From 861e6440574e7a4e6711acea7f7a2c4af4cac745 Mon Sep 17 00:00:00 2001
From: Gabriele Gristina <matrix@users.noreply.github.com>
Date: Fri, 24 Dec 2021 17:30:49 +0100
Subject: [PATCH] OpenCL Backend: added workaround to make optimized kernels
 work on Apple Silicon

---
 OpenCL/inc_rp_optimized.cl | 80 +++++++++++++++++------------------
 OpenCL/inc_rp_optimized.h  | 86 ++++++++++++++++++++------------------
 docs/changes.txt           |  1 +
 include/shared.h           |  4 ++
 src/backend.c              | 21 +++++-----
 src/shared.c               | 18 ++++++++
 6 files changed, 119 insertions(+), 91 deletions(-)

diff --git a/OpenCL/inc_rp_optimized.cl b/OpenCL/inc_rp_optimized.cl
index 38239c0ec..8ebe588ea 100644
--- a/OpenCL/inc_rp_optimized.cl
+++ b/OpenCL/inc_rp_optimized.cl
@@ -1046,7 +1046,7 @@ DECLSPEC void exchange_byte_optimized (u32 *buf, const int off_src, const int of
   ptr[off_dst] = tmp;
 }
 
-DECLSPEC u32 rule_op_mangle_lrest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_lrest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   u32 t;
 
@@ -1062,7 +1062,7 @@ DECLSPEC u32 rule_op_mangle_lrest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_urest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_urest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   u32 t;
 
@@ -1078,7 +1078,7 @@ DECLSPEC u32 rule_op_mangle_urest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_lrest_ufirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_lrest_ufirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   u32 t;
 
@@ -1096,7 +1096,7 @@ DECLSPEC u32 rule_op_mangle_lrest_ufirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSE
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_urest_lfirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_urest_lfirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   u32 t;
 
@@ -1114,7 +1114,7 @@ DECLSPEC u32 rule_op_mangle_urest_lfirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSE
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_trest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_trest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   u32 t;
 
@@ -1130,7 +1130,7 @@ DECLSPEC u32 rule_op_mangle_trest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_toggle_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_toggle_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -1163,7 +1163,7 @@ DECLSPEC u32 rule_op_mangle_toggle_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED c
   return (in_len);
 }
 
-DECLSPEC u32 rule_op_mangle_toggle_at_sep (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_toggle_at_sep (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (in_len == 0) return in_len;
 
@@ -1239,14 +1239,14 @@ DECLSPEC u32 rule_op_mangle_toggle_at_sep (MAYBE_UNUSED const u32 p0, MAYBE_UNUS
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_reverse (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_reverse (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   reverse_block_optimized (buf0, buf1, buf0, buf1, in_len);
 
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_dupeword (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_dupeword (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if ((in_len + in_len) >= 32) return in_len;
 
@@ -1259,7 +1259,7 @@ DECLSPEC u32 rule_op_mangle_dupeword (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED co
   return out_len;
 }
 
-DECLSPEC u32 rule_op_mangle_dupeword_times (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_dupeword_times (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (((in_len * p0) + in_len) >= 32) return in_len;
 
@@ -1287,7 +1287,7 @@ DECLSPEC u32 rule_op_mangle_dupeword_times (MAYBE_UNUSED const u32 p0, MAYBE_UNU
   return out_len;
 }
 
-DECLSPEC u32 rule_op_mangle_reflect (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_reflect (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if ((in_len + in_len) >= 32) return in_len;
 
@@ -1305,7 +1305,7 @@ DECLSPEC u32 rule_op_mangle_reflect (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED con
   return out_len;
 }
 
-DECLSPEC u32 rule_op_mangle_append (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_append (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if ((in_len + 1) >= 32) return in_len;
 
@@ -1318,7 +1318,7 @@ DECLSPEC u32 rule_op_mangle_append (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED cons
   return out_len;
 }
 
-DECLSPEC u32 rule_op_mangle_prepend (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_prepend (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if ((in_len + 1) >= 32) return in_len;
 
@@ -1333,7 +1333,7 @@ DECLSPEC u32 rule_op_mangle_prepend (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED con
   return out_len;
 }
 
-DECLSPEC u32 rule_op_mangle_rotate_left (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_rotate_left (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (in_len == 0) return in_len;
 
@@ -1348,7 +1348,7 @@ DECLSPEC u32 rule_op_mangle_rotate_left (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_rotate_right (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_rotate_right (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (in_len == 0) return in_len;
 
@@ -1390,7 +1390,7 @@ DECLSPEC u32 rule_op_mangle_rotate_right (MAYBE_UNUSED const u32 p0, MAYBE_UNUSE
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_delete_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_delete_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (in_len == 0) return in_len;
 
