From 9ce30defcbba758382460bd01e7ce1821e42264b Mon Sep 17 00:00:00 2001
From: tweqx <romla@sfr.fr>
Date: Sat, 21 May 2022 19:32:39 +0200
Subject: [PATCH] Don't apply the salt in the a3 BLAKE2b($pass.$salt) optimized
 OpenCL code

---
 OpenCL/m00610_a3-optimized.cl | 179 ++++------------------------------
 1 file changed, 18 insertions(+), 161 deletions(-)

diff --git a/OpenCL/m00610_a3-optimized.cl b/OpenCL/m00610_a3-optimized.cl
index 7a406b40e..7402791ef 100644
--- a/OpenCL/m00610_a3-optimized.cl
+++ b/OpenCL/m00610_a3-optimized.cl
@@ -20,38 +20,6 @@ DECLSPEC void m00610m (PRIVATE_AS u32 *w, const u32 pw_len, KERN_ATTR_FUNC_VECTO
    * modifiers are taken from args
    */
 
-  /**
-   * salt
-   */
-
-  u32 salt_buf0[4];
-  u32 salt_buf1[4];
-  u32 salt_buf2[4];
-  u32 salt_buf3[4];
-
-  salt_buf0[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 0];
-  salt_buf0[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 1];
-  salt_buf0[2] = salt_bufs[SALT_POS_HOST].salt_buf[ 2];
-  salt_buf0[3] = salt_bufs[SALT_POS_HOST].salt_buf[ 3];
-  salt_buf1[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 4];
-  salt_buf1[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 5];
-  salt_buf1[2] = salt_bufs[SALT_POS_HOST].salt_buf[ 6];
-  salt_buf1[3] = salt_bufs[SALT_POS_HOST].salt_buf[ 7];
-  salt_buf2[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 8];
-  salt_buf2[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 9];
-  salt_buf2[2] = salt_bufs[SALT_POS_HOST].salt_buf[10];
-  salt_buf2[3] = salt_bufs[SALT_POS_HOST].salt_buf[11];
-  salt_buf3[0] = salt_bufs[SALT_POS_HOST].salt_buf[12];
-  salt_buf3[1] = salt_bufs[SALT_POS_HOST].salt_buf[13];
-  salt_buf3[2] = salt_bufs[SALT_POS_HOST].salt_buf[14];
-  salt_buf3[3] = salt_bufs[SALT_POS_HOST].salt_buf[15];
-
-  const u32 salt_len = salt_bufs[SALT_POS_HOST].salt_len;
-
-  const u32 pw_salt_len = pw_len + salt_len;
-
-  switch_buffer_by_offset_le_S (salt_buf0, salt_buf1, salt_buf2, salt_buf3, pw_len);
-
   /**
    * loop
    */
@@ -63,59 +31,20 @@ DECLSPEC void m00610m (PRIVATE_AS u32 *w, const u32 pw_len, KERN_ATTR_FUNC_VECTO
     const u32x w0r = words_buf_r[il_pos / VECT_SIZE];
     const u32x w0x = w0l | w0r;
 
-    u32x w0[4];
-    u32x w1[4];
-    u32x w2[4];
-    u32x w3[4];
-
-    w0[0] = w0x;
-    w0[1] = w[ 1];
-    w0[2] = w[ 2];
-    w0[3] = w[ 3];
-    w1[0] = w[ 4];
-    w1[1] = w[ 5];
-    w1[2] = w[ 6];
-    w1[3] = w[ 7];
-    w2[0] = w[ 8];
-    w2[1] = w[ 9];
-    w2[2] = w[10];
-    w2[3] = w[11];
-    w3[0] = w[12];
-    w3[1] = w[13];
-    w3[2] = w[14];
-    w3[3] = w[15];
-
-    w0[0] |= salt_buf0[0];
-    w0[1] |= salt_buf0[1];
-    w0[2] |= salt_buf0[2];
-    w0[3] |= salt_buf0[3];
-    w1[0] |= salt_buf1[0];
-    w1[1] |= salt_buf1[1];
-    w1[2] |= salt_buf1[2];
-    w1[3] |= salt_buf1[3];
-    w2[0] |= salt_buf2[0];
-    w2[1] |= salt_buf2[1];
-    w2[2] |= salt_buf2[2];
-    w2[3] |= salt_buf2[3];
-    w3[0] |= salt_buf3[0];
-    w3[1] |= salt_buf3[1];
-    w3[2] |= salt_buf3[2];
-    w3[3] |= salt_buf3[3];
-
     /**
      * blake2b
      */
 
     u64x m[16];
 
