Cleanup -m 47xx kernels to latest standard

diff --git a/OpenCL/m04500_a3.cl b/OpenCL/m04500_a3.cl
index 31434d2..15772dd 100644 (file)
--- a/OpenCL/m04500_a3.cl
+++ b/OpenCL/m04500_a3.cl
@@ -73,8 +73,8 @@ static void m04500m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     u32x wb_t = w2[3];
     u32x wc_t = w3[0];
     u32x wd_t = w3[1];
-    u32x we_t = 0;
-    u32x wf_t = pw_len * 8;
+    u32x we_t = w3[2];
+    u32x wf_t = w3[3];
 
     u32x a = SHA1M_A;
     u32x b = SHA1M_B;
@@ -379,8 +379,8 @@ static void m04500s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     u32x wb_t = w2[3];
     u32x wc_t = w3[0];
     u32x wd_t = w3[1];
-    u32x we_t = 0;
-    u32x wf_t = pw_len * 8;
+    u32x we_t = w3[2];
+    u32x wf_t = w3[3];
 
     u32x a = SHA1M_A;
     u32x b = SHA1M_B;

diff --git a/OpenCL/m04700_a0.cl b/OpenCL/m04700_a0.cl
index e3681b7..bc73392 100644 (file)
--- a/OpenCL/m04700_a0.cl
+++ b/OpenCL/m04700_a0.cl
@@ -18,7 +18,6 @@
 #define DGST_R3 1
 
 #include "include/kernel_functions.c"
-#undef _MD5_
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
 #include "include/rp_kernel.h"
@@ -40,7 +39,7 @@
 __kernel void m04700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, 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 u32 combs_mode, const u32 gid_max)
 {
   /**
-   * base
+   * modifier
    */
 
   const u32 gid = get_global_id (0);
@@ -71,18 +70,16 @@ __kernel void m04700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    */
 
   u32 pw_buf0[4];
-
-  pw_buf0[0] = pws[gid].i[ 0];
-  pw_buf0[1] = pws[gid].i[ 1];
-  pw_buf0[2] = pws[gid].i[ 2];
-  pw_buf0[3] = pws[gid].i[ 3];
-
   u32 pw_buf1[4];
 
-  pw_buf1[0] = pws[gid].i[ 4];
-  pw_buf1[1] = pws[gid].i[ 5];
-  pw_buf1[2] = pws[gid].i[ 6];
-  pw_buf1[3] = pws[gid].i[ 7];
+  pw_buf0[0] = pws[gid].i[0];
+  pw_buf0[1] = pws[gid].i[1];
+  pw_buf0[2] = pws[gid].i[2];
+  pw_buf0[3] = pws[gid].i[3];
+  pw_buf1[0] = pws[gid].i[4];
+  pw_buf1[1] = pws[gid].i[5];
+  pw_buf1[2] = pws[gid].i[6];
+  pw_buf1[3] = pws[gid].i[7];
 
   const u32 pw_len = pws[gid].pw_len;
 
@@ -102,6 +99,7 @@ __kernel void m04700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     append_0x80_2x4_VV (w0, w1, out_len);
 
     w3[2] = out_len * 8;
+    w3[3] = 0;
 
     /**
      * md5
@@ -191,22 +189,21 @@ __kernel void m04700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
      */
 
     u32x w0_t = uint_to_hex_lower8_le ((a >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((a >>  0) & 255) << 16;
+              | uint_to_hex_lower8_le ((a >>  0) & 255) << 16;
     u32x w1_t = uint_to_hex_lower8_le ((a >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((a >> 16) & 255) << 16;
+              | uint_to_hex_lower8_le ((a >> 16) & 255) << 16;
     u32x w2_t = uint_to_hex_lower8_le ((b >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((b >>  0) & 255) << 16;
+              | uint_to_hex_lower8_le ((b >>  0) & 255) << 16;
     u32x w3_t = uint_to_hex_lower8_le ((b >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((b >> 16) & 255) << 16;
+              | uint_to_hex_lower8_le ((b >> 16) & 255) << 16;
     u32x w4_t = uint_to_hex_lower8_le ((c >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((c >>  0) & 255) << 16;
+              | uint_to_hex_lower8_le ((c >>  0) & 255) << 16;
     u32x w5_t = uint_to_hex_lower8_le ((c >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((c >> 16) & 255) << 16;
+              | uint_to_hex_lower8_le ((c >> 16) & 255) << 16;
     u32x w6_t = uint_to_hex_lower8_le ((d >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((d >>  0) & 255) << 16;
+              | uint_to_hex_lower8_le ((d >>  0) & 255) << 16;
     u32x w7_t = uint_to_hex_lower8_le ((d >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((d >> 16) & 255) << 16;
-
+              | uint_to_hex_lower8_le ((d >> 16) & 255) << 16;
     u32x w8_t = 0x80000000;
     u32x w9_t = 0;
     u32x wa_t = 0;
@@ -333,7 +330,7 @@ __kernel void m04700_m16 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 __kernel void m04700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, 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 u32 combs_mode, const u32 gid_max)
 {
   /**
-   * base
+   * modifier
    */
 
   const u32 gid = get_global_id (0);
@@ -364,18 +361,16 @@ __kernel void m04700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    */
 
   u32 pw_buf0[4];
-
-  pw_buf0[0] = pws[gid].i[ 0];
-  pw_buf0[1] = pws[gid].i[ 1];
-  pw_buf0[2] = pws[gid].i[ 2];
-  pw_buf0[3] = pws[gid].i[ 3];
-
   u32 pw_buf1[4];
 
-  pw_buf1[0] = pws[gid].i[ 4];
-  pw_buf1[1] = pws[gid].i[ 5];
-  pw_buf1[2] = pws[gid].i[ 6];
-  pw_buf1[3] = pws[gid].i[ 7];
+  pw_buf0[0] = pws[gid].i[0];
+  pw_buf0[1] = pws[gid].i[1];
+  pw_buf0[2] = pws[gid].i[2];
+  pw_buf0[3] = pws[gid].i[3];
+  pw_buf1[0] = pws[gid].i[4];
+  pw_buf1[1] = pws[gid].i[5];
+  pw_buf1[2] = pws[gid].i[6];
+  pw_buf1[3] = pws[gid].i[7];
 
   const u32 pw_len = pws[gid].pw_len;
 
@@ -413,6 +408,7 @@ __kernel void m04700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     append_0x80_2x4_VV (w0, w1, out_len);
 
     w3[2] = out_len * 8;
+    w3[3] = 0;
 
