jobcore/pcre/0001-pcre-JIT-compiler-update-for-Intel-CET.patch

From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: "H.J. Lu" <hjl.tools@gmail.com>
Date: Sun, 20 Dec 2020 13:32:56 -0800
Subject: [PATCH 1/2] JIT compiler update for Intel CET

---
 pcre_jit_compile.c              |  43 ++++---
 sljit/sljitConfig.h             |  17 ++-
 sljit/sljitConfigInternal.h     |  67 ++++++++--
 sljit/sljitExecAllocator.c      |  17 ++-
 sljit/sljitLir.c                |  95 +++++++++------
 sljit/sljitLir.h                |  92 ++++++++++----
 sljit/sljitNativeARM_32.c       |  64 ++++++----
 sljit/sljitNativeARM_64.c       |  75 +++++++-----
 sljit/sljitNativeARM_T2_32.c    |  56 +++++----
 sljit/sljitNativeMIPS_32.c      |  31 ++---
 sljit/sljitNativeMIPS_64.c      |  21 ++--
 sljit/sljitNativeMIPS_common.c  | 184 +++++++++++++++-------------
 sljit/sljitNativePPC_common.c   |  56 +++++----
 sljit/sljitNativeSPARC_common.c |  54 ++++++---
 sljit/sljitNativeTILEGX_64.c    |  44 ++++---
 sljit/sljitNativeX86_32.c       |  50 ++++++--
 sljit/sljitNativeX86_64.c       |  28 ++++-
 sljit/sljitNativeX86_common.c   | 209 ++++++++++++++++++++++++++++++--
 sljit/sljitUtils.c              | 157 +++++++++++++++++-------
 19 files changed, 953 insertions(+), 407 deletions(-)

diff --git a/pcre_jit_compile.c b/pcre_jit_compile.c
index 4dcf8fc..9e67a20 100644
--- a/pcre_jit_compile.c
+++ b/pcre_jit_compile.c
@@ -549,6 +549,8 @@ the start pointers when the end of the capturing group has not yet reached. */
   sljit_emit_op1(compiler, (op), (dst), (dstw), (src), (srcw))
 #define OP2(op, dst, dstw, src1, src1w, src2, src2w) \
   sljit_emit_op2(compiler, (op), (dst), (dstw), (src1), (src1w), (src2), (src2w))
+#define OP_SRC(op, src, srcw) \
+  sljit_emit_op_src(compiler, (op), (src), (srcw))
 #define LABEL() \
   sljit_emit_label(compiler)
 #define JUMP(type) \
@@ -3230,7 +3232,7 @@ jump = JUMP(SLJIT_NOT_ZERO);
 /* Two byte sequence. */
 OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
 OP1(SLJIT_MOV, TMP2, 0, SLJIT_IMM, IN_UCHARS(2));
-sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
+OP_SRC(SLJIT_FAST_RETURN, RETURN_ADDR, 0);
 
 JUMPHERE(jump);
 OP1(MOV_UCHAR, TMP2, 0, SLJIT_MEM1(STR_PTR), IN_UCHARS(1));
@@ -3244,7 +3246,7 @@ jump = JUMP(SLJIT_NOT_ZERO);
 /* Three byte sequence. */
 OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(2));
 OP1(SLJIT_MOV, TMP2, 0, SLJIT_IMM, IN_UCHARS(3));
-sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
+OP_SRC(SLJIT_FAST_RETURN, RETURN_ADDR, 0);
 
 /* Four byte sequence. */
 JUMPHERE(jump);
@@ -3255,7 +3257,7 @@ OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(3));
 OP2(SLJIT_AND, TMP2, 0, TMP2, 0, SLJIT_IMM, 0x3f);
 OP2(SLJIT_OR, TMP1, 0, TMP1, 0, TMP2, 0);
 OP1(SLJIT_MOV, TMP2, 0, SLJIT_IMM, IN_UCHARS(4));
-sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
+OP_SRC(SLJIT_FAST_RETURN, RETURN_ADDR, 0);
 }
 
 static void do_utfreadchar16(compiler_common *common)
@@ -3277,7 +3279,7 @@ OP2(SLJIT_AND | SLJIT_SET_Z, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x800);
 jump = JUMP(SLJIT_NOT_ZERO);
 /* Two byte sequence. */
 OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
-sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
+OP_SRC(SLJIT_FAST_RETURN, RETURN_ADDR, 0);
 
 JUMPHERE(jump);
 OP2(SLJIT_AND | SLJIT_SET_Z, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x400);
@@ -3291,7 +3293,7 @@ OP2(SLJIT_AND, TMP2, 0, TMP2, 0, SLJIT_IMM, 0x3f);
 OP2(SLJIT_OR, TMP1, 0, TMP1, 0, TMP2, 0);
 /* Three byte sequence. */
 OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(2));
-sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
+OP_SRC(SLJIT_FAST_RETURN, RETURN_ADDR, 0);
 }
 
 static void do_utfreadtype8(compiler_common *common)
@@ -3316,18 +3318,18 @@ OP2(SLJIT_SHL, TMP2, 0, TMP2, 0, SLJIT_IMM, 6);
 OP2(SLJIT_AND, TMP1, 0, TMP1, 0, SLJIT_IMM, 0x3f);
 OP2(SLJIT_OR, TMP2, 0, TMP2, 0, TMP1, 0);
 OP1(SLJIT_MOV_U8, TMP1, 0, SLJIT_MEM1(TMP2), common->ctypes);
-sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
+OP_SRC(SLJIT_FAST_RETURN, RETURN_ADDR, 0);
 
 JUMPHERE(compare);
 OP1(SLJIT_MOV, TMP1, 0, SLJIT_IMM, 0);
-sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
+OP_SRC(SLJIT_FAST_RETURN, RETURN_ADDR, 0);
 
 /* We only have types for characters less than 256. */
 JUMPHERE(jump);
 OP1(SLJIT_MOV_U8, TMP2, 0, SLJIT_MEM1(TMP2), (sljit_sw)PRIV(utf8_table4) - 0xc0);
 OP1(SLJIT_MOV, TMP1, 0, SLJIT_IMM, 0);
 OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, TMP2, 0);
-sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
+OP_SRC(SLJIT_FAST_RETURN, RETURN_ADDR, 0);
 }
 
 #endif /* COMPILE_PCRE8 */
@@ -3378,7 +3380,7 @@ OP1(SLJIT_MOV, TMP2, 0, SLJIT_IMM, (sljit_sw)PRIV(ucd_stage2));
 OP1(SLJIT_MOV_U16, TMP2, 0, SLJIT_MEM2(TMP2, TMP1), 1);
 OP1(SLJIT_MOV, TMP1, 0, SLJIT_IMM, (sljit_sw)PRIV(ucd_records) + SLJIT_OFFSETOF(ucd_record, chartype));
 OP1(SLJIT_MOV_U8, TMP1, 0, SLJIT_MEM2(TMP1, TMP2), 3);
-sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
+OP_SRC(SLJIT_FAST_RETURN, RETURN_ADDR, 0);
 }
 #endif
 
@@ -4868,7 +4870,7 @@ JUMPHERE(jump);
 jump = CMP(SLJIT_NOT_ZERO /* SIG_LESS */, TMP2, 0, SLJIT_IMM, 0);
 /* End of reverting values. */
 OP1(SLJIT_MOV, STACK_TOP, 0, TMP3, 0);
-sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
+OP_SRC(SLJIT_FAST_RETURN, RETURN_ADDR, 0);
 
 JUMPHERE(jump);
 OP1(SLJIT_NEG, TMP2, 0, TMP2, 0);
@@ -4984,7 +4986,7 @@ else
 set_jumps(skipread_list, LABEL());
 
 OP2(SLJIT_XOR | SLJIT_SET_Z, SLJIT_UNUSED, 0, TMP2, 0, SLJIT_MEM1(SLJIT_SP), LOCALS1);
-sljit_emit_fast_return(compiler, SLJIT_MEM1(SLJIT_SP), LOCALS0);
+OP_SRC(SLJIT_FAST_RETURN, SLJIT_MEM1(SLJIT_SP), LOCALS0);
 }
 
 static BOOL check_class_ranges(compiler_common *common, const sljit_u8 *bits, BOOL nclass, BOOL invert, jump_list **backtracks)
@@ -5163,7 +5165,7 @@ if (common->utf)
 #endif
 #endif /* SUPPORT_UTF || COMPILE_PCRE16 || COMPILE_PCRE32 */
 OP_FLAGS(SLJIT_OR | SLJIT_SET_Z, TMP2, 0, SLJIT_EQUAL);
-sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
+OP_SRC(SLJIT_FAST_RETURN, RETURN_ADDR, 0);
 }
 
 static void check_hspace(compiler_common *common)
@@ -5202,7 +5204,7 @@ if (common->utf)
 #endif /* SUPPORT_UTF || COMPILE_PCRE16 || COMPILE_PCRE32 */
 OP_FLAGS(SLJIT_OR | SLJIT_SET_Z, TMP2, 0, SLJIT_EQUAL);
 
-sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
+OP_SRC(SLJIT_FAST_RETURN, RETURN_ADDR, 0);
 }
 
 static void check_vspace(compiler_common *common)
@@ -5230,7 +5232,7 @@ if (common->utf)
 #endif /* SUPPORT_UTF || COMPILE_PCRE16 || COMPILE_PCRE32 */
 OP_FLAGS(SLJIT_OR | SLJIT_SET_Z, TMP2, 0, SLJIT_EQUAL);
 
-sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
+OP_SRC(SLJIT_FAST_RETURN, RETURN_ADDR, 0);
 }
 
 static void do_casefulcmp(compiler_common *common)
@@ -5310,7 +5312,7 @@ if (char1_reg == STR_END)
   OP1(SLJIT_MOV, char2_reg, 0, RETURN_ADDR, 0);
   }
 
-sljit_emit_fast_return(compiler, TMP1, 0);
+OP_SRC(SLJIT_FAST_RETURN, TMP1, 0);
 }
 
 static void do_caselesscmp(compiler_common *common)
@@ -5408,7 +5410,7 @@ if (char2_reg == STACK_TOP)
   }
 
 OP1(SLJIT_MOV, char1_reg, 0, SLJIT_MEM1(SLJIT_SP), LOCALS1);
-sljit_emit_fast_return(compiler, TMP1, 0);
+OP_SRC(SLJIT_FAST_RETURN, TMP1, 0);
 }
 
 #if defined SUPPORT_UTF && defined SUPPORT_UCP
@@ -10221,6 +10223,7 @@ else if (has_alternatives)
       return;
     sljit_emit_ijump(compiler, SLJIT_JUMP, SLJIT_MEM1(TMP1), (sljit_sw)next_update_addr);
     add_label_addr(common, next_update_addr++);
+    sljit_emit_op0(compiler, SLJIT_ENDBR);
     }
   else
     {
@@ -10352,7 +10355,10 @@ if (has_alternatives)
     if (opcode != OP_ONCE)
       {
       if (alt_max > 4)
+        {
         add_label_addr(common, next_update_addr++);
+        sljit_emit_op0(compiler, SLJIT_ENDBR);
+        }
       else
         {
         if (alt_count != 2 * sizeof(sljit_uw))
@@ -10960,7 +10966,7 @@ else
   OP1(SLJIT_MOV, TMP1, 0, TMP3, 0);
   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_SP), common->recursive_head_ptr, TMP2, 0);
   }
-sljit_emit_fast_return(compiler, SLJIT_MEM1(STACK_TOP), STACK(-1));
+OP_SRC(SLJIT_FAST_RETURN, SLJIT_MEM1(STACK_TOP), STACK(-1));
 }
 
 #undef COMPILE_BACKTRACKINGPATH
@@ -11308,6 +11314,7 @@ if (common->forced_quit != NULL)
   set_jumps(common->forced_quit, common->forced_quit_label);
 if (minlength_check_failed != NULL)
   SET_LABEL(minlength_check_failed, common->forced_quit_label);
+sljit_emit_op0(compiler, SLJIT_SKIP_FRAMES_BEFORE_RETURN);
 sljit_emit_return(compiler, SLJIT_MOV, SLJIT_RETURN_REG, 0);
 
 if (mode != JIT_COMPILE)
@@ -11433,7 +11440,7 @@ OP1(SLJIT_MOV, TMP2, 0, STACK_LIMIT, 0);
 OP1(SLJIT_MOV, STACK_LIMIT, 0, SLJIT_RETURN_REG, 0);
 OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(SLJIT_SP), LOCALS0);
 OP1(SLJIT_MOV, STACK_TOP, 0, SLJIT_MEM1(SLJIT_SP), LOCALS1);
-sljit_emit_fast_return(compiler, TMP1, 0);
+OP_SRC(SLJIT_FAST_RETURN, TMP1, 0);
 
 /* Allocation failed. */
 JUMPHERE(jump);
diff --git a/sljit/sljitConfig.h b/sljit/sljitConfig.h
index d54b5e6..4560450 100644
--- a/sljit/sljitConfig.h
+++ b/sljit/sljitConfig.h
@@ -27,6 +27,10 @@
 #ifndef _SLJIT_CONFIG_H_
 #define _SLJIT_CONFIG_H_
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 /* --------------------------------------------------------------------- */
 /*  Custom defines                                                       */
 /* --------------------------------------------------------------------- */
@@ -65,12 +69,19 @@
 #define SLJIT_UTIL_GLOBAL_LOCK 1
 #endif
 
-/* Implements a stack like data structure (by using mmap / VirtualAlloc). */
+/* Implements a stack like data structure (by using mmap / VirtualAlloc  */
+/* or a custom allocator). */
 #ifndef SLJIT_UTIL_STACK
 /* Enabled by default */
 #define SLJIT_UTIL_STACK 1
 #endif
 
+/* Uses user provided allocator to allocate the stack (see SLJIT_UTIL_STACK) */
+#ifndef SLJIT_UTIL_SIMPLE_STACK_ALLOCATION
+/* Disabled by default */
+#define SLJIT_UTIL_SIMPLE_STACK_ALLOCATION 0
+#endif
+
 /* Single threaded application. Does not require any locks. */
 #ifndef SLJIT_SINGLE_THREADED
 /* Disabled by default. */
@@ -144,4 +155,8 @@
 
 /* For further configurations, see the beginning of sljitConfigInternal.h */
 
+#ifdef __cplusplus
+} /* extern "C" */
+#endif
+
 #endif
diff --git a/sljit/sljitConfigInternal.h b/sljit/sljitConfigInternal.h
index acba9da..c81b6a4 100644
--- a/sljit/sljitConfigInternal.h
+++ b/sljit/sljitConfigInternal.h
@@ -27,6 +27,20 @@
 #ifndef _SLJIT_CONFIG_INTERNAL_H_
 #define _SLJIT_CONFIG_INTERNAL_H_
 
