 {
   u8 v = 0;
 
@@ -34,7 +34,7 @@ u8 hex_to_u8 (const u8 hex[2])
   return (v);
 }
 
-void make_kn (u32 *k)
+DECLSPEC void make_kn (u32 *k)
 {
   u32 kl[4];
   u32 kr[4];
diff --git a/OpenCL/m02610_a3-optimized.cl b/OpenCL/m02610_a3-optimized.cl
index e83190836..cb5eda61c 100644
--- a/OpenCL/m02610_a3-optimized.cl
+++ b/OpenCL/m02610_a3-optimized.cl
@@ -24,7 +24,7 @@
 #define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m02610m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m02610m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -282,7 +282,7 @@ void m02610m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m02610s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m02610s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m02710_a3-optimized.cl b/OpenCL/m02710_a3-optimized.cl
index 84ce9629b..b3ca17691 100644
--- a/OpenCL/m02710_a3-optimized.cl
+++ b/OpenCL/m02710_a3-optimized.cl
@@ -24,7 +24,7 @@
 #define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m02710m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m02710m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -367,7 +367,7 @@ void m02710m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m02710s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m02710s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m02810_a3-optimized.cl b/OpenCL/m02810_a3-optimized.cl
index 2d31dba90..e46eb876a 100644
--- a/OpenCL/m02810_a3-optimized.cl
+++ b/OpenCL/m02810_a3-optimized.cl
@@ -24,7 +24,7 @@
 #define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m02810m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m02810m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -368,7 +368,7 @@ void m02810m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m02810s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m02810s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m03000_a0.cl b/OpenCL/m03000_a0.cl
index ed8ee83b5..f135be832 100644
--- a/OpenCL/m03000_a0.cl
+++ b/OpenCL/m03000_a0.cl
@@ -345,7 +345,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32 iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32 r = data[0];
   u32 l = data[1];
@@ -387,7 +387,7 @@ void _des_crypt_encrypt (u32 iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __lo
   iv[1] = rotl32 (r, 29);
 }
 
-void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
@@ -456,7 +456,7 @@ void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (
   }
 }
 
-void transform_netntlmv1_key (const u32 w0, const u32x w1, u32x out[2])
+DECLSPEC void transform_netntlmv1_key (const u32 w0, const u32x w1, u32x out[2])
 {
   u32 t[8];
 
diff --git a/OpenCL/m03000_a1.cl b/OpenCL/m03000_a1.cl
index aea2f080b..0d6f666f0 100644
--- a/OpenCL/m03000_a1.cl
+++ b/OpenCL/m03000_a1.cl
@@ -343,7 +343,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32 iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32 r = data[0];
   u32 l = data[1];
@@ -385,7 +385,7 @@ void _des_crypt_encrypt (u32 iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __lo
   iv[1] = rotl32 (r, 29);
 }
 
-void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
@@ -454,7 +454,7 @@ void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (
   }
 }
 
-void transform_netntlmv1_key (const u32 w0, const u32x w1, u32x out[2])
+DECLSPEC void transform_netntlmv1_key (const u32 w0, const u32x w1, u32x out[2])
 {
   u32 t[8];
 
diff --git a/OpenCL/m03000_a3.cl b/OpenCL/m03000_a3.cl
index f2b01d31b..7354e2d5a 100644
--- a/OpenCL/m03000_a3.cl
+++ b/OpenCL/m03000_a3.cl
@@ -54,7 +54,7 @@
 
 #define LUT(a,b,c,d,e) u32 a; asm volatile ("lop3.b32 %0, %1, %2, %3, "#e";" : "=r"(a): "r"(b), "r"(c), "r"(d));
 
-void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(xAA55AA5500550055, a1, a4, a6, 0xC1)
   LUT(xA55AA55AF0F5F0F5, a3, a6, xAA55AA5500550055, 0x9E)
@@ -88,7 +88,7 @@ void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(xEEEEEEEE99999999, a1, a2, a6, 0x97)
   LUT(xFFFFEEEE66666666, a5, a6, xEEEEEEEE99999999, 0x67)
@@ -121,7 +121,7 @@ void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(xA50FA50FA50FA50F, a1, a3, a4, 0xC9)
   LUT(xF0F00F0FF0F0F0F0, a3, a5, a6, 0x4B)
@@ -155,7 +155,7 @@ void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(x55F055F055F055F0, a1, a3, a4, 0x72)
   LUT(xA500F5F0A500F5F0, a3, a5, x55F055F055F055F0, 0xAD)
@@ -182,7 +182,7 @@ void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(xA0A0A0A0FFFFFFFF, a1, a3, a6, 0xAB)
   LUT(xFFFF00005555FFFF, a1, a5, a6, 0xB9)
@@ -216,7 +216,7 @@ void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(x5050F5F55050F5F5, a1, a3, a5, 0xB2)
   LUT(x6363C6C66363C6C6, a1, a2, x5050F5F55050F5F5, 0x66)
@@ -249,7 +249,7 @@ void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(x88AA88AA88AA88AA, a1, a2, a4, 0x0B)
   LUT(xAAAAFF00AAAAFF00, a1, a4, a5, 0x27)
@@ -282,7 +282,7 @@ void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(xEEEE3333EEEE3333, a1, a2, a5, 0x9D)
   LUT(xBBBBBBBBBBBBBBBB, a1, a1, a2, 0x83)
@@ -340,7 +340,7 @@ void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
  * The effort has been sponsored by Rapid7: http://www.rapid7.com
  */
 
-void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x55005500, x5A0F5A0F, x3333FFFF, x66666666, x22226666, x2D2D6969,
         x25202160;
@@ -415,7 +415,7 @@ void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x33CC33CC;
     u32 x55550000, x00AA00FF, x33BB33FF;
@@ -486,7 +486,7 @@ void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x44444444, x0F0FF0F0, x4F4FF4F4, x00FFFF00, x00AAAA00, x4FE55EF4;
     u32 x3C3CC3C3, x3C3C0000, x7373F4F4, x0C840A00;
@@ -557,7 +557,7 @@ void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out3 ^= x21;
 }
 
-void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x5A5A5A5A, x0F0FF0F0;
     u32 x33FF33FF, x33FFCC00, x0C0030F0, x0C0CC0C0, x0CF3C03F, x5EFBDA7F,
@@ -611,7 +611,7 @@ void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x77777777, x77770000, x22225555, x11116666, x1F1F6F6F;
     u32 x70700000, x43433333, x00430033, x55557777, x55167744, x5A19784B;
@@ -684,7 +684,7 @@ void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out2 ^= x11;
 }
 
-void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x33CC33CC;
     u32 x3333FFFF, x11115555, x22DD6699, x22DD9966, x00220099;
@@ -757,7 +757,7 @@ void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out1 ^= x01;
 }
 
-void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x0FF00FF0, x3CC33CC3, x00003CC3, x0F000F00, x5A555A55, x00001841;
     u32 x00000F00, x33333C33, x7B777E77, x0FF0F00F, x74878E78;
@@ -828,7 +828,7 @@ void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out2 ^= x11;
 }
 
-void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x0C0C0C0C, x0000F0F0, x00FFF00F, x00555005, x00515001;
     u32 x33000330, x77555775, x30303030, x3030CFCF, x30104745, x30555745;
@@ -921,7 +921,7 @@ void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
  * The effort has been sponsored by Rapid7: http://www.rapid7.com
  */
 
-void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x55005500, x5A0F5A0F, x3333FFFF, x66666666, x22226666, x2D2D6969,
         x25202160;
@@ -996,7 +996,7 @@ void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x33CC33CC;
     u32 x55550000, x00AA00FF, x33BB33FF;
@@ -1067,7 +1067,7 @@ void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x44444444, x0F0FF0F0, x4F4FF4F4, x00FFFF00, x00AAAA00, x4FE55EF4;
     u32 x3C3CC3C3, x3C3C0000, x7373F4F4, x0C840A00;
@@ -1138,7 +1138,7 @@ void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out3 ^= x21;
 }
 
-void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x5A5A5A5A, x0F0FF0F0;
     u32 x33FF33FF, x33FFCC00, x0C0030F0, x0C0CC0C0, x0CF3C03F, x5EFBDA7F,
@@ -1192,7 +1192,7 @@ void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x77777777, x77770000, x22225555, x11116666, x1F1F6F6F;
     u32 x70700000, x43433333, x00430033, x55557777, x55167744, x5A19784B;
@@ -1265,7 +1265,7 @@ void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out2 ^= x11;
 }
 
-void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x33CC33CC;
     u32 x3333FFFF, x11115555, x22DD6699, x22DD9966, x00220099;
@@ -1338,7 +1338,7 @@ void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out1 ^= x01;
 }
 
-void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x0FF00FF0, x3CC33CC3, x00003CC3, x0F000F00, x5A555A55, x00001841;
     u32 x00000F00, x33333C33, x7B777E77, x0FF0F00F, x74878E78;
@@ -1409,7 +1409,7 @@ void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out2 ^= x11;
 }
 
-void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x0C0C0C0C, x0000F0F0, x00FFF00F, x00555005, x00515001;
     u32 x33000330, x77555775, x30303030, x3030CFCF, x30104745, x30555745;
@@ -1528,7 +1528,7 @@ void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
 #define KEYSET07 { k00 = K31; k01 = K35; k02 = K52; k03 = K43; k04 = K08; k05 = K37; k06 = K51; k07 = K15; k08 = K49; k09 = K30; k10 = K07; k11 = K02; k12 = K50; k13 = K21; k14 = K45; k15 = K44; k16 = K29; k17 = K16; k18 = K42; k19 = K23; k20 = K22; k21 = K14; k22 = K38; k23 = K01; k24 = K10; k25 = K47; k26 = K53; k27 = K11; k28 = K27; k29 = K26; k30 = K05; k31 = K17; k32 = K54; k33 = K41; k34 = K39; k35 = K20; k36 = K48; k37 = K13; k38 = K24; k39 = K19; k40 = K32; k41 = K40; k42 = K34; k43 = K03; k44 = K06; k45 = K18; k46 = K12; k47 = K46; }
 #define KEYSET17 { k00 = K15; k01 = K51; k02 = K36; k03 = K02; k04 = K49; k05 = K21; k06 = K35; k07 = K31; k08 = K08; k09 = K14; k10 = K23; k11 = K43; k12 = K09; k13 = K37; k14 = K29; k15 = K28; k16 = K45; k17 = K00; k18 = K01; k19 = K07; k20 = K38; k21 = K30; k22 = K22; k23 = K42; k24 = K26; k25 = K04; k26 = K41; k27 = K54; k28 = K39; k29 = K10; k30 = K48; k31 = K33; k32 = K11; k33 = K53; k34 = K27; k35 = K32; k36 = K05; k37 = K25; k38 = K40; k39 = K03; k40 = K20; k41 = K24; k42 = K46; k43 = K19; k44 = K18; k45 = K06; k46 = K55; k47 = K34; }
 
-void DES (const u32 K00, const u32 K01, const u32 K02, const u32 K03, const u32 K04, const u32 K05, const u32 K06, const u32 K07, const u32 K08, const u32 K09, const u32 K10, const u32 K11, const u32 K12, const u32 K13, const u32 K14, const u32 K15, const u32 K16, const u32 K17, const u32 K18, const u32 K19, const u32 K20, const u32 K21, const u32 K22, const u32 K23, const u32 K24, const u32 K25, const u32 K26, const u32 K27, const u32 K28, const u32 K29, const u32 K30, const u32 K31, const u32 K32, const u32 K33, const u32 K34, const u32 K35, const u32 K36, const u32 K37, const u32 K38, const u32 K39, const u32 K40, const u32 K41, const u32 K42, const u32 K43, const u32 K44, const u32 K45, const u32 K46, const u32 K47, const u32 K48, const u32 K49, const u32 K50, const u32 K51, const u32 K52, const u32 K53, const u32 K54, const u32 K55, u32 *D00, u32 *D01, u32 *D02, u32 *D03, u32 *D04, u32 *D05, u32 *D06, u32 *D07, u32 *D08, u32 *D09, u32 *D10, u32 *D11, u32 *D12, u32 *D13, u32 *D14, u32 *D15, u32 *D16, u32 *D17, u32 *D18, u32 *D19, u32 *D20, u32 *D21, u32 *D22, u32 *D23, u32 *D24, u32 *D25, u32 *D26, u32 *D27, u32 *D28, u32 *D29, u32 *D30, u32 *D31, u32 *D32, u32 *D33, u32 *D34, u32 *D35, u32 *D36, u32 *D37, u32 *D38, u32 *D39, u32 *D40, u32 *D41, u32 *D42, u32 *D43, u32 *D44, u32 *D45, u32 *D46, u32 *D47, u32 *D48, u32 *D49, u32 *D50, u32 *D51, u32 *D52, u32 *D53, u32 *D54, u32 *D55, u32 *D56, u32 *D57, u32 *D58, u32 *D59, u32 *D60, u32 *D61, u32 *D62, u32 *D63)
+DECLSPEC void DES (const u32 K00, const u32 K01, const u32 K02, const u32 K03, const u32 K04, const u32 K05, const u32 K06, const u32 K07, const u32 K08, const u32 K09, const u32 K10, const u32 K11, const u32 K12, const u32 K13, const u32 K14, const u32 K15, const u32 K16, const u32 K17, const u32 K18, const u32 K19, const u32 K20, const u32 K21, const u32 K22, const u32 K23, const u32 K24, const u32 K25, const u32 K26, const u32 K27, const u32 K28, const u32 K29, const u32 K30, const u32 K31, const u32 K32, const u32 K33, const u32 K34, const u32 K35, const u32 K36, const u32 K37, const u32 K38, const u32 K39, const u32 K40, const u32 K41, const u32 K42, const u32 K43, const u32 K44, const u32 K45, const u32 K46, const u32 K47, const u32 K48, const u32 K49, const u32 K50, const u32 K51, const u32 K52, const u32 K53, const u32 K54, const u32 K55, u32 *D00, u32 *D01, u32 *D02, u32 *D03, u32 *D04, u32 *D05, u32 *D06, u32 *D07, u32 *D08, u32 *D09, u32 *D10, u32 *D11, u32 *D12, u32 *D13, u32 *D14, u32 *D15, u32 *D16, u32 *D17, u32 *D18, u32 *D19, u32 *D20, u32 *D21, u32 *D22, u32 *D23, u32 *D24, u32 *D25, u32 *D26, u32 *D27, u32 *D28, u32 *D29, u32 *D30, u32 *D31, u32 *D32, u32 *D33, u32 *D34, u32 *D35, u32 *D36, u32 *D37, u32 *D38, u32 *D39, u32 *D40, u32 *D41, u32 *D42, u32 *D43, u32 *D44, u32 *D45, u32 *D46, u32 *D47, u32 *D48, u32 *D49, u32 *D50, u32 *D51, u32 *D52, u32 *D53, u32 *D54, u32 *D55, u32 *D56, u32 *D57, u32 *D58, u32 *D59, u32 *D60, u32 *D61, u32 *D62, u32 *D63)
 {
   KXX_DECL u32 k00, k01, k02, k03, k04, k05;
   KXX_DECL u32 k06, k07, k08, k09, k10, k11;
@@ -1634,7 +1634,7 @@ void DES (const u32 K00, const u32 K01, const u32 K02, const u32 K03, const u32
   }
 }
 
-void transpose32c (u32 data[32])
+DECLSPEC void transpose32c (u32 data[32])
 {
   #define swap(x,y,j,m)               \
      t  = ((x) ^ ((y) >> (j))) & (m); \
@@ -1725,7 +1725,7 @@ void transpose32c (u32 data[32])
   swap (data[30], data[31],  1, 0x55555555);
 }
 
-void m03000m (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bs_word_t *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m03000m (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bs_word_t *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * base
@@ -2097,7 +2097,7 @@ void m03000m (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __glo
   }
 }
 
-void m03000s (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bs_word_t *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m03000s (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bs_word_t *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * base
diff --git a/OpenCL/m03100_a0-optimized.cl b/OpenCL/m03100_a0-optimized.cl
index 6890c4a19..530b73e06 100644
--- a/OpenCL/m03100_a0-optimized.cl
+++ b/OpenCL/m03100_a0-optimized.cl
@@ -361,7 +361,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32x tt;
 
@@ -415,7 +415,7 @@ void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __l
   iv[1] = r;
 }
 
-void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32x tt;
 
diff --git a/OpenCL/m03100_a1-optimized.cl b/OpenCL/m03100_a1-optimized.cl
index 636836c30..450148860 100644
--- a/OpenCL/m03100_a1-optimized.cl
+++ b/OpenCL/m03100_a1-optimized.cl
@@ -359,7 +359,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32x tt;
 
@@ -413,7 +413,7 @@ void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __l
   iv[1] = r;
 }
 
-void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32x tt;
 
diff --git a/OpenCL/m03100_a3-optimized.cl b/OpenCL/m03100_a3-optimized.cl
index fb2d246fd..ec5f8702e 100644
--- a/OpenCL/m03100_a3-optimized.cl
+++ b/OpenCL/m03100_a3-optimized.cl
@@ -358,7 +358,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32x tt;
 
@@ -412,7 +412,7 @@ void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __l
   iv[1] = r;
 }
 
-void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32x tt;
 
@@ -484,7 +484,7 @@ void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32
   }
 }
 
-void m03100m (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m03100m (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -665,7 +665,7 @@ void m03100m (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16],
   }
 }
 
-void m03100s (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m03100s (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m03200.cl b/OpenCL/m03200.cl
index 283e2aa75..fe18a00f7 100644
--- a/OpenCL/m03200.cl
+++ b/OpenCL/m03200.cl
@@ -370,7 +370,7 @@ __constant static u32a c_sbox3[256] =
   L ^= P[17];           \
 }
 
-void expand_key (u32 E[18], const u32 W[18], const u32 len)
+DECLSPEC void expand_key (u32 E[18], const u32 W[18], const u32 len)
 {
   u8 *E_cur  = (u8 *) E;
   u8 *E_stop = E_cur + 72;
diff --git a/OpenCL/m03710_a3-optimized.cl b/OpenCL/m03710_a3-optimized.cl
index 0a275dbc2..69b5ced62 100644
--- a/OpenCL/m03710_a3-optimized.cl
+++ b/OpenCL/m03710_a3-optimized.cl
@@ -24,7 +24,7 @@
 #define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m03710m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m03710m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -313,7 +313,7 @@ void m03710m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m03710s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m03710s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m03800_a3-optimized.cl b/OpenCL/m03800_a3-optimized.cl
index 1dede5e57..1692e9633 100644
--- a/OpenCL/m03800_a3-optimized.cl
+++ b/OpenCL/m03800_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m03800m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m03800m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -236,7 +236,7 @@ void m03800m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m03800s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m03800s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m03910_a3-optimized.cl b/OpenCL/m03910_a3-optimized.cl
index d2b2dbf5a..67eab4892 100644
--- a/OpenCL/m03910_a3-optimized.cl
+++ b/OpenCL/m03910_a3-optimized.cl
@@ -24,7 +24,7 @@
 #define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m03910m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m03910m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -368,7 +368,7 @@ void m03910m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m03910s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m03910s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m04010_a3-optimized.cl b/OpenCL/m04010_a3-optimized.cl
index edcc429c8..2bdd9ac39 100644
--- a/OpenCL/m04010_a3-optimized.cl
+++ b/OpenCL/m04010_a3-optimized.cl
@@ -24,7 +24,7 @@
 #define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m04010m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m04010m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -333,7 +333,7 @@ void m04010m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m04010s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m04010s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m04110_a3-optimized.cl b/OpenCL/m04110_a3-optimized.cl
index 83ef5085a..815cdd180 100644
--- a/OpenCL/m04110_a3-optimized.cl
+++ b/OpenCL/m04110_a3-optimized.cl
@@ -24,7 +24,7 @@
 #define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m04110m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m04110m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -361,7 +361,7 @@ void m04110m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m04110s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m04110s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m04310_a3-optimized.cl b/OpenCL/m04310_a3-optimized.cl
index f26d44ca3..bf3206abd 100644
--- a/OpenCL/m04310_a3-optimized.cl
+++ b/OpenCL/m04310_a3-optimized.cl
@@ -24,7 +24,7 @@
 #define uint_to_hex_upper8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m04310m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m04310m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -282,7 +282,7 @@ void m04310m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m04310s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m04310s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m04400_a3-optimized.cl b/OpenCL/m04400_a3-optimized.cl
index db9766240..8862da4d2 100644
--- a/OpenCL/m04400_a3-optimized.cl
+++ b/OpenCL/m04400_a3-optimized.cl
@@ -24,7 +24,7 @@
 #define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m04400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m04400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -285,7 +285,7 @@ void m04400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m04400s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m04400s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m04500_a3-optimized.cl b/OpenCL/m04500_a3-optimized.cl
index 9cfd63825..c8c67b994 100644
--- a/OpenCL/m04500_a3-optimized.cl
+++ b/OpenCL/m04500_a3-optimized.cl
@@ -24,7 +24,7 @@
 #define uint_to_hex_lower8_le(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m04500m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m04500m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -312,7 +312,7 @@ void m04500m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m04500s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m04500s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m04520_a3-optimized.cl b/OpenCL/m04520_a3-optimized.cl
index 4b902ade1..b08e9e803 100644
--- a/OpenCL/m04520_a3-optimized.cl
+++ b/OpenCL/m04520_a3-optimized.cl
@@ -24,7 +24,7 @@
 #define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m04520m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m04520m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -543,7 +543,7 @@ void m04520m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m04520s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m04520s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m04700_a3-optimized.cl b/OpenCL/m04700_a3-optimized.cl
index 42dced0eb..e4e7c20bc 100644
--- a/OpenCL/m04700_a3-optimized.cl
+++ b/OpenCL/m04700_a3-optimized.cl
@@ -24,7 +24,7 @@
 #define uint_to_hex_lower8_le(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m04700m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m04700m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -282,7 +282,7 @@ void m04700m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m04700s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m04700s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m04800_a3-optimized.cl b/OpenCL/m04800_a3-optimized.cl
index f43840754..5c75cef4c 100644
--- a/OpenCL/m04800_a3-optimized.cl
+++ b/OpenCL/m04800_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m04800m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m04800m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -198,7 +198,7 @@ void m04800m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m04800s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m04800s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m04900_a3-optimized.cl b/OpenCL/m04900_a3-optimized.cl
index 9444cf197..a3914e271 100644
--- a/OpenCL/m04900_a3-optimized.cl
+++ b/OpenCL/m04900_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m04900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m04900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -279,7 +279,7 @@ void m04900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m04900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m04900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m05000_a3-optimized.cl b/OpenCL/m05000_a3-optimized.cl
index e7da770e3..3afe605b6 100644
--- a/OpenCL/m05000_a3-optimized.cl
+++ b/OpenCL/m05000_a3-optimized.cl
@@ -28,7 +28,7 @@ __constant static u64a keccakf_rndc[24] =
 #define KECCAK_ROUNDS 24
 #endif
 
-void m05000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m05000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -214,7 +214,7 @@ void m05000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m05000s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m05000s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m05100_a3-optimized.cl b/OpenCL/m05100_a3-optimized.cl
index 7ac911ae6..6259810d4 100644
--- a/OpenCL/m05100_a3-optimized.cl
+++ b/OpenCL/m05100_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m05100m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m05100m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -147,7 +147,7 @@ void m05100m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m05100s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m05100s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m05300_a0-optimized.cl b/OpenCL/m05300_a0-optimized.cl
index 0acf08b72..1f9cc410c 100644
--- a/OpenCL/m05300_a0-optimized.cl
+++ b/OpenCL/m05300_a0-optimized.cl
@@ -15,7 +15,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md5.cl"
 
-void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
+DECLSPEC void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -66,7 +66,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
+DECLSPEC void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m05300_a1-optimized.cl b/OpenCL/m05300_a1-optimized.cl
index d52d9d2cf..44cb86601 100644
--- a/OpenCL/m05300_a1-optimized.cl
+++ b/OpenCL/m05300_a1-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md5.cl"
 
-void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
+DECLSPEC void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -64,7 +64,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
+DECLSPEC void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m05300_a3-optimized.cl b/OpenCL/m05300_a3-optimized.cl
index 3d26cb40e..b1b3977f8 100644
--- a/OpenCL/m05300_a3-optimized.cl
+++ b/OpenCL/m05300_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md5.cl"
 
-void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
+DECLSPEC void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -64,7 +64,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
+DECLSPEC void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -98,7 +98,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, digest);
 }
 
-void m05300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global ikepsk_t *ikepsk_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *s_msg_buf, __local u32 *s_nr_buf)
+DECLSPEC void m05300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global ikepsk_t *ikepsk_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *s_msg_buf, __local u32 *s_nr_buf)
 {
   /**
    * modifier
@@ -245,7 +245,7 @@ void m05300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m05300s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global ikepsk_t *ikepsk_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *s_msg_buf, __local u32 *s_nr_buf)
+DECLSPEC void m05300s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global ikepsk_t *ikepsk_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *s_msg_buf, __local u32 *s_nr_buf)
 {
   /**
    * modifier
diff --git a/OpenCL/m05400_a0-optimized.cl b/OpenCL/m05400_a0-optimized.cl
index 218aa3bd4..7d3219e26 100644
--- a/OpenCL/m05400_a0-optimized.cl
+++ b/OpenCL/m05400_a0-optimized.cl
@@ -15,7 +15,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha1.cl"
 