     /**
      * md5
@@ -502,22 +498,21 @@ __kernel void m04700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
      */
 
     u32x w0_t = uint_to_hex_lower8_le ((a >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((a >>  0) & 255) << 16;
+              | uint_to_hex_lower8_le ((a >>  0) & 255) << 16;
     u32x w1_t = uint_to_hex_lower8_le ((a >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((a >> 16) & 255) << 16;
+              | uint_to_hex_lower8_le ((a >> 16) & 255) << 16;
     u32x w2_t = uint_to_hex_lower8_le ((b >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((b >>  0) & 255) << 16;
+              | uint_to_hex_lower8_le ((b >>  0) & 255) << 16;
     u32x w3_t = uint_to_hex_lower8_le ((b >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((b >> 16) & 255) << 16;
+              | uint_to_hex_lower8_le ((b >> 16) & 255) << 16;
     u32x w4_t = uint_to_hex_lower8_le ((c >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((c >>  0) & 255) << 16;
+              | uint_to_hex_lower8_le ((c >>  0) & 255) << 16;
     u32x w5_t = uint_to_hex_lower8_le ((c >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((c >> 16) & 255) << 16;
+              | uint_to_hex_lower8_le ((c >> 16) & 255) << 16;
     u32x w6_t = uint_to_hex_lower8_le ((d >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((d >>  0) & 255) << 16;
+              | uint_to_hex_lower8_le ((d >>  0) & 255) << 16;
     u32x w7_t = uint_to_hex_lower8_le ((d >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((d >> 16) & 255) << 16;
-
+              | uint_to_hex_lower8_le ((d >> 16) & 255) << 16;
     u32x w8_t = 0x80000000;
     u32x w9_t = 0;
     u32x wa_t = 0;

diff --git a/OpenCL/m04700_a1.cl b/OpenCL/m04700_a1.cl
index 8d1b0d4..b71ef75 100644 (file)
--- a/OpenCL/m04700_a1.cl
+++ b/OpenCL/m04700_a1.cl
@@ -5,6 +5,8 @@
 
 #define _SHA1_MD5_
 
+#define NEW_SIMD_CODE
+
 #include "include/constants.h"
 #include "include/kernel_vendor.h"
 
@@ -14,19 +16,26 @@
 #define DGST_R3 1
 
 #include "include/kernel_functions.c"
-#undef _MD5_
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
-
-#define COMPARE_S "OpenCL/check_single_comp4.c"
-#define COMPARE_M "OpenCL/check_multi_comp4.c"
-
-#define uint_to_hex_lower8_le(i) l_bin2asc[(i)]
+#include "OpenCL/simd.c"
+
+#if   VECT_SIZE == 1
+#define uint_to_hex_lower8_le(i) (u32x) (l_bin2asc[(i)])
+#elif VECT_SIZE == 2
+#define uint_to_hex_lower8_le(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1])
+#elif VECT_SIZE == 4
+#define uint_to_hex_lower8_le(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3])
+#elif VECT_SIZE == 8
+#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])
+#elif VECT_SIZE == 16
+#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
 
 __kernel void m04700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, 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 u32 combs_mode, const u32 gid_max)
 {
   /**
-   * base
+   * modifier
    */
 
   const u32 gid = get_global_id (0);
@@ -56,111 +65,88 @@ __kernel void m04700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * base
    */
 
-  u32 wordl0[4];
-
-  wordl0[0] = pws[gid].i[ 0];
-  wordl0[1] = pws[gid].i[ 1];
-  wordl0[2] = pws[gid].i[ 2];
-  wordl0[3] = pws[gid].i[ 3];
-
-  u32 wordl1[4];
-
-  wordl1[0] = pws[gid].i[ 4];
-  wordl1[1] = pws[gid].i[ 5];
-  wordl1[2] = pws[gid].i[ 6];
-  wordl1[3] = pws[gid].i[ 7];
-
-  u32 wordl2[4];
+  u32 pw_buf0[4];
+  u32 pw_buf1[4];
 
-  wordl2[0] = 0;
-  wordl2[1] = 0;
-  wordl2[2] = 0;
-  wordl2[3] = 0;
-
-  u32 wordl3[4];
-
-  wordl3[0] = 0;
-  wordl3[1] = 0;
-  wordl3[2] = 0;
-  wordl3[3] = 0;
+  pw_buf0[0] = pws[gid].i[0];
+  pw_buf0[1] = pws[gid].i[1];
+  pw_buf0[2] = pws[gid].i[2];
+  pw_buf0[3] = pws[gid].i[3];
+  pw_buf1[0] = pws[gid].i[4];
+  pw_buf1[1] = pws[gid].i[5];
+  pw_buf1[2] = pws[gid].i[6];
+  pw_buf1[3] = pws[gid].i[7];
 
   const u32 pw_l_len = pws[gid].pw_len;
 
-  if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
-  {
-    append_0x80_2x4 (wordl0, wordl1, pw_l_len);
-
-    switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
-  }
-
   /**
    * loop
    */
 
-  for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
+  for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
   {
-    const u32 pw_r_len = combs_buf[il_pos].pw_len;
-
-    const u32 pw_len = pw_l_len + pw_r_len;
-
-    u32 wordr0[4];
+    const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
 
-    wordr0[0] = combs_buf[il_pos].i[0];
-    wordr0[1] = combs_buf[il_pos].i[1];
-    wordr0[2] = combs_buf[il_pos].i[2];
-    wordr0[3] = combs_buf[il_pos].i[3];
+    const u32x pw_len = pw_l_len + pw_r_len;
 
-    u32 wordr1[4];
-
-    wordr1[0] = combs_buf[il_pos].i[4];
-    wordr1[1] = combs_buf[il_pos].i[5];
-    wordr1[2] = combs_buf[il_pos].i[6];
-    wordr1[3] = combs_buf[il_pos].i[7];
-
-    u32 wordr2[4];
-
-    wordr2[0] = 0;
-    wordr2[1] = 0;
-    wordr2[2] = 0;
-    wordr2[3] = 0;
-
-    u32 wordr3[4];
+    /**
+     * concat password candidate
+     */
 
-    wordr3[0] = 0;
-    wordr3[1] = 0;
-    wordr3[2] = 0;
-    wordr3[3] = 0;
+    u32x wordl0[4] = { 0 };
+    u32x wordl1[4] = { 0 };
+    u32x wordl2[4] = { 0 };
+    u32x wordl3[4] = { 0 };
+
+    wordl0[0] = pw_buf0[0];
+    wordl0[1] = pw_buf0[1];
+    wordl0[2] = pw_buf0[2];
+    wordl0[3] = pw_buf0[3];
+    wordl1[0] = pw_buf1[0];
+    wordl1[1] = pw_buf1[1];
+    wordl1[2] = pw_buf1[2];
+    wordl1[3] = pw_buf1[3];
+
+    u32x wordr0[4] = { 0 };
+    u32x wordr1[4] = { 0 };
+    u32x wordr2[4] = { 0 };
+    u32x wordr3[4] = { 0 };
+
+    wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
+    wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
+    wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
+    wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
+    wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
+    wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
+    wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
+    wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
 
     if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
     {
-      append_0x80_2x4 (wordr0, wordr1, pw_r_len);
-
-      switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+      switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+    }
+    else
+    {
+      switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
     }
 