@@ -1401,7 +1401,7 @@ DECLSPEC u32 rule_op_mangle_delete_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSE
   return in_len1;
 }
 
-DECLSPEC u32 rule_op_mangle_delete_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_delete_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (in_len == 0) return in_len;
 
@@ -1421,7 +1421,7 @@ DECLSPEC u32 rule_op_mangle_delete_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED
   return in_len1;
 }
 
-DECLSPEC u32 rule_op_mangle_delete_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_delete_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -1498,7 +1498,7 @@ DECLSPEC u32 rule_op_mangle_delete_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED c
   return out_len;
 }
 
-DECLSPEC u32 rule_op_mangle_extract (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_extract (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -1513,7 +1513,7 @@ DECLSPEC u32 rule_op_mangle_extract (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED con
   return out_len;
 }
 
-DECLSPEC u32 rule_op_mangle_omit (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_omit (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -1601,7 +1601,7 @@ DECLSPEC u32 rule_op_mangle_omit (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const
   return out_len;
 }
 
-DECLSPEC u32 rule_op_mangle_insert (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_insert (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (p0 > in_len) return in_len;
 
@@ -1675,7 +1675,7 @@ DECLSPEC u32 rule_op_mangle_insert (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED cons
   return out_len;
 }
 
-DECLSPEC u32 rule_op_mangle_overstrike (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_overstrike (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -1710,7 +1710,7 @@ DECLSPEC u32 rule_op_mangle_overstrike (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_truncate_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_truncate_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -1743,7 +1743,7 @@ DECLSPEC u32 replace_on_register (const u32 in, const u32 r, const u32 p1)
   return out;
 }
 
-DECLSPEC u32 rule_op_mangle_replace (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_replace (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   const u32 r0 = search_on_register (buf0[0], p0);
   const u32 r1 = search_on_register (buf0[1], p0);
@@ -1770,7 +1770,7 @@ DECLSPEC u32 rule_op_mangle_replace (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED con
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_purgechar (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_purgechar (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   const u32 r0 = search_on_register (buf0[0], p0);
   const u32 r1 = search_on_register (buf0[1], p0);
@@ -1824,7 +1824,7 @@ DECLSPEC u32 rule_op_mangle_purgechar (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED c
   return out_len;
 }
 
-DECLSPEC u32 rule_op_mangle_dupechar_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_dupechar_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if ( in_len       ==  0) return in_len;
   if ((in_len + p0) >= 32) return in_len;
@@ -1859,7 +1859,7 @@ DECLSPEC u32 rule_op_mangle_dupechar_first (MAYBE_UNUSED const u32 p0, MAYBE_UNU
   return out_len;
 }
 
-DECLSPEC u32 rule_op_mangle_dupechar_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_dupechar_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if ( in_len       ==  0) return in_len;
   if ((in_len + p0) >= 32) return in_len;
@@ -1905,7 +1905,7 @@ DECLSPEC u32 rule_op_mangle_dupechar_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUS
   return out_len;
 }
 
-DECLSPEC u32 rule_op_mangle_dupechar_all (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_dupechar_all (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if ( in_len           ==  0) return in_len;
   if ((in_len + in_len) >= 32) return in_len;
@@ -1938,7 +1938,7 @@ DECLSPEC u32 rule_op_mangle_dupechar_all (MAYBE_UNUSED const u32 p0, MAYBE_UNUSE
   return out_len;
 }
 
-DECLSPEC u32 rule_op_mangle_switch_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_switch_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (in_len < 2) return in_len;
 
@@ -1947,7 +1947,7 @@ DECLSPEC u32 rule_op_mangle_switch_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSE
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_switch_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_switch_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (in_len < 2) return in_len;
 
@@ -1976,7 +1976,7 @@ DECLSPEC u32 rule_op_mangle_switch_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_switch_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_switch_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
   if (p1 >= in_len) return in_len;
@@ -2006,7 +2006,7 @@ DECLSPEC u32 rule_op_mangle_switch_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED c
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_chr_shiftl (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_chr_shiftl (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -2040,7 +2040,7 @@ DECLSPEC u32 rule_op_mangle_chr_shiftl (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_chr_shiftr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_chr_shiftr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -2074,7 +2074,7 @@ DECLSPEC u32 rule_op_mangle_chr_shiftr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_chr_incr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_chr_incr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -2110,7 +2110,7 @@ DECLSPEC u32 rule_op_mangle_chr_incr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED co
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_chr_decr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_chr_decr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (p0 >= in_len) return in_len;
 
@@ -2146,7 +2146,7 @@ DECLSPEC u32 rule_op_mangle_chr_decr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED co
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_replace_np1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_replace_np1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if ((p0 + 1) >= in_len) return in_len;
 