+#if (defined SLJIT_VERBOSE && SLJIT_VERBOSE) \
+	|| (defined SLJIT_DEBUG && SLJIT_DEBUG && (!defined(SLJIT_ASSERT) || !defined(SLJIT_UNREACHABLE)))
+#include <stdio.h>
+#endif
+
+#if (defined SLJIT_DEBUG && SLJIT_DEBUG \
+	&& (!defined(SLJIT_ASSERT) || !defined(SLJIT_UNREACHABLE) || !defined(SLJIT_HALT_PROCESS)))
+#include <stdlib.h>
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 /*
    SLJIT defines the following architecture dependent types and macros:
 
@@ -191,6 +205,24 @@
 #define SLJIT_CONFIG_SPARC 1
 #endif
 
+/***********************************************************/
+/* Intel Control-flow Enforcement Technology (CET) spport. */
+/***********************************************************/
+
+#ifdef SLJIT_CONFIG_X86
+#if defined(__CET__)
+#define SLJIT_CONFIG_X86_CET 1
+#endif
+#if (defined SLJIT_CONFIG_X86_CET && SLJIT_CONFIG_X86_CET)
+#if defined(__GNUC__)
+#if !defined (__SHSTK__)
+#error "-mshstk is needed to compile with -fcf-protection"
+#endif
+#include <x86intrin.h>
+#endif
+#endif
+#endif
+
 /**********************************/
 /* External function definitions. */
 /**********************************/
@@ -287,7 +319,7 @@
 #if __has_builtin(__builtin___clear_cache)
 
 #define SLJIT_CACHE_FLUSH(from, to) \
-	__builtin___clear_cache((char*)from, (char*)to)
+	__builtin___clear_cache((char*)(from), (char*)(to))
 
 #endif /* __has_builtin(__builtin___clear_cache) */
 #endif /* (!defined SLJIT_CACHE_FLUSH && defined __has_builtin) */
@@ -318,7 +350,7 @@
 #elif (defined(__GNUC__) && (__GNUC__ >= 5 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3)))
 
 #define SLJIT_CACHE_FLUSH(from, to) \
-	__builtin___clear_cache((char*)from, (char*)to)
+	__builtin___clear_cache((char*)(from), (char*)(to))
 
 #elif defined __ANDROID__
 
@@ -451,6 +483,25 @@ typedef double sljit_f64;
 #define SLJIT_BIG_ENDIAN 1
 #endif
 
+#ifndef SLJIT_MIPS_REV
+
+/* Auto detecting mips revision. */
+#if (defined __mips_isa_rev) && (__mips_isa_rev >= 6)
+#define SLJIT_MIPS_REV 6
+#elif (defined __mips_isa_rev && __mips_isa_rev >= 1) \
+	|| (defined __clang__ && defined _MIPS_ARCH_OCTEON) \
+	|| (defined __clang__ && defined _MIPS_ARCH_P5600)
+/* clang either forgets to define (clang-7) __mips_isa_rev at all
+ * or sets it to zero (clang-8,-9) for -march=octeon (MIPS64 R2+)
+ * and -march=p5600 (MIPS32 R5).
+ * It also sets the __mips macro to 64 or 32 for -mipsN when N <= 5
+ * (should be set to N exactly) so we cannot rely on this too.
+ */
+#define SLJIT_MIPS_REV 1
+#endif
+
+#endif /* !SLJIT_MIPS_REV */
+
 #elif (defined SLJIT_CONFIG_SPARC_32 && SLJIT_CONFIG_SPARC_32)
 
 #define SLJIT_BIG_ENDIAN 1
@@ -679,24 +730,16 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_sw sljit_exec_offset(void* ptr);
 /* Debug and verbose related macros. */
 /*************************************/
 
-#if (defined SLJIT_VERBOSE && SLJIT_VERBOSE)
-#include <stdio.h>
-#endif
-
 #if (defined SLJIT_DEBUG && SLJIT_DEBUG)
 
 #if !defined(SLJIT_ASSERT) || !defined(SLJIT_UNREACHABLE)
 
 /* SLJIT_HALT_PROCESS must halt the process. */
 #ifndef SLJIT_HALT_PROCESS
-#include <stdlib.h>
-
 #define SLJIT_HALT_PROCESS() \
 	abort();
 #endif /* !SLJIT_HALT_PROCESS */
 
-#include <stdio.h>
-
 #endif /* !SLJIT_ASSERT || !SLJIT_UNREACHABLE */
 
 /* Feel free to redefine these two macros. */
@@ -742,4 +785,8 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_sw sljit_exec_offset(void* ptr);
 
 #endif /* !SLJIT_COMPILE_ASSERT */
 
+#ifdef __cplusplus
+} /* extern "C" */
+#endif
+
 #endif
diff --git a/sljit/sljitExecAllocator.c b/sljit/sljitExecAllocator.c
index 92ddb94..7653907 100644
--- a/sljit/sljitExecAllocator.c
+++ b/sljit/sljitExecAllocator.c
@@ -106,10 +106,10 @@ static SLJIT_INLINE void free_chunk(void *chunk, sljit_uw size)
 
 static SLJIT_INLINE int get_map_jit_flag()
 {
+/* On macOS systems, returns MAP_JIT if it is defined _and_ we're running on a version
+   of macOS where it's OK to have more than one JIT block.
+   On non-macOS systems, returns MAP_JIT if it is defined. */
 #if TARGET_OS_OSX
-	/* On macOS systems, returns MAP_JIT if it is defined _and_ we're running on a version
-	   of macOS where it's OK to have more than one JIT block. On non-macOS systems, returns
-	   MAP_JIT if it is defined. */
 	static int map_jit_flag = -1;
 
 	/* The following code is thread safe because multiple initialization
@@ -124,12 +124,19 @@ static SLJIT_INLINE int get_map_jit_flag()
 		/* Kernel version for 10.14.0 (Mojave) */
 		if (atoi(name.release) >= 18) {
 			/* Only use MAP_JIT if a hardened runtime is used, because MAP_JIT is incompatible with fork(). */
-			void *ptr = mmap(NULL, getpagesize(), PROT_WRITE|PROT_EXEC, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
+
+			/* mirroring page size detection from sljit_allocate_stack */
+			long page_size = sysconf(_SC_PAGESIZE);
+			/* Should never happen */
+			if (page_size < 0)
+				page_size = 4096;
+
+			void *ptr = mmap(NULL, page_size, PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANON, -1, 0);
 
 			if (ptr == MAP_FAILED) {
 				map_jit_flag = MAP_JIT;
 			} else {
-				munmap(ptr, getpagesize());
+				munmap(ptr, page_size);
 			}
 		}
 	}
diff --git a/sljit/sljitLir.c b/sljit/sljitLir.c
index 9bab0c3..86772cc 100644
--- a/sljit/sljitLir.c
+++ b/sljit/sljitLir.c
@@ -926,7 +926,8 @@ static void sljit_verbose_fparam(struct sljit_compiler *compiler, sljit_s32 p, s
 
 static const char* op0_names[] = {
 	(char*)"breakpoint", (char*)"nop", (char*)"lmul.uw", (char*)"lmul.sw",
-	(char*)"divmod.u", (char*)"divmod.s", (char*)"div.u", (char*)"div.s"
+	(char*)"divmod.u", (char*)"divmod.s", (char*)"div.u", (char*)"div.s",
+	(char*)"endbr", (char*)"skip_frames_before_return"
 };
 
 static const char* op1_names[] = {
@@ -943,6 +944,12 @@ static const char* op2_names[] = {
 	(char*)"shl", (char*)"lshr", (char*)"ashr",
 };
 
+static const char* op_src_names[] = {
+	(char*)"fast_return", (char*)"skip_frames_before_fast_return",
+	(char*)"prefetch_l1", (char*)"prefetch_l2",
+	(char*)"prefetch_l3", (char*)"prefetch_once",
+};
+
 static const char* fop1_names[] = {
 	(char*)"mov", (char*)"conv", (char*)"conv", (char*)"conv",
 	(char*)"conv", (char*)"conv", (char*)"cmp", (char*)"neg",
@@ -1152,37 +1159,21 @@ static SLJIT_INLINE CHECK_RETURN_TYPE check_sljit_emit_fast_enter(struct sljit_c
 	CHECK_RETURN_OK;
 }
 
-static SLJIT_INLINE CHECK_RETURN_TYPE check_sljit_emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src, sljit_sw srcw)
-{
-#if (defined SLJIT_ARGUMENT_CHECKS && SLJIT_ARGUMENT_CHECKS)
-	FUNCTION_CHECK_SRC(src, srcw);
-	CHECK_ARGUMENT(src != SLJIT_IMM);
-	compiler->last_flags = 0;
-#endif
-#if (defined SLJIT_VERBOSE && SLJIT_VERBOSE)
-	if (SLJIT_UNLIKELY(!!compiler->verbose)) {
-		fprintf(compiler->verbose, "  fast_return ");
-		sljit_verbose_param(compiler, src, srcw);
-		fprintf(compiler->verbose, "\n");
-	}
-#endif
-	CHECK_RETURN_OK;
-}
-
 static SLJIT_INLINE CHECK_RETURN_TYPE check_sljit_emit_op0(struct sljit_compiler *compiler, sljit_s32 op)
 {
 #if (defined SLJIT_ARGUMENT_CHECKS && SLJIT_ARGUMENT_CHECKS)
 	CHECK_ARGUMENT((op >= SLJIT_BREAKPOINT && op <= SLJIT_LMUL_SW)
-		|| ((op & ~SLJIT_I32_OP) >= SLJIT_DIVMOD_UW && (op & ~SLJIT_I32_OP) <= SLJIT_DIV_SW));
-	CHECK_ARGUMENT(op < SLJIT_LMUL_UW || compiler->scratches >= 2);
-	if (op >= SLJIT_LMUL_UW)
+		|| ((op & ~SLJIT_I32_OP) >= SLJIT_DIVMOD_UW && (op & ~SLJIT_I32_OP) <= SLJIT_DIV_SW)
+		|| (op >= SLJIT_ENDBR && op <= SLJIT_SKIP_FRAMES_BEFORE_RETURN));
+	CHECK_ARGUMENT(GET_OPCODE(op) < SLJIT_LMUL_UW || GET_OPCODE(op) >= SLJIT_ENDBR || compiler->scratches >= 2);
+	if ((GET_OPCODE(op) >= SLJIT_LMUL_UW && GET_OPCODE(op) <= SLJIT_DIV_SW) || op == SLJIT_SKIP_FRAMES_BEFORE_RETURN)
 		compiler->last_flags = 0;
 #endif
 #if (defined SLJIT_VERBOSE && SLJIT_VERBOSE)
 	if (SLJIT_UNLIKELY(!!compiler->verbose))
 	{
 		fprintf(compiler->verbose, "  %s", op0_names[GET_OPCODE(op) - SLJIT_OP0_BASE]);
-		if (GET_OPCODE(op) >= SLJIT_DIVMOD_UW) {
+		if (GET_OPCODE(op) >= SLJIT_DIVMOD_UW && GET_OPCODE(op) <= SLJIT_DIV_SW) {
 			fprintf(compiler->verbose, (op & SLJIT_I32_OP) ? "32" : "w");
 		}
 		fprintf(compiler->verbose, "\n");
@@ -1224,7 +1215,7 @@ static SLJIT_INLINE CHECK_RETURN_TYPE check_sljit_emit_op1(struct sljit_compiler
 		break;
 	}
 
-	FUNCTION_CHECK_DST(dst, dstw, 1);
+	FUNCTION_CHECK_DST(dst, dstw, HAS_FLAGS(op));
 	FUNCTION_CHECK_SRC(src, srcw);
 
 	if (GET_OPCODE(op) >= SLJIT_NOT) {
@@ -1304,7 +1295,7 @@ static SLJIT_INLINE CHECK_RETURN_TYPE check_sljit_emit_op2(struct sljit_compiler
 		break;
 	}
 
-	FUNCTION_CHECK_DST(dst, dstw, 1);
+	FUNCTION_CHECK_DST(dst, dstw, HAS_FLAGS(op));
 	FUNCTION_CHECK_SRC(src1, src1w);
 	FUNCTION_CHECK_SRC(src2, src2w);
 	compiler->last_flags = GET_FLAG_TYPE(op) | (op & (SLJIT_I32_OP | SLJIT_SET_Z));
@@ -1325,6 +1316,33 @@ static SLJIT_INLINE CHECK_RETURN_TYPE check_sljit_emit_op2(struct sljit_compiler
 	CHECK_RETURN_OK;
 }
 
+static SLJIT_INLINE CHECK_RETURN_TYPE check_sljit_emit_op_src(struct sljit_compiler *compiler, sljit_s32 op,
+	sljit_s32 src, sljit_sw srcw)
+{
+#if (defined SLJIT_ARGUMENT_CHECKS && SLJIT_ARGUMENT_CHECKS)
+	CHECK_ARGUMENT(op >= SLJIT_FAST_RETURN && op <= SLJIT_PREFETCH_ONCE);
+	FUNCTION_CHECK_SRC(src, srcw);
+
+	if (op == SLJIT_FAST_RETURN || op == SLJIT_SKIP_FRAMES_BEFORE_FAST_RETURN)
+	{
+		CHECK_ARGUMENT(src != SLJIT_IMM);
+		compiler->last_flags = 0;
+	}
+	else if (op >= SLJIT_PREFETCH_L1 && op <= SLJIT_PREFETCH_ONCE)
+	{
+		CHECK_ARGUMENT(src & SLJIT_MEM);
+	}
+#endif
+#if (defined SLJIT_VERBOSE && SLJIT_VERBOSE)
+	if (SLJIT_UNLIKELY(!!compiler->verbose)) {
+		fprintf(compiler->verbose, "  %s ", op_src_names[op - SLJIT_OP_SRC_BASE]);
+		sljit_verbose_param(compiler, src, srcw);
+		fprintf(compiler->verbose, "\n");
+	}
+#endif
+	CHECK_RETURN_OK;
+}
+
 static SLJIT_INLINE CHECK_RETURN_TYPE check_sljit_get_register_index(sljit_s32 reg)
 {
 	SLJIT_UNUSED_ARG(reg);
@@ -2016,7 +2034,7 @@ static SLJIT_INLINE sljit_s32 emit_mov_before_return(struct sljit_compiler *comp
 #if (defined SLJIT_CONFIG_X86 && SLJIT_CONFIG_X86) \
 		|| (defined SLJIT_CONFIG_PPC && SLJIT_CONFIG_PPC) \
 		|| (defined SLJIT_CONFIG_SPARC_32 && SLJIT_CONFIG_SPARC_32) \
-		|| ((defined SLJIT_CONFIG_MIPS && SLJIT_CONFIG_MIPS) && !(defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1))
+		|| ((defined SLJIT_CONFIG_MIPS && SLJIT_CONFIG_MIPS) && !(defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 1))
 
 static SLJIT_INLINE sljit_s32 sljit_emit_cmov_generic(struct sljit_compiler *compiler, sljit_s32 type,
 	sljit_s32 dst_reg,
@@ -2381,15 +2399,6 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_enter(struct sljit_compiler *
 	return SLJIT_ERR_UNSUPPORTED;
 }
 
-SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src, sljit_sw srcw)
-{
-	SLJIT_UNUSED_ARG(compiler);
-	SLJIT_UNUSED_ARG(src);
-	SLJIT_UNUSED_ARG(srcw);
-	SLJIT_UNREACHABLE();
-	return SLJIT_ERR_UNSUPPORTED;
-}
-
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compiler, sljit_s32 op)
 {
 	SLJIT_UNUSED_ARG(compiler);
@@ -2429,6 +2438,17 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2(struct sljit_compiler *compile
 	return SLJIT_ERR_UNSUPPORTED;
 }
 
+SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_src(struct sljit_compiler *compiler, sljit_s32 op,
+	sljit_s32 src, sljit_sw srcw)
+{
+	SLJIT_UNUSED_ARG(compiler);
+	SLJIT_UNUSED_ARG(op);
+	SLJIT_UNUSED_ARG(src);
+	SLJIT_UNUSED_ARG(srcw);
+	SLJIT_UNREACHABLE();
+	return SLJIT_ERR_UNSUPPORTED;
+}
+
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_get_register_index(sljit_s32 reg)
 {
 	SLJIT_UNREACHABLE();
@@ -2549,6 +2569,13 @@ SLJIT_API_FUNC_ATTRIBUTE void sljit_set_target(struct sljit_jump *jump, sljit_uw
 	SLJIT_UNREACHABLE();
 }
 
+SLJIT_API_FUNC_ATTRIBUTE void sljit_set_put_label(struct sljit_put_label *put_label, struct sljit_label *label)
+{
+	SLJIT_UNUSED_ARG(put_label);
+	SLJIT_UNUSED_ARG(label);
+	SLJIT_UNREACHABLE();
+}
+
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_ijump(struct sljit_compiler *compiler, sljit_s32 type, sljit_s32 src, sljit_sw srcw)
 {
 	SLJIT_UNUSED_ARG(compiler);
diff --git a/sljit/sljitLir.h b/sljit/sljitLir.h
index 836d25c..aef90e6 100644
--- a/sljit/sljitLir.h
+++ b/sljit/sljitLir.h
@@ -80,6 +80,10 @@ of sljitConfigInternal.h */
 
 #include "sljitConfigInternal.h"
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 /* --------------------------------------------------------------------- */
 /*  Error codes                                                          */
 /* --------------------------------------------------------------------- */
@@ -154,10 +158,10 @@ of sljitConfigInternal.h */
 */
 
 /* When SLJIT_UNUSED is specified as the destination of sljit_emit_op1
-   or sljit_emit_op2 operations the result is discarded. If no status
-   flags are set, no instructions are emitted for these operations. Data
-   prefetch is a special exception, see SLJIT_MOV operation. Other SLJIT
-   operations do not support SLJIT_UNUSED as a destination operand. */
+   or sljit_emit_op2 operations the result is discarded. Some status
+   flags must be set when the destination is SLJIT_UNUSED, because the
+   operation would have no effect otherwise. Other SLJIT operations do
+   not support SLJIT_UNUSED as a destination operand. */
 #define SLJIT_UNUSED		0
 
 /* Scratch registers. */
@@ -571,6 +575,8 @@ static SLJIT_INLINE sljit_uw sljit_get_generated_code_size(struct sljit_compiler
 #define SLJIT_HAS_CLZ			2
 /* [Emulated] Conditional move is supported. */
 #define SLJIT_HAS_CMOV			3
+/* [Emulated] Conditional move is supported. */
+#define SLJIT_HAS_PREFETCH		4
 
 #if (defined SLJIT_CONFIG_X86 && SLJIT_CONFIG_X86)
 /* [Not emulated] SSE2 support is available on x86. */
@@ -658,10 +664,10 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_return(struct sljit_compiler *comp
 	sljit_s32 src, sljit_sw srcw);
 
 /* Generating entry and exit points for fast call functions (see SLJIT_FAST_CALL).
-   Both sljit_emit_fast_enter and sljit_emit_fast_return functions preserve the
+   Both sljit_emit_fast_enter and SLJIT_FAST_RETURN operations preserve the
    values of all registers and stack frame. The return address is stored in the
    dst argument of sljit_emit_fast_enter, and this return address can be passed
-   to sljit_emit_fast_return to continue the execution after the fast call.
+   to SLJIT_FAST_RETURN to continue the execution after the fast call.
 
    Fast calls are cheap operations (usually only a single call instruction is
    emitted) but they do not preserve any registers. However the callee function
@@ -669,16 +675,15 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_return(struct sljit_compiler *comp
    efficiently exploited by various optimizations. Registers can be saved
    manually by the callee function if needed.
 
-   Although returning to different address by sljit_emit_fast_return is possible,
+   Although returning to different address by SLJIT_FAST_RETURN is possible,
    this address usually cannot be predicted by the return address predictor of
-   modern CPUs which may reduce performance. Furthermore using sljit_emit_ijump
-   to return is also inefficient since return address prediction is usually
-   triggered by a specific form of ijump.
+   modern CPUs which may reduce performance. Furthermore certain security
+   enhancement technologies such as Intel Control-flow Enforcement Technology
+   (CET) may disallow returning to a different address.
 
    Flags: - (does not modify flags). */
 
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_enter(struct sljit_compiler *compiler, sljit_s32 dst, sljit_sw dstw);
-SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src, sljit_sw srcw);
 
 /*
    Source and destination operands for arithmetical instructions
@@ -887,6 +892,14 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_return(struct sljit_compiler
          the behaviour is undefined. */
 #define SLJIT_DIV_SW			(SLJIT_OP0_BASE + 7)
 #define SLJIT_DIV_S32			(SLJIT_DIV_SW | SLJIT_I32_OP)
+/* Flags: - (does not modify flags)
+   ENDBR32 instruction for x86-32 and ENDBR64 instruction for x86-64
+   when Intel Control-flow Enforcement Technology (CET) is enabled.
+   No instruction for other architectures.  */
+#define SLJIT_ENDBR			(SLJIT_OP0_BASE + 8)
+/* Flags: - (may destroy flags)
+   Skip stack frames before return.  */
+#define SLJIT_SKIP_FRAMES_BEFORE_RETURN	(SLJIT_OP0_BASE + 9)
 
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compiler, sljit_s32 op);
 
@@ -904,15 +917,6 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compile
    U32 - unsigned int (32 bit) data transfer
    S32 - signed int (32 bit) data transfer
    P   - pointer (sljit_p) data transfer
-
-   If the destination of a MOV instruction is SLJIT_UNUSED and the source
-   operand is a memory address the compiler emits a prefetch instruction
-   if this instruction is supported by the current CPU. Higher data sizes
-   bring the data closer to the core: a MOV with word size loads the data
-   into a higher level cache than a byte size. Otherwise the type does not
-   affect the prefetch instruction. Furthermore a prefetch instruction
-   never fails, so it can be used to prefetch a data from an address and
-   check whether that address is NULL afterwards.
 */
 
 /* Flags: - (does not modify flags) */
@@ -1017,8 +1021,46 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2(struct sljit_compiler *compile
 	sljit_s32 src1, sljit_sw src1w,
 	sljit_s32 src2, sljit_sw src2w);
 
+/* Starting index of opcodes for sljit_emit_op2. */
+#define SLJIT_OP_SRC_BASE		128
+
+/* Note: src cannot be an immedate value
+   Flags: - (does not modify flags) */
+#define SLJIT_FAST_RETURN		(SLJIT_OP_SRC_BASE + 0)
+/* Skip stack frames before fast return.
+   Note: src cannot be an immedate value
+   Flags: may destroy flags. */
+#define SLJIT_SKIP_FRAMES_BEFORE_FAST_RETURN	(SLJIT_OP_SRC_BASE + 1)
+/* Prefetch value into the level 1 data cache
+   Note: if the target CPU does not support data prefetch,
+         no instructions are emitted.
+   Note: this instruction never fails, even if the memory address is invalid.
+   Flags: - (does not modify flags) */
+#define SLJIT_PREFETCH_L1		(SLJIT_OP_SRC_BASE + 2)
+/* Prefetch value into the level 2 data cache
+   Note: same as SLJIT_PREFETCH_L1 if the target CPU
+         does not support this instruction form.
+   Note: this instruction never fails, even if the memory address is invalid.
+   Flags: - (does not modify flags) */
+#define SLJIT_PREFETCH_L2		(SLJIT_OP_SRC_BASE + 3)
+/* Prefetch value into the level 3 data cache
+   Note: same as SLJIT_PREFETCH_L2 if the target CPU
+         does not support this instruction form.
+   Note: this instruction never fails, even if the memory address is invalid.
+   Flags: - (does not modify flags) */
+#define SLJIT_PREFETCH_L3		(SLJIT_OP_SRC_BASE + 4)
+/* Prefetch a value which is only used once (and can be discarded afterwards)
+   Note: same as SLJIT_PREFETCH_L1 if the target CPU
+         does not support this instruction form.
+   Note: this instruction never fails, even if the memory address is invalid.
+   Flags: - (does not modify flags) */
+#define SLJIT_PREFETCH_ONCE		(SLJIT_OP_SRC_BASE + 5)
+
+SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_src(struct sljit_compiler *compiler, sljit_s32 op,
+	sljit_s32 src, sljit_sw srcw);
+
 /* Starting index of opcodes for sljit_emit_fop1. */
-#define SLJIT_FOP1_BASE			128
+#define SLJIT_FOP1_BASE			160
 
 /* Flags: - (does not modify flags) */
 #define SLJIT_MOV_F64			(SLJIT_FOP1_BASE + 0)
@@ -1057,7 +1099,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fop1(struct sljit_compiler *compil
 	sljit_s32 src, sljit_sw srcw);
 
 /* Starting index of opcodes for sljit_emit_fop2. */
-#define SLJIT_FOP2_BASE			160
+#define SLJIT_FOP2_BASE			192
 
 /* Flags: - (does not modify flags) */
 #define SLJIT_ADD_F64			(SLJIT_FOP2_BASE + 0)
@@ -1161,7 +1203,7 @@ SLJIT_API_FUNC_ATTRIBUTE struct sljit_label* sljit_emit_label(struct sljit_compi
 
 /* Unconditional jump types. */
 #define SLJIT_JUMP			24
-	/* Fast calling method. See sljit_emit_fast_enter / sljit_emit_fast_return. */
+	/* Fast calling method. See sljit_emit_fast_enter / SLJIT_FAST_RETURN. */
 #define SLJIT_FAST_CALL			25
 	/* Called function must be declared with the SLJIT_FUNC attribute. */
 #define SLJIT_CALL			26
@@ -1490,4 +1532,8 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_custom(struct sljit_compiler *c
 SLJIT_API_FUNC_ATTRIBUTE void sljit_set_current_flags(struct sljit_compiler *compiler,
 	sljit_s32 current_flags);
 
+#ifdef __cplusplus
+} /* extern "C" */
+#endif
+
 #endif /* _SLJIT_LIR_H_ */
diff --git a/sljit/sljitNativeARM_32.c b/sljit/sljitNativeARM_32.c
index 8da0d09..24ef02a 100644
--- a/sljit/sljitNativeARM_32.c
+++ b/sljit/sljitNativeARM_32.c
@@ -872,6 +872,9 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_has_cpu_feature(sljit_s32 feature_type)
 
 	case SLJIT_HAS_CLZ:
 	case SLJIT_HAS_CMOV:
+#if (defined SLJIT_CONFIG_ARM_V7 && SLJIT_CONFIG_ARM_V7)
+	case SLJIT_HAS_PREFETCH:
+#endif
 		return 1;
 
 	default:
@@ -1678,6 +1681,9 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compile
 						| (saved_reg_list[0] << 12) /* ldr rX, [sp], #8/16 */);
 		}
 		return SLJIT_SUCCESS;
+	case SLJIT_ENDBR:
+	case SLJIT_SKIP_FRAMES_BEFORE_RETURN:
+		return SLJIT_SUCCESS;
 	}
 
 	return SLJIT_SUCCESS;
@@ -1692,14 +1698,6 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op1(struct sljit_compiler *compile
 	ADJUST_LOCAL_OFFSET(dst, dstw);
 	ADJUST_LOCAL_OFFSET(src, srcw);
 
-	if (dst == SLJIT_UNUSED && !HAS_FLAGS(op)) {
-#if (defined SLJIT_CONFIG_ARM_V7 && SLJIT_CONFIG_ARM_V7)
-		if (op <= SLJIT_MOV_P && (src & SLJIT_MEM))
-			return emit_op_mem(compiler, PRELOAD | LOAD_DATA, TMP_PC, src, srcw, TMP_REG1);
-#endif
-		return SLJIT_SUCCESS;
-	}
-
 	switch (GET_OPCODE(op)) {
 	case SLJIT_MOV:
 	case SLJIT_MOV_U32:
@@ -1781,6 +1779,40 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2(struct sljit_compiler *compile
 	return SLJIT_SUCCESS;
 }
 
+SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_src(struct sljit_compiler *compiler, sljit_s32 op,
+	sljit_s32 src, sljit_sw srcw)
+{
+	CHECK_ERROR();
+	CHECK(check_sljit_emit_op_src(compiler, op, src, srcw));
+	ADJUST_LOCAL_OFFSET(src, srcw);
+
+	switch (op) {
+	case SLJIT_FAST_RETURN:
+		SLJIT_ASSERT(reg_map[TMP_REG2] == 14);
+
+		if (FAST_IS_REG(src))
+			FAIL_IF(push_inst(compiler, MOV | RD(TMP_REG2) | RM(src)));
+		else
+			FAIL_IF(emit_op_mem(compiler, WORD_SIZE | LOAD_DATA, TMP_REG2, src, srcw, TMP_REG1));
+
+		return push_inst(compiler, BX | RM(TMP_REG2));
+	case SLJIT_SKIP_FRAMES_BEFORE_FAST_RETURN:
+		return SLJIT_SUCCESS;
+	case SLJIT_PREFETCH_L1:
+	case SLJIT_PREFETCH_L2:
+	case SLJIT_PREFETCH_L3:
+	case SLJIT_PREFETCH_ONCE:
+#if (defined SLJIT_CONFIG_ARM_V7 && SLJIT_CONFIG_ARM_V7)
+		SLJIT_ASSERT(src & SLJIT_MEM);
+		return emit_op_mem(compiler, PRELOAD | LOAD_DATA, TMP_PC, src, srcw, TMP_REG1);
+#else /* !SLJIT_CONFIG_ARM_V7 */
+		return SLJIT_SUCCESS;
+#endif /* SLJIT_CONFIG_ARM_V7 */
+	}
+
+	return SLJIT_SUCCESS;
+}
+
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_get_register_index(sljit_s32 reg)
 {
 	CHECK_REG_INDEX(check_sljit_get_register_index(reg));
@@ -2043,22 +2075,6 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_enter(struct sljit_compiler *
 	return emit_op_mem(compiler, WORD_SIZE, TMP_REG2, dst, dstw, TMP_REG1);
 }
 
-SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src, sljit_sw srcw)
-{
-	CHECK_ERROR();
-	CHECK(check_sljit_emit_fast_return(compiler, src, srcw));
-	ADJUST_LOCAL_OFFSET(src, srcw);
-
-	SLJIT_ASSERT(reg_map[TMP_REG2] == 14);
-
-	if (FAST_IS_REG(src))
-		FAIL_IF(push_inst(compiler, MOV | RD(TMP_REG2) | RM(src)));
-	else
-		FAIL_IF(emit_op_mem(compiler, WORD_SIZE | LOAD_DATA, TMP_REG2, src, srcw, TMP_REG1));
-
-	return push_inst(compiler, BX | RM(TMP_REG2));
-}
-
 /* --------------------------------------------------------------------- */
 /*  Conditional instructions                                             */
 /* --------------------------------------------------------------------- */
diff --git a/sljit/sljitNativeARM_64.c b/sljit/sljitNativeARM_64.c
index e15b345..b86fc64 100644
--- a/sljit/sljitNativeARM_64.c
+++ b/sljit/sljitNativeARM_64.c
@@ -396,6 +396,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_has_cpu_feature(sljit_s32 feature_type)
 
 	case SLJIT_HAS_CLZ:
 	case SLJIT_HAS_CMOV:
+	case SLJIT_HAS_PREFETCH:
 		return 1;
 
 	default:
@@ -1154,6 +1155,9 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compile
 	case SLJIT_DIV_UW:
 	case SLJIT_DIV_SW:
 		return push_inst(compiler, ((op == SLJIT_DIV_UW ? UDIV : SDIV) ^ inv_bits) | RD(SLJIT_R0) | RN(SLJIT_R0) | RM(SLJIT_R1));
+	case SLJIT_ENDBR:
+	case SLJIT_SKIP_FRAMES_BEFORE_RETURN:
+		return SLJIT_SUCCESS;
 	}
 
 	return SLJIT_SUCCESS;
@@ -1171,23 +1175,6 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op1(struct sljit_compiler *compile
 	ADJUST_LOCAL_OFFSET(dst, dstw);
 	ADJUST_LOCAL_OFFSET(src, srcw);
 
-	if (dst == SLJIT_UNUSED && !HAS_FLAGS(op)) {
-		if (op <= SLJIT_MOV_P && (src & SLJIT_MEM)) {
-			SLJIT_ASSERT(reg_map[1] == 0 && reg_map[3] == 2 && reg_map[5] == 4);
-
-			if (op >= SLJIT_MOV_U8 && op <= SLJIT_MOV_S8)
-				dst = 5;
-			else if (op >= SLJIT_MOV_U16 && op <= SLJIT_MOV_S16)
-				dst = 3;
-			else
-				dst = 1;
-
-			/* Signed word sized load is the prefetch instruction. */
-			return emit_op_mem(compiler, WORD_SIZE | SIGNED, dst, src, srcw, TMP_REG1);
-		}
-		return SLJIT_SUCCESS;
-	}
-
 	dst_r = SLOW_IS_REG(dst) ? dst : TMP_REG1;
 
 	op = GET_OPCODE(op);
@@ -1327,6 +1314,46 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2(struct sljit_compiler *compile
 	return SLJIT_SUCCESS;
 }
 
+SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_src(struct sljit_compiler *compiler, sljit_s32 op,
+	sljit_s32 src, sljit_sw srcw)
+{
+	CHECK_ERROR();
+	CHECK(check_sljit_emit_op_src(compiler, op, src, srcw));
+	ADJUST_LOCAL_OFFSET(src, srcw);
+
+	switch (op) {
+	case SLJIT_FAST_RETURN:
+		if (FAST_IS_REG(src))
+			FAIL_IF(push_inst(compiler, ORR | RD(TMP_LR) | RN(TMP_ZERO) | RM(src)));
+		else
+			FAIL_IF(emit_op_mem(compiler, WORD_SIZE, TMP_LR, src, srcw, TMP_REG1));
+
+		return push_inst(compiler, RET | RN(TMP_LR));
+	case SLJIT_SKIP_FRAMES_BEFORE_FAST_RETURN:
+		return SLJIT_SUCCESS;
+	case SLJIT_PREFETCH_L1:
+	case SLJIT_PREFETCH_L2:
+	case SLJIT_PREFETCH_L3:
+	case SLJIT_PREFETCH_ONCE:
+		SLJIT_ASSERT(reg_map[1] == 0 && reg_map[3] == 2 && reg_map[5] == 4);
+
+		/* The reg_map[op] should provide the appropriate constant. */
+		if (op == SLJIT_PREFETCH_L1)
+			op = 1;
+		else if (op == SLJIT_PREFETCH_L2)
+			op = 3;
+		else if (op == SLJIT_PREFETCH_L3)
+			op = 5;
+		else
+			op = 2;
+
+		/* Signed word sized load is the prefetch instruction. */
+		return emit_op_mem(compiler, WORD_SIZE | SIGNED, op, src, srcw, TMP_REG1);
+	}
+
+	return SLJIT_SUCCESS;
+}
+
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_get_register_index(sljit_s32 reg)
 {
 	CHECK_REG_INDEX(check_sljit_get_register_index(reg));
@@ -1578,20 +1605,6 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_enter(struct sljit_compiler *
 	return emit_op_mem(compiler, WORD_SIZE | STORE, TMP_LR, dst, dstw, TMP_REG1);
 }
 
-SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src, sljit_sw srcw)
-{
-	CHECK_ERROR();
-	CHECK(check_sljit_emit_fast_return(compiler, src, srcw));
-	ADJUST_LOCAL_OFFSET(src, srcw);
-
-	if (FAST_IS_REG(src))
-		FAIL_IF(push_inst(compiler, ORR | RD(TMP_LR) | RN(TMP_ZERO) | RM(src)));
-	else
-		FAIL_IF(emit_op_mem(compiler, WORD_SIZE, TMP_LR, src, srcw, TMP_REG1));
-
-	return push_inst(compiler, RET | RN(TMP_LR));
-}
-
 /* --------------------------------------------------------------------- */
 /*  Conditional instructions                                             */
 /* --------------------------------------------------------------------- */
diff --git a/sljit/sljitNativeARM_T2_32.c b/sljit/sljitNativeARM_T2_32.c
index cdfe4a4..a26f48f 100644
--- a/sljit/sljitNativeARM_T2_32.c
+++ b/sljit/sljitNativeARM_T2_32.c
@@ -480,6 +480,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_has_cpu_feature(sljit_s32 feature_type)
 
 	case SLJIT_HAS_CLZ:
 	case SLJIT_HAS_CMOV:
+	case SLJIT_HAS_PREFETCH:
 		return 1;
 
 	default:
@@ -1328,6 +1329,9 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compile
 		}
 		return SLJIT_SUCCESS;
 #endif /* __ARM_FEATURE_IDIV || __ARM_ARCH_EXT_IDIV__ */
+	case SLJIT_ENDBR:
+	case SLJIT_SKIP_FRAMES_BEFORE_RETURN:
+		return SLJIT_SUCCESS;
 	}
 
 	return SLJIT_SUCCESS;
@@ -1345,13 +1349,6 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op1(struct sljit_compiler *compile
 	ADJUST_LOCAL_OFFSET(dst, dstw);
 	ADJUST_LOCAL_OFFSET(src, srcw);
 
-	if (dst == SLJIT_UNUSED && !HAS_FLAGS(op)) {
-		/* Since TMP_PC has index 15, IS_2_LO_REGS and IS_3_LO_REGS checks always fail. */
-		if (op <= SLJIT_MOV_P && (src & SLJIT_MEM))
-			return emit_op_mem(compiler, PRELOAD, TMP_PC, src, srcw, TMP_REG1);
-		return SLJIT_SUCCESS;
-	}
-
 	dst_r = SLOW_IS_REG(dst) ? dst : TMP_REG1;
 
 	op = GET_OPCODE(op);
@@ -1475,6 +1472,35 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2(struct sljit_compiler *compile
 	return emit_op_mem(compiler, WORD_SIZE | STORE, dst_reg, dst, dstw, TMP_REG2);
 }
 
+SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_src(struct sljit_compiler *compiler, sljit_s32 op,
+	sljit_s32 src, sljit_sw srcw)
+{
+	CHECK_ERROR();
+	CHECK(check_sljit_emit_op_src(compiler, op, src, srcw));
+	ADJUST_LOCAL_OFFSET(src, srcw);
+
+	switch (op) {
+	case SLJIT_FAST_RETURN:
+		SLJIT_ASSERT(reg_map[TMP_REG2] == 14);
+
+		if (FAST_IS_REG(src))
+			FAIL_IF(push_inst16(compiler, MOV | SET_REGS44(TMP_REG2, src)));
+		else
+			FAIL_IF(emit_op_mem(compiler, WORD_SIZE, TMP_REG2, src, srcw, TMP_REG2));
+
+		return push_inst16(compiler, BX | RN3(TMP_REG2));
+	case SLJIT_SKIP_FRAMES_BEFORE_FAST_RETURN:
+		return SLJIT_SUCCESS;
+	case SLJIT_PREFETCH_L1:
+	case SLJIT_PREFETCH_L2:
+	case SLJIT_PREFETCH_L3:
+	case SLJIT_PREFETCH_ONCE:
+		return emit_op_mem(compiler, PRELOAD, TMP_PC, src, srcw, TMP_REG1);
+	}
+
+	return SLJIT_SUCCESS;
+}
+
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_get_register_index(sljit_s32 reg)
 {
 	CHECK_REG_INDEX(check_sljit_get_register_index(reg));
@@ -1728,22 +1754,6 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_enter(struct sljit_compiler *
 	return emit_op_mem(compiler, WORD_SIZE | STORE, TMP_REG2, dst, dstw, TMP_REG1);
 }
 
-SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src, sljit_sw srcw)
-{
-	CHECK_ERROR();
-	CHECK(check_sljit_emit_fast_return(compiler, src, srcw));
-	ADJUST_LOCAL_OFFSET(src, srcw);
-
-	SLJIT_ASSERT(reg_map[TMP_REG2] == 14);
-
-	if (FAST_IS_REG(src))
-		FAIL_IF(push_inst16(compiler, MOV | SET_REGS44(TMP_REG2, src)));
-	else
-		FAIL_IF(emit_op_mem(compiler, WORD_SIZE, TMP_REG2, src, srcw, TMP_REG2));
-
-	return push_inst16(compiler, BX | RN3(TMP_REG2));
-}
-
 /* --------------------------------------------------------------------- */
 /*  Conditional instructions                                             */
 /* --------------------------------------------------------------------- */
diff --git a/sljit/sljitNativeMIPS_32.c b/sljit/sljitNativeMIPS_32.c
index 16dec05..777627b 100644
--- a/sljit/sljitNativeMIPS_32.c
+++ b/sljit/sljitNativeMIPS_32.c
@@ -86,12 +86,12 @@ static SLJIT_INLINE sljit_s32 emit_single_op(struct sljit_compiler *compiler, sl
 		SLJIT_ASSERT(src1 == TMP_REG1 && !(flags & SRC2_IMM));
 		if ((flags & (REG_DEST | REG2_SOURCE)) == (REG_DEST | REG2_SOURCE)) {
 			if (op == SLJIT_MOV_S8) {
-#if (defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 1)
 				return push_inst(compiler, SEB | T(src2) | D(dst), DR(dst));
-#else
+#else /* SLJIT_MIPS_REV < 1 */
 				FAIL_IF(push_inst(compiler, SLL | T(src2) | D(dst) | SH_IMM(24), DR(dst)));
 				return push_inst(compiler, SRA | T(dst) | D(dst) | SH_IMM(24), DR(dst));
-#endif
+#endif /* SLJIT_MIPS_REV >= 1 */
 			}
 			return push_inst(compiler, ANDI | S(src2) | T(dst) | IMM(0xff), DR(dst));
 		}
@@ -105,12 +105,12 @@ static SLJIT_INLINE sljit_s32 emit_single_op(struct sljit_compiler *compiler, sl
 		SLJIT_ASSERT(src1 == TMP_REG1 && !(flags & SRC2_IMM));
 		if ((flags & (REG_DEST | REG2_SOURCE)) == (REG_DEST | REG2_SOURCE)) {
 			if (op == SLJIT_MOV_S16) {
-#if (defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 1)
 				return push_inst(compiler, SEH | T(src2) | D(dst), DR(dst));
-#else
+#else /* SLJIT_MIPS_REV < 1 */
 				FAIL_IF(push_inst(compiler, SLL | T(src2) | D(dst) | SH_IMM(16), DR(dst)));
 				return push_inst(compiler, SRA | T(dst) | D(dst) | SH_IMM(16), DR(dst));
-#endif
+#endif /* SLJIT_MIPS_REV >= 1 */
 			}
 			return push_inst(compiler, ANDI | S(src2) | T(dst) | IMM(0xffff), DR(dst));
 		}
@@ -129,12 +129,12 @@ static SLJIT_INLINE sljit_s32 emit_single_op(struct sljit_compiler *compiler, sl
 
 	case SLJIT_CLZ:
 		SLJIT_ASSERT(src1 == TMP_REG1 && !(flags & SRC2_IMM));
-#if (defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 1)
 		if (op & SLJIT_SET_Z)
 			FAIL_IF(push_inst(compiler, CLZ | S(src2) | TA(EQUAL_FLAG) | DA(EQUAL_FLAG), EQUAL_FLAG));
 		if (!(flags & UNUSED_DEST))
 			FAIL_IF(push_inst(compiler, CLZ | S(src2) | T(dst) | D(dst), DR(dst)));
-#else
+#else /* SLJIT_MIPS_REV < 1 */
 		if (SLJIT_UNLIKELY(flags & UNUSED_DEST)) {
 			FAIL_IF(push_inst(compiler, SRL | T(src2) | DA(EQUAL_FLAG) | SH_IMM(31), EQUAL_FLAG));
 			return push_inst(compiler, XORI | SA(EQUAL_FLAG) | TA(EQUAL_FLAG) | IMM(1), EQUAL_FLAG);
@@ -149,7 +149,7 @@ static SLJIT_INLINE sljit_s32 emit_single_op(struct sljit_compiler *compiler, sl
 		FAIL_IF(push_inst(compiler, ADDIU | S(dst) | T(dst) | IMM(1), DR(dst)));
 		FAIL_IF(push_inst(compiler, BGEZ | S(TMP_REG1) | IMM(-2), UNMOVABLE_INS));
 		FAIL_IF(push_inst(compiler, SLL | T(TMP_REG1) | D(TMP_REG1) | SH_IMM(1), UNMOVABLE_INS));
-#endif
+#endif /* SLJIT_MIPS_REV >= 1 */
 		return SLJIT_SUCCESS;
 