-    u32 w0[4];
+    u32x w0[4];
+    u32x w1[4];
+    u32x w2[4];
+    u32x w3[4];
 
     w0[0] = wordl0[0] | wordr0[0];
     w0[1] = wordl0[1] | wordr0[1];
     w0[2] = wordl0[2] | wordr0[2];
     w0[3] = wordl0[3] | wordr0[3];
-
-    u32 w1[4];
-
     w1[0] = wordl1[0] | wordr1[0];
     w1[1] = wordl1[1] | wordr1[1];
     w1[2] = wordl1[2] | wordr1[2];
     w1[3] = wordl1[3] | wordr1[3];
-
-    u32 w2[4];
-
     w2[0] = wordl2[0] | wordr2[0];
     w2[1] = wordl2[1] | wordr2[1];
     w2[2] = wordl2[2] | wordr2[2];
     w2[3] = wordl2[3] | wordr2[3];
-
-    u32 w3[4];
-
     w3[0] = wordl3[0] | wordr3[0];
     w3[1] = wordl3[1] | wordr3[1];
     w3[2] = pw_len * 8;
@@ -170,10 +156,11 @@ __kernel void m04700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
      * md5
      */
 
-    u32 a = MD5M_A;
-    u32 b = MD5M_B;
-    u32 c = MD5M_C;
-    u32 d = MD5M_D;
+    u32x a = MD5M_A;
+    u32x b = MD5M_B;
+    u32x c = MD5M_C;
+    u32x d = MD5M_D;
+    u32x e = 0;
 
     MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
     MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
@@ -252,33 +239,30 @@ __kernel void m04700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
      * sha1
      */
 
-    u32 w0_t = uint_to_hex_lower8_le ((a >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((a >>  0) & 255) << 16;
-    u32 w1_t = uint_to_hex_lower8_le ((a >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((a >> 16) & 255) << 16;
-    u32 w2_t = uint_to_hex_lower8_le ((b >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((b >>  0) & 255) << 16;
-    u32 w3_t = uint_to_hex_lower8_le ((b >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((b >> 16) & 255) << 16;
-    u32 w4_t = uint_to_hex_lower8_le ((c >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((c >>  0) & 255) << 16;
-    u32 w5_t = uint_to_hex_lower8_le ((c >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((c >> 16) & 255) << 16;
-    u32 w6_t = uint_to_hex_lower8_le ((d >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((d >>  0) & 255) << 16;
-    u32 w7_t = uint_to_hex_lower8_le ((d >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((d >> 16) & 255) << 16;
-
-    u32 w8_t = 0x80000000;
-    u32 w9_t = 0;
-    u32 wa_t = 0;
-    u32 wb_t = 0;
-    u32 wc_t = 0;
-    u32 wd_t = 0;
-    u32 we_t = 0;
-    u32 wf_t = 32 * 8;
-
-    u32 e;
+    u32x w0_t = uint_to_hex_lower8_le ((a >>  8) & 255) <<  0
+              | uint_to_hex_lower8_le ((a >>  0) & 255) << 16;
+    u32x w1_t = uint_to_hex_lower8_le ((a >> 24) & 255) <<  0
+              | uint_to_hex_lower8_le ((a >> 16) & 255) << 16;
+    u32x w2_t = uint_to_hex_lower8_le ((b >>  8) & 255) <<  0
+              | uint_to_hex_lower8_le ((b >>  0) & 255) << 16;
+    u32x w3_t = uint_to_hex_lower8_le ((b >> 24) & 255) <<  0
+              | uint_to_hex_lower8_le ((b >> 16) & 255) << 16;
+    u32x w4_t = uint_to_hex_lower8_le ((c >>  8) & 255) <<  0
+              | uint_to_hex_lower8_le ((c >>  0) & 255) << 16;
+    u32x w5_t = uint_to_hex_lower8_le ((c >> 24) & 255) <<  0
+              | uint_to_hex_lower8_le ((c >> 16) & 255) << 16;
+    u32x w6_t = uint_to_hex_lower8_le ((d >>  8) & 255) <<  0
+              | uint_to_hex_lower8_le ((d >>  0) & 255) << 16;
+    u32x w7_t = uint_to_hex_lower8_le ((d >> 24) & 255) <<  0
+              | uint_to_hex_lower8_le ((d >> 16) & 255) << 16;
+    u32x w8_t = 0x80000000;
+    u32x w9_t = 0;
+    u32x wa_t = 0;
+    u32x wb_t = 0;
+    u32x wc_t = 0;
+    u32x wd_t = 0;
+    u32x we_t = 0;
+    u32x wf_t = 32 * 8;
 
     a = SHA1M_A;
     b = SHA1M_B;
@@ -382,12 +366,7 @@ __kernel void m04700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, we_t);
     wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wf_t);
 
-    const u32 r0 = d;
-    const u32 r1 = e;
-    const u32 r2 = c;
-    const u32 r3 = b;
-
-    #include COMPARE_M
+    COMPARE_M_SIMD (d, e, c, b);
   }
 }
 
@@ -402,7 +381,7 @@ __kernel void m04700_m16 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 __kernel void m04700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, 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 u32 combs_mode, const u32 gid_max)
 {
   /**
-   * base
+   * modifier
    */
 
   const u32 gid = get_global_id (0);
@@ -432,43 +411,20 @@ __kernel void m04700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * base
    */
 
-  u32 wordl0[4];
-
-  wordl0[0] = pws[gid].i[ 0];
-  wordl0[1] = pws[gid].i[ 1];
-  wordl0[2] = pws[gid].i[ 2];
-  wordl0[3] = pws[gid].i[ 3];
-
-  u32 wordl1[4];
-
-  wordl1[0] = pws[gid].i[ 4];
-  wordl1[1] = pws[gid].i[ 5];
-  wordl1[2] = pws[gid].i[ 6];
-  wordl1[3] = pws[gid].i[ 7];
-
-  u32 wordl2[4];
-
-  wordl2[0] = 0;
-  wordl2[1] = 0;
-  wordl2[2] = 0;
-  wordl2[3] = 0;
-
-  u32 wordl3[4];
+  u32 pw_buf0[4];
+  u32 pw_buf1[4];
 
-  wordl3[0] = 0;
-  wordl3[1] = 0;
-  wordl3[2] = 0;
-  wordl3[3] = 0;
+  pw_buf0[0] = pws[gid].i[0];
+  pw_buf0[1] = pws[gid].i[1];
+  pw_buf0[2] = pws[gid].i[2];
+  pw_buf0[3] = pws[gid].i[3];
+  pw_buf1[0] = pws[gid].i[4];
+  pw_buf1[1] = pws[gid].i[5];
+  pw_buf1[2] = pws[gid].i[6];
+  pw_buf1[3] = pws[gid].i[7];
 
   const u32 pw_l_len = pws[gid].pw_len;
 
