Use pure kernel rule engine for --stdout

include/rp_kernel_on_cpu.h

@@ -6,7 +6,6 @@
 #ifndef _RP_KERNEL_ON_CPU_H
 #define _RP_KERNEL_ON_CPU_H
 
-u32 apply_rule (const u32 name, const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len);
-u32 apply_rules (u32 *cmds, u32 buf0[4], u32 buf1[4], const u32 len);
+int apply_rules (const u32 *cmds, u32 in_buf[64], const int in_len, u32 out_buf[64]);
 
 #endif // _RP_KERNEL_ON_CPU_H

include/rp_kernel_on_cpu_optimized.h

@@ -0,0 +1,12 @@
+/**
+ * Author......: See docs/credits.txt
+ * License.....: MIT
+ */
+
+#ifndef _RP_KERNEL_ON_CPU_OPTIMIZED_H
+#define _RP_KERNEL_ON_CPU_OPTIMIZED_H
+
+u32 apply_rule_optimized (const u32 name, const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len);
+u32 apply_rules_optimized (u32 *cmds, u32 buf0[4], u32 buf1[4], const u32 len);
+
+#endif // _RP_KERNEL_ON_CPU_OPTIMIZED_H

src/Makefile

@@ -273,7 +273,7 @@ LFLAGS_CROSS_WIN64      += -lpsapi
 ## Objects
 ##
 
-OBJS_ALL                := affinity autotune benchmark bitmap bitops combinator common convert cpt cpu_aes cpu_crc32 cpu_des cpu_md4 cpu_md5 cpu_sha1 cpu_sha256 debugfile dictstat dispatch dynloader event ext_ADL ext_nvapi ext_nvml ext_OpenCL ext_sysfs ext_xnvctrl ext_lzma lzma_sdk/Alloc lzma_sdk/Lzma2Dec lzma_sdk/LzmaDec filehandling folder hashcat hashes hlfmt hwmon induct interface locking logfile loopback memory monitor mpsp opencl outfile_check outfile pidfile potfile restore rp rp_cpu rp_kernel_on_cpu selftest shared status stdout straight terminal thread timer tuningdb usage user_options weak_hash wordlist
+OBJS_ALL                := affinity autotune benchmark bitmap bitops combinator common convert cpt cpu_aes cpu_crc32 cpu_des cpu_md4 cpu_md5 cpu_sha1 cpu_sha256 debugfile dictstat dispatch dynloader event ext_ADL ext_nvapi ext_nvml ext_OpenCL ext_sysfs ext_xnvctrl ext_lzma lzma_sdk/Alloc lzma_sdk/Lzma2Dec lzma_sdk/LzmaDec filehandling folder hashcat hashes hlfmt hwmon induct interface locking logfile loopback memory monitor mpsp opencl outfile_check outfile pidfile potfile restore rp rp_cpu rp_kernel_on_cpu rp_kernel_on_cpu_optimized selftest shared status stdout straight terminal thread timer tuningdb usage user_options weak_hash wordlist
 
 NATIVE_STATIC_OBJS      := $(foreach OBJ,$(OBJS_ALL),obj/$(OBJ).NATIVE.STATIC.o)
 NATIVE_SHARED_OBJS      := $(foreach OBJ,$(OBJS_ALL),obj/$(OBJ).NATIVE.SHARED.o)

src/hashes.c

@@ -21,7 +21,6 @@
 #include "outfile.h"
 #include "potfile.h"
 #include "rp.h"
-#include "rp_kernel_on_cpu.h"
 #include "shared.h"
 #include "thread.h"
 #include "timer.h"

src/outfile.c

@@ -13,6 +13,7 @@
 #include "mpsp.h"
 #include "rp.h"
 #include "rp_kernel_on_cpu.h"
+#include "rp_kernel_on_cpu_optimized.h"
 #include "opencl.h"
 #include "shared.h"
 #include "outfile.h"
@@ -42,16 +43,21 @@ int build_plain (hashcat_ctx_t *hashcat_ctx, hc_device_param_t *device_param, pl
 
     if (rc == -1) return -1;
 
-    for (int i = 0; i < 64; i++)
-    {
-      plain_buf[i] = pw.i[i];
-    }
-
-    plain_len = (int) pw.pw_len;
-
     const u32 off = device_param->innerloop_pos + il_pos;
 
-    plain_len = (int) apply_rules (straight_ctx->kernel_rules_buf[off].cmds, &plain_buf[0], &plain_buf[4], (u32) plain_len);
+    if (hashconfig->opti_type & OPTI_TYPE_OPTIMIZED_KERNEL)
+    {
+      for (int i = 0; i < 8; i++)
+      {
+        plain_buf[i] = pw.i[i];
+      }
+
+      plain_len = (int) apply_rules_optimized (straight_ctx->kernel_rules_buf[off].cmds, &plain_buf[0], &plain_buf[4], (u32) pw.pw_len);
+    }
+    else
+    {
+      plain_len = (int) apply_rules (straight_ctx->kernel_rules_buf[off].cmds, pw.i, pw.pw_len, plain_buf);
+    }
 
     if (plain_len > (int) hashconfig->pw_max) plain_len = (int) hashconfig->pw_max;
   }

src/stdout.c

@@ -8,6 +8,7 @@
 #include "event.h"
 #include "locking.h"
 #include "rp_kernel_on_cpu.h"
+#include "rp_kernel_on_cpu_optimized.h"
 #include "mpsp.h"
 #include "opencl.h"
 #include "shared.h"
@@ -112,18 +113,23 @@ int process_stdout (hashcat_ctx_t *hashcat_ctx, hc_device_param_t *device_param,
         return -1;
       }
 
-      const u32 pos = device_param->innerloop_pos;
-
       for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
       {
-        for (int i = 0; i < 64; i++)
+        const u32 off = device_param->innerloop_pos + il_pos;
+
+        if (hashconfig->opti_type & OPTI_TYPE_OPTIMIZED_KERNEL)
         {
-          plain_buf[i] = pw.i[i];
+          for (int i = 0; i < 8; i++)
+          {
+            plain_buf[i] = pw.i[i];
+          }
+
+          plain_len = apply_rules_optimized (straight_ctx->kernel_rules_buf[off].cmds, &plain_buf[0], &plain_buf[4], pw.pw_len);
+        }
+        else
+        {
+          plain_len = (int) apply_rules (straight_ctx->kernel_rules_buf[off].cmds, pw.i, pw.pw_len, plain_buf);
         }
-
-        plain_len = pw.pw_len;
-
-        plain_len = apply_rules (straight_ctx->kernel_rules_buf[pos + il_pos].cmds, &plain_buf[0], &plain_buf[4], plain_len);
 
         if (plain_len > hashconfig->pw_max) plain_len = hashconfig->pw_max;