@@ -2186,7 +2186,7 @@ DECLSPEC u32 rule_op_mangle_replace_np1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_replace_nm1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_replace_nm1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (p0 == 0) return in_len;
 
@@ -2228,7 +2228,7 @@ DECLSPEC u32 rule_op_mangle_replace_nm1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED
   return in_len;
 }
 
-DECLSPEC u32 rule_op_mangle_dupeblock_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_dupeblock_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (p0 > in_len) return in_len;
 
@@ -2266,7 +2266,7 @@ DECLSPEC u32 rule_op_mangle_dupeblock_first (MAYBE_UNUSED const u32 p0, MAYBE_UN
   return out_len;
 }
 
-DECLSPEC u32 rule_op_mangle_dupeblock_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_dupeblock_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (p0 > in_len) return in_len;
 
@@ -2309,7 +2309,7 @@ DECLSPEC u32 toggle_on_register (const u32 in, const u32 r)
   return out;
 }
 
-DECLSPEC u32 rule_op_mangle_title_sep (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
+DECLSPEC HC_INLINE u32 rule_op_mangle_title_sep (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len)
 {
   if (in_len == 0) return in_len;
 
diff --git a/OpenCL/inc_rp_optimized.h b/OpenCL/inc_rp_optimized.h
index d391d1cfe..3df2bf8d2 100644
--- a/OpenCL/inc_rp_optimized.h
+++ b/OpenCL/inc_rp_optimized.h
@@ -14,6 +14,12 @@
 #define MAYBE_UNUSED
 #endif
 
+#ifdef IS_APPLE_SILICON
+#define HC_INLINE __attribute__ ((noinline))
+#else
+#define HC_INLINE
+#endif
+
 #define RULE_OP_MANGLE_NOOP             ':'
 #define RULE_OP_MANGLE_LREST            'l'
 #define RULE_OP_MANGLE_UREST            'u'
@@ -80,49 +86,49 @@ DECLSPEC void append_block1_optimized (const u32 offset, u32 *buf0, u32 *buf1, c
 DECLSPEC void append_block8_optimized (const u32 offset, u32 *buf0, u32 *buf1, const u32 *src_l0, const u32 *src_l1, const u32 *src_r0, const u32 *src_r1);
 DECLSPEC void reverse_block_optimized (u32 *in0, u32 *in1, u32 *out0, u32 *out1, const u32 len);
 DECLSPEC void exchange_byte_optimized (u32 *buf, const int off_src, const int off_dst);
-DECLSPEC u32 rule_op_mangle_lrest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_urest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_lrest_ufirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_urest_lfirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_trest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_toggle_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_toggle_at_sep (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_reverse (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_dupeword (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_dupeword_times (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_reflect (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_append (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_prepend (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_rotate_left (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_rotate_right (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_delete_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_delete_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_delete_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_extract (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_omit (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_insert (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_overstrike (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_truncate_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_lrest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_urest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_lrest_ufirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_urest_lfirst (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_trest (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_toggle_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_toggle_at_sep (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_reverse (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_dupeword (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_dupeword_times (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_reflect (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_append (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_prepend (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_rotate_left (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_rotate_right (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_delete_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_delete_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_delete_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_extract (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_omit (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_insert (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_overstrike (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_truncate_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
 DECLSPEC u32 search_on_register (const u32 in, const u32 p0);
 DECLSPEC u32 replace_on_register (const u32 in, const u32 r, const u32 p1);
-DECLSPEC u32 rule_op_mangle_replace (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_purgechar (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_dupechar_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_dupechar_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_dupechar_all (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_switch_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_switch_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_switch_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_chr_shiftl (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_chr_shiftr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_chr_incr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_chr_decr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_replace_np1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_replace_nm1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_dupeblock_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
-DECLSPEC u32 rule_op_mangle_dupeblock_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_replace (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_purgechar (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_dupechar_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_dupechar_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_dupechar_all (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_switch_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_switch_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_switch_at (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_chr_shiftl (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_chr_shiftr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_chr_incr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_chr_decr (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_replace_np1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_replace_nm1 (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_dupeblock_first (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_dupeblock_last (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
 DECLSPEC u32 toggle_on_register (const u32 in, const u32 r);
-DECLSPEC u32 rule_op_mangle_title_sep (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
+DECLSPEC HC_INLINE u32 rule_op_mangle_title_sep (MAYBE_UNUSED const u32 p0, MAYBE_UNUSED const u32 p1, MAYBE_UNUSED u32 *buf0, MAYBE_UNUSED u32 *buf1, const u32 in_len);
 DECLSPEC u32 apply_rule_optimized (const u32 name, const u32 p0, const u32 p1, u32 *buf0, u32 *buf1, const u32 in_len);
 DECLSPEC u32 apply_rules_optimized (CONSTANT_AS const u32 *cmds, u32 *buf0, u32 *buf1, const u32 len);
 DECLSPEC u32x apply_rules_vect_optimized (const u32 *pw_buf0, const u32 *pw_buf1, const u32 pw_len, CONSTANT_AS const kernel_rule_t *kernel_rules, const u32 il_pos, u32x *buf0, u32x *buf1);
diff --git a/docs/changes.txt b/docs/changes.txt
index 139ecf8af..623f76e8d 100644
--- a/docs/changes.txt
+++ b/docs/changes.txt
@@ -34,6 +34,7 @@
 - OpenCL Runtime: Added support to use Apple Silicon compute devices
 - OpenCL Runtime: Set default device-type to GPU with Apple Silicon compute devices
 - Backend Info: Added local memory size to output
+- OpenCL Backend: added workaround to make optimized kernels work on Apple Silicon
 