-void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
+DECLSPEC void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -68,7 +68,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m05400_a1-optimized.cl b/OpenCL/m05400_a1-optimized.cl
index 6e0911aa1..7e4a2bc0c 100644
--- a/OpenCL/m05400_a1-optimized.cl
+++ b/OpenCL/m05400_a1-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha1.cl"
 
-void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
+DECLSPEC void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -66,7 +66,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m05400_a3-optimized.cl b/OpenCL/m05400_a3-optimized.cl
index e4dd1d345..fc74e5f4b 100644
--- a/OpenCL/m05400_a3-optimized.cl
+++ b/OpenCL/m05400_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha1.cl"
 
-void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
+DECLSPEC void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -66,7 +66,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -102,7 +102,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, digest);
 }
 
-void m05400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global ikepsk_t *ikepsk_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *s_msg_buf, __local u32 *s_nr_buf)
+DECLSPEC void m05400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global ikepsk_t *ikepsk_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *s_msg_buf, __local u32 *s_nr_buf)
 {
   /**
    * modifier
@@ -249,7 +249,7 @@ void m05400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m05400s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global ikepsk_t *ikepsk_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *s_msg_buf, __local u32 *s_nr_buf)
+DECLSPEC void m05400s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global ikepsk_t *ikepsk_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *s_msg_buf, __local u32 *s_nr_buf)
 {
   /**
    * modifier
diff --git a/OpenCL/m05500_a0-optimized.cl b/OpenCL/m05500_a0-optimized.cl
index 560f6ed5e..1615df056 100644
--- a/OpenCL/m05500_a0-optimized.cl
+++ b/OpenCL/m05500_a0-optimized.cl
@@ -342,7 +342,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32x r = data[0];
   u32x l = data[1];
@@ -384,7 +384,7 @@ void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __l
   iv[1] = r;
 }
 
-void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32x tt;
 
@@ -453,7 +453,7 @@ void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32
   }
 }
 
-void transform_netntlmv1_key (const u32x w0, const u32x w1, u32x out[2])
+DECLSPEC void transform_netntlmv1_key (const u32x w0, const u32x w1, u32x out[2])
 {
   u32x t[8];
 
diff --git a/OpenCL/m05500_a0.cl b/OpenCL/m05500_a0.cl
index fda2805b2..7e5ed9e06 100644
--- a/OpenCL/m05500_a0.cl
+++ b/OpenCL/m05500_a0.cl
@@ -343,7 +343,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32 r = data[0];
   u32 l = data[1];
@@ -385,7 +385,7 @@ void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local
   iv[1] = r;
 }
 
-void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
@@ -454,7 +454,7 @@ void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_
   }
 }
 
-void transform_netntlmv1_key (const u32 w0, const u32 w1, u32 out[2])
+DECLSPEC void transform_netntlmv1_key (const u32 w0, const u32 w1, u32 out[2])
 {
   u32 t[8];
 
diff --git a/OpenCL/m05500_a1-optimized.cl b/OpenCL/m05500_a1-optimized.cl
index 45c20539f..e2d2ee1f9 100644
--- a/OpenCL/m05500_a1-optimized.cl
+++ b/OpenCL/m05500_a1-optimized.cl
@@ -340,7 +340,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32x r = data[0];
   u32x l = data[1];
@@ -382,7 +382,7 @@ void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __l
   iv[1] = r;
 }
 
-void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32x tt;
 
@@ -451,7 +451,7 @@ void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32
   }
 }
 
-void transform_netntlmv1_key (const u32x w0, const u32x w1, u32x out[2])
+DECLSPEC void transform_netntlmv1_key (const u32x w0, const u32x w1, u32x out[2])
 {
   u32x t[8];
 
diff --git a/OpenCL/m05500_a1.cl b/OpenCL/m05500_a1.cl
index 0d06309a1..a25445e92 100644
--- a/OpenCL/m05500_a1.cl
+++ b/OpenCL/m05500_a1.cl
@@ -341,7 +341,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32 r = data[0];
   u32 l = data[1];
@@ -383,7 +383,7 @@ void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local
   iv[1] = r;
 }
 
-void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
@@ -452,7 +452,7 @@ void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_
   }
 }
 
-void transform_netntlmv1_key (const u32 w0, const u32 w1, u32 out[2])
+DECLSPEC void transform_netntlmv1_key (const u32 w0, const u32 w1, u32 out[2])
 {
   u32 t[8];
 
diff --git a/OpenCL/m05500_a3-optimized.cl b/OpenCL/m05500_a3-optimized.cl
index f044b9a76..6e5078f5a 100644
--- a/OpenCL/m05500_a3-optimized.cl
+++ b/OpenCL/m05500_a3-optimized.cl
@@ -340,7 +340,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32x r = data[0];
   u32x l = data[1];
@@ -382,7 +382,7 @@ void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __l
   iv[1] = r;
 }
 
-void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32x tt;
 
@@ -451,7 +451,7 @@ void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32
   }
 }
 
-void transform_netntlmv1_key (const u32x w0, const u32x w1, u32x out[2])
+DECLSPEC void transform_netntlmv1_key (const u32x w0, const u32x w1, u32x out[2])
 {
   u32x t[8];
 
@@ -486,7 +486,7 @@ void transform_netntlmv1_key (const u32x w0, const u32x w1, u32x out[2])
          | ((k[7] & 0xff) << 24);
 }
 
-void m05500m (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m05500m (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -643,7 +643,7 @@ void m05500m (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16],
   }
 }
 
-void m05500s (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m05500s (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m05500_a3.cl b/OpenCL/m05500_a3.cl
index 7c4d79717..563c12888 100644
--- a/OpenCL/m05500_a3.cl
+++ b/OpenCL/m05500_a3.cl
@@ -341,7 +341,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32x r = data[0];
   u32x l = data[1];
@@ -383,7 +383,7 @@ void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __l
   iv[1] = r;
 }
 
-void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32x tt;
 
@@ -452,7 +452,7 @@ void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32
   }
 }
 
-void transform_netntlmv1_key (const u32x w0, const u32x w1, u32x out[2])
+DECLSPEC void transform_netntlmv1_key (const u32x w0, const u32x w1, u32x out[2])
 {
   u32x t[8];
 
diff --git a/OpenCL/m05600_a0-optimized.cl b/OpenCL/m05600_a0-optimized.cl
index 377f8571a..228f8fb52 100644
--- a/OpenCL/m05600_a0-optimized.cl
+++ b/OpenCL/m05600_a0-optimized.cl
@@ -16,7 +16,7 @@
 #include "inc_hash_md4.cl"
 #include "inc_hash_md5.cl"
 
-void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
+DECLSPEC void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -67,7 +67,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
+DECLSPEC void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m05600_a1-optimized.cl b/OpenCL/m05600_a1-optimized.cl
index 5449b50f5..c2bf6bfa9 100644
--- a/OpenCL/m05600_a1-optimized.cl
+++ b/OpenCL/m05600_a1-optimized.cl
@@ -14,7 +14,7 @@
 #include "inc_hash_md4.cl"
 #include "inc_hash_md5.cl"
 
-void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
+DECLSPEC void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -65,7 +65,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
+DECLSPEC void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m05600_a3-optimized.cl b/OpenCL/m05600_a3-optimized.cl
index 9d19eb6f2..bce2dcadd 100644
--- a/OpenCL/m05600_a3-optimized.cl
+++ b/OpenCL/m05600_a3-optimized.cl
@@ -14,7 +14,7 @@
 #include "inc_hash_md4.cl"
 #include "inc_hash_md5.cl"
 
-void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
+DECLSPEC void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -65,7 +65,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
+DECLSPEC void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -99,7 +99,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
   md5_transform_vector (w0, w1, w2, w3, digest);
 }
 
-void m05600m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global netntlm_t *netntlm_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *s_userdomain_buf, __local u32 *s_chall_buf)
+DECLSPEC void m05600m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global netntlm_t *netntlm_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *s_userdomain_buf, __local u32 *s_chall_buf)
 {
   /**
    * modifier
@@ -305,7 +305,7 @@ void m05600m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m05600s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global netntlm_t *netntlm_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *s_userdomain_buf, __local u32 *s_chall_buf)
+DECLSPEC void m05600s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global netntlm_t *netntlm_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *s_userdomain_buf, __local u32 *s_chall_buf)
 {
   /**
    * modifier
diff --git a/OpenCL/m05800-optimized.cl b/OpenCL/m05800-optimized.cl
index eac541f45..48c78601d 100644
--- a/OpenCL/m05800-optimized.cl
+++ b/OpenCL/m05800-optimized.cl
@@ -2069,7 +2069,7 @@ __constant static u32a c_pc_len[1024] =
   4
 };
 
-void append_word (u32 w0[4], u32 w1[4], const u32 append[4], const u32 offset)
+DECLSPEC void append_word (u32 w0[4], u32 w1[4], const u32 append[4], const u32 offset)
 {
   switch (offset)
   {
@@ -2103,7 +2103,7 @@ void append_word (u32 w0[4], u32 w1[4], const u32 append[4], const u32 offset)
   }
 }
 
-void append_salt (u32 w0[4], u32 w1[4], u32 w2[4], const u32 append[5], const u32 offset)
+DECLSPEC void append_salt (u32 w0[4], u32 w1[4], u32 w2[4], const u32 append[5], const u32 offset)
 {
   u32 tmp0;
   u32 tmp1;
diff --git a/OpenCL/m05800.cl b/OpenCL/m05800.cl
index 83b2fd9cc..bc2a8cef0 100644
--- a/OpenCL/m05800.cl
+++ b/OpenCL/m05800.cl
@@ -2069,7 +2069,7 @@ __constant static u32a c_pc_len[1024] =
   4
 };
 
-void append_word (u32 w0[4], u32 w1[4], const u32 append[4], const u32 offset)
+DECLSPEC void append_word (u32 w0[4], u32 w1[4], const u32 append[4], const u32 offset)
 {
   switch (offset)
   {
@@ -2103,7 +2103,7 @@ void append_word (u32 w0[4], u32 w1[4], const u32 append[4], const u32 offset)
   }
 }
 
-void append_salt (u32 w0[4], u32 w1[4], u32 w2[4], const u32 append[5], const u32 offset)
+DECLSPEC void append_salt (u32 w0[4], u32 w1[4], u32 w2[4], const u32 append[5], const u32 offset)
 {
   u32 tmp0;
   u32 tmp1;
diff --git a/OpenCL/m06000_a0-optimized.cl b/OpenCL/m06000_a0-optimized.cl
index cd9db9af1..de974669c 100644
--- a/OpenCL/m06000_a0-optimized.cl
+++ b/OpenCL/m06000_a0-optimized.cl
@@ -15,7 +15,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_ripemd160.cl"
 
-void ripemd160_transform_transport_vector (const u32x w[16], u32x dgst[5])
+DECLSPEC void ripemd160_transform_transport_vector (const u32x w[16], u32x dgst[5])
 {
   ripemd160_transform_vector (w + 0, w + 4, w + 8, w + 12, dgst);
 }
diff --git a/OpenCL/m06000_a1-optimized.cl b/OpenCL/m06000_a1-optimized.cl
index d2f0e2bcc..dd091749d 100644
--- a/OpenCL/m06000_a1-optimized.cl
+++ b/OpenCL/m06000_a1-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_ripemd160.cl"
 
-void ripemd160_transform_transport_vector (const u32x w[16], u32x dgst[5])
+DECLSPEC void ripemd160_transform_transport_vector (const u32x w[16], u32x dgst[5])
 {
   ripemd160_transform_vector (w + 0, w + 4, w + 8, w + 12, dgst);
 }
diff --git a/OpenCL/m06000_a3-optimized.cl b/OpenCL/m06000_a3-optimized.cl
index a13e41b87..91760e142 100644
--- a/OpenCL/m06000_a3-optimized.cl
+++ b/OpenCL/m06000_a3-optimized.cl
@@ -13,12 +13,12 @@
 #include "inc_simd.cl"
 #include "inc_hash_ripemd160.cl"
 
-void ripemd160_transform_transport_vector (const u32x w[16], u32x dgst[5])
+DECLSPEC void ripemd160_transform_transport_vector (const u32x w[16], u32x dgst[5])
 {
   ripemd160_transform_vector (w + 0, w + 4, w + 8, w + 12, dgst);
 }
 
-void m06000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m06000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -76,7 +76,7 @@ void m06000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m06000s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m06000s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m06100_a0-optimized.cl b/OpenCL/m06100_a0-optimized.cl
index dcb16a7b9..98ec3a87e 100644
--- a/OpenCL/m06100_a0-optimized.cl
+++ b/OpenCL/m06100_a0-optimized.cl
@@ -15,7 +15,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_whirlpool.cl"
 
-void whirlpool_transform_transport_vector (const u32x w[16], u32x digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void whirlpool_transform_transport_vector (const u32x w[16], u32x digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   whirlpool_transform_vector (w + 0, w + 4, w + 8, w + 12, digest, s_Ch, s_Cl);
 }
diff --git a/OpenCL/m06100_a1-optimized.cl b/OpenCL/m06100_a1-optimized.cl
index 6da1a8819..fdb7333f5 100644
--- a/OpenCL/m06100_a1-optimized.cl
+++ b/OpenCL/m06100_a1-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_whirlpool.cl"
 
-void whirlpool_transform_transport_vector (const u32x w[16], u32x digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void whirlpool_transform_transport_vector (const u32x w[16], u32x digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   whirlpool_transform_vector (w + 0, w + 4, w + 8, w + 12, digest, s_Ch, s_Cl);
 }
diff --git a/OpenCL/m06100_a3-optimized.cl b/OpenCL/m06100_a3-optimized.cl
index 291285cc6..062eeaba0 100644
--- a/OpenCL/m06100_a3-optimized.cl
+++ b/OpenCL/m06100_a3-optimized.cl
@@ -13,12 +13,12 @@
 #include "inc_simd.cl"
 #include "inc_hash_whirlpool.cl"
 
-void whirlpool_transform_transport_vector (const u32x w[16], u32x digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void whirlpool_transform_transport_vector (const u32x w[16], u32x digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   whirlpool_transform_vector (w + 0, w + 4, w + 8, w + 12, digest, s_Ch, s_Cl);
 }
 
-void m06100m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, SHM_TYPE u32 (*s_Cl)[256], SHM_TYPE u32 (*s_Ch)[256])
+DECLSPEC void m06100m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, SHM_TYPE u32 (*s_Cl)[256], SHM_TYPE u32 (*s_Ch)[256])
 {
   /**
    * modifier
@@ -87,7 +87,7 @@ void m06100m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m06100s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, SHM_TYPE u32 (*s_Cl)[256], SHM_TYPE u32 (*s_Ch)[256])
+DECLSPEC void m06100s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, SHM_TYPE u32 (*s_Cl)[256], SHM_TYPE u32 (*s_Ch)[256])
 {
   /**
    * modifier
diff --git a/OpenCL/m06211.cl b/OpenCL/m06211.cl
index 6416df58d..6b9edd1cb 100644
--- a/OpenCL/m06211.cl
+++ b/OpenCL/m06211.cl
@@ -20,7 +20,7 @@
 #include "inc_truecrypt_crc32.cl"
 #include "inc_truecrypt_xts.cl"
 
-u32 u8add (const u32 a, const u32 b)
+DECLSPEC u32 u8add (const u32 a, const u32 b)
 {
   const u32 a1 = (a >>  0) & 0xff;
   const u32 a2 = (a >>  8) & 0xff;
@@ -45,7 +45,7 @@ u32 u8add (const u32 a, const u32 b)
   return r;
 }
 
-void hmac_ripemd160_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_ripemd160_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m06212.cl b/OpenCL/m06212.cl
index 61351ef4f..8505cb968 100644
--- a/OpenCL/m06212.cl
+++ b/OpenCL/m06212.cl
@@ -20,7 +20,7 @@
 #include "inc_truecrypt_crc32.cl"
 #include "inc_truecrypt_xts.cl"
 
-u32 u8add (const u32 a, const u32 b)
+DECLSPEC u32 u8add (const u32 a, const u32 b)
 {
   const u32 a1 = (a >>  0) & 0xff;
   const u32 a2 = (a >>  8) & 0xff;
@@ -45,7 +45,7 @@ u32 u8add (const u32 a, const u32 b)
   return r;
 }
 
-void hmac_ripemd160_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_ripemd160_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m06213.cl b/OpenCL/m06213.cl
index 0ee3441b8..b564fef4d 100644
--- a/OpenCL/m06213.cl
+++ b/OpenCL/m06213.cl
@@ -20,7 +20,7 @@
 #include "inc_truecrypt_crc32.cl"
 #include "inc_truecrypt_xts.cl"
 
-u32 u8add (const u32 a, const u32 b)
+DECLSPEC u32 u8add (const u32 a, const u32 b)
 {
   const u32 a1 = (a >>  0) & 0xff;
   const u32 a2 = (a >>  8) & 0xff;
@@ -45,7 +45,7 @@ u32 u8add (const u32 a, const u32 b)
   return r;
 }
 
-void hmac_ripemd160_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_ripemd160_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m06221.cl b/OpenCL/m06221.cl
index 34588733d..f4a3289c8 100644
--- a/OpenCL/m06221.cl
+++ b/OpenCL/m06221.cl
@@ -20,7 +20,7 @@
 #include "inc_truecrypt_crc32.cl"
 #include "inc_truecrypt_xts.cl"
 
-u32 u8add (const u32 a, const u32 b)
+DECLSPEC u32 u8add (const u32 a, const u32 b)
 {
   const u32 a1 = (a >>  0) & 0xff;
   const u32 a2 = (a >>  8) & 0xff;
@@ -45,7 +45,7 @@ u32 u8add (const u32 a, const u32 b)
   return r;
 }
 
-void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m06222.cl b/OpenCL/m06222.cl
index 1f2f40030..e2f3a5e3c 100644
--- a/OpenCL/m06222.cl
+++ b/OpenCL/m06222.cl
@@ -20,7 +20,7 @@
 #include "inc_truecrypt_crc32.cl"
 #include "inc_truecrypt_xts.cl"
 
-u32 u8add (const u32 a, const u32 b)
+DECLSPEC u32 u8add (const u32 a, const u32 b)
 {
   const u32 a1 = (a >>  0) & 0xff;
   const u32 a2 = (a >>  8) & 0xff;
@@ -45,7 +45,7 @@ u32 u8add (const u32 a, const u32 b)
   return r;
 }
 
-void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m06223.cl b/OpenCL/m06223.cl
index e540d2ec2..5403969e2 100644
--- a/OpenCL/m06223.cl
+++ b/OpenCL/m06223.cl
@@ -20,7 +20,7 @@
 #include "inc_truecrypt_crc32.cl"
 #include "inc_truecrypt_xts.cl"
 
-u32 u8add (const u32 a, const u32 b)
+DECLSPEC u32 u8add (const u32 a, const u32 b)
 {
   const u32 a1 = (a >>  0) & 0xff;
   const u32 a2 = (a >>  8) & 0xff;
@@ -45,7 +45,7 @@ u32 u8add (const u32 a, const u32 b)
   return r;
 }
 
-void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m06231.cl b/OpenCL/m06231.cl
index c978bd38a..e2bfc5353 100644
--- a/OpenCL/m06231.cl
+++ b/OpenCL/m06231.cl
@@ -20,7 +20,7 @@
 #include "inc_truecrypt_crc32.cl"
 #include "inc_truecrypt_xts.cl"
 
-u32 u8add (const u32 a, const u32 b)
+DECLSPEC u32 u8add (const u32 a, const u32 b)
 {
   const u32 a1 = (a >>  0) & 0xff;
   const u32 a2 = (a >>  8) & 0xff;
@@ -45,7 +45,7 @@ u32 u8add (const u32 a, const u32 b)
   return r;
 }
 
-void hmac_whirlpool_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[16], u32x opad[16], u32x digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void hmac_whirlpool_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[16], u32x opad[16], u32x digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   digest[ 0] = ipad[ 0];
   digest[ 1] = ipad[ 1];
diff --git a/OpenCL/m06232.cl b/OpenCL/m06232.cl
index e2e2a1a3e..912592e3c 100644
--- a/OpenCL/m06232.cl
+++ b/OpenCL/m06232.cl
@@ -20,7 +20,7 @@
 #include "inc_truecrypt_crc32.cl"
 #include "inc_truecrypt_xts.cl"
 
-u32 u8add (const u32 a, const u32 b)
+DECLSPEC u32 u8add (const u32 a, const u32 b)
 {
   const u32 a1 = (a >>  0) & 0xff;
   const u32 a2 = (a >>  8) & 0xff;
@@ -45,7 +45,7 @@ u32 u8add (const u32 a, const u32 b)
   return r;
 }
 
-void hmac_whirlpool_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[16], u32x opad[16], u32x digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void hmac_whirlpool_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[16], u32x opad[16], u32x digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   digest[ 0] = ipad[ 0];
   digest[ 1] = ipad[ 1];
diff --git a/OpenCL/m06233.cl b/OpenCL/m06233.cl
index 601b0fb45..0d7111b56 100644
--- a/OpenCL/m06233.cl
+++ b/OpenCL/m06233.cl
@@ -20,7 +20,7 @@
 #include "inc_truecrypt_crc32.cl"
 #include "inc_truecrypt_xts.cl"
 
-u32 u8add (const u32 a, const u32 b)
+DECLSPEC u32 u8add (const u32 a, const u32 b)
 {
   const u32 a1 = (a >>  0) & 0xff;
   const u32 a2 = (a >>  8) & 0xff;
@@ -45,7 +45,7 @@ u32 u8add (const u32 a, const u32 b)
   return r;
 }
 
-void hmac_whirlpool_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[16], u32x opad[16], u32x digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
+DECLSPEC void hmac_whirlpool_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[16], u32x opad[16], u32x digest[16], SHM_TYPE u32 (*s_Ch)[256], SHM_TYPE u32 (*s_Cl)[256])
 {
   digest[ 0] = ipad[ 0];
   digest[ 1] = ipad[ 1];
diff --git a/OpenCL/m06300-optimized.cl b/OpenCL/m06300-optimized.cl
index 675cde5b6..b9a20ae8b 100644
--- a/OpenCL/m06300-optimized.cl
+++ b/OpenCL/m06300-optimized.cl
@@ -13,7 +13,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void memcat16 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[4])
+DECLSPEC void memcat16 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[4])
 {
   u32 tmp0;
   u32 tmp1;
@@ -126,7 +126,7 @@ void memcat16 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const
   }
 }
 
-void memcat16_x80 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[4])
+DECLSPEC void memcat16_x80 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[4])
 {
   u32 tmp0;
   u32 tmp1;
@@ -241,7 +241,7 @@ void memcat16_x80 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], c
   }
 }
 
-void memcat8 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[2])
+DECLSPEC void memcat8 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 offset, const u32 append[2])
 {
   u32 tmp0;
   u32 tmp1;
@@ -330,7 +330,7 @@ void memcat8 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const
   }
 }
 
-void append_1st (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 block_len, const u32 append)
+DECLSPEC void append_1st (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 block_len, const u32 append)
 {
   switch (block_len)
   {
diff --git a/OpenCL/m06400.cl b/OpenCL/m06400.cl
index c6120914d..5e62bb6bb 100644
--- a/OpenCL/m06400.cl
+++ b/OpenCL/m06400.cl
@@ -16,7 +16,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m06500.cl b/OpenCL/m06500.cl
index e9202f191..82f8f9f69 100644
--- a/OpenCL/m06500.cl
+++ b/OpenCL/m06500.cl
@@ -16,7 +16,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m06600.cl b/OpenCL/m06600.cl
index 635c60c48..b44cf6945 100644
--- a/OpenCL/m06600.cl
+++ b/OpenCL/m06600.cl
@@ -17,7 +17,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m06700.cl b/OpenCL/m06700.cl
index c68c34357..062402af8 100644
--- a/OpenCL/m06700.cl
+++ b/OpenCL/m06700.cl
@@ -16,7 +16,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m06800.cl b/OpenCL/m06800.cl
index a860515ec..a71aa467c 100644
--- a/OpenCL/m06800.cl
+++ b/OpenCL/m06800.cl
@@ -17,7 +17,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m06900_a3-optimized.cl b/OpenCL/m06900_a3-optimized.cl
index bcde422b6..b6df5fede 100644
--- a/OpenCL/m06900_a3-optimized.cl
+++ b/OpenCL/m06900_a3-optimized.cl
@@ -695,7 +695,7 @@ __constant static u32a c_tables[4][256] =
   R (k, h, s, 6, t);      \
 }
 