-  if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
-  {
-    append_0x80_2x4 (wordl0, wordl1, pw_l_len);
-
-    switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
-  }
-
   /**
    * digest
    */
@@ -485,76 +441,76 @@ __kernel void m04700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * reverse
    */
 
-  const u32 e_rev = rotl32 (search[1], 2u);
+  const u32 e_rev = rotl32_S (search[1], 2u);
 
   /**
    * loop
    */
 
-  for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
+  for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
   {
-    const u32 pw_r_len = combs_buf[il_pos].pw_len;
-
-    const u32 pw_len = pw_l_len + pw_r_len;
-
-    u32 wordr0[4];
-
-    wordr0[0] = combs_buf[il_pos].i[0];
-    wordr0[1] = combs_buf[il_pos].i[1];
-    wordr0[2] = combs_buf[il_pos].i[2];
-    wordr0[3] = combs_buf[il_pos].i[3];
+    const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
 
-    u32 wordr1[4];
+    const u32x pw_len = pw_l_len + pw_r_len;
 
-    wordr1[0] = combs_buf[il_pos].i[4];
-    wordr1[1] = combs_buf[il_pos].i[5];
-    wordr1[2] = combs_buf[il_pos].i[6];
-    wordr1[3] = combs_buf[il_pos].i[7];
-
-    u32 wordr2[4];
-
-    wordr2[0] = 0;
-    wordr2[1] = 0;
-    wordr2[2] = 0;
-    wordr2[3] = 0;
-
-    u32 wordr3[4];
+    /**
+     * concat password candidate
+     */
 
-    wordr3[0] = 0;
-    wordr3[1] = 0;
-    wordr3[2] = 0;
-    wordr3[3] = 0;
+    u32x wordl0[4] = { 0 };
+    u32x wordl1[4] = { 0 };
+    u32x wordl2[4] = { 0 };
+    u32x wordl3[4] = { 0 };
+
+    wordl0[0] = pw_buf0[0];
+    wordl0[1] = pw_buf0[1];
+    wordl0[2] = pw_buf0[2];
+    wordl0[3] = pw_buf0[3];
+    wordl1[0] = pw_buf1[0];
+    wordl1[1] = pw_buf1[1];
+    wordl1[2] = pw_buf1[2];
+    wordl1[3] = pw_buf1[3];
+
+    u32x wordr0[4] = { 0 };
+    u32x wordr1[4] = { 0 };
+    u32x wordr2[4] = { 0 };
+    u32x wordr3[4] = { 0 };
+
+    wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
+    wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
+    wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
+    wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
+    wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
+    wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
+    wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
+    wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
 
     if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
     {
-      append_0x80_2x4 (wordr0, wordr1, pw_r_len);
-
-      switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+      switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+    }
+    else
+    {
+      switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
     }
 
-    u32 w0[4];
+    u32x w0[4];
+    u32x w1[4];
+    u32x w2[4];
+    u32x w3[4];
 
     w0[0] = wordl0[0] | wordr0[0];
     w0[1] = wordl0[1] | wordr0[1];
     w0[2] = wordl0[2] | wordr0[2];
     w0[3] = wordl0[3] | wordr0[3];
-
-    u32 w1[4];
-
     w1[0] = wordl1[0] | wordr1[0];
     w1[1] = wordl1[1] | wordr1[1];
     w1[2] = wordl1[2] | wordr1[2];
     w1[3] = wordl1[3] | wordr1[3];
-
-    u32 w2[4];
-
     w2[0] = wordl2[0] | wordr2[0];
     w2[1] = wordl2[1] | wordr2[1];
     w2[2] = wordl2[2] | wordr2[2];
     w2[3] = wordl2[3] | wordr2[3];
-
-    u32 w3[4];
-
     w3[0] = wordl3[0] | wordr3[0];
     w3[1] = wordl3[1] | wordr3[1];
     w3[2] = pw_len * 8;
@@ -564,10 +520,11 @@ __kernel void m04700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
      * md5
      */
 
-    u32 a = MD5M_A;
-    u32 b = MD5M_B;
-    u32 c = MD5M_C;
-    u32 d = MD5M_D;
+    u32x a = MD5M_A;
+    u32x b = MD5M_B;
+    u32x c = MD5M_C;
+    u32x d = MD5M_D;
+    u32x e = 0;
 
     MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
     MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
@@ -646,33 +603,30 @@ __kernel void m04700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
      * sha1
      */
 
-    u32 w0_t = uint_to_hex_lower8_le ((a >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((a >>  0) & 255) << 16;
-    u32 w1_t = uint_to_hex_lower8_le ((a >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((a >> 16) & 255) << 16;
-    u32 w2_t = uint_to_hex_lower8_le ((b >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((b >>  0) & 255) << 16;
-    u32 w3_t = uint_to_hex_lower8_le ((b >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((b >> 16) & 255) << 16;
-    u32 w4_t = uint_to_hex_lower8_le ((c >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((c >>  0) & 255) << 16;
-    u32 w5_t = uint_to_hex_lower8_le ((c >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((c >> 16) & 255) << 16;
-    u32 w6_t = uint_to_hex_lower8_le ((d >>  8) & 255) <<  0
-               | uint_to_hex_lower8_le ((d >>  0) & 255) << 16;
-    u32 w7_t = uint_to_hex_lower8_le ((d >> 24) & 255) <<  0
-               | uint_to_hex_lower8_le ((d >> 16) & 255) << 16;
-
-    u32 w8_t = 0x80000000;
-    u32 w9_t = 0;
-    u32 wa_t = 0;
-    u32 wb_t = 0;
-    u32 wc_t = 0;
-    u32 wd_t = 0;
-    u32 we_t = 0;
-    u32 wf_t = 32 * 8;
-
-    u32 e;
+    u32x w0_t = uint_to_hex_lower8_le ((a >>  8) & 255) <<  0
+              | uint_to_hex_lower8_le ((a >>  0) & 255) << 16;
+    u32x w1_t = uint_to_hex_lower8_le ((a >> 24) & 255) <<  0
+              | uint_to_hex_lower8_le ((a >> 16) & 255) << 16;
+    u32x w2_t = uint_to_hex_lower8_le ((b >>  8) & 255) <<  0
+              | uint_to_hex_lower8_le ((b >>  0) & 255) << 16;
+    u32x w3_t = uint_to_hex_lower8_le ((b >> 24) & 255) <<  0
+              | uint_to_hex_lower8_le ((b >> 16) & 255) << 16;
+    u32x w4_t = uint_to_hex_lower8_le ((c >>  8) & 255) <<  0
+              | uint_to_hex_lower8_le ((c >>  0) & 255) << 16;
+    u32x w5_t = uint_to_hex_lower8_le ((c >> 24) & 255) <<  0
+              | uint_to_hex_lower8_le ((c >> 16) & 255) << 16;
+    u32x w6_t = uint_to_hex_lower8_le ((d >>  8) & 255) <<  0
+              | uint_to_hex_lower8_le ((d >>  0) & 255) << 16;
+    u32x w7_t = uint_to_hex_lower8_le ((d >> 24) & 255) <<  0
+              | uint_to_hex_lower8_le ((d >> 16) & 255) << 16;
+    u32x w8_t = 0x80000000;
+    u32x w9_t = 0;
+    u32x wa_t = 0;
+    u32x wb_t = 0;
+    u32x wc_t = 0;
+    u32x wd_t = 0;
+    u32x we_t = 0;
+    u32x wf_t = 32 * 8;
 
     a = SHA1M_A;
     b = SHA1M_B;
@@ -772,19 +726,14 @@ __kernel void m04700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wa_t);
     wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wb_t);
 