-    m[ 0] = hl32_to_64 (w0[1], w0[0]);
-    m[ 1] = hl32_to_64 (w0[3], w0[2]);
-    m[ 2] = hl32_to_64 (w1[1], w1[0]);
-    m[ 3] = hl32_to_64 (w1[3], w1[2]);
-    m[ 4] = hl32_to_64 (w2[1], w2[0]);
-    m[ 5] = hl32_to_64 (w2[3], w2[2]);
-    m[ 6] = hl32_to_64 (w3[1], w3[0]);
-    m[ 7] = hl32_to_64 (w3[3], w3[2]);
+    m[ 0] = hl32_to_64 (w[ 1], w0x  );
+    m[ 1] = hl32_to_64 (w[ 3], w[ 2]);
+    m[ 2] = hl32_to_64 (w[ 5], w[ 4]);
+    m[ 3] = hl32_to_64 (w[ 7], w[ 6]);
+    m[ 4] = hl32_to_64 (w[ 9], w[ 8]);
+    m[ 5] = hl32_to_64 (w[11], w[10]);
+    m[ 6] = hl32_to_64 (w[13], w[12]);
+    m[ 7] = hl32_to_64 (w[15], w[14]);
     m[ 8] = 0;
     m[ 9] = 0;
     m[10] = 0;
@@ -136,7 +65,7 @@ DECLSPEC void m00610m (PRIVATE_AS u32 *w, const u32 pw_len, KERN_ATTR_FUNC_VECTO
     h[6] = BLAKE2B_IV_06;
     h[7] = BLAKE2B_IV_07;
 
-    blake2b_transform_vector (h, m, pw_salt_len, BLAKE2B_FINAL);
+    blake2b_transform_vector (h, m, pw_len, BLAKE2B_FINAL);
 
     const u32x r0 = h32_from_64 (h[0]);
     const u32x r1 = l32_from_64 (h[0]);
@@ -165,38 +94,6 @@ DECLSPEC void m00610s (PRIVATE_AS u32 *w, const u32 pw_len, KERN_ATTR_FUNC_VECTO
     digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R3]
   };
 
-  /**
-   * salt
-   */
-
-  u32 salt_buf0[4];
-  u32 salt_buf1[4];
-  u32 salt_buf2[4];
-  u32 salt_buf3[4];
-
-  salt_buf0[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 0];
-  salt_buf0[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 1];
-  salt_buf0[2] = salt_bufs[SALT_POS_HOST].salt_buf[ 2];
-  salt_buf0[3] = salt_bufs[SALT_POS_HOST].salt_buf[ 3];
-  salt_buf1[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 4];
-  salt_buf1[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 5];
-  salt_buf1[2] = salt_bufs[SALT_POS_HOST].salt_buf[ 6];
-  salt_buf1[3] = salt_bufs[SALT_POS_HOST].salt_buf[ 7];
-  salt_buf2[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 8];
-  salt_buf2[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 9];
-  salt_buf2[2] = salt_bufs[SALT_POS_HOST].salt_buf[10];
-  salt_buf2[3] = salt_bufs[SALT_POS_HOST].salt_buf[11];
-  salt_buf3[0] = salt_bufs[SALT_POS_HOST].salt_buf[12];
-  salt_buf3[1] = salt_bufs[SALT_POS_HOST].salt_buf[13];
-  salt_buf3[2] = salt_bufs[SALT_POS_HOST].salt_buf[14];
-  salt_buf3[3] = salt_bufs[SALT_POS_HOST].salt_buf[15];
-
-  const u32 salt_len = salt_bufs[SALT_POS_HOST].salt_len;
-
-  const u32 pw_salt_len = pw_len + salt_len;
-
-  switch_buffer_by_offset_le_S (salt_buf0, salt_buf1, salt_buf2, salt_buf3, pw_len);
-
   /**
    * loop
    */
@@ -208,59 +105,20 @@ DECLSPEC void m00610s (PRIVATE_AS u32 *w, const u32 pw_len, KERN_ATTR_FUNC_VECTO
     const u32x w0r = words_buf_r[il_pos / VECT_SIZE];
     const u32x w0x = w0l | w0r;
 