-void m06900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 (*s_tables)[256])
+DECLSPEC void m06900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 (*s_tables)[256])
 {
   /**
    * modifier
@@ -869,7 +869,7 @@ void m06900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m06900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 (*s_tables)[256])
+DECLSPEC void m06900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 (*s_tables)[256])
 {
   /**
    * modifier
diff --git a/OpenCL/m07000_a3-optimized.cl b/OpenCL/m07000_a3-optimized.cl
index 76f6055f7..af1e5e931 100644
--- a/OpenCL/m07000_a3-optimized.cl
+++ b/OpenCL/m07000_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m07000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m07000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -279,7 +279,7 @@ void m07000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m07000s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m07000s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m07100.cl b/OpenCL/m07100.cl
index 143306840..e63e3eb47 100644
--- a/OpenCL/m07100.cl
+++ b/OpenCL/m07100.cl
@@ -16,7 +16,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m07300_a0-optimized.cl b/OpenCL/m07300_a0-optimized.cl
index 123cc24db..c8354dd00 100644
--- a/OpenCL/m07300_a0-optimized.cl
+++ b/OpenCL/m07300_a0-optimized.cl
@@ -15,7 +15,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha1.cl"
 
-void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
+DECLSPEC void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -68,7 +68,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m07300_a1-optimized.cl b/OpenCL/m07300_a1-optimized.cl
index 03b4bd73f..536628cdf 100644
--- a/OpenCL/m07300_a1-optimized.cl
+++ b/OpenCL/m07300_a1-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha1.cl"
 
-void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
+DECLSPEC void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -66,7 +66,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m07300_a3-optimized.cl b/OpenCL/m07300_a3-optimized.cl
index 991bba6a7..673bf099f 100644
--- a/OpenCL/m07300_a3-optimized.cl
+++ b/OpenCL/m07300_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha1.cl"
 
-void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
+DECLSPEC void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -66,7 +66,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, opad);
 }
 
-void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -102,7 +102,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
   sha1_transform_vector (w0, w1, w2, w3, digest);
 }
 
-void m07300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global rakp_t *rakp_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m07300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global rakp_t *rakp_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -212,7 +212,7 @@ void m07300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m07300s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global rakp_t *rakp_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m07300s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global rakp_t *rakp_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m07400-optimized.cl b/OpenCL/m07400-optimized.cl
index 76e308528..9ab079cd9 100644
--- a/OpenCL/m07400-optimized.cl
+++ b/OpenCL/m07400-optimized.cl
@@ -13,7 +13,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void sha256_transform_transport (const u32 w[16], u32 digest[8])
+DECLSPEC void sha256_transform_transport (const u32 w[16], u32 digest[8])
 {
   u32 w0[4];
   u32 w1[4];
@@ -40,7 +40,7 @@ void sha256_transform_transport (const u32 w[16], u32 digest[8])
   sha256_transform (w0, w1, w2, w3, digest);
 }
 
-void init_ctx (u32 digest[8])
+DECLSPEC void init_ctx (u32 digest[8])
 {
   digest[0] = SHA256M_A;
   digest[1] = SHA256M_B;
@@ -52,7 +52,7 @@ void init_ctx (u32 digest[8])
   digest[7] = SHA256M_H;
 }
 
-void bzero16 (u32 block[16])
+DECLSPEC void bzero16 (u32 block[16])
 {
   block[ 0] = 0;
   block[ 1] = 0;
@@ -72,7 +72,7 @@ void bzero16 (u32 block[16])
   block[15] = 0;
 }
 
-void bswap8 (u32 block[16])
+DECLSPEC void bswap8 (u32 block[16])
 {
   block[ 0] = swap32 (block[ 0]);
   block[ 1] = swap32 (block[ 1]);
@@ -84,7 +84,7 @@ void bswap8 (u32 block[16])
   block[ 7] = swap32 (block[ 7]);
 }
 
-u32 memcat16 (u32 block[16], const u32 offset, const u32 append[4], const u32 append_len)
+DECLSPEC u32 memcat16 (u32 block[16], const u32 offset, const u32 append[4], const u32 append_len)
 {
   u32 tmp0;
   u32 tmp1;
@@ -225,7 +225,7 @@ u32 memcat16 (u32 block[16], const u32 offset, const u32 append[4], const u32 ap
   return new_len;
 }
 
-u32 memcat16c (u32 block[16], const u32 offset, const u32 append[4], const u32 append_len, u32 digest[8])
+DECLSPEC u32 memcat16c (u32 block[16], const u32 offset, const u32 append[4], const u32 append_len, u32 digest[8])
 {
   u32 tmp0;
   u32 tmp1;
@@ -392,7 +392,7 @@ u32 memcat16c (u32 block[16], const u32 offset, const u32 append[4], const u32 a
   return new_len;
 }
 
-u32 memcat20 (u32 block[32], const u32 offset, const u32 append[4], const u32 append_len)
+DECLSPEC u32 memcat20 (u32 block[32], const u32 offset, const u32 append[4], const u32 append_len)
 {
   u32 tmp0;
   u32 tmp1;
@@ -541,7 +541,7 @@ u32 memcat20 (u32 block[32], const u32 offset, const u32 append[4], const u32 ap
   return offset + append_len;
 }
 
-u32 memcat20_x80 (u32 block[32], const u32 offset, const u32 append[4], const u32 append_len)
+DECLSPEC u32 memcat20_x80 (u32 block[32], const u32 offset, const u32 append[4], const u32 append_len)
 {
   u32 tmp0;
   u32 tmp1;
diff --git a/OpenCL/m07500_a0-optimized.cl b/OpenCL/m07500_a0-optimized.cl
index 40503b18c..f6fa1debc 100644
--- a/OpenCL/m07500_a0-optimized.cl
+++ b/OpenCL/m07500_a0-optimized.cl
@@ -25,7 +25,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -34,7 +34,7 @@ void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -87,7 +87,7 @@ void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -140,7 +140,7 @@ u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[
   return j;
 }
 
-int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[8])
+DECLSPEC int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[8])
 {
   rc4_init_16 (rc4_key, data);
 
@@ -170,7 +170,7 @@ int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[
   return 1;
 }
 
-void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
+DECLSPEC void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -221,7 +221,7 @@ void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32
   md5_transform (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
+DECLSPEC void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -255,7 +255,7 @@ void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32
   md5_transform (w0, w1, w2, w3, digest);
 }
 
-void kerb_prepare (const u32 w0[4], const u32 w1[4], const u32 pw_len, const u32 checksum[4], u32 digest[4])
+DECLSPEC void kerb_prepare (const u32 w0[4], const u32 w1[4], const u32 pw_len, const u32 checksum[4], u32 digest[4])
 {
   /**
    * pads
diff --git a/OpenCL/m07500_a0.cl b/OpenCL/m07500_a0.cl
index 4b6912401..e2f919857 100644
--- a/OpenCL/m07500_a0.cl
+++ b/OpenCL/m07500_a0.cl
@@ -24,7 +24,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -33,7 +33,7 @@ void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -86,7 +86,7 @@ void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -139,7 +139,7 @@ u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[
   return j;
 }
 
-int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[8])
+DECLSPEC int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[8])
 {
   rc4_init_16 (rc4_key, data);
 
@@ -169,7 +169,7 @@ int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[
   return 1;
 }
 
-void kerb_prepare (const u32 K[4], const u32 checksum[4], u32 digest[4])
+DECLSPEC void kerb_prepare (const u32 K[4], const u32 checksum[4], u32 digest[4])
 {
   // K1=MD5_HMAC(K,1); with 1 encoded as little indian on 4 bytes (01000000 in hexa);
 
diff --git a/OpenCL/m07500_a1-optimized.cl b/OpenCL/m07500_a1-optimized.cl
index de0435870..73a98b9f7 100644
--- a/OpenCL/m07500_a1-optimized.cl
+++ b/OpenCL/m07500_a1-optimized.cl
@@ -23,7 +23,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -32,7 +32,7 @@ void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -85,7 +85,7 @@ void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -138,7 +138,7 @@ u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[
   return j;
 }
 
-int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[8])
+DECLSPEC int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[8])
 {
   rc4_init_16 (rc4_key, data);
 
@@ -168,7 +168,7 @@ int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[
   return 1;
 }
 
-void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
+DECLSPEC void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -219,7 +219,7 @@ void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32
   md5_transform (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
+DECLSPEC void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -253,7 +253,7 @@ void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32
   md5_transform (w0, w1, w2, w3, digest);
 }
 
-void kerb_prepare (const u32 w0[4], const u32 w1[4], const u32 pw_len, const u32 checksum[4], u32 digest[4])
+DECLSPEC void kerb_prepare (const u32 w0[4], const u32 w1[4], const u32 pw_len, const u32 checksum[4], u32 digest[4])
 {
   /**
    * pads
diff --git a/OpenCL/m07500_a1.cl b/OpenCL/m07500_a1.cl
index fb5689e7c..fd3fcd2bc 100644
--- a/OpenCL/m07500_a1.cl
+++ b/OpenCL/m07500_a1.cl
@@ -22,7 +22,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -31,7 +31,7 @@ void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -84,7 +84,7 @@ void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -137,7 +137,7 @@ u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[
   return j;
 }
 
-int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[8])
+DECLSPEC int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[8])
 {
   rc4_init_16 (rc4_key, data);
 
@@ -167,7 +167,7 @@ int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[
   return 1;
 }
 
-void kerb_prepare (const u32 K[4], const u32 checksum[4], u32 digest[4])
+DECLSPEC void kerb_prepare (const u32 K[4], const u32 checksum[4], u32 digest[4])
 {
   // K1=MD5_HMAC(K,1); with 1 encoded as little indian on 4 bytes (01000000 in hexa);
 
diff --git a/OpenCL/m07500_a3-optimized.cl b/OpenCL/m07500_a3-optimized.cl
index 85569e4bd..ec5356f2c 100644
--- a/OpenCL/m07500_a3-optimized.cl
+++ b/OpenCL/m07500_a3-optimized.cl
@@ -23,7 +23,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -32,7 +32,7 @@ void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -85,7 +85,7 @@ void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -138,7 +138,7 @@ u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[
   return j;
 }
 
-int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[8])
+DECLSPEC int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[8])
 {
   rc4_init_16 (rc4_key, data);
 
@@ -168,7 +168,7 @@ int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[
   return 1;
 }
 
-void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
+DECLSPEC void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -219,7 +219,7 @@ void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32
   md5_transform (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
+DECLSPEC void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -253,7 +253,7 @@ void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32
   md5_transform (w0, w1, w2, w3, digest);
 }
 
-void kerb_prepare (const u32 w0[4], const u32 w1[4], const u32 pw_len, const u32 checksum[4], u32 digest[4])
+DECLSPEC void kerb_prepare (const u32 w0[4], const u32 w1[4], const u32 pw_len, const u32 checksum[4], u32 digest[4])
 {
   /**
    * pads
@@ -382,7 +382,7 @@ void kerb_prepare (const u32 w0[4], const u32 w1[4], const u32 pw_len, const u32
   hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
 }
 
-void m07500 (SCR_TYPE RC4_KEY *rc4_key, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global krb5pa_t *krb5pa_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m07500 (SCR_TYPE RC4_KEY *rc4_key, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global krb5pa_t *krb5pa_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m07500_a3.cl b/OpenCL/m07500_a3.cl
index 1ac2984e9..90f7a24d6 100644
--- a/OpenCL/m07500_a3.cl
+++ b/OpenCL/m07500_a3.cl
@@ -22,7 +22,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -31,7 +31,7 @@ void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -84,7 +84,7 @@ void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -137,7 +137,7 @@ u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[
   return j;
 }
 
-int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[8])
+DECLSPEC int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[8])
 {
   rc4_init_16 (rc4_key, data);
 
@@ -167,7 +167,7 @@ int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[
   return 1;
 }
 
-void kerb_prepare (const u32 K[4], const u32 checksum[4], u32 digest[4])
+DECLSPEC void kerb_prepare (const u32 K[4], const u32 checksum[4], u32 digest[4])
 {
   // K1=MD5_HMAC(K,1); with 1 encoded as little indian on 4 bytes (01000000 in hexa);
 
diff --git a/OpenCL/m07700_a0-optimized.cl b/OpenCL/m07700_a0-optimized.cl
index cd2d2519f..0d9c78089 100644
--- a/OpenCL/m07700_a0-optimized.cl
+++ b/OpenCL/m07700_a0-optimized.cl
@@ -56,7 +56,7 @@ __constant static u32a bcodeArray[48] =
   0xe2, 0xb7, 0x33, 0x71, 0x8b, 0x9f, 0x5d, 0x01, 0x44, 0x70, 0xae, 0x11, 0xef, 0x28, 0xf0, 0x0d
 };
 
-u32 sapb_trans (const u32 in)
+DECLSPEC u32 sapb_trans (const u32 in)
 {
   u32 out = 0;
 
@@ -68,7 +68,7 @@ u32 sapb_trans (const u32 in)
   return out;
 }
 
-u32 walld0rf_magic (const u32 w0[4], const u32 pw_len, const u32 salt_buf0[4], const u32 salt_len, const u32 a, const u32 b, const u32 c, const u32 d, u32 t[16])
+DECLSPEC u32 walld0rf_magic (const u32 w0[4], const u32 pw_len, const u32 salt_buf0[4], const u32 salt_len, const u32 a, const u32 b, const u32 c, const u32 d, u32 t[16])
 {
   t[ 0] = 0;
   t[ 1] = 0;
diff --git a/OpenCL/m07700_a1-optimized.cl b/OpenCL/m07700_a1-optimized.cl
index ab491fedd..643906c60 100644
--- a/OpenCL/m07700_a1-optimized.cl
+++ b/OpenCL/m07700_a1-optimized.cl
@@ -54,7 +54,7 @@ __constant static u32a bcodeArray[48] =
   0xe2, 0xb7, 0x33, 0x71, 0x8b, 0x9f, 0x5d, 0x01, 0x44, 0x70, 0xae, 0x11, 0xef, 0x28, 0xf0, 0x0d
 };
 
-u32 sapb_trans (const u32 in)
+DECLSPEC u32 sapb_trans (const u32 in)
 {
   u32 out = 0;
 
@@ -66,7 +66,7 @@ u32 sapb_trans (const u32 in)
   return out;
 }
 
-u32 walld0rf_magic (const u32 w0[4], const u32 pw_len, const u32 salt_buf0[4], const u32 salt_len, const u32 a, const u32 b, const u32 c, const u32 d, u32 t[16])
+DECLSPEC u32 walld0rf_magic (const u32 w0[4], const u32 pw_len, const u32 salt_buf0[4], const u32 salt_len, const u32 a, const u32 b, const u32 c, const u32 d, u32 t[16])
 {
   t[ 0] = 0;
   t[ 1] = 0;
diff --git a/OpenCL/m07700_a3-optimized.cl b/OpenCL/m07700_a3-optimized.cl
index e944ca4c0..279d10ea7 100644
--- a/OpenCL/m07700_a3-optimized.cl
+++ b/OpenCL/m07700_a3-optimized.cl
@@ -45,7 +45,7 @@ __constant static u32a bcodeArray[48] =
   0xe2, 0xb7, 0x33, 0x71, 0x8b, 0x9f, 0x5d, 0x01, 0x44, 0x70, 0xae, 0x11, 0xef, 0x28, 0xf0, 0x0d
 };
 
-u32 sapb_trans (const u32 in)
+DECLSPEC u32 sapb_trans (const u32 in)
 {
   u32 out = 0;
 
@@ -57,7 +57,7 @@ u32 sapb_trans (const u32 in)
   return out;
 }
 
-u32 walld0rf_magic (const u32 w0[4], const u32 pw_len, const u32 salt_buf0[4], const u32 salt_len, const u32 a, const u32 b, const u32 c, const u32 d, u32 t[16])
+DECLSPEC u32 walld0rf_magic (const u32 w0[4], const u32 pw_len, const u32 salt_buf0[4], const u32 salt_len, const u32 a, const u32 b, const u32 c, const u32 d, u32 t[16])
 {
   t[ 0] = 0;
   t[ 1] = 0;
@@ -136,7 +136,7 @@ u32 walld0rf_magic (const u32 w0[4], const u32 pw_len, const u32 salt_buf0[4], c
   return sum20;
 }
 
-void m07700m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m07700m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -261,7 +261,7 @@ void m07700m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m07700s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m07700s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m07800_a0-optimized.cl b/OpenCL/m07800_a0-optimized.cl
index 149f64ec9..17cc0c416 100644
--- a/OpenCL/m07800_a0-optimized.cl
+++ b/OpenCL/m07800_a0-optimized.cl
@@ -28,7 +28,7 @@ __constant static u32a theMagicArray[64] =
   0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
 };
 
-u32 GETSHIFTEDINT_CONST (__constant u32 *a, const int n)
+DECLSPEC u32 GETSHIFTEDINT_CONST (__constant u32 *a, const int n)
 {
   const int d = n / 4;
   const int m = n & 3;
@@ -40,7 +40,7 @@ u32 GETSHIFTEDINT_CONST (__constant u32 *a, const int n)
   return h32_from_64_S (tmp);
 }
 
-void SETSHIFTEDINT (u32 *a, const int n, const u32 v)
+DECLSPEC void SETSHIFTEDINT (u32 *a, const int n, const u32 v)
 {
   const int d = n / 4;
   const int m = n & 3;
diff --git a/OpenCL/m07800_a1-optimized.cl b/OpenCL/m07800_a1-optimized.cl
index 1a5206ec5..60bd07866 100644
--- a/OpenCL/m07800_a1-optimized.cl
+++ b/OpenCL/m07800_a1-optimized.cl
@@ -26,7 +26,7 @@ __constant static u32a theMagicArray[64] =
   0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
 };
 
-u32 GETSHIFTEDINT_CONST (__constant u32 *a, const int n)
+DECLSPEC u32 GETSHIFTEDINT_CONST (__constant u32 *a, const int n)
 {
   const int d = n / 4;
   const int m = n & 3;
@@ -38,7 +38,7 @@ u32 GETSHIFTEDINT_CONST (__constant u32 *a, const int n)
   return h32_from_64_S (tmp);
 }
 
-void SETSHIFTEDINT (u32 *a, const int n, const u32 v)
+DECLSPEC void SETSHIFTEDINT (u32 *a, const int n, const u32 v)
 {
   const int d = n / 4;
   const int m = n & 3;
diff --git a/OpenCL/m07800_a3-optimized.cl b/OpenCL/m07800_a3-optimized.cl
index 00d7045bd..ecccf276e 100644
--- a/OpenCL/m07800_a3-optimized.cl
+++ b/OpenCL/m07800_a3-optimized.cl
@@ -26,7 +26,7 @@ __constant static u32a theMagicArray[64] =
   0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
 };
 
-u32 GETSHIFTEDINT_CONST (__constant u32 *a, const int n)
+DECLSPEC u32 GETSHIFTEDINT_CONST (__constant u32 *a, const int n)
 {
   const int d = n / 4;
   const int m = n & 3;
@@ -38,7 +38,7 @@ u32 GETSHIFTEDINT_CONST (__constant u32 *a, const int n)
   return h32_from_64_S (tmp);
 }
 
-void SETSHIFTEDINT (u32 *a, const int n, const u32 v)
+DECLSPEC void SETSHIFTEDINT (u32 *a, const int n, const u32 v)
 {
   const int d = n / 4;
   const int m = n & 3;
@@ -51,7 +51,7 @@ void SETSHIFTEDINT (u32 *a, const int n, const u32 v)
   a[d + 1]  = l32_from_64_S (tmp);
 }
 
-void m07800m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m07800m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -267,7 +267,7 @@ void m07800m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m07800s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m07800s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m08000_a0-optimized.cl b/OpenCL/m08000_a0-optimized.cl
index ea4dfecea..08d3cf4d3 100644
--- a/OpenCL/m08000_a0-optimized.cl
+++ b/OpenCL/m08000_a0-optimized.cl
@@ -39,7 +39,7 @@ __constant static u32a k_sha256[64] =
 
 #define SHA256_EXPAND_S(x,y,z,w) (SHA256_S1_S (x) + y + SHA256_S0_S (z) + w)
 
-void sha256_transform (u32x digest[8], const u32x w[16])
+DECLSPEC void sha256_transform (u32x digest[8], const u32x w[16])
 {
   u32x a = digest[0];
   u32x b = digest[1];
@@ -127,7 +127,7 @@ void sha256_transform (u32x digest[8], const u32x w[16])
   digest[7] += h;
 }
 
-void sha256_transform_z (u32x digest[8])
+DECLSPEC void sha256_transform_z (u32x digest[8])
 {
   u32x a = digest[0];
   u32x b = digest[1];
@@ -178,7 +178,7 @@ void sha256_transform_z (u32x digest[8])
   digest[7] += h;
 }
 
-void sha256_transform_s (u32x digest[8], __local u32 *w)
+DECLSPEC void sha256_transform_s (u32x digest[8], __local u32 *w)
 {
   u32x a = digest[0];
   u32x b = digest[1];
diff --git a/OpenCL/m08000_a1-optimized.cl b/OpenCL/m08000_a1-optimized.cl
index f5fa0abdf..6c4544f88 100644
--- a/OpenCL/m08000_a1-optimized.cl
+++ b/OpenCL/m08000_a1-optimized.cl
@@ -37,7 +37,7 @@ __constant static u32a k_sha256[64] =
 
 #define SHA256_EXPAND_S(x,y,z,w) (SHA256_S1_S (x) + y + SHA256_S0_S (z) + w)
 
-void sha256_transform (u32x digest[8], const u32x w[16])
+DECLSPEC void sha256_transform (u32x digest[8], const u32x w[16])
 {
   u32x a = digest[0];
   u32x b = digest[1];
@@ -125,7 +125,7 @@ void sha256_transform (u32x digest[8], const u32x w[16])
   digest[7] += h;
 }
 
-void sha256_transform_z (u32x digest[8])
+DECLSPEC void sha256_transform_z (u32x digest[8])
 {
   u32x a = digest[0];
   u32x b = digest[1];
@@ -176,7 +176,7 @@ void sha256_transform_z (u32x digest[8])
   digest[7] += h;
 }
 
-void sha256_transform_s (u32x digest[8], __local u32 *w)
+DECLSPEC void sha256_transform_s (u32x digest[8], __local u32 *w)
 {
   u32x a = digest[0];
   u32x b = digest[1];
diff --git a/OpenCL/m08000_a3-optimized.cl b/OpenCL/m08000_a3-optimized.cl
index ab8d6bbb2..52f10529d 100644
--- a/OpenCL/m08000_a3-optimized.cl
+++ b/OpenCL/m08000_a3-optimized.cl
@@ -37,7 +37,7 @@ __constant static u32a k_sha256[64] =
 
 #define SHA256_EXPAND_S(x,y,z,w) (SHA256_S1_S (x) + y + SHA256_S0_S (z) + w)
 
-void sha256_transform (u32x digest[8], const u32x w[16])
+DECLSPEC void sha256_transform (u32x digest[8], const u32x w[16])
 {
   u32x a = digest[0];
   u32x b = digest[1];
@@ -125,7 +125,7 @@ void sha256_transform (u32x digest[8], const u32x w[16])
   digest[7] += h;
 }
 
-void sha256_transform_z (u32x digest[8])
+DECLSPEC void sha256_transform_z (u32x digest[8])
 {
   u32x a = digest[0];
   u32x b = digest[1];
@@ -176,7 +176,7 @@ void sha256_transform_z (u32x digest[8])
   digest[7] += h;
 }
 
-void sha256_transform_s (u32x digest[8], __local u32 *w)
+DECLSPEC void sha256_transform_s (u32x digest[8], __local u32 *w)
 {
   u32x a = digest[0];
   u32x b = digest[1];
@@ -227,7 +227,7 @@ void sha256_transform_s (u32x digest[8], __local u32 *w)
   digest[7] += h;
 }
 
-void m08000m (__local u32 *w_s1, __local u32 *w_s2, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u64 gid_max)
+DECLSPEC void m08000m (__local u32 *w_s1, __local u32 *w_s2, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u64 gid_max)
 {
   /**
    * modifier
@@ -349,7 +349,7 @@ void m08000m (__local u32 *w_s1, __local u32 *w_s2, u32 w[16], const u32 pw_len,
   }
 }
 
-void m08000s (__local u32 *w_s1, __local u32 *w_s2, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u64 gid_max)
+DECLSPEC void m08000s (__local u32 *w_s1, __local u32 *w_s2, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u64 gid_max)
 {
   /**
    * modifier
diff --git a/OpenCL/m08100_a3-optimized.cl b/OpenCL/m08100_a3-optimized.cl
index 1ddc7042c..b3004b4ac 100644
--- a/OpenCL/m08100_a3-optimized.cl
+++ b/OpenCL/m08100_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m08100m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m08100m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -173,7 +173,7 @@ void m08100m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m08100s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m08100s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m08200.cl b/OpenCL/m08200.cl
index 32c5f2d6a..c1551437d 100644
--- a/OpenCL/m08200.cl
+++ b/OpenCL/m08200.cl
@@ -17,7 +17,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m08300_a3-optimized.cl b/OpenCL/m08300_a3-optimized.cl
index 167546007..31161f91c 100644
--- a/OpenCL/m08300_a3-optimized.cl
+++ b/OpenCL/m08300_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha1.cl"
 