-    if (allx (e != e_rev)) continue;
+    if (MATCHES_NONE_VS (e, e_rev)) continue;
 
     wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wc_t);
     wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wd_t);
     we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, we_t);
     wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wf_t);
 
-    const u32 r0 = d;
-    const u32 r1 = e;
-    const u32 r2 = c;
-    const u32 r3 = b;
-
-    #include COMPARE_S
+    COMPARE_S_SIMD (d, e, c, b);
   }
 }
 

diff --git a/OpenCL/m04700_a3.cl b/OpenCL/m04700_a3.cl
index 9c9db46..ba73946 100644 (file)
--- a/OpenCL/m04700_a3.cl
+++ b/OpenCL/m04700_a3.cl
@@ -18,7 +18,6 @@
 #define DGST_R3 1
 
 #include "include/kernel_functions.c"
-#undef _MD5_
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
 #include "OpenCL/simd.c"
@@ -56,33 +55,22 @@ static void m04700m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
     const u32x w0lr = w0l | w0r;
 
-    u32x w0_t[4];
-
-    w0_t[0] = w0lr;
-    w0_t[1] = w0[1];
-    w0_t[2] = w0[2];
-    w0_t[3] = w0[3];
-
-    u32x w1_t[4];
-
-    w1_t[0] = w1[0];
-    w1_t[1] = w1[1];
-    w1_t[2] = w1[2];
-    w1_t[3] = w1[3];
-
-    u32x w2_t[4];
-
-    w2_t[0] = w2[0];
-    w2_t[1] = w2[1];
-    w2_t[2] = w2[2];
-    w2_t[3] = w2[3];
-
-    u32x w3_t[4];
-
-    w3_t[0] = w3[0];
-    w3_t[1] = w3[1];
-    w3_t[2] = w3[2];
-    w3_t[3] = w3[3];
+    u32x w0_t = w0lr;
+    u32x w1_t = w0[1];
+    u32x w2_t = w0[2];
+    u32x w3_t = w0[3];
+    u32x w4_t = w1[0];
+    u32x w5_t = w1[1];
+    u32x w6_t = w1[2];
+    u32x w7_t = w1[3];
+    u32x w8_t = w2[0];
+    u32x w9_t = w2[1];
+    u32x wa_t = w2[2];
+    u32x wb_t = w2[3];
+    u32x wc_t = w3[0];
+    u32x wd_t = w3[1];
+    u32x we_t = w3[2];
+    u32x wf_t = w3[3];
 
     /**
      * md5
@@ -92,74 +80,75 @@ static void m04700m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     u32x b = MD5M_B;
     u32x c = MD5M_C;
     u32x d = MD5M_D;
-
-    MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
-    MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
-    MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
-    MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
-    MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
-    MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
-    MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
-    MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
-    MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
-    MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
-    MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
-    MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
-    MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
-    MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
-    MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
-    MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);
-
-    MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
-    MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
-    MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
-    MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
-    MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
-    MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
-    MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
-    MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
-    MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
-    MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
-    MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
-    MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
-    MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
-    MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
-    MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
-    MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
-
-    MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
-    MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
-    MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
-    MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
-
-    MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
-    MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
-    MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
-    MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
-    MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
-    MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
-    MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
-    MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
-    MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
-    MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
-    MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
-    MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
-    MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
-    MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
-    MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
-    MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);
+    u32x e = 0;
+
+    MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
+    MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
+    MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
+    MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
+
+    MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
+    MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
+    MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
+    MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
+
+    MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
+    MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
+    MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
+    MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
+
+    MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
+    MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
+    MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
+    MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
 
     a += MD5M_A;
     b += MD5M_B;
@@ -170,34 +159,30 @@ static void m04700m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
      * sha1
      */
 
-    w0_t[0] = uint_to_hex_lower8_le ((a >>  8) & 255) <<  0
-            | uint_to_hex_lower8_le ((a >>  0) & 255) << 16;
-    w0_t[1] = uint_to_hex_lower8_le ((a >> 24) & 255) <<  0
-            | uint_to_hex_lower8_le ((a >> 16) & 255) << 16;
-    w0_t[2] = uint_to_hex_lower8_le ((b >>  8) & 255) <<  0
-            | uint_to_hex_lower8_le ((b >>  0) & 255) << 16;
-    w0_t[3] = uint_to_hex_lower8_le ((b >> 24) & 255) <<  0
-            | uint_to_hex_lower8_le ((b >> 16) & 255) << 16;
-    w1_t[0] = uint_to_hex_lower8_le ((c >>  8) & 255) <<  0
-            | uint_to_hex_lower8_le ((c >>  0) & 255) << 16;
-    w1_t[1] = uint_to_hex_lower8_le ((c >> 24) & 255) <<  0
-            | uint_to_hex_lower8_le ((c >> 16) & 255) << 16;
-    w1_t[2] = uint_to_hex_lower8_le ((d >>  8) & 255) <<  0
-            | uint_to_hex_lower8_le ((d >>  0) & 255) << 16;
-    w1_t[3] = uint_to_hex_lower8_le ((d >> 24) & 255) <<  0
-            | uint_to_hex_lower8_le ((d >> 16) & 255) << 16;
-
-    w2_t[0] = 0x80000000;
-    w2_t[1] = 0;
-    w2_t[2] = 0;
-    w2_t[3] = 0;
-
-    w3_t[0] = 0;
-    w3_t[1] = 0;
-    w3_t[2] = 0;
-    w3_t[3] = 32 * 8;
-
-    u32x e;
+    w0_t = uint_to_hex_lower8_le ((a >>  8) & 255) <<  0
+         | uint_to_hex_lower8_le ((a >>  0) & 255) << 16;
+    w1_t = uint_to_hex_lower8_le ((a >> 24) & 255) <<  0
+         | uint_to_hex_lower8_le ((a >> 16) & 255) << 16;
+    w2_t = uint_to_hex_lower8_le ((b >>  8) & 255) <<  0
+         | uint_to_hex_lower8_le ((b >>  0) & 255) << 16;
+    w3_t = uint_to_hex_lower8_le ((b >> 24) & 255) <<  0
+         | uint_to_hex_lower8_le ((b >> 16) & 255) << 16;
+    w4_t = uint_to_hex_lower8_le ((c >>  8) & 255) <<  0
+         | uint_to_hex_lower8_le ((c >>  0) & 255) << 16;
+    w5_t = uint_to_hex_lower8_le ((c >> 24) & 255) <<  0
+         | uint_to_hex_lower8_le ((c >> 16) & 255) << 16;
+    w6_t = uint_to_hex_lower8_le ((d >>  8) & 255) <<  0
+         | uint_to_hex_lower8_le ((d >>  0) & 255) << 16;
+    w7_t = uint_to_hex_lower8_le ((d >> 24) & 255) <<  0
+         | uint_to_hex_lower8_le ((d >> 16) & 255) << 16;
+    w8_t = 0x80000000;
+    w9_t = 0;
+    wa_t = 0;
+    wb_t = 0;
+    wc_t = 0;
+    wd_t = 0;
+    we_t = 0;
+    wf_t = 32 * 8;
 