-    u32x w0[4];
-    u32x w1[4];
-    u32x w2[4];
-    u32x w3[4];
-
-    w0[0] = w0x;
-    w0[1] = w[ 1];
-    w0[2] = w[ 2];
-    w0[3] = w[ 3];
-    w1[0] = w[ 4];
-    w1[1] = w[ 5];
-    w1[2] = w[ 6];
-    w1[3] = w[ 7];
-    w2[0] = w[ 8];
-    w2[1] = w[ 9];
-    w2[2] = w[10];
-    w2[3] = w[11];
-    w3[0] = w[12];
-    w3[1] = w[13];
-    w3[2] = w[14];
-    w3[3] = w[15];
-
-    w0[0] |= salt_buf0[0];
-    w0[1] |= salt_buf0[1];
-    w0[2] |= salt_buf0[2];
-    w0[3] |= salt_buf0[3];
-    w1[0] |= salt_buf1[0];
-    w1[1] |= salt_buf1[1];
-    w1[2] |= salt_buf1[2];
-    w1[3] |= salt_buf1[3];
-    w2[0] |= salt_buf2[0];
-    w2[1] |= salt_buf2[1];
-    w2[2] |= salt_buf2[2];
-    w2[3] |= salt_buf2[3];
-    w3[0] |= salt_buf3[0];
-    w3[1] |= salt_buf3[1];
-    w3[2] |= salt_buf3[2];
-    w3[3] |= salt_buf3[3];
-
     /**
      * blake2b
      */
 
     u64x m[16];
 
-    m[ 0] = hl32_to_64 (w0[1], w0[0]);
-    m[ 1] = hl32_to_64 (w0[3], w0[2]);
-    m[ 2] = hl32_to_64 (w1[1], w1[0]);
-    m[ 3] = hl32_to_64 (w1[3], w1[2]);
-    m[ 4] = hl32_to_64 (w2[1], w2[0]);
-    m[ 5] = hl32_to_64 (w2[3], w2[2]);
-    m[ 6] = hl32_to_64 (w3[1], w3[0]);
-    m[ 7] = hl32_to_64 (w3[3], w3[2]);
+    m[ 0] = hl32_to_64 (w[ 1], w0x  );
+    m[ 1] = hl32_to_64 (w[ 3], w[ 2]);
+    m[ 2] = hl32_to_64 (w[ 5], w[ 4]);
+    m[ 3] = hl32_to_64 (w[ 7], w[ 6]);
+    m[ 4] = hl32_to_64 (w[ 9], w[ 8]);
+    m[ 5] = hl32_to_64 (w[11], w[10]);
+    m[ 6] = hl32_to_64 (w[13], w[12]);
+    m[ 7] = hl32_to_64 (w[15], w[14]);
     m[ 8] = 0;
     m[ 9] = 0;
     m[10] = 0;
@@ -281,7 +139,7 @@ DECLSPEC void m00610s (PRIVATE_AS u32 *w, const u32 pw_len, KERN_ATTR_FUNC_VECTO
     h[6] = BLAKE2B_IV_06;
     h[7] = BLAKE2B_IV_07;
 
-    blake2b_transform_vector (h, m, pw_salt_len, BLAKE2B_FINAL);
+    blake2b_transform_vector (h, m, pw_len, BLAKE2B_FINAL);
 
     const u32x r0 = h32_from_64 (h[0]);
     const u32x r1 = l32_from_64 (h[0]);
@@ -531,4 +389,3 @@ KERNEL_FQ void m00610_s16 (KERN_ATTR_VECTOR ())
 
   m00610s (w, pw_len, pws, rules_buf, combs_buf, words_buf_r, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, kernel_param, gid, lid, lsz);
 }
-