-void m08300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m08300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -221,7 +221,7 @@ void m08300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m08300s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m08300s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m08400_a3-optimized.cl b/OpenCL/m08400_a3-optimized.cl
index 2c15ae55f..a4e10a6da 100644
--- a/OpenCL/m08400_a3-optimized.cl
+++ b/OpenCL/m08400_a3-optimized.cl
@@ -25,7 +25,7 @@
 #define uint_to_hex_lower8_le(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m08400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m08400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -235,7 +235,7 @@ void m08400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m08400s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m08400s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m08500_a0.cl b/OpenCL/m08500_a0.cl
index 2e6e92505..0060a62c9 100644
--- a/OpenCL/m08500_a0.cl
+++ b/OpenCL/m08500_a0.cl
@@ -394,7 +394,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX1(i,S) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32 iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32 r = data[0];
   u32 l = data[1];
@@ -436,7 +436,7 @@ void _des_crypt_encrypt (u32 iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __lo
   iv[1] = r;
 }
 
-void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
@@ -508,7 +508,7 @@ void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (
   }
 }
 
-void transform_racf_key (const u32 w0, const u32x w1, u32x key[2])
+DECLSPEC void transform_racf_key (const u32 w0, const u32x w1, u32x key[2])
 {
   key[0] = BOX1 (((w0 >>  0) & 0xff), c_ascii_to_ebcdic_pc) <<  0
          | BOX1 (((w0 >>  8) & 0xff), c_ascii_to_ebcdic_pc) <<  8
diff --git a/OpenCL/m08500_a1.cl b/OpenCL/m08500_a1.cl
index 6f59374b1..a4a7e12bb 100644
--- a/OpenCL/m08500_a1.cl
+++ b/OpenCL/m08500_a1.cl
@@ -392,7 +392,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX1(i,S) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32 iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32 r = data[0];
   u32 l = data[1];
@@ -434,7 +434,7 @@ void _des_crypt_encrypt (u32 iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __lo
   iv[1] = r;
 }
 
-void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
@@ -506,7 +506,7 @@ void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (
   }
 }
 
-void transform_racf_key (const u32 w0, const u32x w1, u32x key[2])
+DECLSPEC void transform_racf_key (const u32 w0, const u32x w1, u32x key[2])
 {
   key[0] = BOX1 (((w0 >>  0) & 0xff), c_ascii_to_ebcdic_pc) <<  0
          | BOX1 (((w0 >>  8) & 0xff), c_ascii_to_ebcdic_pc) <<  8
diff --git a/OpenCL/m08500_a3.cl b/OpenCL/m08500_a3.cl
index 82cb3f59c..8d362f440 100644
--- a/OpenCL/m08500_a3.cl
+++ b/OpenCL/m08500_a3.cl
@@ -392,7 +392,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX1(i,S) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32x r = data[0];
   u32x l = data[1];
@@ -434,7 +434,7 @@ void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __l
   iv[1] = r;
 }
 
-void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32x tt;
 
@@ -506,7 +506,7 @@ void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32
   }
 }
 
-void transform_racf_key (const u32x w0, const u32x w1, u32x key[2])
+DECLSPEC void transform_racf_key (const u32x w0, const u32x w1, u32x key[2])
 {
   key[0] = BOX1 (((w0 >>  0) & 0xff), c_ascii_to_ebcdic_pc) <<  0
          | BOX1 (((w0 >>  8) & 0xff), c_ascii_to_ebcdic_pc) <<  8
@@ -519,7 +519,7 @@ void transform_racf_key (const u32x w0, const u32x w1, u32x key[2])
          | BOX1 (((w1 >> 24) & 0xff), c_ascii_to_ebcdic_pc) << 24;
 }
 
-void m08500m (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m08500m (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -582,7 +582,7 @@ void m08500m (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16],
   }
 }
 
-void m08500s (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m08500s (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m08600_a0.cl b/OpenCL/m08600_a0.cl
index a2b0f90ca..85dd1f040 100644
--- a/OpenCL/m08600_a0.cl
+++ b/OpenCL/m08600_a0.cl
@@ -63,7 +63,7 @@ __constant static u32a lotus_magic_table[256] =
 #define BOX1(S,i) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
 #endif
 
-void lotus_mix (u32 *in, __local u32 *s_lotus_magic_table)
+DECLSPEC void lotus_mix (u32 *in, __local u32 *s_lotus_magic_table)
 {
   u32 p = 0;
 
@@ -86,7 +86,7 @@ void lotus_mix (u32 *in, __local u32 *s_lotus_magic_table)
   }
 }
 
-void lotus_transform_password (const u32 in[4], u32x out[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void lotus_transform_password (const u32 in[4], u32x out[4], __local u32 *s_lotus_magic_table)
 {
   u32 t = out[3] >> 24;
 
@@ -104,7 +104,7 @@ void lotus_transform_password (const u32 in[4], u32x out[4], __local u32 *s_lotu
   }
 }
 
-void pad (u32 w[4], const u32 len)
+DECLSPEC void pad (u32 w[4], const u32 len)
 {
   const u32 val = 16 - len;
 
@@ -183,7 +183,7 @@ void pad (u32 w[4], const u32 len)
   }
 }
 
-void mdtransform_norecalc (u32 state[4], const u32x block[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void mdtransform_norecalc (u32 state[4], const u32x block[4], __local u32 *s_lotus_magic_table)
 {
   u32 x[12];
 
@@ -208,14 +208,14 @@ void mdtransform_norecalc (u32 state[4], const u32x block[4], __local u32 *s_lot
   state[3] = x[3];
 }
 
-void mdtransform (u32 state[4], u32x checksum[4], const u32x block[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void mdtransform (u32 state[4], u32x checksum[4], const u32x block[4], __local u32 *s_lotus_magic_table)
 {
   mdtransform_norecalc (state, block, s_lotus_magic_table);
 
   lotus_transform_password (block, checksum, s_lotus_magic_table);
 }
 
-void domino_big_md (const u32 saved_key[4], const u32 size, u32x state[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void domino_big_md (const u32 saved_key[4], const u32 size, u32x state[4], __local u32 *s_lotus_magic_table)
 {
   u32 checksum[4];
 
diff --git a/OpenCL/m08600_a1.cl b/OpenCL/m08600_a1.cl
index 7155873cb..5d82c9295 100644
--- a/OpenCL/m08600_a1.cl
+++ b/OpenCL/m08600_a1.cl
@@ -61,7 +61,7 @@ __constant static u32a lotus_magic_table[256] =
 #define BOX1(S,i) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
 #endif
 
-void lotus_mix (u32 *in, __local u32 *s_lotus_magic_table)
+DECLSPEC void lotus_mix (u32 *in, __local u32 *s_lotus_magic_table)
 {
   u32 p = 0;
 
@@ -84,7 +84,7 @@ void lotus_mix (u32 *in, __local u32 *s_lotus_magic_table)
   }
 }
 
-void lotus_transform_password (const u32 in[4], u32x out[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void lotus_transform_password (const u32 in[4], u32x out[4], __local u32 *s_lotus_magic_table)
 {
   u32 t = out[3] >> 24;
 
@@ -102,7 +102,7 @@ void lotus_transform_password (const u32 in[4], u32x out[4], __local u32 *s_lotu
   }
 }
 
-void pad (u32 w[4], const u32 len)
+DECLSPEC void pad (u32 w[4], const u32 len)
 {
   const u32 val = 16 - len;
 
@@ -181,7 +181,7 @@ void pad (u32 w[4], const u32 len)
   }
 }
 
-void mdtransform_norecalc (u32 state[4], const u32x block[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void mdtransform_norecalc (u32 state[4], const u32x block[4], __local u32 *s_lotus_magic_table)
 {
   u32 x[12];
 
@@ -206,14 +206,14 @@ void mdtransform_norecalc (u32 state[4], const u32x block[4], __local u32 *s_lot
   state[3] = x[3];
 }
 
-void mdtransform (u32 state[4], u32x checksum[4], const u32x block[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void mdtransform (u32 state[4], u32x checksum[4], const u32x block[4], __local u32 *s_lotus_magic_table)
 {
   mdtransform_norecalc (state, block, s_lotus_magic_table);
 
   lotus_transform_password (block, checksum, s_lotus_magic_table);
 }
 
-void domino_big_md (const u32 saved_key[4], const u32 size, u32x state[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void domino_big_md (const u32 saved_key[4], const u32 size, u32x state[4], __local u32 *s_lotus_magic_table)
 {
   u32 checksum[4];
 
diff --git a/OpenCL/m08600_a3.cl b/OpenCL/m08600_a3.cl
index 7c571a3da..c856e861f 100644
--- a/OpenCL/m08600_a3.cl
+++ b/OpenCL/m08600_a3.cl
@@ -60,7 +60,7 @@ __constant static u32a lotus_magic_table[256] =
 #define BOX1(S,i) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
 #endif
 
-void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
+DECLSPEC void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
 {
   u32x p = 0;
 
@@ -83,7 +83,7 @@ void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
   }
 }
 
-void lotus_transform_password (const u32x in[4], u32x out[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void lotus_transform_password (const u32x in[4], u32x out[4], __local u32 *s_lotus_magic_table)
 {
   u32x t = out[3] >> 24;
 
@@ -101,7 +101,7 @@ void lotus_transform_password (const u32x in[4], u32x out[4], __local u32 *s_lot
   }
 }
 
-void pad (u32 w[4], const u32 len)
+DECLSPEC void pad (u32 w[4], const u32 len)
 {
   const u32 val = 16 - len;
 
@@ -180,7 +180,7 @@ void pad (u32 w[4], const u32 len)
   }
 }
 
-void mdtransform_norecalc (u32x state[4], const u32x block[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void mdtransform_norecalc (u32x state[4], const u32x block[4], __local u32 *s_lotus_magic_table)
 {
   u32x x[12];
 
@@ -205,14 +205,14 @@ void mdtransform_norecalc (u32x state[4], const u32x block[4], __local u32 *s_lo
   state[3] = x[3];
 }
 
-void mdtransform (u32x state[4], u32x checksum[4], const u32x block[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void mdtransform (u32x state[4], u32x checksum[4], const u32x block[4], __local u32 *s_lotus_magic_table)
 {
   mdtransform_norecalc (state, block, s_lotus_magic_table);
 
   lotus_transform_password (block, checksum, s_lotus_magic_table);
 }
 
-void domino_big_md (const u32x saved_key[4], const u32 size, u32x state[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void domino_big_md (const u32x saved_key[4], const u32 size, u32x state[4], __local u32 *s_lotus_magic_table)
 {
   u32x checksum[4];
 
@@ -226,7 +226,7 @@ void domino_big_md (const u32x saved_key[4], const u32 size, u32x state[4], __lo
   mdtransform_norecalc (state, checksum, s_lotus_magic_table);
 }
 
-void m08600m (__local u32 *s_lotus_magic_table, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m08600m (__local u32 *s_lotus_magic_table, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -273,7 +273,7 @@ void m08600m (__local u32 *s_lotus_magic_table, u32 w[16], const u32 pw_len, __g
   }
 }
 
-void m08600s (__local u32 *s_lotus_magic_table, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m08600s (__local u32 *s_lotus_magic_table, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m08700_a0-optimized.cl b/OpenCL/m08700_a0-optimized.cl
index 6da04c1e2..abe670124 100644
--- a/OpenCL/m08700_a0-optimized.cl
+++ b/OpenCL/m08700_a0-optimized.cl
@@ -75,7 +75,7 @@ __constant static u32a lotus_magic_table[256] =
 #define BOX1(S,i) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
 #endif
 
-void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
+DECLSPEC void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
 {
   u32x p = 0;
 
@@ -98,7 +98,7 @@ void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
   }
 }
 
-void lotus_transform_password (u32x in[4], u32x out[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void lotus_transform_password (u32x in[4], u32x out[4], __local u32 *s_lotus_magic_table)
 {
   u32x t = out[3] >> 24;
 
@@ -116,7 +116,7 @@ void lotus_transform_password (u32x in[4], u32x out[4], __local u32 *s_lotus_mag
   }
 }
 
-void pad (u32 w[4], const u32 len)
+DECLSPEC void pad (u32 w[4], const u32 len)
 {
   const u32 val = 16 - len;
 
@@ -195,7 +195,7 @@ void pad (u32 w[4], const u32 len)
   }
 }
 
-void mdtransform_norecalc (u32x state[4], u32x block[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void mdtransform_norecalc (u32x state[4], u32x block[4], __local u32 *s_lotus_magic_table)
 {
   u32x x[12];
 
@@ -220,14 +220,14 @@ void mdtransform_norecalc (u32x state[4], u32x block[4], __local u32 *s_lotus_ma
   state[3] = x[3];
 }
 
-void mdtransform (u32x state[4], u32x checksum[4], u32x block[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void mdtransform (u32x state[4], u32x checksum[4], u32x block[4], __local u32 *s_lotus_magic_table)
 {
   mdtransform_norecalc (state, block, s_lotus_magic_table);
 
   lotus_transform_password (block, checksum, s_lotus_magic_table);
 }
 
-void domino_big_md (const u32x saved_key[16], const u32 size, u32x state[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void domino_big_md (const u32x saved_key[16], const u32 size, u32x state[4], __local u32 *s_lotus_magic_table)
 {
   u32x checksum[4];
 
diff --git a/OpenCL/m08700_a1-optimized.cl b/OpenCL/m08700_a1-optimized.cl
index 13f1f0a2c..450757348 100644
--- a/OpenCL/m08700_a1-optimized.cl
+++ b/OpenCL/m08700_a1-optimized.cl
@@ -73,7 +73,7 @@ __constant static u32a lotus_magic_table[256] =
 #define BOX1(S,i) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
 #endif
 
-void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
+DECLSPEC void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
 {
   u32x p = 0;
 
@@ -96,7 +96,7 @@ void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
   }
 }
 
-void lotus_transform_password (u32x in[4], u32x out[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void lotus_transform_password (u32x in[4], u32x out[4], __local u32 *s_lotus_magic_table)
 {
   u32x t = out[3] >> 24;
 
@@ -114,7 +114,7 @@ void lotus_transform_password (u32x in[4], u32x out[4], __local u32 *s_lotus_mag
   }
 }
 
-void pad (u32 w[4], const u32 len)
+DECLSPEC void pad (u32 w[4], const u32 len)
 {
   const u32 val = 16 - len;
 
@@ -193,7 +193,7 @@ void pad (u32 w[4], const u32 len)
   }
 }
 
-void mdtransform_norecalc (u32x state[4], u32x block[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void mdtransform_norecalc (u32x state[4], u32x block[4], __local u32 *s_lotus_magic_table)
 {
   u32x x[12];
 
@@ -218,14 +218,14 @@ void mdtransform_norecalc (u32x state[4], u32x block[4], __local u32 *s_lotus_ma
   state[3] = x[3];
 }
 
-void mdtransform (u32x state[4], u32x checksum[4], u32x block[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void mdtransform (u32x state[4], u32x checksum[4], u32x block[4], __local u32 *s_lotus_magic_table)
 {
   mdtransform_norecalc (state, block, s_lotus_magic_table);
 
   lotus_transform_password (block, checksum, s_lotus_magic_table);
 }
 
-void domino_big_md (const u32x saved_key[16], const u32 size, u32x state[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void domino_big_md (const u32x saved_key[16], const u32 size, u32x state[4], __local u32 *s_lotus_magic_table)
 {
   u32x checksum[4];
 
diff --git a/OpenCL/m08700_a3-optimized.cl b/OpenCL/m08700_a3-optimized.cl
index b54293536..9d9baa16b 100644
--- a/OpenCL/m08700_a3-optimized.cl
+++ b/OpenCL/m08700_a3-optimized.cl
@@ -74,7 +74,7 @@ __constant static u32a lotus_magic_table[256] =
 #define BOX1(S,i) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
 #endif
 
-void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
+DECLSPEC void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
 {
   u32x p = 0;
 
@@ -97,7 +97,7 @@ void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
   }
 }
 
-void lotus_transform_password (u32x in[4], u32x out[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void lotus_transform_password (u32x in[4], u32x out[4], __local u32 *s_lotus_magic_table)
 {
   u32x t = out[3] >> 24;
 
@@ -115,7 +115,7 @@ void lotus_transform_password (u32x in[4], u32x out[4], __local u32 *s_lotus_mag
   }
 }
 
-void pad (u32 w[4], const u32 len)
+DECLSPEC void pad (u32 w[4], const u32 len)
 {
   const u32 val = 16 - len;
 
@@ -194,7 +194,7 @@ void pad (u32 w[4], const u32 len)
   }
 }
 
-void mdtransform_norecalc (u32x state[4], u32x block[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void mdtransform_norecalc (u32x state[4], u32x block[4], __local u32 *s_lotus_magic_table)
 {
   u32x x[12];
 
@@ -219,14 +219,14 @@ void mdtransform_norecalc (u32x state[4], u32x block[4], __local u32 *s_lotus_ma
   state[3] = x[3];
 }
 
-void mdtransform (u32x state[4], u32x checksum[4], u32x block[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void mdtransform (u32x state[4], u32x checksum[4], u32x block[4], __local u32 *s_lotus_magic_table)
 {
   mdtransform_norecalc (state, block, s_lotus_magic_table);
 
   lotus_transform_password (block, checksum, s_lotus_magic_table);
 }
 
-void domino_big_md (const u32x saved_key[16], const u32 size, u32x state[4], __local u32 *s_lotus_magic_table)
+DECLSPEC void domino_big_md (const u32x saved_key[16], const u32 size, u32x state[4], __local u32 *s_lotus_magic_table)
 {
   u32x checksum[4];
 
@@ -265,7 +265,7 @@ void domino_big_md (const u32x saved_key[16], const u32 size, u32x state[4], __l
   mdtransform_norecalc (state, checksum, s_lotus_magic_table);
 }
 
-void m08700m (__local u32 *s_lotus_magic_table, __local u32 *l_bin2asc, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m08700m (__local u32 *s_lotus_magic_table, __local u32 *l_bin2asc, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -398,7 +398,7 @@ void m08700m (__local u32 *s_lotus_magic_table, __local u32 *l_bin2asc, u32 w[16
   }
 }
 
-void m08700s (__local u32 *s_lotus_magic_table, __local u32 *l_bin2asc, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m08700s (__local u32 *s_lotus_magic_table, __local u32 *l_bin2asc, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m08800.cl b/OpenCL/m08800.cl
index f5da5fe81..c41c8219b 100644
--- a/OpenCL/m08800.cl
+++ b/OpenCL/m08800.cl
@@ -15,7 +15,7 @@
 #include "inc_hash_sha256.cl"
 #include "inc_cipher_aes.cl"
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m08900.cl b/OpenCL/m08900.cl
index 0233643f9..e72919601 100644
--- a/OpenCL/m08900.cl
+++ b/OpenCL/m08900.cl
@@ -13,7 +13,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-uint4 swap32_4 (uint4 v)
+DECLSPEC uint4 swap32_4 (uint4 v)
 {
   return (rotate ((v & 0x00FF00FF), 24u) | rotate ((v & 0xFF00FF00),  8u));
 }
@@ -73,7 +73,7 @@ uint4 swap32_4 (uint4 v)
   R3 = R3 + X3;         \
 }
 
-void salsa_r (uint4 *TI)
+DECLSPEC void salsa_r (uint4 *TI)
 {
   uint4 R0 = TI[STATE_CNT4 - 4];
   uint4 R1 = TI[STATE_CNT4 - 3];
@@ -125,7 +125,7 @@ void salsa_r (uint4 *TI)
   }
 }
 
-void scrypt_smix (uint4 *X, uint4 *T, __global uint4 *V0, __global uint4 *V1, __global uint4 *V2, __global uint4 *V3)
+DECLSPEC void scrypt_smix (uint4 *X, uint4 *T, __global uint4 *V0, __global uint4 *V1, __global uint4 *V2, __global uint4 *V3)
 {
   #define Coord(xd4,y,z) (((xd4) * ySIZE * zSIZE) + ((y) * zSIZE) + (z))
   #define CO Coord(xd4,y,z)
diff --git a/OpenCL/m09000.cl b/OpenCL/m09000.cl
index 87002b801..8ef6ede41 100644
--- a/OpenCL/m09000.cl
+++ b/OpenCL/m09000.cl
@@ -371,7 +371,7 @@ __constant static u32a c_pbox[18] =
   L ^= P[17];           \
 }
 
-void sha1_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5])
+DECLSPEC void sha1_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5])
 {
   u32 A = digest[0];
   u32 B = digest[1];
diff --git a/OpenCL/m09100.cl b/OpenCL/m09100.cl
index 72a2119d4..5ea051966 100644
--- a/OpenCL/m09100.cl
+++ b/OpenCL/m09100.cl
@@ -70,7 +70,7 @@ __constant static u8a lotus_magic_table[256] =
 
 #define BOX1(S,i) (S)[(i)]
 
-void lotus_mix (u32 *in, const __local u8 *s_lotus_magic_table)
+DECLSPEC void lotus_mix (u32 *in, const __local u8 *s_lotus_magic_table)
 {
   u8 p = 0;
 
@@ -93,7 +93,7 @@ void lotus_mix (u32 *in, const __local u8 *s_lotus_magic_table)
   }
 }
 
-void lotus_transform_password (const u32 in[4], u32 out[4], const __local u8 *s_lotus_magic_table)
+DECLSPEC void lotus_transform_password (const u32 in[4], u32 out[4], const __local u8 *s_lotus_magic_table)
 {
   u8 t = (u8) (out[3] >> 24);
 
@@ -111,7 +111,7 @@ void lotus_transform_password (const u32 in[4], u32 out[4], const __local u8 *s_
   }
 }
 
-void pad (u32 w[4], const u32 len)
+DECLSPEC void pad (u32 w[4], const u32 len)
 {
   const u32 val = 16 - len;
 
@@ -190,7 +190,7 @@ void pad (u32 w[4], const u32 len)
   }
 }
 
-void mdtransform_norecalc (u32 state[4], const u32 block[4], const __local u8 *s_lotus_magic_table)
+DECLSPEC void mdtransform_norecalc (u32 state[4], const u32 block[4], const __local u8 *s_lotus_magic_table)
 {
   u32 x[12];
 
@@ -215,14 +215,14 @@ void mdtransform_norecalc (u32 state[4], const u32 block[4], const __local u8 *s
   state[3] = x[3];
 }
 
-void mdtransform (u32 state[4], u32 checksum[4], const u32 block[4], const __local u8 *s_lotus_magic_table)
+DECLSPEC void mdtransform (u32 state[4], u32 checksum[4], const u32 block[4], const __local u8 *s_lotus_magic_table)
 {
   mdtransform_norecalc (state, block, s_lotus_magic_table);
 
   lotus_transform_password (block, checksum, s_lotus_magic_table);
 }
 
-void domino_big_md (const u32 saved_key[16], const u32 size, u32 state[4], const __local u8 *s_lotus_magic_table)
+DECLSPEC void domino_big_md (const u32 saved_key[16], const u32 size, u32 state[4], const __local u8 *s_lotus_magic_table)
 {
   u32 checksum[4];
 
@@ -261,7 +261,7 @@ void domino_big_md (const u32 saved_key[16], const u32 size, u32 state[4], const
   mdtransform_norecalc (state, checksum, s_lotus_magic_table);
 }
 
-void base64_encode (u8 *base64_hash, const u32 len, const u8 *base64_plain)
+DECLSPEC void base64_encode (u8 *base64_hash, const u32 len, const u8 *base64_plain)
 {
   u8 *out_ptr = (u8 *) base64_hash;
   u8 *in_ptr  = (u8 *) base64_plain;
@@ -285,7 +285,7 @@ void base64_encode (u8 *base64_hash, const u32 len, const u8 *base64_plain)
   }
 }
 
-void lotus6_base64_encode (u8 base64_hash[24], const u32 salt0, const u32 salt1, const u32 a, const u32 b, const u32 c)
+DECLSPEC void lotus6_base64_encode (u8 base64_hash[24], const u32 salt0, const u32 salt1, const u32 a, const u32 b, const u32 c)
 {
   const uchar4 salt0c = as_uchar4 (salt0);
   const uchar4 salt1c = as_uchar4 (salt1);
@@ -350,7 +350,7 @@ void lotus6_base64_encode (u8 base64_hash[24], const u32 salt0, const u32 salt1,
   base64_hash[21] = ')';
 }
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m09700_a0-optimized.cl b/OpenCL/m09700_a0-optimized.cl
index ce83a0f21..6f8c7c785 100644
--- a/OpenCL/m09700_a0-optimized.cl
+++ b/OpenCL/m09700_a0-optimized.cl
@@ -24,7 +24,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -33,7 +33,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -86,7 +86,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -139,7 +139,7 @@ u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4
   return j;
 }
 
-void gen336 (u32 digest_pre[4], u32 salt_buf[4], u32 digest[4])
+DECLSPEC void gen336 (u32 digest_pre[4], u32 salt_buf[4], u32 digest[4])
 {
   u32 digest_t0[2];
   u32 digest_t1[2];
diff --git a/OpenCL/m09700_a1-optimized.cl b/OpenCL/m09700_a1-optimized.cl
index eec6bc3f6..a76e77601 100644
--- a/OpenCL/m09700_a1-optimized.cl
+++ b/OpenCL/m09700_a1-optimized.cl
@@ -22,7 +22,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -31,7 +31,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -84,7 +84,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -137,7 +137,7 @@ u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4
   return j;
 }
 