     a = SHA1M_A;
     b = SHA1M_B;
@@ -208,98 +193,98 @@ static void m04700m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     #undef K
     #define K SHA1C00
 
-    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w0_t[0]);
-    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t[1]);
-    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w0_t[2]);
-    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w0_t[3]);
-    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w1_t[0]);
-    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w1_t[1]);
-    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w1_t[2]);
-    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w1_t[3]);
-    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w2_t[0]);
-    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w2_t[1]);
-    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w2_t[2]);
-    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w2_t[3]);
-    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w3_t[0]);
-    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w3_t[1]);
-    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w3_t[2]);
-    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w3_t[3]);
-    w0_t[0] = rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t[0]);
-    w0_t[1] = rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F0o, d, e, a, b, c, w0_t[1]);
-    w0_t[2] = rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F0o, c, d, e, a, b, w0_t[2]);
-    w0_t[3] = rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F0o, b, c, d, e, a, w0_t[3]);
+    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w0_t);
+    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w1_t);
+    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w2_t);
+    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w3_t);
+    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w4_t);
+    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w5_t);
+    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w6_t);
+    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w7_t);
+    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w8_t);
+    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w9_t);
+    SHA1_STEP (SHA1_F0o, a, b, c, d, e, wa_t);
+    SHA1_STEP (SHA1_F0o, e, a, b, c, d, wb_t);
+    SHA1_STEP (SHA1_F0o, d, e, a, b, c, wc_t);
+    SHA1_STEP (SHA1_F0o, c, d, e, a, b, wd_t);
+    SHA1_STEP (SHA1_F0o, b, c, d, e, a, we_t);
+    SHA1_STEP (SHA1_F0o, a, b, c, d, e, wf_t);
+    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t);
+    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F0o, d, e, a, b, c, w1_t);
+    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F0o, c, d, e, a, b, w2_t);
+    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F0o, b, c, d, e, a, w3_t);
 
     #undef K
     #define K SHA1C01
 
-    w1_t[0] = rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t[0]);
-    w1_t[1] = rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w1_t[1]);
-    w1_t[2] = rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w1_t[2]);
-    w1_t[3] = rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t[3]);
-    w2_t[0] = rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t[0]);
-    w2_t[1] = rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w2_t[1]);
-    w2_t[2] = rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w2_t[2]);
-    w2_t[3] = rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w2_t[3]);
-    w3_t[0] = rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w3_t[0]);
-    w3_t[1] = rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w3_t[1]);
-    w3_t[2] = rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t[2]);
-    w3_t[3] = rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w3_t[3]);
-    w0_t[0] = rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t[0]);
-    w0_t[1] = rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w0_t[1]);
-    w0_t[2] = rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t[2]);
-    w0_t[3] = rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w0_t[3]);
-    w1_t[0] = rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w1_t[0]);
-    w1_t[1] = rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w1_t[1]);
-    w1_t[2] = rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t[2]);
-    w1_t[3] = rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w1_t[3]);
+    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w4_t);
+    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w5_t);
+    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w6_t);
+    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w7_t);
+    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w8_t);
+    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w9_t);
+    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wa_t);
+    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wb_t);
+    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wc_t);
+    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wd_t);
+    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, we_t);
+    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wf_t);
+    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t);
+    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t);
+    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t);
+    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t);
+    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w4_t);
+    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w5_t);
+    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w6_t);
+    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w7_t);
 
     #undef K
     #define K SHA1C02
 
-    w2_t[0] = rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w2_t[0]);
-    w2_t[1] = rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w2_t[1]);
-    w2_t[2] = rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w2_t[2]);
-    w2_t[3] = rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w2_t[3]);
-    w3_t[0] = rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w3_t[0]);
-    w3_t[1] = rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w3_t[1]);
-    w3_t[2] = rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w3_t[2]);
-    w3_t[3] = rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w3_t[3]);
-    w0_t[0] = rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w0_t[0]);
-    w0_t[1] = rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w0_t[1]);
-    w0_t[2] = rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w0_t[2]);
-    w0_t[3] = rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w0_t[3]);
-    w1_t[0] = rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w1_t[0]);
-    w1_t[1] = rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w1_t[1]);
-    w1_t[2] = rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w1_t[2]);
-    w1_t[3] = rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w1_t[3]);
-    w2_t[0] = rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w2_t[0]);
-    w2_t[1] = rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w2_t[1]);
-    w2_t[2] = rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w2_t[2]);
-    w2_t[3] = rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w2_t[3]);
+    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w8_t);
+    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w9_t);
+    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wa_t);
+    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wb_t);
+    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wc_t);
+    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, wd_t);
+    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, we_t);
+    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wf_t);
+    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w0_t);
+    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w1_t);
+    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w2_t);
+    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w3_t);
+    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w4_t);
+    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w5_t);
+    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w6_t);
+    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w7_t);
+    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w8_t);
+    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w9_t);
+    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wa_t);
+    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wb_t);
 