 	case SLJIT_ADD:
@@ -368,21 +368,22 @@ static SLJIT_INLINE sljit_s32 emit_single_op(struct sljit_compiler *compiler, sl
 		SLJIT_ASSERT(!(flags & SRC2_IMM));
 
 		if (GET_FLAG_TYPE(op) != SLJIT_MUL_OVERFLOW) {
-#if (defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1) || (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 1)
 			return push_inst(compiler, MUL | S(src1) | T(src2) | D(dst), DR(dst));
-#else /* !SLJIT_MIPS_R1 && !SLJIT_MIPS_R6 */
+#else /* SLJIT_MIPS_REV < 1 */
 			FAIL_IF(push_inst(compiler, MULT | S(src1) | T(src2), MOVABLE_INS));
 			return push_inst(compiler, MFLO | D(dst), DR(dst));
-#endif /* SLJIT_MIPS_R1 || SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 1 */
 		}
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 		FAIL_IF(push_inst(compiler, MUL | S(src1) | T(src2) | D(dst), DR(dst)));
 		FAIL_IF(push_inst(compiler, MUH | S(src1) | T(src2) | DA(EQUAL_FLAG), EQUAL_FLAG));
-#else /* !SLJIT_MIPS_R6 */
+#else /* SLJIT_MIPS_REV < 6 */
 		FAIL_IF(push_inst(compiler, MULT | S(src1) | T(src2), MOVABLE_INS));
 		FAIL_IF(push_inst(compiler, MFHI | DA(EQUAL_FLAG), EQUAL_FLAG));
 		FAIL_IF(push_inst(compiler, MFLO | D(dst), DR(dst)));
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
 		FAIL_IF(push_inst(compiler, SRA | T(dst) | DA(OTHER_FLAG) | SH_IMM(31), OTHER_FLAG));
 		return push_inst(compiler, SUBU | SA(EQUAL_FLAG) | TA(OTHER_FLAG) | DA(OTHER_FLAG), OTHER_FLAG);
 
diff --git a/sljit/sljitNativeMIPS_64.c b/sljit/sljitNativeMIPS_64.c
index a6a2bcc..479244d 100644
--- a/sljit/sljitNativeMIPS_64.c
+++ b/sljit/sljitNativeMIPS_64.c
@@ -220,12 +220,12 @@ static SLJIT_INLINE sljit_s32 emit_single_op(struct sljit_compiler *compiler, sl
 
 	case SLJIT_CLZ:
 		SLJIT_ASSERT(src1 == TMP_REG1 && !(flags & SRC2_IMM));
-#if (defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 1)
 		if (op & SLJIT_SET_Z)
 			FAIL_IF(push_inst(compiler, SELECT_OP(DCLZ, CLZ) | S(src2) | TA(EQUAL_FLAG) | DA(EQUAL_FLAG), EQUAL_FLAG));
 		if (!(flags & UNUSED_DEST))
 			FAIL_IF(push_inst(compiler, SELECT_OP(DCLZ, CLZ) | S(src2) | T(dst) | D(dst), DR(dst)));
-#else
+#else /* SLJIT_MIPS_REV < 1 */
 		if (SLJIT_UNLIKELY(flags & UNUSED_DEST)) {
 			FAIL_IF(push_inst(compiler, SELECT_OP(DSRL32, SRL) | T(src2) | DA(EQUAL_FLAG) | SH_IMM(31), EQUAL_FLAG));
 			return push_inst(compiler, XORI | SA(EQUAL_FLAG) | TA(EQUAL_FLAG) | IMM(1), EQUAL_FLAG);
@@ -240,7 +240,7 @@ static SLJIT_INLINE sljit_s32 emit_single_op(struct sljit_compiler *compiler, sl
 		FAIL_IF(push_inst(compiler, SELECT_OP(DADDIU, ADDIU) | S(dst) | T(dst) | IMM(1), DR(dst)));
 		FAIL_IF(push_inst(compiler, BGEZ | S(TMP_REG1) | IMM(-2), UNMOVABLE_INS));
 		FAIL_IF(push_inst(compiler, SELECT_OP(DSLL, SLL) | T(TMP_REG1) | D(TMP_REG1) | SH_IMM(1), UNMOVABLE_INS));
-#endif
+#endif /* SLJIT_MIPS_REV >= 1 */
 		return SLJIT_SUCCESS;
 
 	case SLJIT_ADD:
@@ -459,26 +459,27 @@ static SLJIT_INLINE sljit_s32 emit_single_op(struct sljit_compiler *compiler, sl
 		SLJIT_ASSERT(!(flags & SRC2_IMM));
 
 		if (GET_FLAG_TYPE(op) != SLJIT_MUL_OVERFLOW) {
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 			return push_inst(compiler, SELECT_OP(DMUL, MUL) | S(src1) | T(src2) | D(dst), DR(dst));
-#elif (defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1)
+#elif (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 1)
 			if (op & SLJIT_I32_OP)
 				return push_inst(compiler, MUL | S(src1) | T(src2) | D(dst), DR(dst));
 			FAIL_IF(push_inst(compiler, DMULT | S(src1) | T(src2), MOVABLE_INS));
 			return push_inst(compiler, MFLO | D(dst), DR(dst));
-#else /* !SLJIT_MIPS_R6 && !SLJIT_MIPS_R1 */
+#else /* SLJIT_MIPS_REV < 1 */
 			FAIL_IF(push_inst(compiler, SELECT_OP(DMULT, MULT) | S(src1) | T(src2), MOVABLE_INS));
 			return push_inst(compiler, MFLO | D(dst), DR(dst));
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
 		}
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 		FAIL_IF(push_inst(compiler, SELECT_OP(DMUL, MUL) | S(src1) | T(src2) | D(dst), DR(dst)));
 		FAIL_IF(push_inst(compiler, SELECT_OP(DMUH, MUH) | S(src1) | T(src2) | DA(EQUAL_FLAG), EQUAL_FLAG));
-#else /* !SLJIT_MIPS_R6 */
+#else /* SLJIT_MIPS_REV < 6 */
 		FAIL_IF(push_inst(compiler, SELECT_OP(DMULT, MULT) | S(src1) | T(src2), MOVABLE_INS));
 		FAIL_IF(push_inst(compiler, MFHI | DA(EQUAL_FLAG), EQUAL_FLAG));
 		FAIL_IF(push_inst(compiler, MFLO | D(dst), DR(dst)));
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
 		FAIL_IF(push_inst(compiler, SELECT_OP(DSRA32, SRA) | T(dst) | DA(OTHER_FLAG) | SH_IMM(31), OTHER_FLAG));
 		return push_inst(compiler, SELECT_OP(DSUBU, SUBU) | SA(EQUAL_FLAG) | TA(OTHER_FLAG) | DA(OTHER_FLAG), OTHER_FLAG);
 
diff --git a/sljit/sljitNativeMIPS_common.c b/sljit/sljitNativeMIPS_common.c
index 7d1d087..7628be6 100644
--- a/sljit/sljitNativeMIPS_common.c
+++ b/sljit/sljitNativeMIPS_common.c
@@ -25,15 +25,16 @@
  */
 
 /* Latest MIPS architecture. */
-/* Automatically detect SLJIT_MIPS_R1 */
 
-#if (defined __mips_isa_rev) && (__mips_isa_rev >= 6)
-#define SLJIT_MIPS_R6 1
+#ifndef __mips_hard_float
+/* Disable automatic detection, covers both -msoft-float and -mno-float */
+#undef SLJIT_IS_FPU_AVAILABLE
+#define SLJIT_IS_FPU_AVAILABLE 0
 #endif
 
 SLJIT_API_FUNC_ATTRIBUTE const char* sljit_get_platform_name(void)
 {
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 
 #if (defined SLJIT_CONFIG_MIPS_32 && SLJIT_CONFIG_MIPS_32)
 	return "MIPS32-R6" SLJIT_CPUINFO;
@@ -41,7 +42,7 @@ SLJIT_API_FUNC_ATTRIBUTE const char* sljit_get_platform_name(void)
 	return "MIPS64-R6" SLJIT_CPUINFO;
 #endif /* SLJIT_CONFIG_MIPS_32 */
 
-#elif (defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1)
+#elif (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 1)
 
 #if (defined SLJIT_CONFIG_MIPS_32 && SLJIT_CONFIG_MIPS_32)
 	return "MIPS32-R1" SLJIT_CPUINFO;
@@ -49,9 +50,9 @@ SLJIT_API_FUNC_ATTRIBUTE const char* sljit_get_platform_name(void)
 	return "MIPS64-R1" SLJIT_CPUINFO;
 #endif /* SLJIT_CONFIG_MIPS_32 */
 
-#else /* SLJIT_MIPS_R1 */
+#else /* SLJIT_MIPS_REV < 1 */
 	return "MIPS III" SLJIT_CPUINFO;
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
 }
 
 /* Length of an instruction word
@@ -117,11 +118,11 @@ static const sljit_u8 freg_map[SLJIT_NUMBER_OF_FLOAT_REGISTERS + 4] = {
 #define FR(dr)		(freg_map[dr])
 #define HI(opcode)	((opcode) << 26)
 #define LO(opcode)	(opcode)
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 /* CMP.cond.fmt */
 /* S = (20 << 21) D = (21 << 21) */
 #define CMP_FMT_S	(20 << 21)
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
 /* S = (16 << 21) D = (17 << 21) */
 #define FMT_S		(16 << 21)
 #define FMT_D		(17 << 21)
@@ -134,13 +135,13 @@ static const sljit_u8 freg_map[SLJIT_NUMBER_OF_FLOAT_REGISTERS + 4] = {
 #define ANDI		(HI(12))
 #define B		(HI(4))
 #define BAL		(HI(1) | (17 << 16))
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 #define BC1EQZ		(HI(17) | (9 << 21) | FT(TMP_FREG3))
 #define BC1NEZ		(HI(17) | (13 << 21) | FT(TMP_FREG3))
-#else /* !SLJIT_MIPS_R6 */
+#else /* SLJIT_MIPS_REV < 6 */
 #define BC1F		(HI(17) | (8 << 21))
 #define BC1T		(HI(17) | (8 << 21) | (1 << 16))
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
 #define BEQ		(HI(4))
 #define BGEZ		(HI(1) | (1 << 16))
 #define BGTZ		(HI(7))
@@ -149,23 +150,23 @@ static const sljit_u8 freg_map[SLJIT_NUMBER_OF_FLOAT_REGISTERS + 4] = {
 #define BNE		(HI(5))
 #define BREAK		(HI(0) | LO(13))
 #define CFC1		(HI(17) | (2 << 21))
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 #define C_UEQ_S		(HI(17) | CMP_FMT_S | LO(3))
 #define C_ULE_S		(HI(17) | CMP_FMT_S | LO(7))
 #define C_ULT_S		(HI(17) | CMP_FMT_S | LO(5))
 #define C_UN_S		(HI(17) | CMP_FMT_S | LO(1))
 #define C_FD		(FD(TMP_FREG3))
-#else /* !SLJIT_MIPS_R6 */
+#else /* SLJIT_MIPS_REV < 6 */
 #define C_UEQ_S		(HI(17) | FMT_S | LO(51))
 #define C_ULE_S		(HI(17) | FMT_S | LO(55))
 #define C_ULT_S		(HI(17) | FMT_S | LO(53))
 #define C_UN_S		(HI(17) | FMT_S | LO(49))
 #define C_FD		(0)
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
 #define CVT_S_S		(HI(17) | FMT_S | LO(32))
 #define DADDIU		(HI(25))
 #define DADDU		(HI(0) | LO(45))
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 #define DDIV		(HI(0) | (2 << 6) | LO(30))
 #define DDIVU		(HI(0) | (2 << 6) | LO(31))
 #define DMOD		(HI(0) | (3 << 6) | LO(30))
@@ -176,14 +177,14 @@ static const sljit_u8 freg_map[SLJIT_NUMBER_OF_FLOAT_REGISTERS + 4] = {
 #define DMUHU		(HI(0) | (3 << 6) | LO(29))
 #define DMUL		(HI(0) | (2 << 6) | LO(28))
 #define DMULU		(HI(0) | (2 << 6) | LO(29))
-#else /* !SLJIT_MIPS_R6 */
+#else /* SLJIT_MIPS_REV < 6 */
 #define DDIV		(HI(0) | LO(30))
 #define DDIVU		(HI(0) | LO(31))
 #define DIV		(HI(0) | LO(26))
 #define DIVU		(HI(0) | LO(27))
 #define DMULT		(HI(0) | LO(28))
 #define DMULTU		(HI(0) | LO(29))
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
 #define DIV_S		(HI(17) | FMT_S | LO(3))
 #define DSLL		(HI(0) | LO(56))
 #define DSLL32		(HI(0) | LO(60))
@@ -198,33 +199,33 @@ static const sljit_u8 freg_map[SLJIT_NUMBER_OF_FLOAT_REGISTERS + 4] = {
 #define J		(HI(2))
 #define JAL		(HI(3))
 #define JALR		(HI(0) | LO(9))
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 #define JR		(HI(0) | LO(9))
-#else /* !SLJIT_MIPS_R6 */
+#else /* SLJIT_MIPS_REV < 6 */
 #define JR		(HI(0) | LO(8))
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
 #define LD		(HI(55))
 #define LUI		(HI(15))
 #define LW		(HI(35))
 #define MFC1		(HI(17))
-#if !(defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
-#define MFHI		(HI(0) | LO(16))
-#define MFLO		(HI(0) | LO(18))
-#else /* SLJIT_MIPS_R6 */
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 #define MOD		(HI(0) | (3 << 6) | LO(26))
 #define MODU		(HI(0) | (3 << 6) | LO(27))
-#endif /* !SLJIT_MIPS_R6 */
+#else /* SLJIT_MIPS_REV < 6 */
+#define MFHI		(HI(0) | LO(16))
+#define MFLO		(HI(0) | LO(18))
+#endif /* SLJIT_MIPS_REV >= 6 */
 #define MOV_S		(HI(17) | FMT_S | LO(6))
 #define MTC1		(HI(17) | (4 << 21))
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 #define MUH		(HI(0) | (3 << 6) | LO(24))
 #define MUHU		(HI(0) | (3 << 6) | LO(25))
 #define MUL		(HI(0) | (2 << 6) | LO(24))
 #define MULU		(HI(0) | (2 << 6) | LO(25))
-#else /* !SLJIT_MIPS_R6 */
+#else /* SLJIT_MIPS_REV < 6 */
 #define MULT		(HI(0) | LO(24))
 #define MULTU		(HI(0) | LO(25))
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
 #define MUL_S		(HI(17) | FMT_S | LO(2))
 #define NEG_S		(HI(17) | FMT_S | LO(7))
 #define NOP		(HI(0) | LO(0))
@@ -251,23 +252,23 @@ static const sljit_u8 freg_map[SLJIT_NUMBER_OF_FLOAT_REGISTERS + 4] = {
 #define XOR		(HI(0) | LO(38))
 #define XORI		(HI(14))
 