-void gen336 (u32 digest_pre[4], u32 salt_buf[4], u32 digest[4])
+DECLSPEC void gen336 (u32 digest_pre[4], u32 salt_buf[4], u32 digest[4])
 {
   u32 digest_t0[2];
   u32 digest_t1[2];
diff --git a/OpenCL/m09700_a3-optimized.cl b/OpenCL/m09700_a3-optimized.cl
index bf6e0933e..d96d773a3 100644
--- a/OpenCL/m09700_a3-optimized.cl
+++ b/OpenCL/m09700_a3-optimized.cl
@@ -19,7 +19,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -28,7 +28,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -81,7 +81,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -134,7 +134,7 @@ u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4
   return j;
 }
 
-void m09700m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m09700m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -544,7 +544,7 @@ void m09700m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3
   }
 }
 
-void m09700s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m09700s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m09710_a0-optimized.cl b/OpenCL/m09710_a0-optimized.cl
index 5eb8de77b..4c4ebb19a 100644
--- a/OpenCL/m09710_a0-optimized.cl
+++ b/OpenCL/m09710_a0-optimized.cl
@@ -24,7 +24,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -33,7 +33,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -86,7 +86,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
diff --git a/OpenCL/m09710_a1-optimized.cl b/OpenCL/m09710_a1-optimized.cl
index 3d1535c59..cfec1563e 100644
--- a/OpenCL/m09710_a1-optimized.cl
+++ b/OpenCL/m09710_a1-optimized.cl
@@ -22,7 +22,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -31,7 +31,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -84,7 +84,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
diff --git a/OpenCL/m09710_a3-optimized.cl b/OpenCL/m09710_a3-optimized.cl
index b2f1a3a57..f401f069c 100644
--- a/OpenCL/m09710_a3-optimized.cl
+++ b/OpenCL/m09710_a3-optimized.cl
@@ -22,7 +22,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -31,7 +31,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -84,7 +84,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -137,7 +137,7 @@ u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4
   return j;
 }
 
-void m09710m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m09710m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -248,7 +248,7 @@ void m09710m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3
   }
 }
 
-void m09710s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m09710s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m09720_a0-optimized.cl b/OpenCL/m09720_a0-optimized.cl
index 6bef77f66..9382aa225 100644
--- a/OpenCL/m09720_a0-optimized.cl
+++ b/OpenCL/m09720_a0-optimized.cl
@@ -15,7 +15,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md5.cl"
 
-void gen336 (u32x digest_pre[4], u32 salt_buf[4], u32x digest[4])
+DECLSPEC void gen336 (u32x digest_pre[4], u32 salt_buf[4], u32x digest[4])
 {
   u32x digest_t0[2];
   u32x digest_t1[2];
diff --git a/OpenCL/m09720_a1-optimized.cl b/OpenCL/m09720_a1-optimized.cl
index fb6baafe9..0e731a195 100644
--- a/OpenCL/m09720_a1-optimized.cl
+++ b/OpenCL/m09720_a1-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md5.cl"
 
-void gen336 (u32x digest_pre[4], u32 salt_buf[4], u32x digest[4])
+DECLSPEC void gen336 (u32x digest_pre[4], u32 salt_buf[4], u32x digest[4])
 {
   u32x digest_t0[2];
   u32x digest_t1[2];
diff --git a/OpenCL/m09720_a3-optimized.cl b/OpenCL/m09720_a3-optimized.cl
index cca7e9d85..666c698d8 100644
--- a/OpenCL/m09720_a3-optimized.cl
+++ b/OpenCL/m09720_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md5.cl"
 
-void gen336 (u32x digest_pre[4], u32 salt_buf[4], u32x digest[4])
+DECLSPEC void gen336 (u32x digest_pre[4], u32 salt_buf[4], u32x digest[4])
 {
   u32x digest_t0[2];
   u32x digest_t1[2];
@@ -357,7 +357,7 @@ void gen336 (u32x digest_pre[4], u32 salt_buf[4], u32x digest[4])
   md5_transform_vector (w0_t, w1_t, w2_t, w3_t, digest);
 }
 
-void m09720m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m09720m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -447,7 +447,7 @@ void m09720m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m09720s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m09720s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m09800_a0-optimized.cl b/OpenCL/m09800_a0-optimized.cl
index 2ad8209bf..817172658 100644
--- a/OpenCL/m09800_a0-optimized.cl
+++ b/OpenCL/m09800_a0-optimized.cl
@@ -24,7 +24,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -33,7 +33,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -86,7 +86,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
diff --git a/OpenCL/m09800_a1-optimized.cl b/OpenCL/m09800_a1-optimized.cl
index d76b6b275..9151a94cb 100644
--- a/OpenCL/m09800_a1-optimized.cl
+++ b/OpenCL/m09800_a1-optimized.cl
@@ -22,7 +22,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -31,7 +31,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -84,7 +84,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
diff --git a/OpenCL/m09800_a3-optimized.cl b/OpenCL/m09800_a3-optimized.cl
index 688f48ebb..787ead666 100644
--- a/OpenCL/m09800_a3-optimized.cl
+++ b/OpenCL/m09800_a3-optimized.cl
@@ -19,7 +19,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -28,7 +28,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -81,7 +81,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -134,7 +134,7 @@ u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4
   return j;
 }
 
-void m09800m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice34_t *oldoffice34_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m09800m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice34_t *oldoffice34_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -302,7 +302,7 @@ void m09800m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3
   }
 }
 
-void m09800s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice34_t *oldoffice34_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m09800s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice34_t *oldoffice34_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m09810_a0-optimized.cl b/OpenCL/m09810_a0-optimized.cl
index 75423f4b4..510f4ea61 100644
--- a/OpenCL/m09810_a0-optimized.cl
+++ b/OpenCL/m09810_a0-optimized.cl
@@ -24,7 +24,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -33,7 +33,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -86,7 +86,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
diff --git a/OpenCL/m09810_a1-optimized.cl b/OpenCL/m09810_a1-optimized.cl
index a3278e8fe..412c266ec 100644
--- a/OpenCL/m09810_a1-optimized.cl
+++ b/OpenCL/m09810_a1-optimized.cl
@@ -22,7 +22,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -31,7 +31,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -84,7 +84,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
diff --git a/OpenCL/m09810_a3-optimized.cl b/OpenCL/m09810_a3-optimized.cl
index 059eb1b23..e415e2616 100644
--- a/OpenCL/m09810_a3-optimized.cl
+++ b/OpenCL/m09810_a3-optimized.cl
@@ -22,7 +22,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -31,7 +31,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -84,7 +84,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -137,7 +137,7 @@ u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4
   return j;
 }
 
-void m09810m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice34_t *oldoffice34_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m09810m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice34_t *oldoffice34_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -233,7 +233,7 @@ void m09810m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3
   }
 }
 
-void m09810s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice34_t *oldoffice34_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m09810s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice34_t *oldoffice34_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m09820_a3-optimized.cl b/OpenCL/m09820_a3-optimized.cl
index b1459720d..05ab576e0 100644
--- a/OpenCL/m09820_a3-optimized.cl
+++ b/OpenCL/m09820_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha1.cl"
 
-void m09820m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice34_t *oldoffice34_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m09820m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice34_t *oldoffice34_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -115,7 +115,7 @@ void m09820m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m09820s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice34_t *oldoffice34_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m09820s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice34_t *oldoffice34_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m09900_a3-optimized.cl b/OpenCL/m09900_a3-optimized.cl
index a08f49796..dd78ccc4e 100644
--- a/OpenCL/m09900_a3-optimized.cl
+++ b/OpenCL/m09900_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m09900m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m09900m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -293,7 +293,7 @@ void m09900m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m09900s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m09900s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m10100_a3-optimized.cl b/OpenCL/m10100_a3-optimized.cl
index fc66ec069..3f5004ac9 100644
--- a/OpenCL/m10100_a3-optimized.cl
+++ b/OpenCL/m10100_a3-optimized.cl
@@ -28,7 +28,7 @@
   (v1) ^= (v2);               \
   (v2)  = rotl64 ((v2), 32)
 
-void m10100m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m10100m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -118,7 +118,7 @@ void m10100m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m10100s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m10100s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m10400_a0-optimized.cl b/OpenCL/m10400_a0-optimized.cl
index 3acc03bc3..494f0b83d 100644
--- a/OpenCL/m10400_a0-optimized.cl
+++ b/OpenCL/m10400_a0-optimized.cl
@@ -36,7 +36,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -45,7 +45,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -83,7 +83,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   j += rc4_key->S[255] + d0; swap (rc4_key, 255, j);
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 *in, u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 *in, u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
diff --git a/OpenCL/m10400_a1-optimized.cl b/OpenCL/m10400_a1-optimized.cl
index e90bb3395..e25207bef 100644
--- a/OpenCL/m10400_a1-optimized.cl
+++ b/OpenCL/m10400_a1-optimized.cl
@@ -34,7 +34,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -43,7 +43,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -81,7 +81,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   j += rc4_key->S[255] + d0; swap (rc4_key, 255, j);
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 *in, u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 *in, u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
diff --git a/OpenCL/m10400_a3-optimized.cl b/OpenCL/m10400_a3-optimized.cl
index 20aa97767..a8cab300f 100644
--- a/OpenCL/m10400_a3-optimized.cl
+++ b/OpenCL/m10400_a3-optimized.cl
@@ -34,7 +34,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -43,7 +43,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -81,7 +81,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   j += rc4_key->S[255] + d0; swap (rc4_key, 255, j);
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 *in, u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 *in, u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -134,7 +134,7 @@ u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 *in, u32 ou
   return j;
 }
 
-void m10400m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pdf_t *pdf_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m10400m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pdf_t *pdf_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -300,7 +300,7 @@ void m10400m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3
   }
 }
 
-void m10400s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pdf_t *pdf_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m10400s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pdf_t *pdf_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m10410_a0-optimized.cl b/OpenCL/m10410_a0-optimized.cl
index 4b4112c77..3bbd282eb 100644
--- a/OpenCL/m10410_a0-optimized.cl
+++ b/OpenCL/m10410_a0-optimized.cl
@@ -36,7 +36,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -45,7 +45,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -83,7 +83,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   j += rc4_key->S[255] + d0; swap (rc4_key, 255, j);
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 *in, u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 *in, u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
diff --git a/OpenCL/m10410_a1-optimized.cl b/OpenCL/m10410_a1-optimized.cl
index 73f402b77..2d68cc5b1 100644
--- a/OpenCL/m10410_a1-optimized.cl
+++ b/OpenCL/m10410_a1-optimized.cl
@@ -34,7 +34,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -43,7 +43,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -81,7 +81,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   j += rc4_key->S[255] + d0; swap (rc4_key, 255, j);
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 *in, u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 *in, u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
diff --git a/OpenCL/m10410_a3-optimized.cl b/OpenCL/m10410_a3-optimized.cl
index d064c488a..63b828b33 100644
--- a/OpenCL/m10410_a3-optimized.cl
+++ b/OpenCL/m10410_a3-optimized.cl
@@ -34,7 +34,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -43,7 +43,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -81,7 +81,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   j += rc4_key->S[255] + d0; swap (rc4_key, 255, j);
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 *in, u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 *in, u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -134,7 +134,7 @@ u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 *in, u32 ou
   return j;
 }
 
-void m10410m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pdf_t *pdf_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m10410m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pdf_t *pdf_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -177,7 +177,7 @@ void m10410m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3
   }
 }
 
-void m10410s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pdf_t *pdf_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m10410s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pdf_t *pdf_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m10420_a3-optimized.cl b/OpenCL/m10420_a3-optimized.cl
index 333035c16..fdfffecbb 100644
--- a/OpenCL/m10420_a3-optimized.cl
+++ b/OpenCL/m10420_a3-optimized.cl
@@ -25,7 +25,7 @@ __constant static u32a padding[8] =
   0x7a695364
 };
 
-void m10420m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pdf_t *pdf_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m10420m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pdf_t *pdf_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -178,7 +178,7 @@ void m10420m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m10420s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pdf_t *pdf_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m10420s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pdf_t *pdf_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m10500.cl b/OpenCL/m10500.cl
index df17517ec..898e2d485 100644
--- a/OpenCL/m10500.cl
+++ b/OpenCL/m10500.cl
@@ -33,7 +33,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -42,7 +42,7 @@ void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -98,7 +98,7 @@ void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+DECLSPEC u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
diff --git a/OpenCL/m10700-optimized.cl b/OpenCL/m10700-optimized.cl
index b4e2f10a1..e2871f390 100644
--- a/OpenCL/m10700-optimized.cl
+++ b/OpenCL/m10700-optimized.cl
@@ -36,7 +36,7 @@ typedef struct
 
 } ctx_t;
 
-void orig_sha256_transform (const u32 *w0, const u32 *w1, const u32 *w2, const u32 *w3, u32 *digest)
+DECLSPEC void orig_sha256_transform (const u32 *w0, const u32 *w1, const u32 *w2, const u32 *w3, u32 *digest)
 {
   u32 t0[4];
   u32 t1[4];
@@ -63,7 +63,7 @@ void orig_sha256_transform (const u32 *w0, const u32 *w1, const u32 *w2, const u
   sha256_transform (t0, t1, t2, t3, digest);
 }
 
-void orig_sha384_transform (const u64 *w0, const u64 *w1, const u64 *w2, const u64 *w3, u64 *digest)
+DECLSPEC void orig_sha384_transform (const u64 *w0, const u64 *w1, const u64 *w2, const u64 *w3, u64 *digest)
 {
   u32 t0[4];
   u32 t1[4];
@@ -110,7 +110,7 @@ void orig_sha384_transform (const u64 *w0, const u64 *w1, const u64 *w2, const u
   sha384_transform (t0, t1, t2, t3, t4, t5, t6, t7, digest);
 }
 
-void orig_sha512_transform (const u64 *w0, const u64 *w1, const u64 *w2, const u64 *w3, u64 *digest)
+DECLSPEC void orig_sha512_transform (const u64 *w0, const u64 *w1, const u64 *w2, const u64 *w3, u64 *digest)
 {
   u32 t0[4];
   u32 t1[4];
@@ -176,7 +176,7 @@ void orig_sha512_transform (const u64 *w0, const u64 *w1, const u64 *w2, const u
 #define WORDMAXSZ4  (WORDMAXSZ  / 4)
 #define AESSZ4      (AESSZ      / 4)
 
-void make_sc (u32 *sc, const u32 *pw, const u32 pw_len, const u32 *bl, const u32 bl_len)
+DECLSPEC void make_sc (u32 *sc, const u32 *pw, const u32 pw_len, const u32 *bl, const u32 bl_len)
 {
   const u32 bd = bl_len / 4;
 
@@ -221,7 +221,7 @@ void make_sc (u32 *sc, const u32 *pw, const u32 pw_len, const u32 *bl, const u32
   }
 }
 
-void make_pt_with_offset (u32 *pt, const u32 offset, const u32 *sc, const u32 pwbl_len)
+DECLSPEC void make_pt_with_offset (u32 *pt, const u32 offset, const u32 *sc, const u32 pwbl_len)
 {
   const u32 m = offset % pwbl_len;
 
@@ -245,7 +245,7 @@ void make_pt_with_offset (u32 *pt, const u32 offset, const u32 *sc, const u32 pw
   #endif
 }
 
-void make_w_with_offset (ctx_t *ctx, const u32 W_len, const u32 offset, const u32 *sc, const u32 pwbl_len, u32 *iv, const u32 *ks, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void make_w_with_offset (ctx_t *ctx, const u32 W_len, const u32 offset, const u32 *sc, const u32 pwbl_len, u32 *iv, const u32 *ks, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   for (u32 k = 0, wk = 0; k < W_len; k += AESSZ, wk += AESSZ4)
   {
@@ -267,7 +267,7 @@ void make_w_with_offset (ctx_t *ctx, const u32 W_len, const u32 offset, const u3
   }
 }
 
-u32 do_round (const u32 *pw, const u32 pw_len, ctx_t *ctx, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC u32 do_round (const u32 *pw, const u32 pw_len, ctx_t *ctx, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   // make scratch buffer
 
diff --git a/OpenCL/m10700.cl b/OpenCL/m10700.cl
index 15dfa86d3..c9b44adde 100644
--- a/OpenCL/m10700.cl
+++ b/OpenCL/m10700.cl
@@ -22,7 +22,7 @@
 #define PUTCHAR_BE(a,p,c) ((u8 *)(a))[(p) ^ 3] = (u8) (c)
 #define GETCHAR_BE(a,p)   ((u8 *)(a))[(p) ^ 3]
 
-void aes128_encrypt_cbc (const u32 *aes_ks, u32 *aes_iv, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void aes128_encrypt_cbc (const u32 *aes_ks, u32 *aes_iv, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   u32 data[4];
 
@@ -49,7 +49,7 @@ void aes128_encrypt_cbc (const u32 *aes_ks, u32 *aes_iv, const u32 *in, u32 *out
   out[3] = swap32_S (out[3]);
 }
 
-u32 sha256_update_aes_64 (sha256_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC u32 sha256_update_aes_64 (sha256_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   u32 ex = 0;
 
@@ -134,7 +134,7 @@ u32 sha256_update_aes_64 (sha256_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u3
   return ex;
 }
 
-void sha256_update_aes (sha256_ctx_t *ctx, const u32 *w, const int len, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void sha256_update_aes (sha256_ctx_t *ctx, const u32 *w, const int len, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   u32 w0[4];
   u32 w1[4];
@@ -186,7 +186,7 @@ void sha256_update_aes (sha256_ctx_t *ctx, const u32 *w, const int len, const u3
   sha256_update_aes_64 (ctx, w0, w1, w2, w3, len - pos1, aes_ks, aes_iv, s_te0, s_te1, s_te2, s_te3, s_te4);
 }
 
-void sha256_final_aes (sha256_ctx_t *ctx, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void sha256_final_aes (sha256_ctx_t *ctx, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   int pos = ctx->len & 63;
 
@@ -222,7 +222,7 @@ void sha256_final_aes (sha256_ctx_t *ctx, const u32 *aes_ks, u32 *aes_iv, SHM_TY
   sha256_transform (ctx->w0, ctx->w1, ctx->w2, ctx->w3, ctx->h);
 }
 
-void sha384_update_aes_128 (sha384_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const int len, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void sha384_update_aes_128 (sha384_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const int len, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   const int pos = ctx->len & 127;
 
@@ -357,7 +357,7 @@ void sha384_update_aes_128 (sha384_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4],
   }
 }
 
-void sha384_update_aes (sha384_ctx_t *ctx, const u32 *w, const int len, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void sha384_update_aes (sha384_ctx_t *ctx, const u32 *w, const int len, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   u32 w0[4];
   u32 w1[4];
@@ -445,7 +445,7 @@ void sha384_update_aes (sha384_ctx_t *ctx, const u32 *w, const int len, const u3
   sha384_update_aes_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, len - pos1, aes_ks, aes_iv, s_te0, s_te1, s_te2, s_te3, s_te4);
 }
 
-void sha384_final_aes (sha384_ctx_t *ctx, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void sha384_final_aes (sha384_ctx_t *ctx, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   int pos = ctx->len & 127;
 
@@ -500,7 +500,7 @@ void sha384_final_aes (sha384_ctx_t *ctx, const u32 *aes_ks, u32 *aes_iv, SHM_TY
   sha384_transform (ctx->w0, ctx->w1, ctx->w2, ctx->w3, ctx->w4, ctx->w5, ctx->w6, ctx->w7, ctx->h);
 }
 
-void sha512_update_aes_128 (sha512_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const int len, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void sha512_update_aes_128 (sha512_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const int len, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   const int pos = ctx->len & 127;
 
@@ -635,7 +635,7 @@ void sha512_update_aes_128 (sha512_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4],
   }
 }
 
-void sha512_update_aes (sha512_ctx_t *ctx, const u32 *w, const int len, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void sha512_update_aes (sha512_ctx_t *ctx, const u32 *w, const int len, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   u32 w0[4];
   u32 w1[4];
@@ -723,7 +723,7 @@ void sha512_update_aes (sha512_ctx_t *ctx, const u32 *w, const int len, const u3
   sha512_update_aes_128 (ctx, w0, w1, w2, w3, w4, w5, w6, w7, len - pos1, aes_ks, aes_iv, s_te0, s_te1, s_te2, s_te3, s_te4);
 }
 
-void sha512_final_aes (sha512_ctx_t *ctx, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC void sha512_final_aes (sha512_ctx_t *ctx, const u32 *aes_ks, u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   int pos = ctx->len & 127;
 
@@ -778,7 +778,7 @@ void sha512_final_aes (sha512_ctx_t *ctx, const u32 *aes_ks, u32 *aes_iv, SHM_TY
   sha512_transform (ctx->w0, ctx->w1, ctx->w2, ctx->w3, ctx->w4, ctx->w5, ctx->w6, ctx->w7, ctx->h);
 }
 
-int find_sum (const u32 *w, const u32 pw_len, u32 *bb, const u32 *aes_ks, const u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC int find_sum (const u32 *w, const u32 pw_len, u32 *bb, const u32 *aes_ks, const u32 *aes_iv, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   u32 data[4];
 
@@ -819,7 +819,7 @@ int find_sum (const u32 *w, const u32 pw_len, u32 *bb, const u32 *aes_ks, const
   return sum;
 }
 
-u32 do_round (const u32 *w, const u32 pw_len, pdf17l8_tmp_t *tmp, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC u32 do_round (const u32 *w, const u32 pw_len, pdf17l8_tmp_t *tmp, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   // get previous hash (already padded)
 
diff --git a/OpenCL/m10800_a0-optimized.cl b/OpenCL/m10800_a0-optimized.cl
index e49902848..670c623c5 100644
--- a/OpenCL/m10800_a0-optimized.cl
+++ b/OpenCL/m10800_a0-optimized.cl
@@ -38,7 +38,7 @@ __constant static u64a k_sha384[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha384_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha384_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
diff --git a/OpenCL/m10800_a1-optimized.cl b/OpenCL/m10800_a1-optimized.cl
index 72acdc33b..9c4d62b24 100644
--- a/OpenCL/m10800_a1-optimized.cl
+++ b/OpenCL/m10800_a1-optimized.cl
@@ -36,7 +36,7 @@ __constant static u64a k_sha384[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha384_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha384_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
diff --git a/OpenCL/m10800_a3-optimized.cl b/OpenCL/m10800_a3-optimized.cl
index 92c24ce87..6f3c9d3ee 100644
--- a/OpenCL/m10800_a3-optimized.cl
+++ b/OpenCL/m10800_a3-optimized.cl
@@ -36,7 +36,7 @@ __constant static u64a k_sha384[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha384_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
+DECLSPEC void sha384_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
@@ -135,7 +135,7 @@ void sha384_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], con
   digest[7] = 0;
 }
 
-void m10800m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m10800m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -200,7 +200,7 @@ void m10800m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m10800s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m10800s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m10900.cl b/OpenCL/m10900.cl
index 39e6b5726..4015b6ce1 100644
--- a/OpenCL/m10900.cl
+++ b/OpenCL/m10900.cl
@@ -16,7 +16,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m11000_a3-optimized.cl b/OpenCL/m11000_a3-optimized.cl
index f696fa90c..2f7f8f86e 100644
--- a/OpenCL/m11000_a3-optimized.cl
+++ b/OpenCL/m11000_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m11000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m11000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -277,7 +277,7 @@ void m11000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m11000s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m11000s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m11100_a3-optimized.cl b/OpenCL/m11100_a3-optimized.cl
index a1bb9c82e..8853b92f7 100644
--- a/OpenCL/m11100_a3-optimized.cl
+++ b/OpenCL/m11100_a3-optimized.cl
@@ -24,7 +24,7 @@
 #define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m11100m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m11100m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -329,7 +329,7 @@ void m11100m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m11100s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m11100s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m11200_a3-optimized.cl b/OpenCL/m11200_a3-optimized.cl
index 22f58798c..a91c85e50 100644
--- a/OpenCL/m11200_a3-optimized.cl
+++ b/OpenCL/m11200_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m11200m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m11200m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -442,7 +442,7 @@ void m11200m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m11200s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m11200s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m11300.cl b/OpenCL/m11300.cl
index fec461b73..14b851742 100644
--- a/OpenCL/m11300.cl
+++ b/OpenCL/m11300.cl
@@ -14,7 +14,7 @@
 #include "inc_hash_sha512.cl"
 #include "inc_cipher_aes.cl"
 
-void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m11500_a0-optimized.cl b/OpenCL/m11500_a0-optimized.cl
index 8b7eaae95..a02752fd8 100644
--- a/OpenCL/m11500_a0-optimized.cl
+++ b/OpenCL/m11500_a0-optimized.cl
@@ -83,7 +83,7 @@ __constant static u32a crc32tab[0x100] =
   0xb40bbe37, 0xc30c8ea1, 0x5a05df1b, 0x2d02ef8d
 };
 
-u32x round_crc32 (u32x a, const u32x v)
+DECLSPEC u32x round_crc32 (u32x a, const u32x v)
 {
   const u32x k = (a ^ v) & 0xff;
 