     #undef K
     #define K SHA1C03
 
-    w3_t[0] = rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t[0]);
-    w3_t[1] = rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w3_t[1]);
-    w3_t[2] = rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t[2]);
-    w3_t[3] = rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w3_t[3]);
-    w0_t[0] = rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t[0]);
-    w0_t[1] = rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w0_t[1]);
-    w0_t[2] = rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w0_t[2]);
-    w0_t[3] = rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t[3]);
-    w1_t[0] = rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t[0]);
-    w1_t[1] = rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w1_t[1]);
-    w1_t[2] = rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t[2]);
-    w1_t[3] = rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w1_t[3]);
-    w2_t[0] = rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w2_t[0]);
-    w2_t[1] = rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w2_t[1]);
-    w2_t[2] = rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t[2]);
-    w2_t[3] = rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w2_t[3]);
-    w3_t[0] = rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w3_t[0]);
-    w3_t[1] = rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t[1]);
-    w3_t[2] = rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w3_t[2]);
-    w3_t[3] = rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w3_t[3]);
+    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wc_t);
+    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wd_t);
+    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, we_t);
+    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wf_t);
+    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t);
+    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t);
+    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w2_t);
+    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t);
+    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w4_t);
+    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w5_t);
+    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w6_t);
+    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w7_t);
+    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w8_t);
+    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w9_t);
+    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wa_t);
+    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wb_t);
+    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wc_t);
+    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wd_t);
+    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, we_t);
+    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wf_t);
 
     COMPARE_M_SIMD (d, e, c, b);
   }
@@ -344,33 +329,22 @@ static void m04700s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
     const u32x w0lr = w0l | w0r;
 
-    u32x w0_t[4];
-
-    w0_t[0] = w0lr;
-    w0_t[1] = w0[1];
-    w0_t[2] = w0[2];
-    w0_t[3] = w0[3];
-
-    u32x w1_t[4];
-
-    w1_t[0] = w1[0];
-    w1_t[1] = w1[1];
-    w1_t[2] = w1[2];
-    w1_t[3] = w1[3];
-
-    u32x w2_t[4];
-
-    w2_t[0] = w2[0];
-    w2_t[1] = w2[1];
-    w2_t[2] = w2[2];
-    w2_t[3] = w2[3];
-
-    u32x w3_t[4];
-
-    w3_t[0] = w3[0];
-    w3_t[1] = w3[1];
-    w3_t[2] = w3[2];
-    w3_t[3] = w3[3];
+    u32x w0_t = w0lr;
+    u32x w1_t = w0[1];
+    u32x w2_t = w0[2];
+    u32x w3_t = w0[3];
+    u32x w4_t = w1[0];
+    u32x w5_t = w1[1];
+    u32x w6_t = w1[2];
+    u32x w7_t = w1[3];
+    u32x w8_t = w2[0];
+    u32x w9_t = w2[1];
+    u32x wa_t = w2[2];
+    u32x wb_t = w2[3];
+    u32x wc_t = w3[0];
+    u32x wd_t = w3[1];
+    u32x we_t = w3[2];
+    u32x wf_t = w3[3];
 
     /**
      * md5
@@ -380,74 +354,75 @@ static void m04700s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     u32x b = MD5M_B;
     u32x c = MD5M_C;
     u32x d = MD5M_D;
-
-    MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
-    MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
-    MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
-    MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
-    MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
-    MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
-    MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
-    MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
-    MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
-    MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
-    MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
-    MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
-    MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
-    MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
-    MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
-    MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);
-
-    MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
-    MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
-    MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
-    MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
-    MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
-    MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
-    MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
-    MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
-    MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
-    MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
-    MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
-    MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
-    MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
-    MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
-    MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
-    MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
-
-    MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
-    MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
-    MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
-    MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
-
-    MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
-    MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
-    MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
-    MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
-    MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
-    MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
-    MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
-    MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
-    MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
-    MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
-    MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
-    MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
-    MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
-    MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
-    MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
-    MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);
+    u32x e = 0;
+
+    MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
+    MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
+    MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
+    MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
+
+    MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
+    MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
+    MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
+    MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
+
+    MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
+    MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
+    MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
+    MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
+
+    MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
+    MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
+    MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
+    MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
 
     a += MD5M_A;
     b += MD5M_B;
@@ -458,34 +433,30 @@ static void m04700s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
      * sha1
      */
 
-    w0_t[0] = uint_to_hex_lower8_le ((a >>  8) & 255) <<  0
-            | uint_to_hex_lower8_le ((a >>  0) & 255) << 16;
-    w0_t[1] = uint_to_hex_lower8_le ((a >> 24) & 255) <<  0
-            | uint_to_hex_lower8_le ((a >> 16) & 255) << 16;
-    w0_t[2] = uint_to_hex_lower8_le ((b >>  8) & 255) <<  0
-            | uint_to_hex_lower8_le ((b >>  0) & 255) << 16;
-    w0_t[3] = uint_to_hex_lower8_le ((b >> 24) & 255) <<  0
-            | uint_to_hex_lower8_le ((b >> 16) & 255) << 16;
-    w1_t[0] = uint_to_hex_lower8_le ((c >>  8) & 255) <<  0
-            | uint_to_hex_lower8_le ((c >>  0) & 255) << 16;
-    w1_t[1] = uint_to_hex_lower8_le ((c >> 24) & 255) <<  0
-            | uint_to_hex_lower8_le ((c >> 16) & 255) << 16;
-    w1_t[2] = uint_to_hex_lower8_le ((d >>  8) & 255) <<  0
-            | uint_to_hex_lower8_le ((d >>  0) & 255) << 16;
-    w1_t[3] = uint_to_hex_lower8_le ((d >> 24) & 255) <<  0
-            | uint_to_hex_lower8_le ((d >> 16) & 255) << 16;
-
-    w2_t[0] = 0x80000000;
-    w2_t[1] = 0;
-    w2_t[2] = 0;
-    w2_t[3] = 0;
-
-    w3_t[0] = 0;
-    w3_t[1] = 0;
-    w3_t[2] = 0;
-    w3_t[3] = 32 * 8;
-
-    u32x e;
+    w0_t = uint_to_hex_lower8_le ((a >>  8) & 255) <<  0
+         | uint_to_hex_lower8_le ((a >>  0) & 255) << 16;
+    w1_t = uint_to_hex_lower8_le ((a >> 24) & 255) <<  0
+         | uint_to_hex_lower8_le ((a >> 16) & 255) << 16;
+    w2_t = uint_to_hex_lower8_le ((b >>  8) & 255) <<  0
+         | uint_to_hex_lower8_le ((b >>  0) & 255) << 16;
+    w3_t = uint_to_hex_lower8_le ((b >> 24) & 255) <<  0
+         | uint_to_hex_lower8_le ((b >> 16) & 255) << 16;
+    w4_t = uint_to_hex_lower8_le ((c >>  8) & 255) <<  0
+         | uint_to_hex_lower8_le ((c >>  0) & 255) << 16;
+    w5_t = uint_to_hex_lower8_le ((c >> 24) & 255) <<  0
+         | uint_to_hex_lower8_le ((c >> 16) & 255) << 16;
+    w6_t = uint_to_hex_lower8_le ((d >>  8) & 255) <<  0
+         | uint_to_hex_lower8_le ((d >>  0) & 255) << 16;
+    w7_t = uint_to_hex_lower8_le ((d >> 24) & 255) <<  0
+         | uint_to_hex_lower8_le ((d >> 16) & 255) << 16;
+    w8_t = 0x80000000;
+    w9_t = 0;
+    wa_t = 0;
+    wb_t = 0;
+    wc_t = 0;
+    wd_t = 0;
+    we_t = 0;
+    wf_t = 32 * 8;
 
     a = SHA1M_A;
     b = SHA1M_B;
@@ -496,107 +467,106 @@ static void m04700s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     #undef K
     #define K SHA1C00
 