-#if (defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1) || (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 1)
 #define CLZ		(HI(28) | LO(32))
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 #define DCLZ		(LO(18))
-#else /* !SLJIT_MIPS_R6 */
+#else /* SLJIT_MIPS_REV < 6 */
 #define DCLZ		(HI(28) | LO(36))
 #define MOVF		(HI(0) | (0 << 16) | LO(1))
 #define MOVN		(HI(0) | LO(11))
 #define MOVT		(HI(0) | (1 << 16) | LO(1))
 #define MOVZ		(HI(0) | LO(10))
 #define MUL		(HI(28) | LO(2))
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
 #define PREF		(HI(51))
 #define PREFX		(HI(19) | LO(15))
 #define SEB		(HI(31) | (16 << 6) | LO(32))
 #define SEH		(HI(31) | (24 << 6) | LO(32))
-#endif
+#endif /* SLJIT_MIPS_REV >= 1 */
 
 #if (defined SLJIT_CONFIG_MIPS_32 && SLJIT_CONFIG_MIPS_32)
 #define ADDU_W		ADDU
@@ -303,10 +304,10 @@ static SLJIT_INLINE sljit_ins invert_branch(sljit_s32 flags)
 {
 	if (flags & IS_BIT26_COND)
 		return (1 << 26);
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 	if (flags & IS_BIT23_COND)
 		return (1 << 23);
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
 	return (1 << 16);
 }
 
@@ -684,11 +685,12 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_has_cpu_feature(sljit_s32 feature_type)
 #error "FIR check is not implemented for this architecture"
 #endif
 
-#if (defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 1)
 	case SLJIT_HAS_CLZ:
 	case SLJIT_HAS_CMOV:
+	case SLJIT_HAS_PREFETCH:
 		return 1;
-#endif
+#endif /* SLJIT_MIPS_REV >= 1 */
 
 	default:
 		return fir;
@@ -1230,7 +1232,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compile
 		return push_inst(compiler, NOP, UNMOVABLE_INS);
 	case SLJIT_LMUL_UW:
 	case SLJIT_LMUL_SW:
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 #if (defined SLJIT_CONFIG_MIPS_64 && SLJIT_CONFIG_MIPS_64)
 		FAIL_IF(push_inst(compiler, (op == SLJIT_LMUL_UW ? DMULU : DMUL) | S(SLJIT_R0) | T(SLJIT_R1) | D(TMP_REG3), DR(TMP_REG3)));
 		FAIL_IF(push_inst(compiler, (op == SLJIT_LMUL_UW ? DMUHU : DMUH) | S(SLJIT_R0) | T(SLJIT_R1) | D(TMP_REG1), DR(TMP_REG1)));
@@ -1240,7 +1242,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compile
 #endif /* SLJIT_CONFIG_MIPS_64 */
 		FAIL_IF(push_inst(compiler, ADDU_W | S(TMP_REG3) | TA(0) | D(SLJIT_R0), DR(SLJIT_R0)));
 		return push_inst(compiler, ADDU_W | S(TMP_REG1) | TA(0) | D(SLJIT_R1), DR(SLJIT_R1));
-#else /* !SLJIT_MIPS_R6 */
+#else /* SLJIT_MIPS_REV < 6 */
 #if (defined SLJIT_CONFIG_MIPS_64 && SLJIT_CONFIG_MIPS_64)
 		FAIL_IF(push_inst(compiler, (op == SLJIT_LMUL_UW ? DMULTU : DMULT) | S(SLJIT_R0) | T(SLJIT_R1), MOVABLE_INS));
 #else /* !SLJIT_CONFIG_MIPS_64 */
@@ -1248,13 +1250,13 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compile
 #endif /* SLJIT_CONFIG_MIPS_64 */
 		FAIL_IF(push_inst(compiler, MFLO | D(SLJIT_R0), DR(SLJIT_R0)));
 		return push_inst(compiler, MFHI | D(SLJIT_R1), DR(SLJIT_R1));
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
 	case SLJIT_DIVMOD_UW:
 	case SLJIT_DIVMOD_SW:
 	case SLJIT_DIV_UW:
 	case SLJIT_DIV_SW:
 		SLJIT_COMPILE_ASSERT((SLJIT_DIVMOD_UW & 0x2) == 0 && SLJIT_DIV_UW - 0x2 == SLJIT_DIVMOD_UW, bad_div_opcode_assignments);
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 #if (defined SLJIT_CONFIG_MIPS_64 && SLJIT_CONFIG_MIPS_64)
 		if (int_op) {
 			FAIL_IF(push_inst(compiler, ((op | 0x2) == SLJIT_DIV_UW ? DIVU : DIV) | S(SLJIT_R0) | T(SLJIT_R1) | D(TMP_REG3), DR(TMP_REG3)));
@@ -1270,11 +1272,11 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compile
 #endif /* SLJIT_CONFIG_MIPS_64 */
 		FAIL_IF(push_inst(compiler, ADDU_W | S(TMP_REG3) | TA(0) | D(SLJIT_R0), DR(SLJIT_R0)));
 		return (op >= SLJIT_DIV_UW) ? SLJIT_SUCCESS : push_inst(compiler, ADDU_W | S(TMP_REG1) | TA(0) | D(SLJIT_R1), DR(SLJIT_R1));
-#else /* !SLJIT_MIPS_R6 */
-#if !(defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1)
+#else /* SLJIT_MIPS_REV < 6 */
+#if !(defined SLJIT_MIPS_REV)
 		FAIL_IF(push_inst(compiler, NOP, UNMOVABLE_INS));
 		FAIL_IF(push_inst(compiler, NOP, UNMOVABLE_INS));
-#endif /* !SLJIT_MIPS_R1 */
+#endif /* !SLJIT_MIPS_REV */
 #if (defined SLJIT_CONFIG_MIPS_64 && SLJIT_CONFIG_MIPS_64)
 		if (int_op)
 			FAIL_IF(push_inst(compiler, ((op | 0x2) == SLJIT_DIV_UW ? DIVU : DIV) | S(SLJIT_R0) | T(SLJIT_R1), MOVABLE_INS));
@@ -1285,13 +1287,16 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compile
 #endif /* SLJIT_CONFIG_MIPS_64 */
 		FAIL_IF(push_inst(compiler, MFLO | D(SLJIT_R0), DR(SLJIT_R0)));
 		return (op >= SLJIT_DIV_UW) ? SLJIT_SUCCESS : push_inst(compiler, MFHI | D(SLJIT_R1), DR(SLJIT_R1));
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
+	case SLJIT_ENDBR:
+	case SLJIT_SKIP_FRAMES_BEFORE_RETURN:
+		return SLJIT_SUCCESS;
 	}
 
 	return SLJIT_SUCCESS;
 }
 
-#if (defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 1)
 static sljit_s32 emit_prefetch(struct sljit_compiler *compiler,
         sljit_s32 src, sljit_sw srcw)
 {
@@ -1312,7 +1317,7 @@ static sljit_s32 emit_prefetch(struct sljit_compiler *compiler,
 
 	return push_inst(compiler, PREFX | S(src & REG_MASK) | T(OFFS_REG(src)), MOVABLE_INS);
 }
-#endif
+#endif /* SLJIT_MIPS_REV >= 1 */
 
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op1(struct sljit_compiler *compiler, sljit_s32 op,
 	sljit_s32 dst, sljit_sw dstw,
@@ -1329,14 +1334,6 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op1(struct sljit_compiler *compile
 	ADJUST_LOCAL_OFFSET(dst, dstw);
 	ADJUST_LOCAL_OFFSET(src, srcw);
 
-	if (dst == SLJIT_UNUSED && !HAS_FLAGS(op)) {
-#if (defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1)
-		if (op <= SLJIT_MOV_P && (src & SLJIT_MEM))
-			return emit_prefetch(compiler, src, srcw);
-#endif
-		return SLJIT_SUCCESS;
-	}
-
 #if (defined SLJIT_CONFIG_MIPS_64 && SLJIT_CONFIG_MIPS_64)
 	if ((op & SLJIT_I32_OP) && GET_OPCODE(op) >= SLJIT_NOT)
 		flags |= INT_DATA | SIGNED_DATA;
@@ -1463,6 +1460,38 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2(struct sljit_compiler *compile
 #endif
 }
 
+SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_src(struct sljit_compiler *compiler, sljit_s32 op,
+	sljit_s32 src, sljit_sw srcw)
+{
+	CHECK_ERROR();
+	CHECK(check_sljit_emit_op_src(compiler, op, src, srcw));
+	ADJUST_LOCAL_OFFSET(src, srcw);
+
+	switch (op) {
+	case SLJIT_FAST_RETURN:
+		if (FAST_IS_REG(src))
+			FAIL_IF(push_inst(compiler, ADDU_W | S(src) | TA(0) | DA(RETURN_ADDR_REG), RETURN_ADDR_REG));
+		else
+			FAIL_IF(emit_op_mem(compiler, WORD_DATA | LOAD_DATA, RETURN_ADDR_REG, src, srcw));
+
+		FAIL_IF(push_inst(compiler, JR | SA(RETURN_ADDR_REG), UNMOVABLE_INS));
+		return push_inst(compiler, NOP, UNMOVABLE_INS);
+	case SLJIT_SKIP_FRAMES_BEFORE_FAST_RETURN:
+		return SLJIT_SUCCESS;
+	case SLJIT_PREFETCH_L1:
+	case SLJIT_PREFETCH_L2:
+	case SLJIT_PREFETCH_L3:
+	case SLJIT_PREFETCH_ONCE:
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 1)
+		return emit_prefetch(compiler, src, srcw);
+#else /* SLJIT_MIPS_REV < 1 */
+		return SLJIT_SUCCESS;
+#endif /* SLJIT_MIPS_REV >= 1 */
+	}
+
+	return SLJIT_SUCCESS;
+}
+
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_get_register_index(sljit_s32 reg)
 {
 	CHECK_REG_INDEX(check_sljit_get_register_index(reg));
@@ -1732,25 +1761,12 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_enter(struct sljit_compiler *
 	ADJUST_LOCAL_OFFSET(dst, dstw);
 
 	if (FAST_IS_REG(dst))
-		return push_inst(compiler, ADDU_W | SA(RETURN_ADDR_REG) | TA(0) | D(dst), DR(dst));
+		return push_inst(compiler, ADDU_W | SA(RETURN_ADDR_REG) | TA(0) | D(dst), UNMOVABLE_INS);
 
 	/* Memory. */
-	return emit_op_mem(compiler, WORD_DATA, RETURN_ADDR_REG, dst, dstw);
-}
-
-SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src, sljit_sw srcw)
-{
-	CHECK_ERROR();
-	CHECK(check_sljit_emit_fast_return(compiler, src, srcw));
-	ADJUST_LOCAL_OFFSET(src, srcw);
-
-	if (FAST_IS_REG(src))
-		FAIL_IF(push_inst(compiler, ADDU_W | S(src) | TA(0) | DA(RETURN_ADDR_REG), RETURN_ADDR_REG));
-	else
-		FAIL_IF(emit_op_mem(compiler, WORD_DATA | LOAD_DATA, RETURN_ADDR_REG, src, srcw));
-
-	FAIL_IF(push_inst(compiler, JR | SA(RETURN_ADDR_REG), UNMOVABLE_INS));
-	return push_inst(compiler, NOP, UNMOVABLE_INS);
+	FAIL_IF(emit_op_mem(compiler, WORD_DATA, RETURN_ADDR_REG, dst, dstw));
+	compiler->delay_slot = UNMOVABLE_INS;
+	return SLJIT_SUCCESS;
 }
 
 /* --------------------------------------------------------------------- */
@@ -1790,7 +1806,7 @@ SLJIT_API_FUNC_ATTRIBUTE struct sljit_label* sljit_emit_label(struct sljit_compi
 	flags = IS_BIT26_COND; \
 	delay_check = src;
 
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 
 #define BR_T() \
 	inst = BC1NEZ; \
@@ -1801,7 +1817,7 @@ SLJIT_API_FUNC_ATTRIBUTE struct sljit_label* sljit_emit_label(struct sljit_compi
 	flags = IS_BIT23_COND; \
 	delay_check = FCSR_FCC;
 
-#else /* !SLJIT_MIPS_R6 */
+#else /* SLJIT_MIPS_REV < 6 */
 
 #define BR_T() \
 	inst = BC1T | JUMP_LENGTH; \
@@ -1812,7 +1828,7 @@ SLJIT_API_FUNC_ATTRIBUTE struct sljit_label* sljit_emit_label(struct sljit_compi
 	flags = IS_BIT16_COND; \
 	delay_check = FCSR_FCC;
 
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
 
 SLJIT_API_FUNC_ATTRIBUTE struct sljit_jump* sljit_emit_jump(struct sljit_compiler *compiler, sljit_s32 type)
 {
@@ -2123,11 +2139,11 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_flags(struct sljit_compiler *co
 	case SLJIT_GREATER_EQUAL_F64:
 	case SLJIT_UNORDERED_F64:
 	case SLJIT_ORDERED_F64:
-#if (defined SLJIT_MIPS_R6 && SLJIT_MIPS_R6)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 6)
 		FAIL_IF(push_inst(compiler, MFC1 | TA(dst_ar) | FS(TMP_FREG3), dst_ar));
-#else /* !SLJIT_MIPS_R6 */
+#else /* SLJIT_MIPS_REV < 6 */
 		FAIL_IF(push_inst(compiler, CFC1 | TA(dst_ar) | DA(FCSR_REG), dst_ar));
-#endif /* SLJIT_MIPS_R6 */
+#endif /* SLJIT_MIPS_REV >= 6 */
 		FAIL_IF(push_inst(compiler, SRL | TA(dst_ar) | DA(dst_ar) | SH_IMM(23), dst_ar));
 		FAIL_IF(push_inst(compiler, ANDI | SA(dst_ar) | TA(dst_ar) | IMM(1), dst_ar));
 		src_ar = dst_ar;
@@ -2167,14 +2183,14 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_cmov(struct sljit_compiler *compil
 	sljit_s32 dst_reg,
 	sljit_s32 src, sljit_sw srcw)
 {
-#if (defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 1)
 	sljit_ins ins;
-#endif
+#endif /* SLJIT_MIPS_REV >= 1 */
 