@@ -106,7 +106,7 @@ u32x round_crc32 (u32x a, const u32x v)
   return a;
 }
 
-u32x crc32 (const u32x w[16], const u32 pw_len, const u32 iv)
+DECLSPEC u32x crc32 (const u32x w[16], const u32 pw_len, const u32 iv)
 {
   u32x a = iv ^ ~0;
 
diff --git a/OpenCL/m11500_a1-optimized.cl b/OpenCL/m11500_a1-optimized.cl
index c94d5ced0..bc050b839 100644
--- a/OpenCL/m11500_a1-optimized.cl
+++ b/OpenCL/m11500_a1-optimized.cl
@@ -81,7 +81,7 @@ __constant static u32a crc32tab[0x100] =
   0xb40bbe37, 0xc30c8ea1, 0x5a05df1b, 0x2d02ef8d
 };
 
-u32x round_crc32 (u32x a, const u32x v)
+DECLSPEC u32x round_crc32 (u32x a, const u32x v)
 {
   const u32x k = (a ^ v) & 0xff;
 
@@ -104,7 +104,7 @@ u32x round_crc32 (u32x a, const u32x v)
   return a;
 }
 
-u32x crc32 (const u32x w[16], const u32 pw_len, const u32 iv)
+DECLSPEC u32x crc32 (const u32x w[16], const u32 pw_len, const u32 iv)
 {
   u32x a = iv ^ ~0;
 
diff --git a/OpenCL/m11500_a3-optimized.cl b/OpenCL/m11500_a3-optimized.cl
index ff009e2aa..3ea2b83c8 100644
--- a/OpenCL/m11500_a3-optimized.cl
+++ b/OpenCL/m11500_a3-optimized.cl
@@ -81,7 +81,7 @@ __constant static u32a crc32tab[0x100] =
   0xb40bbe37, 0xc30c8ea1, 0x5a05df1b, 0x2d02ef8d
 };
 
-u32x round_crc32 (u32x a, const u32x v)
+DECLSPEC u32x round_crc32 (u32x a, const u32x v)
 {
   const u32x k = (a ^ v) & 0xff;
 
@@ -104,7 +104,7 @@ u32x round_crc32 (u32x a, const u32x v)
   return a;
 }
 
-u32x crc32 (const u32x w[16], const u32 pw_len, const u32 iv)
+DECLSPEC u32x crc32 (const u32x w[16], const u32 pw_len, const u32 iv)
 {
   u32x a = iv ^ ~0;
 
@@ -124,7 +124,7 @@ u32x crc32 (const u32x w[16], const u32 pw_len, const u32 iv)
   return ~a;
 }
 
-void m11500m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m11500m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -182,7 +182,7 @@ void m11500m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m11500s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m11500s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m11600.cl b/OpenCL/m11600.cl
index b8ecd55a0..ed3e6c2bc 100644
--- a/OpenCL/m11600.cl
+++ b/OpenCL/m11600.cl
@@ -10,7 +10,7 @@
 #include "inc_common.cl"
 #include "inc_hash_sha256.cl"
 
-void memcat8c_be (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 len, const u32 append, u32 digest[8])
+DECLSPEC void memcat8c_be (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 len, const u32 append, u32 digest[8])
 {
   const u32 func_len = len & 63;
 
diff --git a/OpenCL/m11700_a0-optimized.cl b/OpenCL/m11700_a0-optimized.cl
index 113dc0655..e69353d29 100644
--- a/OpenCL/m11700_a0-optimized.cl
+++ b/OpenCL/m11700_a0-optimized.cl
@@ -2233,7 +2233,7 @@ __constant static u64a sbob_rc64[12][8] =
   },
 };
 
-void streebog_g (u64x h[8], const u64x m[8], __local u64 (*s_sbob_sl64)[256])
+DECLSPEC void streebog_g (u64x h[8], const u64x m[8], __local u64 (*s_sbob_sl64)[256])
 {
   u64x k[8];
   u64x s[8];
diff --git a/OpenCL/m11700_a1-optimized.cl b/OpenCL/m11700_a1-optimized.cl
index d06115300..cb624b403 100644
--- a/OpenCL/m11700_a1-optimized.cl
+++ b/OpenCL/m11700_a1-optimized.cl
@@ -2231,7 +2231,7 @@ __constant static u64a sbob_rc64[12][8] =
   },
 };
 
-void streebog_g (u64x h[8], const u64x m[8], __local u64 (*s_sbob_sl64)[256])
+DECLSPEC void streebog_g (u64x h[8], const u64x m[8], __local u64 (*s_sbob_sl64)[256])
 {
   u64x k[8];
   u64x s[8];
diff --git a/OpenCL/m11700_a3-optimized.cl b/OpenCL/m11700_a3-optimized.cl
index 54e1858f9..5d2926766 100644
--- a/OpenCL/m11700_a3-optimized.cl
+++ b/OpenCL/m11700_a3-optimized.cl
@@ -2231,7 +2231,7 @@ __constant static u64a sbob_rc64[12][8] =
   },
 };
 
-void streebog_g (u64x h[8], const u64x m[8], __local u64 (*s_sbob_sl64)[256])
+DECLSPEC void streebog_g (u64x h[8], const u64x m[8], __local u64 (*s_sbob_sl64)[256])
 {
   u64x k[8];
   u64x s[8];
@@ -2290,7 +2290,7 @@ void streebog_g (u64x h[8], const u64x m[8], __local u64 (*s_sbob_sl64)[256])
   }
 }
 
-void m11700m (__local u64 (*s_sbob_sl64)[256], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m11700m (__local u64 (*s_sbob_sl64)[256], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -2373,7 +2373,7 @@ void m11700m (__local u64 (*s_sbob_sl64)[256], u32 w[16], const u32 pw_len, __gl
   }
 }
 
-void m11700s (__local u64 (*s_sbob_sl64)[256], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m11700s (__local u64 (*s_sbob_sl64)[256], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m11800_a0-optimized.cl b/OpenCL/m11800_a0-optimized.cl
index 9647d7e81..96e077763 100644
--- a/OpenCL/m11800_a0-optimized.cl
+++ b/OpenCL/m11800_a0-optimized.cl
@@ -2233,7 +2233,7 @@ __constant static u64a sbob_rc64[12][8] =
   },
 };
 
-void streebog_g (u64x h[8], const u64x m[8], __local u64 (*s_sbob_sl64)[256])
+DECLSPEC void streebog_g (u64x h[8], const u64x m[8], __local u64 (*s_sbob_sl64)[256])
 {
   u64x k[8];
   u64x s[8];
diff --git a/OpenCL/m11800_a1-optimized.cl b/OpenCL/m11800_a1-optimized.cl
index c3d838592..e9a539f5d 100644
--- a/OpenCL/m11800_a1-optimized.cl
+++ b/OpenCL/m11800_a1-optimized.cl
@@ -2231,7 +2231,7 @@ __constant static u64a sbob_rc64[12][8] =
   },
 };
 
-void streebog_g (u64x h[8], const u64x m[8], __local u64 (*s_sbob_sl64)[256])
+DECLSPEC void streebog_g (u64x h[8], const u64x m[8], __local u64 (*s_sbob_sl64)[256])
 {
   u64x k[8];
   u64x s[8];
diff --git a/OpenCL/m11800_a3-optimized.cl b/OpenCL/m11800_a3-optimized.cl
index db609fae2..518523601 100644
--- a/OpenCL/m11800_a3-optimized.cl
+++ b/OpenCL/m11800_a3-optimized.cl
@@ -2231,7 +2231,7 @@ __constant static u64a sbob_rc64[12][8] =
   },
 };
 
-void streebog_g (u64x h[8], const u64x m[8], __local u64 (*s_sbob_sl64)[256])
+DECLSPEC void streebog_g (u64x h[8], const u64x m[8], __local u64 (*s_sbob_sl64)[256])
 {
   u64x k[8];
   u64x s[8];
@@ -2290,7 +2290,7 @@ void streebog_g (u64x h[8], const u64x m[8], __local u64 (*s_sbob_sl64)[256])
   }
 }
 
-void m11800m (__local u64 (*s_sbob_sl64)[256], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m11800m (__local u64 (*s_sbob_sl64)[256], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -2373,7 +2373,7 @@ void m11800m (__local u64 (*s_sbob_sl64)[256], u32 w[16], const u32 pw_len, __gl
   }
 }
 
-void m11800s (__local u64 (*s_sbob_sl64)[256], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m11800s (__local u64 (*s_sbob_sl64)[256], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m11900.cl b/OpenCL/m11900.cl
index b49a96e78..ee8de4ebe 100644
--- a/OpenCL/m11900.cl
+++ b/OpenCL/m11900.cl
@@ -16,7 +16,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_md5_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_md5_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m12000.cl b/OpenCL/m12000.cl
index 1830d34a8..13c3dffb6 100644
--- a/OpenCL/m12000.cl
+++ b/OpenCL/m12000.cl
@@ -16,7 +16,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m12300.cl b/OpenCL/m12300.cl
index 6446aec01..5d2c2e6d3 100644
--- a/OpenCL/m12300.cl
+++ b/OpenCL/m12300.cl
@@ -16,7 +16,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m12400.cl b/OpenCL/m12400.cl
index 7ad9d767e..56a012773 100644
--- a/OpenCL/m12400.cl
+++ b/OpenCL/m12400.cl
@@ -348,7 +348,7 @@ __constant static u32a c_skb[8][64] =
 
 #define BOX(i,n,S) (S)[(n)][(i)]
 
-void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
@@ -417,7 +417,7 @@ void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_
   }
 }
 
-void _des_crypt_encrypt (u32 iv[2], u32 mask, u32 rounds, u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32 mask, u32 rounds, u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32 tt;
 
diff --git a/OpenCL/m12500.cl b/OpenCL/m12500.cl
index e2ff654c8..35e370e94 100644
--- a/OpenCL/m12500.cl
+++ b/OpenCL/m12500.cl
@@ -23,7 +23,7 @@
 
 #define MIN(a,b) (((a) < (b)) ? (a) : (b))
 
-void sha1_transform (const u32 w[16], u32 digest[5])
+DECLSPEC void sha1_transform (const u32 w[16], u32 digest[5])
 {
   u32 A = digest[0];
   u32 B = digest[1];
diff --git a/OpenCL/m12600_a3-optimized.cl b/OpenCL/m12600_a3-optimized.cl
index 6e43f7814..a8cdeaa8b 100644
--- a/OpenCL/m12600_a3-optimized.cl
+++ b/OpenCL/m12600_a3-optimized.cl
@@ -24,7 +24,7 @@
 #define uint_to_hex_upper8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m12600m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m12600m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -315,7 +315,7 @@ void m12600m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m12600s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m12600s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m12700.cl b/OpenCL/m12700.cl
index 251e0a0cd..bf0793dfb 100644
--- a/OpenCL/m12700.cl
+++ b/OpenCL/m12700.cl
@@ -17,7 +17,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m12800.cl b/OpenCL/m12800.cl
index 66d1717bf..55ea9916e 100644
--- a/OpenCL/m12800.cl
+++ b/OpenCL/m12800.cl
@@ -17,7 +17,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m12900.cl b/OpenCL/m12900.cl
index b023841aa..1c6de8a21 100644
--- a/OpenCL/m12900.cl
+++ b/OpenCL/m12900.cl
@@ -16,7 +16,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m13000.cl b/OpenCL/m13000.cl
index 46909572f..1c5e84e9c 100644
--- a/OpenCL/m13000.cl
+++ b/OpenCL/m13000.cl
@@ -16,7 +16,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m13100_a0-optimized.cl b/OpenCL/m13100_a0-optimized.cl
index 3d0e8bb36..4556b8fbb 100644
--- a/OpenCL/m13100_a0-optimized.cl
+++ b/OpenCL/m13100_a0-optimized.cl
@@ -25,7 +25,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -34,7 +34,7 @@ void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -87,7 +87,7 @@ void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, __global u32 *in, u32 out[4])
+DECLSPEC u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const __global u32 *in, u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -140,7 +140,7 @@ u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, __global u32 *in, u32 out
   return j;
 }
 
-void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
+DECLSPEC void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -191,7 +191,7 @@ void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32
   md5_transform (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
+DECLSPEC void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -225,7 +225,7 @@ void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32
   md5_transform (w0, w1, w2, w3, digest);
 }
 
-int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global u32 *edata2, const u32 edata2_len, const u32 K2[4], const u32 checksum[4])
+DECLSPEC int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global const u32 *edata2, const u32 edata2_len, const u32 K2[4], const u32 checksum[4])
 {
   rc4_init_16 (rc4_key, data);
 
@@ -429,7 +429,7 @@ int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global u32 *eda
   return 1;
 }
 
-void kerb_prepare (const u32 w0[4], const u32 w1[4], const u32 pw_len, const u32 checksum[4], u32 digest[4], u32 K2[4])
+DECLSPEC void kerb_prepare (const u32 w0[4], const u32 w1[4], const u32 pw_len, const u32 checksum[4], u32 digest[4], u32 K2[4])
 {
   /**
    * pads
diff --git a/OpenCL/m13100_a0.cl b/OpenCL/m13100_a0.cl
index 768cf47e9..990c57013 100644
--- a/OpenCL/m13100_a0.cl
+++ b/OpenCL/m13100_a0.cl
@@ -24,7 +24,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -33,7 +33,7 @@ void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -86,7 +86,7 @@ void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, __global u32 *in, u32 out[4])
+DECLSPEC u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const __global u32 *in, u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -139,7 +139,7 @@ u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, __global u32 *in, u32 out
   return j;
 }
 
-int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global const u32 *edata2, const u32 edata2_len, const u32 K2[4], const u32 checksum[4])
+DECLSPEC int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global const u32 *edata2, const u32 edata2_len, const u32 K2[4], const u32 checksum[4])
 {
   rc4_init_16 (rc4_key, data);
 
@@ -273,7 +273,7 @@ int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global const u3
   return 1;
 }
 
-void kerb_prepare (const u32 K[4], const u32 checksum[4], u32 digest[4], u32 K2[4])
+DECLSPEC void kerb_prepare (const u32 K[4], const u32 checksum[4], u32 digest[4], u32 K2[4])
 {
   // K1=MD5_HMAC(K,1); with 1 encoded as little indian on 4 bytes (01000000 in hexa);
 
diff --git a/OpenCL/m13100_a1-optimized.cl b/OpenCL/m13100_a1-optimized.cl
index f2314427d..d55a75e0e 100644
--- a/OpenCL/m13100_a1-optimized.cl
+++ b/OpenCL/m13100_a1-optimized.cl
@@ -23,7 +23,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -32,7 +32,7 @@ void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -85,7 +85,7 @@ void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, __global u32 *in, u32 out[4])
+DECLSPEC u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const __global u32 *in, u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -138,7 +138,7 @@ u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, __global u32 *in, u32 out
   return j;
 }
 
-void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
+DECLSPEC void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -189,7 +189,7 @@ void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32
   md5_transform (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
+DECLSPEC void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -223,7 +223,7 @@ void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32
   md5_transform (w0, w1, w2, w3, digest);
 }
 
-int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global u32 *edata2, const u32 edata2_len, const u32 K2[4], const u32 checksum[4])
+DECLSPEC int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global const u32 *edata2, const u32 edata2_len, const u32 K2[4], const u32 checksum[4])
 {
   rc4_init_16 (rc4_key, data);
 
@@ -427,7 +427,7 @@ int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global u32 *eda
   return 1;
 }
 
-void kerb_prepare (const u32 w0[4], const u32 w1[4], const u32 pw_len, const u32 checksum[4], u32 digest[4], u32 K2[4])
+DECLSPEC void kerb_prepare (const u32 w0[4], const u32 w1[4], const u32 pw_len, const u32 checksum[4], u32 digest[4], u32 K2[4])
 {
   /**
    * pads
diff --git a/OpenCL/m13100_a1.cl b/OpenCL/m13100_a1.cl
index 53ecaf5f5..486441888 100644
--- a/OpenCL/m13100_a1.cl
+++ b/OpenCL/m13100_a1.cl
@@ -22,7 +22,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -31,7 +31,7 @@ void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -84,7 +84,7 @@ void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, __global u32 *in, u32 out[4])
+DECLSPEC u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const __global u32 *in, u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -137,7 +137,7 @@ u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, __global u32 *in, u32 out
   return j;
 }
 
-int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global const u32 *edata2, const u32 edata2_len, const u32 K2[4], const u32 checksum[4])
+DECLSPEC int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global const u32 *edata2, const u32 edata2_len, const u32 K2[4], const u32 checksum[4])
 {
   rc4_init_16 (rc4_key, data);
 
@@ -271,7 +271,7 @@ int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global const u3
   return 1;
 }
 
-void kerb_prepare (const u32 K[4], const u32 checksum[4], u32 digest[4], u32 K2[4])
+DECLSPEC void kerb_prepare (const u32 K[4], const u32 checksum[4], u32 digest[4], u32 K2[4])
 {
   // K1=MD5_HMAC(K,1); with 1 encoded as little indian on 4 bytes (01000000 in hexa);
 
diff --git a/OpenCL/m13100_a3-optimized.cl b/OpenCL/m13100_a3-optimized.cl
index b397fa378..0ffbc8df2 100644
--- a/OpenCL/m13100_a3-optimized.cl
+++ b/OpenCL/m13100_a3-optimized.cl
@@ -23,7 +23,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -32,7 +32,7 @@ void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -85,7 +85,7 @@ void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, __global u32 *in, u32 out[4])
+DECLSPEC u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const __global u32 *in, u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -138,7 +138,7 @@ u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, __global u32 *in, u32 out
   return j;
 }
 
-void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
+DECLSPEC void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -189,7 +189,7 @@ void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32
   md5_transform (w0, w1, w2, w3, opad);
 }
 
-void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
+DECLSPEC void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -223,7 +223,7 @@ void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32
   md5_transform (w0, w1, w2, w3, digest);
 }
 
-int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global u32 *edata2, const u32 edata2_len, const u32 K2[4], const u32 checksum[4])
+DECLSPEC int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global const u32 *edata2, const u32 edata2_len, const u32 K2[4], const u32 checksum[4])
 {
   rc4_init_16 (rc4_key, data);
 
@@ -427,7 +427,7 @@ int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global u32 *eda
   return 1;
 }
 
-void kerb_prepare (const u32 w0[4], const u32 w1[4], const u32 pw_len, const u32 checksum[4], u32 digest[4], u32 K2[4])
+DECLSPEC void kerb_prepare (const u32 w0[4], const u32 w1[4], const u32 pw_len, const u32 checksum[4], u32 digest[4], u32 K2[4])
 {
   /**
    * pads
@@ -563,7 +563,7 @@ void kerb_prepare (const u32 w0[4], const u32 w1[4], const u32 pw_len, const u32
   hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
 }
 
-void m13100 (SCR_TYPE RC4_KEY *rc4_key, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global krb5tgs_t *krb5tgs_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m13100 (SCR_TYPE RC4_KEY *rc4_key, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global krb5tgs_t *krb5tgs_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m13100_a3.cl b/OpenCL/m13100_a3.cl
index f053436ab..d7c59e50f 100644
--- a/OpenCL/m13100_a3.cl
+++ b/OpenCL/m13100_a3.cl
@@ -22,7 +22,7 @@ typedef struct
 
 } RC4_KEY;
 
-void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
+DECLSPEC void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -31,7 +31,7 @@ void swap (SCR_TYPE RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
+DECLSPEC void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
@@ -84,7 +84,7 @@ void rc4_init_16 (SCR_TYPE RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, __global u32 *in, u32 out[4])
+DECLSPEC u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, const __global u32 *in, u32 out[4])
 {
   #ifdef _unroll
   #pragma unroll
@@ -137,7 +137,7 @@ u8 rc4_next_16 (SCR_TYPE RC4_KEY *rc4_key, u8 i, u8 j, __global u32 *in, u32 out
   return j;
 }
 
-int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global const u32 *edata2, const u32 edata2_len, const u32 K2[4], const u32 checksum[4])
+DECLSPEC int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global const u32 *edata2, const u32 edata2_len, const u32 K2[4], const u32 checksum[4])
 {
   rc4_init_16 (rc4_key, data);
 
@@ -271,7 +271,7 @@ int decrypt_and_check (SCR_TYPE RC4_KEY *rc4_key, u32 data[4], __global const u3
   return 1;
 }
 
-void kerb_prepare (const u32 K[4], const u32 checksum[4], u32 digest[4], u32 K2[4])
+DECLSPEC void kerb_prepare (const u32 K[4], const u32 checksum[4], u32 digest[4], u32 K2[4])
 {
   // K1=MD5_HMAC(K,1); with 1 encoded as little indian on 4 bytes (01000000 in hexa);
 
diff --git a/OpenCL/m13300_a3-optimized.cl b/OpenCL/m13300_a3-optimized.cl
index 30db8b488..89e15963f 100644
--- a/OpenCL/m13300_a3-optimized.cl
+++ b/OpenCL/m13300_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m13300m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m13300m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -263,7 +263,7 @@ void m13300m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m13300s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m13300s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m13500_a3-optimized.cl b/OpenCL/m13500_a3-optimized.cl
index cbb32770e..02c057ac5 100644
--- a/OpenCL/m13500_a3-optimized.cl
+++ b/OpenCL/m13500_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m13500m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pstoken_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m13500m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pstoken_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -406,7 +406,7 @@ void m13500m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m13500s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pstoken_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m13500s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global pstoken_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m13600.cl b/OpenCL/m13600.cl
index 1d462c631..ada5b01c2 100644
--- a/OpenCL/m13600.cl
+++ b/OpenCL/m13600.cl
@@ -16,7 +16,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m13751.cl b/OpenCL/m13751.cl
index cd9b209d9..b839f7b88 100644
--- a/OpenCL/m13751.cl
+++ b/OpenCL/m13751.cl
@@ -20,7 +20,7 @@
 #include "inc_truecrypt_crc32.cl"
 #include "inc_truecrypt_xts.cl"
 
-u32 u8add (const u32 a, const u32 b)
+DECLSPEC u32 u8add (const u32 a, const u32 b)
 {
   const u32 a1 = (a >>  0) & 0xff;
   const u32 a2 = (a >>  8) & 0xff;
@@ -45,7 +45,7 @@ u32 u8add (const u32 a, const u32 b)
   return r;
 }
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m13752.cl b/OpenCL/m13752.cl
index 39e37fb52..72fb0df41 100644
--- a/OpenCL/m13752.cl
+++ b/OpenCL/m13752.cl
@@ -20,7 +20,7 @@
 #include "inc_truecrypt_crc32.cl"
 #include "inc_truecrypt_xts.cl"
 
-u32 u8add (const u32 a, const u32 b)
+DECLSPEC u32 u8add (const u32 a, const u32 b)
 {
   const u32 a1 = (a >>  0) & 0xff;
   const u32 a2 = (a >>  8) & 0xff;
@@ -45,7 +45,7 @@ u32 u8add (const u32 a, const u32 b)
   return r;
 }
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m13753.cl b/OpenCL/m13753.cl
index 599102d9f..4eaafa940 100644
--- a/OpenCL/m13753.cl
+++ b/OpenCL/m13753.cl
@@ -20,7 +20,7 @@
 #include "inc_truecrypt_crc32.cl"
 #include "inc_truecrypt_xts.cl"
 
-u32 u8add (const u32 a, const u32 b)
+DECLSPEC u32 u8add (const u32 a, const u32 b)
 {
   const u32 a1 = (a >>  0) & 0xff;
   const u32 a2 = (a >>  8) & 0xff;
@@ -45,7 +45,7 @@ u32 u8add (const u32 a, const u32 b)
   return r;
 }
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m13800_a0-optimized.cl b/OpenCL/m13800_a0-optimized.cl
index 9175a5a8b..cf8bdfbca 100644
--- a/OpenCL/m13800_a0-optimized.cl
+++ b/OpenCL/m13800_a0-optimized.cl
@@ -16,12 +16,12 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha256.cl"
 
-void sha256_transform_transport_vector (const u32x w[16], u32x digest[8])
+DECLSPEC void sha256_transform_transport_vector (const u32x w[16], u32x digest[8])
 {
   sha256_transform_vector (w + 0, w + 4, w + 8, w + 12, digest);
 }
 
-void memcat64c_be (u32x block[16], const u32 offset, u32x carry[16])
+DECLSPEC void memcat64c_be (u32x block[16], const u32 offset, u32x carry[16])
 {
   const u32 mod = offset & 3;
   const u32 div = offset / 4;
diff --git a/OpenCL/m13800_a1-optimized.cl b/OpenCL/m13800_a1-optimized.cl
index a3ced2b38..7b677ceac 100644
--- a/OpenCL/m13800_a1-optimized.cl
+++ b/OpenCL/m13800_a1-optimized.cl
@@ -14,12 +14,12 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha256.cl"
 
-void sha256_transform_transport_vector (const u32x w[16], u32x digest[8])
+DECLSPEC void sha256_transform_transport_vector (const u32x w[16], u32x digest[8])
 {
   sha256_transform_vector (w + 0, w + 4, w + 8, w + 12, digest);
 }
 
-void memcat64c_be (u32x block[16], const u32 offset, u32x carry[16])
+DECLSPEC void memcat64c_be (u32x block[16], const u32 offset, u32x carry[16])
 {
   const u32 mod = offset & 3;
   const u32 div = offset / 4;
diff --git a/OpenCL/m13800_a3-optimized.cl b/OpenCL/m13800_a3-optimized.cl
index 51c4aea24..d3c7029e0 100644
--- a/OpenCL/m13800_a3-optimized.cl
+++ b/OpenCL/m13800_a3-optimized.cl
@@ -13,12 +13,12 @@
 #include "inc_simd.cl"
 #include "inc_hash_sha256.cl"
 