 * changes v6.2.4 -> v6.2.5
 
diff --git a/include/shared.h b/include/shared.h
index 8291ecfeb..2416d963c 100644
--- a/include/shared.h
+++ b/include/shared.h
@@ -103,4 +103,8 @@ int  generic_salt_encode (MAYBE_UNUSED const hashconfig_t *hashconfig, const u8
 
 int input_tokenizer (const u8 *input_buf, const int input_len, hc_token_t *token);
 
+#if defined (__APPLE__)
+bool is_apple_silicon (void);
+#endif
+
 #endif // _SHARED_H
diff --git a/src/backend.c b/src/backend.c
index 5b99a9371..036b03100 100644
--- a/src/backend.c
+++ b/src/backend.c
@@ -341,22 +341,15 @@ static bool setup_opencl_device_types_filter (hashcat_ctx_t *hashcat_ctx, const
   {
     #if defined (__APPLE__)
 
-    #include <sys/sysctl.h>
-
-    size_t size;
-    cpu_type_t cpu_type = 0;
-    size = sizeof (cpu_type);
-    sysctlbyname ("hw.cputype", &cpu_type, &size, NULL, 0);
-
-    if (cpu_type == 0x100000c)
+    if (is_apple_silicon() == true)
     {
-      // For apple M1* use GPU only, because CPU device it is not recognized by OpenCL
+      // With Apple's M1* use GPU only, because CPU device it is not recognized by OpenCL
 
       opencl_device_types_filter = CL_DEVICE_TYPE_GPU;
     }
     else
     {
-      // For apple use CPU only, because GPU drivers are not reliable
+      // With Apple Intel use CPU only, because GPU drivers are not reliable
       // The user can explicitly enable GPU by setting -D2
 
       //opencl_device_types_filter = CL_DEVICE_TYPE_ALL & ~CL_DEVICE_TYPE_GPU;
@@ -470,7 +463,6 @@ static bool opencl_test_instruction (hashcat_ctx_t *hashcat_ctx, cl_context cont
 
   OCL_PTR *ocl = (OCL_PTR *) backend_ctx->ocl;
 
-
   #ifndef DEBUG
   const int fd_stderr = fileno (stderr);
   const int stderr_bak = dup (fd_stderr);
@@ -11480,6 +11472,13 @@ int backend_session_begin (hashcat_ctx_t *hashcat_ctx)
       {
         build_options_len += snprintf (build_options_buf + build_options_len, build_options_sz - build_options_len, "-D KERNEL_STATIC -I OpenCL -I \"%s\" ", folder_config->cpath_real);
       }
+
+      #if defined (__APPLE__)
+      if (is_apple_silicon() == true)
+      {
+        build_options_len += snprintf (build_options_buf + build_options_len, build_options_sz - build_options_len, "-D IS_APPLE_SILICON ");
+      }
+      #endif
     }
 
     /* currently disabled, hangs NEO drivers since 20.09.
diff --git a/src/shared.c b/src/shared.c
index 26a79f625..2ac8917ca 100644
--- a/src/shared.c
+++ b/src/shared.c
@@ -15,6 +15,10 @@
 #include <sys/cygwin.h>
 #endif
 
+#if defined (__APPLE__)
+#include <sys/sysctl.h>
+#endif
+
 static const char *const PA_000 = "OK";
 static const char *const PA_001 = "Ignored due to comment";
 static const char *const PA_002 = "Ignored due to zero length";
@@ -1369,3 +1373,17 @@ int generic_salt_encode (MAYBE_UNUSED const hashconfig_t *hashconfig, const u8 *
 
   return tmp_len;
 }
+
+#if defined (__APPLE__)
+
+bool is_apple_silicon (void)
+{
+  size_t size;
+  cpu_type_t cpu_type = 0;
+  size = sizeof (cpu_type);
+  sysctlbyname ("hw.cputype", &cpu_type, &size, NULL, 0);
+
+  return (cpu_type == 0x100000c);
+}
+
+#endif // __APPLE__