-    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w0_t[0]);
-    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t[1]);
-    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w0_t[2]);
-    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w0_t[3]);
-    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w1_t[0]);
-    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w1_t[1]);
-    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w1_t[2]);
-    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w1_t[3]);
-    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w2_t[0]);
-    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w2_t[1]);
-    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w2_t[2]);
-    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w2_t[3]);
-    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w3_t[0]);
-    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w3_t[1]);
-    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w3_t[2]);
-    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w3_t[3]);
-    w0_t[0] = rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t[0]);
-    w0_t[1] = rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F0o, d, e, a, b, c, w0_t[1]);
-    w0_t[2] = rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F0o, c, d, e, a, b, w0_t[2]);
-    w0_t[3] = rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F0o, b, c, d, e, a, w0_t[3]);
+    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w0_t);
+    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w1_t);
+    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w2_t);
+    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w3_t);
+    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w4_t);
+    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w5_t);
+    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w6_t);
+    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w7_t);
+    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w8_t);
+    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w9_t);
+    SHA1_STEP (SHA1_F0o, a, b, c, d, e, wa_t);
+    SHA1_STEP (SHA1_F0o, e, a, b, c, d, wb_t);
+    SHA1_STEP (SHA1_F0o, d, e, a, b, c, wc_t);
+    SHA1_STEP (SHA1_F0o, c, d, e, a, b, wd_t);
+    SHA1_STEP (SHA1_F0o, b, c, d, e, a, we_t);
+    SHA1_STEP (SHA1_F0o, a, b, c, d, e, wf_t);
+    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t);
+    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F0o, d, e, a, b, c, w1_t);
+    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F0o, c, d, e, a, b, w2_t);
+    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F0o, b, c, d, e, a, w3_t);
 
     #undef K
     #define K SHA1C01
 
-    w1_t[0] = rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t[0]);
-    w1_t[1] = rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w1_t[1]);
-    w1_t[2] = rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w1_t[2]);
-    w1_t[3] = rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t[3]);
-    w2_t[0] = rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t[0]);
-    w2_t[1] = rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w2_t[1]);
-    w2_t[2] = rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w2_t[2]);
-    w2_t[3] = rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w2_t[3]);
-    w3_t[0] = rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w3_t[0]);
-    w3_t[1] = rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w3_t[1]);
-    w3_t[2] = rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t[2]);
-    w3_t[3] = rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w3_t[3]);
-    w0_t[0] = rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t[0]);
-    w0_t[1] = rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w0_t[1]);
-    w0_t[2] = rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t[2]);
-    w0_t[3] = rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w0_t[3]);
-    w1_t[0] = rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w1_t[0]);
-    w1_t[1] = rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w1_t[1]);
-    w1_t[2] = rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t[2]);
-    w1_t[3] = rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w1_t[3]);
+    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w4_t);
+    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w5_t);
+    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w6_t);
+    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w7_t);
+    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w8_t);
+    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w9_t);
+    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wa_t);
+    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wb_t);
+    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wc_t);
+    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wd_t);
+    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, we_t);
+    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wf_t);
+    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t);
+    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t);
+    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t);
+    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t);
+    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w4_t);
+    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w5_t);
+    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w6_t);
+    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w7_t);
 
     #undef K
     #define K SHA1C02
 
-    w2_t[0] = rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w2_t[0]);
-    w2_t[1] = rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w2_t[1]);
-    w2_t[2] = rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w2_t[2]);
-    w2_t[3] = rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w2_t[3]);
-    w3_t[0] = rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w3_t[0]);
-    w3_t[1] = rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w3_t[1]);
-    w3_t[2] = rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w3_t[2]);
-    w3_t[3] = rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w3_t[3]);
-    w0_t[0] = rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w0_t[0]);
-    w0_t[1] = rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w0_t[1]);
-    w0_t[2] = rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w0_t[2]);
-    w0_t[3] = rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w0_t[3]);
-    w1_t[0] = rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w1_t[0]);
-    w1_t[1] = rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w1_t[1]);
-    w1_t[2] = rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w1_t[2]);
-    w1_t[3] = rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w1_t[3]);
-    w2_t[0] = rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w2_t[0]);
-    w2_t[1] = rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w2_t[1]);
-    w2_t[2] = rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w2_t[2]);
-    w2_t[3] = rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w2_t[3]);
+    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w8_t);
+    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w9_t);
+    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wa_t);
+    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wb_t);
+    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wc_t);
+    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, wd_t);
+    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, we_t);
+    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wf_t);
+    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w0_t);
+    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w1_t);
+    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w2_t);
+    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w3_t);
+    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w4_t);
+    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w5_t);
+    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w6_t);
+    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w7_t);
+    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w8_t);
+    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w9_t);
+    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wa_t);
+    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wb_t);
 
     #undef K
     #define K SHA1C03
 
-    w3_t[0] = rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t[0]);
-    w3_t[1] = rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w3_t[1]);
-    w3_t[2] = rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t[2]);
-    w3_t[3] = rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w3_t[3]);
-    w0_t[0] = rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t[0]);
-    w0_t[1] = rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w0_t[1]);
-    w0_t[2] = rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w0_t[2]);
-    w0_t[3] = rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t[3]);
-    w1_t[0] = rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t[0]);
-    w1_t[1] = rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w1_t[1]);
-    w1_t[2] = rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t[2]);
-    w1_t[3] = rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w1_t[3]);
-    w2_t[0] = rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w2_t[0]);
-    w2_t[1] = rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w2_t[1]);
-    w2_t[2] = rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t[2]);
-    w2_t[3] = rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w2_t[3]);
+    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wc_t);
+    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wd_t);
+    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, we_t);
+    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wf_t);
+    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t);
+    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t);
+    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w2_t);
+    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t);
+    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w4_t);
+    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w5_t);
+    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w6_t);
+    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w7_t);
+    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w8_t);
+    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w9_t);
+    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wa_t);
+    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wb_t);
 
     if (MATCHES_NONE_VS (e, e_rev)) continue;
 
-    w3_t[0] = rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w3_t[0]);
-    w3_t[1] = rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t[1]);
-    w3_t[2] = rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w3_t[2]);
-    w3_t[3] = rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w3_t[3]);
+    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wc_t);
+    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wd_t);
+    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, we_t);
+    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wf_t);
 
     COMPARE_S_SIMD (d, e, c, b);
   }
 }
 
-
 __kernel void m04700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, 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 u32 combs_mode, const u32 gid_max)
 {
   /**