-void sha256_transform_transport_vector (const u32x w[16], u32x digest[8])
+DECLSPEC void sha256_transform_transport_vector (const u32x w[16], u32x digest[8])
 {
   sha256_transform_vector (w + 0, w + 4, w + 8, w + 12, digest);
 }
 
-void memcat64c_be (u32x block[16], const u32 offset, u32x carry[16])
+DECLSPEC void memcat64c_be (u32x block[16], const u32 offset, u32x carry[16])
 {
   const u32 mod = offset & 3;
   const u32 div = offset / 4;
@@ -393,7 +393,7 @@ void memcat64c_be (u32x block[16], const u32 offset, u32x carry[16])
   }
 }
 
-void m13800m (__local u32 *s_esalt, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global win8phone_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m13800m (__local u32 *s_esalt, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global win8phone_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -541,7 +541,7 @@ void m13800m (__local u32 *s_esalt, u32 w[16], const u32 pw_len, __global pw_t *
   }
 }
 
-void m13800s (__local u32 *s_esalt, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global win8phone_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m13800s (__local u32 *s_esalt, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global win8phone_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m13900_a3-optimized.cl b/OpenCL/m13900_a3-optimized.cl
index a200dfdb6..777bb87b7 100644
--- a/OpenCL/m13900_a3-optimized.cl
+++ b/OpenCL/m13900_a3-optimized.cl
@@ -25,7 +25,7 @@
 #define uint_to_hex_lower8_le(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void m13900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m13900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -208,7 +208,7 @@ void m13900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m13900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m13900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m14000_a0.cl b/OpenCL/m14000_a0.cl
index 7a851db23..4d0993908 100644
--- a/OpenCL/m14000_a0.cl
+++ b/OpenCL/m14000_a0.cl
@@ -380,7 +380,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX1(i,S) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32 r = rotl32_S (data[0], 3u);
   u32 l = rotl32_S (data[1], 3u);
@@ -424,7 +424,7 @@ void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local
   iv[1] = rotl32_S (r, 29u);
 }
 
-void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
diff --git a/OpenCL/m14000_a1.cl b/OpenCL/m14000_a1.cl
index bb3ed9c95..254b7e0aa 100644
--- a/OpenCL/m14000_a1.cl
+++ b/OpenCL/m14000_a1.cl
@@ -370,7 +370,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX1(i,S) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32 r = rotl32_S (data[0], 3u);
   u32 l = rotl32_S (data[1], 3u);
@@ -414,7 +414,7 @@ void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local
   iv[1] = rotl32_S (r, 29u);
 }
 
-void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
diff --git a/OpenCL/m14000_a3.cl b/OpenCL/m14000_a3.cl
index 9dcc7fd36..f44a6a9c4 100644
--- a/OpenCL/m14000_a3.cl
+++ b/OpenCL/m14000_a3.cl
@@ -54,7 +54,7 @@
 
 #define LUT(a,b,c,d,e) u32 a; asm volatile ("lop3.b32 %0, %1, %2, %3, "#e";" : "=r"(a): "r"(b), "r"(c), "r"(d));
 
-void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(xAA55AA5500550055, a1, a4, a6, 0xC1)
   LUT(xA55AA55AF0F5F0F5, a3, a6, xAA55AA5500550055, 0x9E)
@@ -88,7 +88,7 @@ void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(xEEEEEEEE99999999, a1, a2, a6, 0x97)
   LUT(xFFFFEEEE66666666, a5, a6, xEEEEEEEE99999999, 0x67)
@@ -121,7 +121,7 @@ void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(xA50FA50FA50FA50F, a1, a3, a4, 0xC9)
   LUT(xF0F00F0FF0F0F0F0, a3, a5, a6, 0x4B)
@@ -155,7 +155,7 @@ void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(x55F055F055F055F0, a1, a3, a4, 0x72)
   LUT(xA500F5F0A500F5F0, a3, a5, x55F055F055F055F0, 0xAD)
@@ -182,7 +182,7 @@ void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(xA0A0A0A0FFFFFFFF, a1, a3, a6, 0xAB)
   LUT(xFFFF00005555FFFF, a1, a5, a6, 0xB9)
@@ -216,7 +216,7 @@ void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(x5050F5F55050F5F5, a1, a3, a5, 0xB2)
   LUT(x6363C6C66363C6C6, a1, a2, x5050F5F55050F5F5, 0x66)
@@ -249,7 +249,7 @@ void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(x88AA88AA88AA88AA, a1, a2, a4, 0x0B)
   LUT(xAAAAFF00AAAAFF00, a1, a4, a5, 0x27)
@@ -282,7 +282,7 @@ void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
   *out4 ^= x4;
 }
 
-void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
   LUT(xEEEE3333EEEE3333, a1, a2, a5, 0x9D)
   LUT(xBBBBBBBBBBBBBBBB, a1, a1, a2, 0x83)
@@ -340,7 +340,7 @@ void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
  * The effort has been sponsored by Rapid7: http://www.rapid7.com
  */
 
-void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x55005500, x5A0F5A0F, x3333FFFF, x66666666, x22226666, x2D2D6969,
         x25202160;
@@ -415,7 +415,7 @@ void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x33CC33CC;
     u32 x55550000, x00AA00FF, x33BB33FF;
@@ -486,7 +486,7 @@ void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x44444444, x0F0FF0F0, x4F4FF4F4, x00FFFF00, x00AAAA00, x4FE55EF4;
     u32 x3C3CC3C3, x3C3C0000, x7373F4F4, x0C840A00;
@@ -557,7 +557,7 @@ void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out3 ^= x21;
 }
 
-void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x5A5A5A5A, x0F0FF0F0;
     u32 x33FF33FF, x33FFCC00, x0C0030F0, x0C0CC0C0, x0CF3C03F, x5EFBDA7F,
@@ -611,7 +611,7 @@ void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x77777777, x77770000, x22225555, x11116666, x1F1F6F6F;
     u32 x70700000, x43433333, x00430033, x55557777, x55167744, x5A19784B;
@@ -684,7 +684,7 @@ void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out2 ^= x11;
 }
 
-void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x33CC33CC;
     u32 x3333FFFF, x11115555, x22DD6699, x22DD9966, x00220099;
@@ -757,7 +757,7 @@ void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out1 ^= x01;
 }
 
-void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x0FF00FF0, x3CC33CC3, x00003CC3, x0F000F00, x5A555A55, x00001841;
     u32 x00000F00, x33333C33, x7B777E77, x0FF0F00F, x74878E78;
@@ -828,7 +828,7 @@ void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out2 ^= x11;
 }
 
-void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x0C0C0C0C, x0000F0F0, x00FFF00F, x00555005, x00515001;
     u32 x33000330, x77555775, x30303030, x3030CFCF, x30104745, x30555745;
@@ -921,7 +921,7 @@ void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
  * The effort has been sponsored by Rapid7: http://www.rapid7.com
  */
 
-void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x55005500, x5A0F5A0F, x3333FFFF, x66666666, x22226666, x2D2D6969,
         x25202160;
@@ -996,7 +996,7 @@ void s1 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x33CC33CC;
     u32 x55550000, x00AA00FF, x33BB33FF;
@@ -1067,7 +1067,7 @@ void s2 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x44444444, x0F0FF0F0, x4F4FF4F4, x00FFFF00, x00AAAA00, x4FE55EF4;
     u32 x3C3CC3C3, x3C3C0000, x7373F4F4, x0C840A00;
@@ -1138,7 +1138,7 @@ void s3 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out3 ^= x21;
 }
 
-void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x5A5A5A5A, x0F0FF0F0;
     u32 x33FF33FF, x33FFCC00, x0C0030F0, x0C0CC0C0, x0CF3C03F, x5EFBDA7F,
@@ -1192,7 +1192,7 @@ void s4 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out4 ^= x31;
 }
 
-void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x77777777, x77770000, x22225555, x11116666, x1F1F6F6F;
     u32 x70700000, x43433333, x00430033, x55557777, x55167744, x5A19784B;
@@ -1265,7 +1265,7 @@ void s5 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out2 ^= x11;
 }
 
-void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x33CC33CC;
     u32 x3333FFFF, x11115555, x22DD6699, x22DD9966, x00220099;
@@ -1338,7 +1338,7 @@ void s6 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out1 ^= x01;
 }
 
-void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x0FF00FF0, x3CC33CC3, x00003CC3, x0F000F00, x5A555A55, x00001841;
     u32 x00000F00, x33333C33, x7B777E77, x0FF0F00F, x74878E78;
@@ -1409,7 +1409,7 @@ void s7 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
     *out2 ^= x11;
 }
 
-void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
+DECLSPEC void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, const u32 a6, u32 *out1, u32 *out2, u32 *out3, u32 *out4)
 {
     u32 x0C0C0C0C, x0000F0F0, x00FFF00F, x00555005, x00515001;
     u32 x33000330, x77555775, x30303030, x3030CFCF, x30104745, x30555745;
@@ -1528,7 +1528,7 @@ void s8 (const u32 a1, const u32 a2, const u32 a3, const u32 a4, const u32 a5, c
 #define KEYSET07 { k00 = K31; k01 = K35; k02 = K52; k03 = K43; k04 = K08; k05 = K37; k06 = K51; k07 = K15; k08 = K49; k09 = K30; k10 = K07; k11 = K02; k12 = K50; k13 = K21; k14 = K45; k15 = K44; k16 = K29; k17 = K16; k18 = K42; k19 = K23; k20 = K22; k21 = K14; k22 = K38; k23 = K01; k24 = K10; k25 = K47; k26 = K53; k27 = K11; k28 = K27; k29 = K26; k30 = K05; k31 = K17; k32 = K54; k33 = K41; k34 = K39; k35 = K20; k36 = K48; k37 = K13; k38 = K24; k39 = K19; k40 = K32; k41 = K40; k42 = K34; k43 = K03; k44 = K06; k45 = K18; k46 = K12; k47 = K46; }
 #define KEYSET17 { k00 = K15; k01 = K51; k02 = K36; k03 = K02; k04 = K49; k05 = K21; k06 = K35; k07 = K31; k08 = K08; k09 = K14; k10 = K23; k11 = K43; k12 = K09; k13 = K37; k14 = K29; k15 = K28; k16 = K45; k17 = K00; k18 = K01; k19 = K07; k20 = K38; k21 = K30; k22 = K22; k23 = K42; k24 = K26; k25 = K04; k26 = K41; k27 = K54; k28 = K39; k29 = K10; k30 = K48; k31 = K33; k32 = K11; k33 = K53; k34 = K27; k35 = K32; k36 = K05; k37 = K25; k38 = K40; k39 = K03; k40 = K20; k41 = K24; k42 = K46; k43 = K19; k44 = K18; k45 = K06; k46 = K55; k47 = K34; }
 
-void DES (const u32 K00, const u32 K01, const u32 K02, const u32 K03, const u32 K04, const u32 K05, const u32 K06, const u32 K07, const u32 K08, const u32 K09, const u32 K10, const u32 K11, const u32 K12, const u32 K13, const u32 K14, const u32 K15, const u32 K16, const u32 K17, const u32 K18, const u32 K19, const u32 K20, const u32 K21, const u32 K22, const u32 K23, const u32 K24, const u32 K25, const u32 K26, const u32 K27, const u32 K28, const u32 K29, const u32 K30, const u32 K31, const u32 K32, const u32 K33, const u32 K34, const u32 K35, const u32 K36, const u32 K37, const u32 K38, const u32 K39, const u32 K40, const u32 K41, const u32 K42, const u32 K43, const u32 K44, const u32 K45, const u32 K46, const u32 K47, const u32 K48, const u32 K49, const u32 K50, const u32 K51, const u32 K52, const u32 K53, const u32 K54, const u32 K55, u32 *D00, u32 *D01, u32 *D02, u32 *D03, u32 *D04, u32 *D05, u32 *D06, u32 *D07, u32 *D08, u32 *D09, u32 *D10, u32 *D11, u32 *D12, u32 *D13, u32 *D14, u32 *D15, u32 *D16, u32 *D17, u32 *D18, u32 *D19, u32 *D20, u32 *D21, u32 *D22, u32 *D23, u32 *D24, u32 *D25, u32 *D26, u32 *D27, u32 *D28, u32 *D29, u32 *D30, u32 *D31, u32 *D32, u32 *D33, u32 *D34, u32 *D35, u32 *D36, u32 *D37, u32 *D38, u32 *D39, u32 *D40, u32 *D41, u32 *D42, u32 *D43, u32 *D44, u32 *D45, u32 *D46, u32 *D47, u32 *D48, u32 *D49, u32 *D50, u32 *D51, u32 *D52, u32 *D53, u32 *D54, u32 *D55, u32 *D56, u32 *D57, u32 *D58, u32 *D59, u32 *D60, u32 *D61, u32 *D62, u32 *D63)
+DECLSPEC void DES (const u32 K00, const u32 K01, const u32 K02, const u32 K03, const u32 K04, const u32 K05, const u32 K06, const u32 K07, const u32 K08, const u32 K09, const u32 K10, const u32 K11, const u32 K12, const u32 K13, const u32 K14, const u32 K15, const u32 K16, const u32 K17, const u32 K18, const u32 K19, const u32 K20, const u32 K21, const u32 K22, const u32 K23, const u32 K24, const u32 K25, const u32 K26, const u32 K27, const u32 K28, const u32 K29, const u32 K30, const u32 K31, const u32 K32, const u32 K33, const u32 K34, const u32 K35, const u32 K36, const u32 K37, const u32 K38, const u32 K39, const u32 K40, const u32 K41, const u32 K42, const u32 K43, const u32 K44, const u32 K45, const u32 K46, const u32 K47, const u32 K48, const u32 K49, const u32 K50, const u32 K51, const u32 K52, const u32 K53, const u32 K54, const u32 K55, u32 *D00, u32 *D01, u32 *D02, u32 *D03, u32 *D04, u32 *D05, u32 *D06, u32 *D07, u32 *D08, u32 *D09, u32 *D10, u32 *D11, u32 *D12, u32 *D13, u32 *D14, u32 *D15, u32 *D16, u32 *D17, u32 *D18, u32 *D19, u32 *D20, u32 *D21, u32 *D22, u32 *D23, u32 *D24, u32 *D25, u32 *D26, u32 *D27, u32 *D28, u32 *D29, u32 *D30, u32 *D31, u32 *D32, u32 *D33, u32 *D34, u32 *D35, u32 *D36, u32 *D37, u32 *D38, u32 *D39, u32 *D40, u32 *D41, u32 *D42, u32 *D43, u32 *D44, u32 *D45, u32 *D46, u32 *D47, u32 *D48, u32 *D49, u32 *D50, u32 *D51, u32 *D52, u32 *D53, u32 *D54, u32 *D55, u32 *D56, u32 *D57, u32 *D58, u32 *D59, u32 *D60, u32 *D61, u32 *D62, u32 *D63)
 {
   KXX_DECL u32 k00, k01, k02, k03, k04, k05;
   KXX_DECL u32 k06, k07, k08, k09, k10, k11;
@@ -1634,7 +1634,7 @@ void DES (const u32 K00, const u32 K01, const u32 K02, const u32 K03, const u32
   }
 }
 
-void transpose32c (u32 data[32])
+DECLSPEC void transpose32c (u32 data[32])
 {
   #define swap(x,y,j,m)               \
      t  = ((x) ^ ((y) >> (j))) & (m); \
@@ -1725,7 +1725,7 @@ void transpose32c (u32 data[32])
   swap (data[30], data[31],  1, 0x55555555);
 }
 
-void m14000m (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bs_word_t *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m14000m (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bs_word_t *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * base
@@ -2233,7 +2233,7 @@ void m14000m (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __glo
   }
 }
 
-void m14000s (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bs_word_t *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m14000s (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bs_word_t *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * base
diff --git a/OpenCL/m14100_a0.cl b/OpenCL/m14100_a0.cl
index 7244acdef..3463e1c02 100644
--- a/OpenCL/m14100_a0.cl
+++ b/OpenCL/m14100_a0.cl
@@ -380,7 +380,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX1(i,S) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32 r = rotl32_S (data[0], 3u);
   u32 l = rotl32_S (data[1], 3u);
@@ -424,7 +424,7 @@ void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local
   iv[1] = rotl32_S (r, 29u);
 }
 
-void _des_crypt_decrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_decrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32 r = rotl32_S (data[0], 3u);
   u32 l = rotl32_S (data[1], 3u);
@@ -468,7 +468,7 @@ void _des_crypt_decrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local
   iv[1] = rotl32_S (r, 29u);
 }
 
-void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
diff --git a/OpenCL/m14100_a1.cl b/OpenCL/m14100_a1.cl
index 1b5811c0b..863aec881 100644
--- a/OpenCL/m14100_a1.cl
+++ b/OpenCL/m14100_a1.cl
@@ -370,7 +370,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX1(i,S) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32 r = rotl32_S (data[0], 3u);
   u32 l = rotl32_S (data[1], 3u);
@@ -414,7 +414,7 @@ void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local
   iv[1] = rotl32_S (r, 29u);
 }
 
-void _des_crypt_decrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_decrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32 r = rotl32_S (data[0], 3u);
   u32 l = rotl32_S (data[1], 3u);
@@ -458,7 +458,7 @@ void _des_crypt_decrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local
   iv[1] = rotl32_S (r, 29u);
 }
 
-void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
diff --git a/OpenCL/m14100_a3.cl b/OpenCL/m14100_a3.cl
index ea0a09504..e36588864 100644
--- a/OpenCL/m14100_a3.cl
+++ b/OpenCL/m14100_a3.cl
@@ -370,7 +370,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX1(i,S) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
 #endif
 
-void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32x r = rotl32 (data[0], 3u);
   u32x l = rotl32 (data[1], 3u);
@@ -414,7 +414,7 @@ void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __l
   iv[1] = rotl32 (r, 29u);
 }
 
-void _des_crypt_decrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_decrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32x r = rotl32 (data[0], 3u);
   u32x l = rotl32 (data[1], 3u);
@@ -458,7 +458,7 @@ void _des_crypt_decrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __l
   iv[1] = rotl32 (r, 29u);
 }
 
-void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32x tt;
 
@@ -530,7 +530,7 @@ void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32
   }
 }
 
-void m14100m (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m14100m (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -615,7 +615,7 @@ void m14100m (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16],
   }
 }
 
-void m14100s (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m14100s (__local u32 (*s_SPtrans)[64], __local u32 (*s_skb)[64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m14400_a0-optimized.cl b/OpenCL/m14400_a0-optimized.cl
index 5d3968fcc..0a00e5b7b 100644
--- a/OpenCL/m14400_a0-optimized.cl
+++ b/OpenCL/m14400_a0-optimized.cl
@@ -25,7 +25,7 @@
 #define uint_to_hex_lower8_le(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void append_4 (const u32 offset, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 src_r0)
+DECLSPEC void append_4 (const u32 offset, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 src_r0)
 {
   u32 tmp[2];
 
@@ -97,7 +97,7 @@ void append_4 (const u32 offset, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], con
   }
 }
 
-void shift_2 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4])
+DECLSPEC void shift_2 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4])
 {
   w3[3] = w3[2] >> 16 | w3[3] << 16;
   w3[2] = w3[1] >> 16 | w3[2] << 16;
diff --git a/OpenCL/m14400_a1-optimized.cl b/OpenCL/m14400_a1-optimized.cl
index 8b075b26c..57b35d5df 100644
--- a/OpenCL/m14400_a1-optimized.cl
+++ b/OpenCL/m14400_a1-optimized.cl
@@ -25,7 +25,7 @@
 #define uint_to_hex_lower8_le(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void append_4 (const u32 offset, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 src_r0)
+DECLSPEC void append_4 (const u32 offset, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 src_r0)
 {
   u32 tmp[2];
 
@@ -97,7 +97,7 @@ void append_4 (const u32 offset, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], con
   }
 }
 
-void shift_2 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4])
+DECLSPEC void shift_2 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4])
 {
   w3[3] = w3[2] >> 16 | w3[3] << 16;
   w3[2] = w3[1] >> 16 | w3[2] << 16;
diff --git a/OpenCL/m14400_a3-optimized.cl b/OpenCL/m14400_a3-optimized.cl
index 02db469b7..b52061d2b 100644
--- a/OpenCL/m14400_a3-optimized.cl
+++ b/OpenCL/m14400_a3-optimized.cl
@@ -25,7 +25,7 @@
 #define uint_to_hex_lower8_le(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
 #endif
 
-void append_4 (const u32 offset, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 src_r0)
+DECLSPEC void append_4 (const u32 offset, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 src_r0)
 {
   u32 tmp[2];
 
@@ -97,7 +97,7 @@ void append_4 (const u32 offset, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], con
   }
 }
 
-void shift_2 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4])
+DECLSPEC void shift_2 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4])
 {
   w3[3] = w3[2] >> 16 | w3[3] << 16;
   w3[2] = w3[1] >> 16 | w3[2] << 16;
@@ -117,7 +117,7 @@ void shift_2 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4])
   w0[0] =           0 | w0[0] << 16;
 }
 
-void m14400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m14400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
@@ -366,7 +366,7 @@ void m14400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m14400s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
+DECLSPEC void m14400s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 *l_bin2asc)
 {
   /**
    * modifier
diff --git a/OpenCL/m14611.cl b/OpenCL/m14611.cl
index d46c6b4ba..477e284d5 100644
--- a/OpenCL/m14611.cl
+++ b/OpenCL/m14611.cl
@@ -25,7 +25,7 @@
 
 #define MAX_ENTROPY 7.0
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m14612.cl b/OpenCL/m14612.cl
index 992429ca2..5d1357a8e 100644
--- a/OpenCL/m14612.cl
+++ b/OpenCL/m14612.cl
@@ -25,7 +25,7 @@
 
 #define MAX_ENTROPY 7.0
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m14613.cl b/OpenCL/m14613.cl
index d328e3cb6..8e08e44a1 100644
--- a/OpenCL/m14613.cl
+++ b/OpenCL/m14613.cl
@@ -25,7 +25,7 @@
 
 #define MAX_ENTROPY 7.0
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m14621.cl b/OpenCL/m14621.cl
index 622945efd..589fdfaa3 100644
--- a/OpenCL/m14621.cl
+++ b/OpenCL/m14621.cl
@@ -25,7 +25,7 @@
 
 #define MAX_ENTROPY 7.0
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m14622.cl b/OpenCL/m14622.cl
index a66ec0e93..0080d18a8 100644
--- a/OpenCL/m14622.cl
+++ b/OpenCL/m14622.cl
@@ -25,7 +25,7 @@
 
 #define MAX_ENTROPY 7.0
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m14623.cl b/OpenCL/m14623.cl
index 79937b817..e8f14a881 100644
--- a/OpenCL/m14623.cl
+++ b/OpenCL/m14623.cl
@@ -25,7 +25,7 @@
 
 #define MAX_ENTROPY 7.0
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m14631.cl b/OpenCL/m14631.cl
index 0d4a7f84f..ea0b1142c 100644
--- a/OpenCL/m14631.cl
+++ b/OpenCL/m14631.cl
@@ -25,7 +25,7 @@
 
 #define MAX_ENTROPY 7.0
 
-void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m14632.cl b/OpenCL/m14632.cl
index 86f273ee9..90918a055 100644
--- a/OpenCL/m14632.cl
+++ b/OpenCL/m14632.cl
@@ -25,7 +25,7 @@
 
 #define MAX_ENTROPY 7.0
 
-void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m14633.cl b/OpenCL/m14633.cl
index ae2e98a85..558820641 100644
--- a/OpenCL/m14633.cl
+++ b/OpenCL/m14633.cl
@@ -25,7 +25,7 @@
 
 #define MAX_ENTROPY 7.0
 
-void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m14641.cl b/OpenCL/m14641.cl
index 240c1fb43..299ffb986 100644
--- a/OpenCL/m14641.cl
+++ b/OpenCL/m14641.cl
@@ -25,7 +25,7 @@
 
 #define MAX_ENTROPY 7.0
 
-void hmac_ripemd160_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_ripemd160_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m14642.cl b/OpenCL/m14642.cl
index 5bad9397d..bf54af928 100644
--- a/OpenCL/m14642.cl
+++ b/OpenCL/m14642.cl
@@ -25,7 +25,7 @@
 
 #define MAX_ENTROPY 7.0
 
-void hmac_ripemd160_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_ripemd160_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m14643.cl b/OpenCL/m14643.cl
index 0bb93ccee..dd7b023d0 100644
--- a/OpenCL/m14643.cl
+++ b/OpenCL/m14643.cl
@@ -25,7 +25,7 @@
 
 #define MAX_ENTROPY 7.0
 
-void hmac_ripemd160_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_ripemd160_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m14700.cl b/OpenCL/m14700.cl
index d764b707c..ea59252ff 100644
--- a/OpenCL/m14700.cl
+++ b/OpenCL/m14700.cl
@@ -17,7 +17,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m14800.cl b/OpenCL/m14800.cl
index f3bca9b66..f6316165d 100644
--- a/OpenCL/m14800.cl
+++ b/OpenCL/m14800.cl
@@ -15,7 +15,7 @@
 #include "inc_hash_sha256.cl"
 #include "inc_cipher_aes.cl"
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -51,7 +51,7 @@ void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[
   sha1_transform_vector (w0, w1, w2, w3, digest);
 }
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m14900_a0-optimized.cl b/OpenCL/m14900_a0-optimized.cl
index a1e4fd029..7450635eb 100644
--- a/OpenCL/m14900_a0-optimized.cl
+++ b/OpenCL/m14900_a0-optimized.cl
@@ -51,7 +51,7 @@ __constant static u8a c_ftable[256] =
   0xbd, 0xa8, 0x3a, 0x01, 0x05, 0x59, 0x2a, 0x46
 };
 
-void g (__local u8 *s_ftable, const u32 *key, const int k, const u32 *wx, u32 *out)
+DECLSPEC void g (__local u8 *s_ftable, const u32 *key, const int k, const u32 *wx, u32 *out)
 {
   const u32 g1 = wx[1];
   const u32 g2 = wx[0];
@@ -64,7 +64,7 @@ void g (__local u8 *s_ftable, const u32 *key, const int k, const u32 *wx, u32 *o
   out[1] = g5;
 }
 
-u32 skip32 (__local u8 *s_ftable, const u32 KP, const u32 *key)
+DECLSPEC u32 skip32 (__local u8 *s_ftable, const u32 KP, const u32 *key)
 {
   u32 wl[2];
   u32 wr[2];
diff --git a/OpenCL/m14900_a1-optimized.cl b/OpenCL/m14900_a1-optimized.cl
index c37e1155a..8482d99a5 100644
--- a/OpenCL/m14900_a1-optimized.cl
+++ b/OpenCL/m14900_a1-optimized.cl
@@ -49,7 +49,7 @@ __constant static u8a c_ftable[256] =
   0xbd, 0xa8, 0x3a, 0x01, 0x05, 0x59, 0x2a, 0x46
 };
 
-void g (__local u8 *s_ftable, const u32 *key, const int k, const u32 *wx, u32 *out)
+DECLSPEC void g (__local u8 *s_ftable, const u32 *key, const int k, const u32 *wx, u32 *out)
 {
   const u32 g1 = wx[1];
   const u32 g2 = wx[0];
@@ -62,7 +62,7 @@ void g (__local u8 *s_ftable, const u32 *key, const int k, const u32 *wx, u32 *o
   out[1] = g5;
 }
 
-u32 skip32 (__local u8 *s_ftable, const u32 KP, const u32 *key)
+DECLSPEC u32 skip32 (__local u8 *s_ftable, const u32 KP, const u32 *key)
 {
   u32 wl[2];
   u32 wr[2];
diff --git a/OpenCL/m14900_a3-optimized.cl b/OpenCL/m14900_a3-optimized.cl
index 56beb4e49..a0187975a 100644
--- a/OpenCL/m14900_a3-optimized.cl
+++ b/OpenCL/m14900_a3-optimized.cl
@@ -49,7 +49,7 @@ __constant static u8a c_ftable[256] =
   0xbd, 0xa8, 0x3a, 0x01, 0x05, 0x59, 0x2a, 0x46
 };
 
-void g (__local u8 *s_ftable, const u32 *key, const int k, const u32 *wx, u32 *out)
+DECLSPEC void g (__local u8 *s_ftable, const u32 *key, const int k, const u32 *wx, u32 *out)
 {
   const u32 g1 = wx[1];
   const u32 g2 = wx[0];
@@ -62,7 +62,7 @@ void g (__local u8 *s_ftable, const u32 *key, const int k, const u32 *wx, u32 *o
   out[1] = g5;
 }
 
-u32 skip32 (__local u8 *s_ftable, const u32 KP, const u32 *key)
+DECLSPEC u32 skip32 (__local u8 *s_ftable, const u32 KP, const u32 *key)
 {
   u32 wl[2];
   u32 wr[2];
@@ -102,7 +102,7 @@ u32 skip32 (__local u8 *s_ftable, const u32 KP, const u32 *key)
   return r;
 }
 
-void m14900m (__local u8 *s_ftable, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m14900m (__local u8 *s_ftable, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -150,7 +150,7 @@ void m14900m (__local u8 *s_ftable, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4],
   }
 }
 
-void m14900s (__local u8 *s_ftable, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m14900s (__local u8 *s_ftable, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global oldoffice01_t *oldoffice01_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m15000_a0-optimized.cl b/OpenCL/m15000_a0-optimized.cl
index ebb39bd78..9578b4271 100644
--- a/OpenCL/m15000_a0-optimized.cl
+++ b/OpenCL/m15000_a0-optimized.cl
@@ -38,7 +38,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x w4[4], const u32x w5[4], const u32x w6[4], const u32x w7[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x w4[4], const u32x w5[4], const u32x w6[4], const u32x w7[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
diff --git a/OpenCL/m15000_a1-optimized.cl b/OpenCL/m15000_a1-optimized.cl
index de2394049..3f08960fb 100644
--- a/OpenCL/m15000_a1-optimized.cl
+++ b/OpenCL/m15000_a1-optimized.cl
@@ -36,7 +36,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x w4[4], const u32x w5[4], const u32x w6[4], const u32x w7[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x w4[4], const u32x w5[4], const u32x w6[4], const u32x w7[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
diff --git a/OpenCL/m15000_a3-optimized.cl b/OpenCL/m15000_a3-optimized.cl
index bda2c0735..87616d213 100644
--- a/OpenCL/m15000_a3-optimized.cl
+++ b/OpenCL/m15000_a3-optimized.cl
@@ -36,7 +36,7 @@ __constant static u64a k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x w4[4], const u32x w5[4], const u32x w6[4], const u32x w7[4], u64x digest[8])
+DECLSPEC void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], const u32x w4[4], const u32x w5[4], const u32x w6[4], const u32x w7[4], u64x digest[8])
 {
   u64x w0_t = hl32_to_64 (w0[0], w0[1]);
   u64x w1_t = hl32_to_64 (w0[2], w0[3]);
@@ -135,7 +135,7 @@ void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], con
   digest[7] = h;
 }
 
-void m15000m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m15000m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -296,7 +296,7 @@ void m15000m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m15000s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m15000s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m15100.cl b/OpenCL/m15100.cl
index b1d072efd..35a4077f6 100644
--- a/OpenCL/m15100.cl
+++ b/OpenCL/m15100.cl
@@ -16,7 +16,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m15300.cl b/OpenCL/m15300.cl
index ed7f51788..13f7b0b46 100644
--- a/OpenCL/m15300.cl
+++ b/OpenCL/m15300.cl
@@ -362,7 +362,7 @@ __constant static u32a c_skb[8][64] =
 #define BOX(i,n,S) (S)[(n)][(i)]
 #define BOX1(i,S) (S)[(i)]
 
-void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32 tt;
 
@@ -416,7 +416,7 @@ void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local
   iv[1] = r;
 }
 
-void _des_crypt_decrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_decrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32 tt;
 
@@ -470,7 +470,7 @@ void _des_crypt_decrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local
   iv[1] = r;
 }
 
-void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
@@ -542,7 +542,7 @@ void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 (*s_
   }
 }
 
-void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
+DECLSPEC void hmac_sha1_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m15400_a0-optimized.cl b/OpenCL/m15400_a0-optimized.cl
index 093b4af49..f4c3b2a4a 100644
--- a/OpenCL/m15400_a0-optimized.cl
+++ b/OpenCL/m15400_a0-optimized.cl
@@ -31,7 +31,7 @@
     x[b] = rotl32(x[b] ^ x[c], 7);    \
   } while (0);
 
-void chacha20_transform (const u32x w0[4], const u32x w1[4], const u32 position[2], const u32 offset, const u32 iv[2], const u32 plain[4], u32x digest[4])
+DECLSPEC void chacha20_transform (const u32x w0[4], const u32x w1[4], const u32 position[2], const u32 offset, const u32 iv[2], const u32 plain[4], u32x digest[4])
 {
   /**
    * Key expansion
diff --git a/OpenCL/m15400_a1-optimized.cl b/OpenCL/m15400_a1-optimized.cl
index 85b29a6cd..68b75bc5c 100644
--- a/OpenCL/m15400_a1-optimized.cl
+++ b/OpenCL/m15400_a1-optimized.cl
@@ -31,7 +31,7 @@
     x[b] = rotl32(x[b] ^ x[c], 7);    \
   } while (0);
 
-void chacha20_transform (const u32x w0[4], const u32x w1[4], const u32 position[2], const u32 offset, const u32 iv[2], const u32 plain[4], u32x digest[4])
+DECLSPEC void chacha20_transform (const u32x w0[4], const u32x w1[4], const u32 position[2], const u32 offset, const u32 iv[2], const u32 plain[4], u32x digest[4])
 {
   /**
    * Key expansion
diff --git a/OpenCL/m15400_a3-optimized.cl b/OpenCL/m15400_a3-optimized.cl
index f65ca0ecb..8a1f883f3 100644
--- a/OpenCL/m15400_a3-optimized.cl
+++ b/OpenCL/m15400_a3-optimized.cl
@@ -29,7 +29,7 @@
     x[b] = rotl32(x[b] ^ x[c], 7);    \
   } while (0);
 
-void chacha20_transform (const u32x w0[4], const u32x w1[4], const u32 position[2], const u32 offset, const u32 iv[2], const u32 plain[4], u32x digest[4])
+DECLSPEC void chacha20_transform (const u32x w0[4], const u32x w1[4], const u32 position[2], const u32 offset, const u32 iv[2], const u32 plain[4], u32x digest[4])
 {
   /**
    * Key expansion
diff --git a/OpenCL/m15500_a3-optimized.cl b/OpenCL/m15500_a3-optimized.cl
index e8958c3ba..bdfd5c061 100644
--- a/OpenCL/m15500_a3-optimized.cl
+++ b/OpenCL/m15500_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m15500m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m15500m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -316,7 +316,7 @@ void m15500m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m15500s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m15500s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m15600.cl b/OpenCL/m15600.cl
index 8a23bd17e..5019f0bc7 100644
--- a/OpenCL/m15600.cl
+++ b/OpenCL/m15600.cl
@@ -66,7 +66,7 @@ __constant static u64a keccakf_rndc[24] =
   st[4 + s] ^= ~bc0 & bc1;      \
 }
 
-void keccak_transform_S (u64 st[25])
+DECLSPEC void keccak_transform_S (u64 st[25])
 {
   const u8 keccakf_rotc[24] =
   {
@@ -147,7 +147,7 @@ void keccak_transform_S (u64 st[25])
   }
 }
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m15700.cl b/OpenCL/m15700.cl
index 35f3729d2..3b7cf527f 100644
--- a/OpenCL/m15700.cl
+++ b/OpenCL/m15700.cl
@@ -13,7 +13,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-uint4 swap32_4 (uint4 v)
+DECLSPEC uint4 swap32_4 (uint4 v)
 {
   return (rotate ((v & 0x00FF00FF), 24u) | rotate ((v & 0xFF00FF00),  8u));
 }
@@ -73,7 +73,7 @@ uint4 swap32_4 (uint4 v)
   R3 = R3 + X3;         \
 }
 
-void salsa_r (uint4 *TI)
+DECLSPEC void salsa_r (uint4 *TI)
 {
   uint4 R0 = TI[STATE_CNT4 - 4];
   uint4 R1 = TI[STATE_CNT4 - 3];
@@ -125,7 +125,7 @@ void salsa_r (uint4 *TI)
   }
 }
 
-void scrypt_smix (uint4 *X, uint4 *T, __global uint4 *V0, __global uint4 *V1, __global uint4 *V2, __global uint4 *V3)
+DECLSPEC void scrypt_smix (uint4 *X, uint4 *T, __global uint4 *V0, __global uint4 *V1, __global uint4 *V2, __global uint4 *V3)
 {
   #define Coord(xd4,y,z) (((xd4) * ySIZE * zSIZE) + ((y) * zSIZE) + (z))
   #define CO Coord(xd4,y,z)
@@ -255,7 +255,7 @@ __constant static u64a keccakf_rndc[24] =
   0x8000000000008080, 0x0000000080000001, 0x8000000080008008
 };
 
-void keccak_transform_S (u64 st[25])
+DECLSPEC void keccak_transform_S (u64 st[25])
 {
   const u8 keccakf_rotc[24] =
   {
diff --git a/OpenCL/m15900.cl b/OpenCL/m15900.cl
index 603a62122..14576f739 100644
--- a/OpenCL/m15900.cl
+++ b/OpenCL/m15900.cl
@@ -19,7 +19,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
+DECLSPEC void hmac_sha512_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], u64x ipad[8], u64x opad[8], u64x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m16000_a0.cl b/OpenCL/m16000_a0.cl
index 65972017e..af8a439d5 100644
--- a/OpenCL/m16000_a0.cl
+++ b/OpenCL/m16000_a0.cl
@@ -354,7 +354,7 @@ __constant static u32a c_tripcode_salt[128] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
@@ -423,7 +423,7 @@ void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (
   }
 }
 
-void _des_crypt_encrypt (u32 iv[2], u32 mask, u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32 mask, u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   const u32 E1 = (mask >> 2) & 0x3f0;
   const u32 E0 = mask & 0x3f;
diff --git a/OpenCL/m16000_a1.cl b/OpenCL/m16000_a1.cl
index a60cc3179..4ae38712b 100644
--- a/OpenCL/m16000_a1.cl
+++ b/OpenCL/m16000_a1.cl
@@ -352,7 +352,7 @@ __constant static u32a c_tripcode_salt[128] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
@@ -421,7 +421,7 @@ void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (
   }
 }
 
-void _des_crypt_encrypt (u32 iv[2], u32 mask, u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32 mask, u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   const u32 E1 = (mask >> 2) & 0x3f0;
   const u32 E0 = mask & 0x3f;
diff --git a/OpenCL/m16000_a3.cl b/OpenCL/m16000_a3.cl
index bbfb6e8fa..118d439a5 100644
--- a/OpenCL/m16000_a3.cl
+++ b/OpenCL/m16000_a3.cl
@@ -352,7 +352,7 @@ __constant static u32a c_tripcode_salt[128] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+DECLSPEC void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32 tt;
 
@@ -421,7 +421,7 @@ void _des_crypt_keysetup (u32 c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (
   }
 }
 
-void _des_crypt_encrypt (u32 iv[2], u32 mask, u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+DECLSPEC void _des_crypt_encrypt (u32 iv[2], u32 mask, u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   const u32 E1 = (mask >> 2) & 0x3f0;
   const u32 E0 = mask & 0x3f;
diff --git a/OpenCL/m16100_a3-optimized.cl b/OpenCL/m16100_a3-optimized.cl
index 148a91be8..49e24217e 100644
--- a/OpenCL/m16100_a3-optimized.cl
+++ b/OpenCL/m16100_a3-optimized.cl
@@ -12,7 +12,7 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-void m16100m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const tacacs_plus_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m16100m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const tacacs_plus_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -265,7 +265,7 @@ void m16100m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
   }
 }
 
-void m16100s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const tacacs_plus_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m16100s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const tacacs_plus_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m16200.cl b/OpenCL/m16200.cl
index dd5614942..3c89cc003 100644
--- a/OpenCL/m16200.cl
+++ b/OpenCL/m16200.cl
@@ -17,7 +17,7 @@
 #define COMPARE_S "inc_comp_single.cl"
 #define COMPARE_M "inc_comp_multi.cl"
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m16300.cl b/OpenCL/m16300.cl
index c0a5de79a..9022882d8 100644
--- a/OpenCL/m16300.cl
+++ b/OpenCL/m16300.cl
@@ -67,7 +67,7 @@ __constant static u64a keccakf_rndc[24] =
   st[4 + s] ^= ~bc0 & bc1;      \
 }
 
-void keccak_transform_S (u64 st[25])
+DECLSPEC void keccak_transform_S (u64 st[25])
 {
   const u8 keccakf_rotc[24] =
   {
@@ -148,7 +148,7 @@ void keccak_transform_S (u64 st[25])
   }
 }
 
-void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
+DECLSPEC void hmac_sha256_run_V (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[8], u32x opad[8], u32x digest[8])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
diff --git a/OpenCL/m16400_a0.cl b/OpenCL/m16400_a0.cl
index ede2fc167..056c16780 100644
--- a/OpenCL/m16400_a0.cl
+++ b/OpenCL/m16400_a0.cl
@@ -15,7 +15,7 @@
 #include "inc_scalar.cl"
 #include "inc_hash_md5.cl"
 
-void cram_md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
+DECLSPEC void cram_md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
 {
   u32 a = digest[0];
   u32 b = digest[1];
@@ -115,7 +115,7 @@ void cram_md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], cons
   digest[3] += d;
 }
 
-void cram_md5_update_64 (md5_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void cram_md5_update_64 (md5_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -145,7 +145,7 @@ void cram_md5_update_64 (md5_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3
   ctx->w3[3] |= w3[3];
 }
 
-void cram_md5_update (md5_ctx_t *ctx, const u32 *w, const int len)
+DECLSPEC void cram_md5_update (md5_ctx_t *ctx, const u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -172,7 +172,7 @@ void cram_md5_update (md5_ctx_t *ctx, const u32 *w, const int len)
   cram_md5_update_64 (ctx, w0, w1, w2, w3, len);
 }
 
-void cram_md5_final (md5_ctx_t *ctx)
+DECLSPEC void cram_md5_final (md5_ctx_t *ctx)
 {
   cram_md5_transform (ctx->w0, ctx->w1, ctx->w2, ctx->w3, ctx->h);
 }
diff --git a/OpenCL/m16400_a1.cl b/OpenCL/m16400_a1.cl
index b25f4cd60..8533a061c 100644
--- a/OpenCL/m16400_a1.cl
+++ b/OpenCL/m16400_a1.cl
@@ -13,7 +13,7 @@
 #include "inc_scalar.cl"
 #include "inc_hash_md5.cl"
 
-void cram_md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
+DECLSPEC void cram_md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
 {
   u32 a = digest[0];
   u32 b = digest[1];
@@ -113,7 +113,7 @@ void cram_md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], cons
   digest[3] += d;
 }
 
-void cram_md5_update_64 (md5_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
+DECLSPEC void cram_md5_update_64 (md5_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -143,7 +143,7 @@ void cram_md5_update_64 (md5_ctx_t *ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3
   ctx->w3[3] |= w3[3];
 }
 
-void cram_md5_update_global (md5_ctx_t *ctx, const __global u32 *w, const int len)
+DECLSPEC void cram_md5_update_global (md5_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
   u32 w1[4];
@@ -170,7 +170,7 @@ void cram_md5_update_global (md5_ctx_t *ctx, const __global u32 *w, const int le
   cram_md5_update_64 (ctx, w0, w1, w2, w3, len);
 }
 
-void cram_md5_final (md5_ctx_t *ctx)
+DECLSPEC void cram_md5_final (md5_ctx_t *ctx)
 {
   cram_md5_transform (ctx->w0, ctx->w1, ctx->w2, ctx->w3, ctx->h);
 }
diff --git a/OpenCL/m16400_a3-optimized.cl b/OpenCL/m16400_a3-optimized.cl
index e307908b8..bd5bdc725 100644
--- a/OpenCL/m16400_a3-optimized.cl
+++ b/OpenCL/m16400_a3-optimized.cl
@@ -30,7 +30,7 @@
   a -= t;                               \
 }
 
-void m16400m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m16400m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -222,7 +222,7 @@ void m16400m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const ke
   }
 }
 
-void m16400s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m16400s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
diff --git a/OpenCL/m16400_a3.cl b/OpenCL/m16400_a3.cl
index 0cc3a6c3d..26e731c8a 100644
--- a/OpenCL/m16400_a3.cl
+++ b/OpenCL/m16400_a3.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md5.cl"
 
-void cram_md5_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
+DECLSPEC void cram_md5_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
 {
   u32x a = digest[0];
   u32x b = digest[1];
@@ -113,7 +113,7 @@ void cram_md5_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w
   digest[3] += d;
 }
 
-void cram_md5_update_vector_64 (md5_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
+DECLSPEC void cram_md5_update_vector_64 (md5_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len)
 {
   #ifdef IS_AMD
   volatile const int pos = ctx->len & 63;
@@ -143,7 +143,7 @@ void cram_md5_update_vector_64 (md5_ctx_vector_t *ctx, u32x w0[4], u32x w1[4], u
   ctx->w3[3] |= w3[3];
 }
 
-void cram_md5_update_vector (md5_ctx_vector_t *ctx, const u32x *w, const int len)
+DECLSPEC void cram_md5_update_vector (md5_ctx_vector_t *ctx, const u32x *w, const int len)
 {
   u32x w0[4];
   u32x w1[4];
@@ -171,7 +171,7 @@ void cram_md5_update_vector (md5_ctx_vector_t *ctx, const u32x *w, const int len
 }
 
 
-void cram_md5_final_vector (md5_ctx_vector_t *ctx)
+DECLSPEC void cram_md5_final_vector (md5_ctx_vector_t *ctx)
 {
   cram_md5_transform_vector (ctx->w0, ctx->w1, ctx->w2, ctx->w3, ctx->h);
 }
diff --git a/OpenCL/m16600_a3-optimized.cl b/OpenCL/m16600_a3-optimized.cl
index 48c2eceb1..37447f145 100644
--- a/OpenCL/m16600_a3-optimized.cl
+++ b/OpenCL/m16600_a3-optimized.cl
@@ -13,7 +13,7 @@
 #include "inc_simd.cl"
 #include "inc_cipher_aes.cl"
 
-void m16600 (SHM_TYPE u32a *s_te0, SHM_TYPE u32a *s_te1, SHM_TYPE u32a *s_te2, SHM_TYPE u32a *s_te3, SHM_TYPE u32a *s_te4, SHM_TYPE u32a *s_td0, SHM_TYPE u32a *s_td1, SHM_TYPE u32a *s_td2, SHM_TYPE u32a *s_td3, SHM_TYPE u32a *s_td4, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const electrum_wallet_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
+DECLSPEC void m16600 (SHM_TYPE u32a *s_te0, SHM_TYPE u32a *s_te1, SHM_TYPE u32a *s_te2, SHM_TYPE u32a *s_te3, SHM_TYPE u32a *s_te4, SHM_TYPE u32a *s_td0, SHM_TYPE u32a *s_td1, SHM_TYPE u32a *s_td2, SHM_TYPE u32a *s_td3, SHM_TYPE u32a *s_td4, u32 w[16], const u32 pw_len, __global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const electrum_wallet_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * base
diff --git a/OpenCL/markov_be.cl b/OpenCL/markov_be.cl
index fd539f2d9..64103d62e 100644
--- a/OpenCL/markov_be.cl
+++ b/OpenCL/markov_be.cl
@@ -9,7 +9,7 @@
 
 #include "inc_types.cl"
 
-void generate_pw (u32 pw_buf[64], __global const cs_t *root_css_buf, __global const cs_t *markov_css_buf, const u32 pw_l_len, const u32 pw_r_len, const u32 mask80, const u32 bits14, const u32 bits15, u64 val)
+DECLSPEC void generate_pw (u32 pw_buf[64], __global const cs_t *root_css_buf, __global const cs_t *markov_css_buf, const u32 pw_l_len, const u32 pw_r_len, const u32 mask80, const u32 bits14, const u32 bits15, u64 val)
 {
   __global const cs_t *cs = &root_css_buf[pw_r_len];
 
diff --git a/OpenCL/markov_le.cl b/OpenCL/markov_le.cl
index a50644276..22b47d2d5 100644
--- a/OpenCL/markov_le.cl
+++ b/OpenCL/markov_le.cl
@@ -9,7 +9,7 @@
 
 #include "inc_types.cl"
 
-void generate_pw (u32 pw_buf[64], __global const cs_t *root_css_buf, __global const cs_t *markov_css_buf, const u32 pw_l_len, const u32 pw_r_len, const u32 mask80, const u32 bits14, const u32 bits15, u64 val)
+DECLSPEC void generate_pw (u32 pw_buf[64], __global const cs_t *root_css_buf, __global const cs_t *markov_css_buf, const u32 pw_l_len, const u32 pw_r_len, const u32 mask80, const u32 bits14, const u32 bits15, u64 val)
 {
   __global const cs_t *cs = &root_css_buf[pw_r_len];
 
diff --git a/docs/changes.txt b/docs/changes.txt
index 14586299e..370e61be6 100644
--- a/docs/changes.txt
+++ b/docs/changes.txt
@@ -57,6 +57,7 @@
 - OpenCL Runtime: Updated rocm detection
 - OpenCL Runtime: Enforce to use OpenCL version 1.2 to restrain OpenCL runtimes to make use of the __generic address space qualifier
 - OpenCL Runtime: Add current timestamp to OpenCL kernel source in order to force OpenCL JiT compiler to recompile and not use the cache
+- OpenCL Kernels: Add general function declaration keyword (static inline)
 - OpenCL Kernels: Replace variables from uXX to uXXa if used in __constant space
 - OpenCL Kernels: Use a special kernel to initialize the password buffer used during autotune measurements, to reduce startup time
 - OpenCL Kernels: Use static declaraction for uXXa variables used in __constant space