 	CHECK_ERROR();
 	CHECK(check_sljit_emit_cmov(compiler, type, dst_reg, src, srcw));
 
-#if (defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1)
+#if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 1)
 
 	if (SLJIT_UNLIKELY(src & SLJIT_IMM)) {
 #if (defined SLJIT_CONFIG_MIPS_64 && SLJIT_CONFIG_MIPS_64)
@@ -2231,9 +2247,9 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_cmov(struct sljit_compiler *compil
 
 	return push_inst(compiler, ins | S(src) | D(dst_reg), DR(dst_reg));
 
-#else
+#else /* SLJIT_MIPS_REV < 1 */
 	return sljit_emit_cmov_generic(compiler, type, dst_reg, src, srcw);
-#endif
+#endif /* SLJIT_MIPS_REV >= 1 */
 }
 
 SLJIT_API_FUNC_ATTRIBUTE struct sljit_const* sljit_emit_const(struct sljit_compiler *compiler, sljit_s32 dst, sljit_sw dstw, sljit_sw init_value)
diff --git a/sljit/sljitNativePPC_common.c b/sljit/sljitNativePPC_common.c
index e827514..17bf9a9 100644
--- a/sljit/sljitNativePPC_common.c
+++ b/sljit/sljitNativePPC_common.c
@@ -627,6 +627,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_has_cpu_feature(sljit_s32 feature_type)
 #endif
 
 	case SLJIT_HAS_CLZ:
+	case SLJIT_HAS_PREFETCH:
 		return 1;
 
 	default:
@@ -1158,6 +1159,9 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compile
 #else
 		return push_inst(compiler, (op == SLJIT_DIV_UW ? DIVWU : DIVW) | D(SLJIT_R0) | A(SLJIT_R0) | B(SLJIT_R1));
 #endif
+	case SLJIT_ENDBR:
+	case SLJIT_SKIP_FRAMES_BEFORE_RETURN:
+		return SLJIT_SUCCESS;
 	}
 
 	return SLJIT_SUCCESS;
@@ -1203,13 +1207,6 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op1(struct sljit_compiler *compile
 	ADJUST_LOCAL_OFFSET(dst, dstw);
 	ADJUST_LOCAL_OFFSET(src, srcw);
 
-	if (dst == SLJIT_UNUSED && !HAS_FLAGS(op)) {
-		if (op <= SLJIT_MOV_P && (src & SLJIT_MEM))
-			return emit_prefetch(compiler, src, srcw);
-
-		return SLJIT_SUCCESS;
-	}
-
 	op = GET_OPCODE(op);
 	if ((src & SLJIT_IMM) && srcw == 0)
 		src = TMP_ZERO;
@@ -1536,6 +1533,35 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2(struct sljit_compiler *compile
 	return SLJIT_SUCCESS;
 }
 
+SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_src(struct sljit_compiler *compiler, sljit_s32 op,
+	sljit_s32 src, sljit_sw srcw)
+{
+	CHECK_ERROR();
+	CHECK(check_sljit_emit_op_src(compiler, op, src, srcw));
+	ADJUST_LOCAL_OFFSET(src, srcw);
+
+	switch (op) {
+	case SLJIT_FAST_RETURN:
+		if (FAST_IS_REG(src))
+			FAIL_IF(push_inst(compiler, MTLR | S(src)));
+		else {
+			FAIL_IF(emit_op(compiler, SLJIT_MOV, WORD_DATA, TMP_REG2, 0, TMP_REG1, 0, src, srcw));
+			FAIL_IF(push_inst(compiler, MTLR | S(TMP_REG2)));
+		}
+
+		return push_inst(compiler, BLR);
+	case SLJIT_SKIP_FRAMES_BEFORE_FAST_RETURN:
+		return SLJIT_SUCCESS;
+	case SLJIT_PREFETCH_L1:
+	case SLJIT_PREFETCH_L2:
+	case SLJIT_PREFETCH_L3:
+	case SLJIT_PREFETCH_ONCE:
+		return emit_prefetch(compiler, src, srcw);
+	}
+
+	return SLJIT_SUCCESS;
+}
+
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_get_register_index(sljit_s32 reg)
 {
 	CHECK_REG_INDEX(check_sljit_get_register_index(reg));
@@ -1854,22 +1880,6 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_enter(struct sljit_compiler *
 	return emit_op(compiler, SLJIT_MOV, WORD_DATA, dst, dstw, TMP_REG1, 0, TMP_REG2, 0);
 }
 
-SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src, sljit_sw srcw)
-{
-	CHECK_ERROR();
-	CHECK(check_sljit_emit_fast_return(compiler, src, srcw));
-	ADJUST_LOCAL_OFFSET(src, srcw);
-
-	if (FAST_IS_REG(src))
-		FAIL_IF(push_inst(compiler, MTLR | S(src)));
-	else {
-		FAIL_IF(emit_op(compiler, SLJIT_MOV, WORD_DATA, TMP_REG2, 0, TMP_REG1, 0, src, srcw));
-		FAIL_IF(push_inst(compiler, MTLR | S(TMP_REG2)));
-	}
-
-	return push_inst(compiler, BLR);
-}
-
 /* --------------------------------------------------------------------- */
 /*  Conditional instructions                                             */
 /* --------------------------------------------------------------------- */
diff --git a/sljit/sljitNativeSPARC_common.c b/sljit/sljitNativeSPARC_common.c
index bfa4ece..4c95350 100644
--- a/sljit/sljitNativeSPARC_common.c
+++ b/sljit/sljitNativeSPARC_common.c
@@ -872,6 +872,9 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compile
 #else
 #error "Implementation required"
 #endif
+	case SLJIT_ENDBR:
+	case SLJIT_SKIP_FRAMES_BEFORE_RETURN:
+		return SLJIT_SUCCESS;
 	}
 
 	return SLJIT_SUCCESS;
@@ -888,9 +891,6 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op1(struct sljit_compiler *compile
 	ADJUST_LOCAL_OFFSET(dst, dstw);
 	ADJUST_LOCAL_OFFSET(src, srcw);
 
-	if (dst == SLJIT_UNUSED && !HAS_FLAGS(op))
-		return SLJIT_SUCCESS;
-
 	op = GET_OPCODE(op);
 	switch (op) {
 	case SLJIT_MOV:
@@ -971,6 +971,33 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2(struct sljit_compiler *compile
 	return SLJIT_SUCCESS;
 }
 
+SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_src(struct sljit_compiler *compiler, sljit_s32 op,
+	sljit_s32 src, sljit_sw srcw)
+{
+	CHECK_ERROR();
+	CHECK(check_sljit_emit_op_src(compiler, op, src, srcw));
+	ADJUST_LOCAL_OFFSET(src, srcw);
+
+	switch (op) {
+	case SLJIT_FAST_RETURN:
+		if (FAST_IS_REG(src))
+			FAIL_IF(push_inst(compiler, OR | D(TMP_LINK) | S1(0) | S2(src), DR(TMP_LINK)));
+		else
+			FAIL_IF(emit_op_mem(compiler, WORD_DATA | LOAD_DATA, TMP_LINK, src, srcw));
+
+		FAIL_IF(push_inst(compiler, JMPL | D(0) | S1(TMP_LINK) | IMM(8), UNMOVABLE_INS));
+		return push_inst(compiler, NOP, UNMOVABLE_INS);
+	case SLJIT_SKIP_FRAMES_BEFORE_FAST_RETURN:
+	case SLJIT_PREFETCH_L1:
+	case SLJIT_PREFETCH_L2:
+	case SLJIT_PREFETCH_L3:
+	case SLJIT_PREFETCH_ONCE:
+		return SLJIT_SUCCESS;
+	}
+
+	return SLJIT_SUCCESS;
+}
+
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_get_register_index(sljit_s32 reg)
 {
 	CHECK_REG_INDEX(check_sljit_get_register_index(reg));
@@ -1215,25 +1242,12 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_enter(struct sljit_compiler *
 	ADJUST_LOCAL_OFFSET(dst, dstw);
 
 	if (FAST_IS_REG(dst))
-		return push_inst(compiler, OR | D(dst) | S1(0) | S2(TMP_LINK), DR(dst));
+		return push_inst(compiler, OR | D(dst) | S1(0) | S2(TMP_LINK), UNMOVABLE_INS);
 
 	/* Memory. */
-	return emit_op_mem(compiler, WORD_DATA, TMP_LINK, dst, dstw);
-}
-
-SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src, sljit_sw srcw)
-{
-	CHECK_ERROR();
-	CHECK(check_sljit_emit_fast_return(compiler, src, srcw));
-	ADJUST_LOCAL_OFFSET(src, srcw);
-
-	if (FAST_IS_REG(src))
-		FAIL_IF(push_inst(compiler, OR | D(TMP_LINK) | S1(0) | S2(src), DR(TMP_LINK)));
-	else
-		FAIL_IF(emit_op_mem(compiler, WORD_DATA | LOAD_DATA, TMP_LINK, src, srcw));
-
-	FAIL_IF(push_inst(compiler, JMPL | D(0) | S1(TMP_LINK) | IMM(8), UNMOVABLE_INS));
-	return push_inst(compiler, NOP, UNMOVABLE_INS);
+	FAIL_IF(emit_op_mem(compiler, WORD_DATA, TMP_LINK, dst, dstw));
+	compiler->delay_slot = UNMOVABLE_INS;
+	return SLJIT_SUCCESS;
 }
 
 /* --------------------------------------------------------------------- */
diff --git a/sljit/sljitNativeTILEGX_64.c b/sljit/sljitNativeTILEGX_64.c
index 003f43a..d69ecd6 100644
--- a/sljit/sljitNativeTILEGX_64.c
+++ b/sljit/sljitNativeTILEGX_64.c
@@ -1564,24 +1564,6 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_enter(struct sljit_compiler *
 	return emit_op_mem(compiler, WORD_DATA, RA, dst, dstw);
 }
 
-SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src, sljit_sw srcw)
-{
-	CHECK_ERROR();
-	CHECK(check_sljit_emit_fast_return(compiler, src, srcw));
-	ADJUST_LOCAL_OFFSET(src, srcw);
-
-	if (FAST_IS_REG(src))
-		FAIL_IF(ADD(RA, reg_map[src], ZERO));
-
-	else if (src & SLJIT_MEM)
-		FAIL_IF(emit_op_mem(compiler, WORD_DATA | LOAD_DATA, RA, src, srcw));
-
-	else if (src & SLJIT_IMM)
-		FAIL_IF(load_immediate(compiler, RA, srcw));
-
-	return JR(RA);
-}
-
 static SLJIT_INLINE sljit_s32 emit_single_op(struct sljit_compiler *compiler, sljit_s32 op, sljit_s32 flags, sljit_s32 dst, sljit_s32 src1, sljit_sw src2)
 {
 	sljit_s32 overflow_ra = 0;
@@ -2184,6 +2166,9 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compile
 	case SLJIT_DIV_UW:
 	case SLJIT_DIV_SW:
 		SLJIT_UNREACHABLE();
+	case SLJIT_ENDBR:
+	case SLJIT_SKIP_FRAMES_BEFORE_RETURN:
+		return SLJIT_SUCCESS;
 	}
 
 	return SLJIT_SUCCESS;
@@ -2293,6 +2278,29 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2(struct sljit_compiler *compile
 	return SLJIT_SUCCESS;
 }
 
+SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_src(struct sljit_compiler *compiler, sljit_s32 op,
+	sljit_s32 src, sljit_sw srcw)
+{
+	CHECK_ERROR();
+	CHECK(check_sljit_emit_op_src(compiler, op, src, srcw));
+	ADJUST_LOCAL_OFFSET(src, srcw);
+
+	switch (op) {
+	case SLJIT_FAST_RETURN:
+		if (FAST_IS_REG(src))
+			FAIL_IF(ADD(RA, reg_map[src], ZERO));
+
+		else
+			FAIL_IF(emit_op_mem(compiler, WORD_DATA | LOAD_DATA, RA, src, srcw));
+
+		return JR(RA);
+	case SLJIT_SKIP_FRAMES_BEFORE_FAST_RETURN:
+		return SLJIT_SUCCESS;
+	}
+
+	return SLJIT_SUCCESS;
+}
+
 SLJIT_API_FUNC_ATTRIBUTE struct sljit_label * sljit_emit_label(struct sljit_compiler *compiler)
 {
 	struct sljit_label *label;
diff --git a/sljit/sljitNativeX86_32.c b/sljit/sljitNativeX86_32.c
index 34a3a3d..79a7e8b 100644
--- a/sljit/sljitNativeX86_32.c
+++ b/sljit/sljitNativeX86_32.c
@@ -76,6 +76,9 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_enter(struct sljit_compiler *compi
 	CHECK(check_sljit_emit_enter(compiler, options, arg_types, scratches, saveds, fscratches, fsaveds, local_size));
 	set_emit_enter(compiler, options, arg_types, scratches, saveds, fscratches, fsaveds, local_size);
 
+	/* Emit ENDBR32 at function entry if needed.  */
+	FAIL_IF(emit_endbranch(compiler));
+
 	args = get_arg_count(arg_types);
 	compiler->args = args;
 
@@ -307,14 +310,11 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_return(struct sljit_compiler *comp
 		SLJIT_SP, 0, SLJIT_SP, 0, SLJIT_IMM, compiler->local_size));
 #endif
 
-	size = 2 + (compiler->scratches > 7 ? (compiler->scratches - 7) : 0) +
+	size = 2 + (compiler->scratches > 9 ? (compiler->scratches - 9) : 0) +
 		(compiler->saveds <= 3 ? compiler->saveds : 3);
 #if (defined SLJIT_X86_32_FASTCALL && SLJIT_X86_32_FASTCALL)
 	if (compiler->args > 2)
 		size += 2;
-#else
-	if (compiler->args > 0)
-		size += 2;
 #endif
 	inst = (sljit_u8*)ensure_buf(compiler, 1 + size);
 	FAIL_IF(!inst);
@@ -367,6 +367,8 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_s32
 	SLJIT_ASSERT((flags & (EX86_PREF_F2 | EX86_PREF_F3)) != (EX86_PREF_F2 | EX86_PREF_F3)
 		&& (flags & (EX86_PREF_F2 | EX86_PREF_66)) != (EX86_PREF_F2 | EX86_PREF_66)
 		&& (flags & (EX86_PREF_F3 | EX86_PREF_66)) != (EX86_PREF_F3 | EX86_PREF_66));
+	/* We don't support (%ebp). */
+	SLJIT_ASSERT(!(b & SLJIT_MEM) || immb || reg_map[b & REG_MASK] != 5);
 
 	size &= 0xf;
 	inst_size = size;
@@ -863,14 +865,10 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_enter(struct sljit_compiler *
 	return SLJIT_SUCCESS;
 }
 
-SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src, sljit_sw srcw)
+static sljit_s32 emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src, sljit_sw srcw)
 {
 	sljit_u8 *inst;
 
-	CHECK_ERROR();
-	CHECK(check_sljit_emit_fast_return(compiler, src, srcw));
-	ADJUST_LOCAL_OFFSET(src, srcw);
-
 	CHECK_EXTRA_REGS(src, srcw, (void)0);
 
 	if (FAST_IS_REG(src)) {
@@ -894,3 +892,37 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_return(struct sljit_compiler
 	RET();
 	return SLJIT_SUCCESS;
 }
+
+static sljit_s32 skip_frames_before_return(struct sljit_compiler *compiler)
+{
+	sljit_s32 size, saved_size;
+	sljit_s32 has_f64_aligment;
+
+	/* Don't adjust shadow stack if it isn't enabled.  */
+	if (!cpu_has_shadow_stack ())
+		return SLJIT_SUCCESS;
+
+	SLJIT_ASSERT(compiler->args >= 0);
+	SLJIT_ASSERT(compiler->local_size > 0);
+
+#if !defined(__APPLE__)
+	has_f64_aligment = compiler->options & SLJIT_F64_ALIGNMENT;
+#else
+	has_f64_aligment = 0;
+#endif
+
+	size = compiler->local_size;
+	saved_size = (1 + (compiler->scratches > 9 ? (compiler->scratches - 9) : 0) + (compiler->saveds <= 3 ? compiler->saveds : 3)) * sizeof(sljit_uw);
+	if (has_f64_aligment) {
+		/* mov TMP_REG1, [esp + local_size].  */
+		EMIT_MOV(compiler, TMP_REG1, 0, SLJIT_MEM1(SLJIT_SP), size);
+		/* mov TMP_REG1, [TMP_REG1+ saved_size].  */
+		EMIT_MOV(compiler, TMP_REG1, 0, SLJIT_MEM1(TMP_REG1), saved_size);
+		/* Move return address to [esp]. */
+		EMIT_MOV(compiler, SLJIT_MEM1(SLJIT_SP), 0, TMP_REG1, 0);
+		size = 0;
+	} else
+		size += saved_size;
+
+	return adjust_shadow_stack(compiler, SLJIT_UNUSED, 0, SLJIT_SP, size);
+}
diff --git a/sljit/sljitNativeX86_64.c b/sljit/sljitNativeX86_64.c
index 5758711..e85b56a 100644
--- a/sljit/sljitNativeX86_64.c
+++ b/sljit/sljitNativeX86_64.c
@@ -135,6 +135,9 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_enter(struct sljit_compiler *compi
 	CHECK(check_sljit_emit_enter(compiler, options, arg_types, scratches, saveds, fscratches, fsaveds, local_size));
 	set_emit_enter(compiler, options, arg_types, scratches, saveds, fscratches, fsaveds, local_size);
 
+	/* Emit ENDBR64 at function entry if needed.  */
+	FAIL_IF(emit_endbranch(compiler));
+
 	compiler->mode32 = 0;
 
 #ifdef _WIN64
@@ -796,14 +799,10 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_enter(struct sljit_compiler *
 	return SLJIT_SUCCESS;
 }
 
-SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src, sljit_sw srcw)
+static sljit_s32 emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src, sljit_sw srcw)
 {
 	sljit_u8 *inst;
 
-	CHECK_ERROR();
-	CHECK(check_sljit_emit_fast_return(compiler, src, srcw));
-	ADJUST_LOCAL_OFFSET(src, srcw);
-
 	if (FAST_IS_REG(src)) {
 		if (reg_map[src] < 8) {
 			inst = (sljit_u8*)ensure_buf(compiler, 1 + 1 + 1);
@@ -898,3 +897,22 @@ static sljit_s32 emit_mov_int(struct sljit_compiler *compiler, sljit_s32 sign,
 
 	return SLJIT_SUCCESS;
 }
+
+static sljit_s32 skip_frames_before_return(struct sljit_compiler *compiler)
+{
+	sljit_s32 tmp, size;
+
+	/* Don't adjust shadow stack if it isn't enabled.  */
+	if (!cpu_has_shadow_stack ())
+		return SLJIT_SUCCESS;
+
+	size = compiler->local_size;
+	tmp = compiler->scratches;
+	if (tmp >= SLJIT_FIRST_SAVED_REG)
+		size += (tmp - SLJIT_FIRST_SAVED_REG + 1) * sizeof(sljit_uw);
+	tmp = compiler->saveds < SLJIT_NUMBER_OF_SAVED_REGISTERS ? (SLJIT_S0 + 1 - compiler->saveds) : SLJIT_FIRST_SAVED_REG;
+	if (SLJIT_S0 >= tmp)
+		size += (SLJIT_S0 - tmp + 1) * sizeof(sljit_uw);
+
+	return adjust_shadow_stack(compiler, SLJIT_UNUSED, 0, SLJIT_SP, size);
+}
diff --git a/sljit/sljitNativeX86_common.c b/sljit/sljitNativeX86_common.c
index 6296da5..eea9510 100644
--- a/sljit/sljitNativeX86_common.c
+++ b/sljit/sljitNativeX86_common.c
@@ -657,6 +657,9 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_has_cpu_feature(sljit_s32 feature_type)
 			get_cpu_features();
 		return cpu_has_cmov;
 
+	case SLJIT_HAS_PREFETCH:
+		return 1;
+
 	case SLJIT_HAS_SSE2:
 #if (defined SLJIT_DETECT_SSE2 && SLJIT_DETECT_SSE2)
 		if (cpu_has_sse2 == -1)
@@ -702,6 +705,165 @@ static SLJIT_INLINE sljit_s32 emit_sse2_store(struct sljit_compiler *compiler,
 static SLJIT_INLINE sljit_s32 emit_sse2_load(struct sljit_compiler *compiler,
 	sljit_s32 single, sljit_s32 dst, sljit_s32 src, sljit_sw srcw);
 
+static sljit_s32 emit_cmp_binary(struct sljit_compiler *compiler,
+	sljit_s32 src1, sljit_sw src1w,
+	sljit_s32 src2, sljit_sw src2w);
+
+static SLJIT_INLINE sljit_s32 emit_endbranch(struct sljit_compiler *compiler)
+{
+#if (defined SLJIT_CONFIG_X86_CET && SLJIT_CONFIG_X86_CET)
+	/* Emit endbr32/endbr64 when CET is enabled.  */
+	sljit_u8 *inst;
+	inst = (sljit_u8*)ensure_buf(compiler, 1 + 4);
+	FAIL_IF(!inst);
+	INC_SIZE(4);
+	*inst++ = 0xf3;
+	*inst++ = 0x0f;
+	*inst++ = 0x1e;
+#if (defined SLJIT_CONFIG_X86_32 && SLJIT_CONFIG_X86_32)
+	*inst = 0xfb;
+#else
+	*inst = 0xfa;
+#endif
+#else
+	(void)compiler;
+#endif
+	return SLJIT_SUCCESS;
+}
+
+static SLJIT_INLINE sljit_s32 emit_rdssp(struct sljit_compiler *compiler, sljit_s32 reg)
+{
+#if (defined SLJIT_CONFIG_X86_CET && SLJIT_CONFIG_X86_CET)
+	sljit_u8 *inst;
+	sljit_s32 size;
+
+#if (defined SLJIT_CONFIG_X86_64 && SLJIT_CONFIG_X86_64)
+	size = 5;
+#else
+	size = 4;
+#endif
+
+	inst = (sljit_u8*)ensure_buf(compiler, 1 + size);
+	FAIL_IF(!inst);
+	INC_SIZE(size);
+	*inst++ = 0xf3;
+#if (defined SLJIT_CONFIG_X86_64 && SLJIT_CONFIG_X86_64)
+	*inst++ = REX_W | (reg_map[reg] <= 7 ? 0 : REX_B);
+#endif
+	*inst++ = 0x0f;
+	*inst++ = 0x1e;
+	*inst = (0x3 << 6) | (0x1 << 3) | (reg_map[reg] & 0x7);
+#else
+	(void)compiler;
+#endif
+	return SLJIT_SUCCESS;
+}
+
+static SLJIT_INLINE sljit_s32 emit_incssp(struct sljit_compiler *compiler, sljit_s32 reg)
+{
+#if (defined SLJIT_CONFIG_X86_CET && SLJIT_CONFIG_X86_CET)
+	sljit_u8 *inst;
+	sljit_s32 size;
+
+#if (defined SLJIT_CONFIG_X86_64 && SLJIT_CONFIG_X86_64)
+	size = 5;
+#else
+	size = 4;
+#endif
+
+	inst = (sljit_u8*)ensure_buf(compiler, 1 + size);
+	FAIL_IF(!inst);
+	INC_SIZE(size);
+	*inst++ = 0xf3;
+#if (defined SLJIT_CONFIG_X86_64 && SLJIT_CONFIG_X86_64)
+	*inst++ = REX_W | (reg_map[reg] <= 7 ? 0 : REX_B);
+#endif
+	*inst++ = 0x0f;
+	*inst++ = 0xae;
+	*inst = (0x3 << 6) | (0x5 << 3) | (reg_map[reg] & 0x7);
+#else
+	(void)compiler;
+#endif
+	return SLJIT_SUCCESS;
+}
+
+static SLJIT_INLINE sljit_s32 cpu_has_shadow_stack(void)
+{
+#if (defined SLJIT_CONFIG_X86_CET && SLJIT_CONFIG_X86_CET)
+	return _get_ssp() != 0;
+#else
+	return 0;
+#endif
+}
+
+static SLJIT_INLINE sljit_s32 adjust_shadow_stack(struct sljit_compiler *compiler,
+	sljit_s32 src, sljit_sw srcw, sljit_s32 base, sljit_sw disp)
+{
+#if (defined SLJIT_CONFIG_X86_CET && SLJIT_CONFIG_X86_CET)
+	sljit_u8 *inst;
+
+	sljit_s32 size_before_rdssp_inst = compiler->size;
+
+	/* Generate "RDSSP TMP_REG1". */
+	FAIL_IF(emit_rdssp(compiler, TMP_REG1));
+
+	/* Load return address on shadow stack into TMP_REG1. */
+#if (defined SLJIT_CONFIG_X86_32 && SLJIT_CONFIG_X86_32)
+	SLJIT_ASSERT(reg_map[TMP_REG1] == 5);
+
+	/* Hand code unsupported "mov 0x0(%ebp),%ebp". */
+	inst = (sljit_u8*)ensure_buf(compiler, 1 + 3);
+	FAIL_IF(!inst);
+	INC_SIZE(3);
+	*inst++ = 0x8b;
+	*inst++ = 0x6d;
+	*inst = 0;
+#else /* !SLJIT_CONFIG_X86_32 */
+	EMIT_MOV(compiler, TMP_REG1, 0, SLJIT_MEM1(TMP_REG1), 0);
+#endif /* SLJIT_CONFIG_X86_32 */
+
+	if (src == SLJIT_UNUSED) {
+		/* Return address is on stack.  */
+		src = SLJIT_MEM1(base);
+		srcw = disp;
+	}
+
+	/* Compare return address against TMP_REG1. */
+	FAIL_IF(emit_cmp_binary (compiler, TMP_REG1, 0, src, srcw));
+
+	/* Generate JZ to skip shadow stack ajdustment when shadow
+	   stack matches normal stack. */
+	inst = (sljit_u8*)ensure_buf(compiler, 1 + 2);
+	FAIL_IF(!inst);
+	INC_SIZE(2);
+	*inst++ = get_jump_code(SLJIT_EQUAL) - 0x10;
+	sljit_uw size_jz_after_cmp_inst = compiler->size;
+	sljit_u8 *jz_after_cmp_inst = inst;
+
+#if (defined SLJIT_CONFIG_X86_64 && SLJIT_CONFIG_X86_64)
+	/* REX_W is not necessary. */
+	compiler->mode32 = 1;
+#endif
+	/* Load 1 into TMP_REG1. */
+	EMIT_MOV(compiler, TMP_REG1, 0, SLJIT_IMM, 1);
+
+	/* Generate "INCSSP TMP_REG1". */
+	FAIL_IF(emit_incssp(compiler, TMP_REG1));
+
+	/* Jump back to "RDSSP TMP_REG1" to check shadow stack again. */
+	inst = (sljit_u8*)ensure_buf(compiler, 1 + 2);
+	FAIL_IF(!inst);
+	INC_SIZE(2);
+	*inst++ = JMP_i8;
+	*inst = size_before_rdssp_inst - compiler->size;
+
+	*jz_after_cmp_inst = compiler->size - size_jz_after_cmp_inst;
+#else /* SLJIT_CONFIG_X86_CET */
+	SLJIT_UNUSED_ARG(compiler);
+#endif /* SLJIT_CONFIG_X86_CET */
+	return SLJIT_SUCCESS;
+}
+
 #if (defined SLJIT_CONFIG_X86_32 && SLJIT_CONFIG_X86_32)
 #include "sljitNativeX86_32.c"
 #else
@@ -905,6 +1067,10 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compile
 			EMIT_MOV(compiler, SLJIT_R1, 0, TMP_REG1, 0);
 #endif
 		break;
+	case SLJIT_ENDBR:
+		return emit_endbranch(compiler);
+	case SLJIT_SKIP_FRAMES_BEFORE_RETURN:
+		return skip_frames_before_return(compiler);
 	}
 
 	return SLJIT_SUCCESS;
@@ -1074,12 +1240,12 @@ static sljit_s32 emit_prefetch(struct sljit_compiler *compiler, sljit_s32 op,
 	*inst++ = GROUP_0F;
 	*inst++ = PREFETCH;
 
-	if (op >= SLJIT_MOV_U8 && op <= SLJIT_MOV_S8)
-		*inst |= (3 << 3);
-	else if (op >= SLJIT_MOV_U16 && op <= SLJIT_MOV_S16)
-		*inst |= (2 << 3);
-	else
+	if (op == SLJIT_PREFETCH_L1)
 		*inst |= (1 << 3);
+	else if (op == SLJIT_PREFETCH_L2)
+		*inst |= (2 << 3);
+	else if (op == SLJIT_PREFETCH_L3)
+		*inst |= (3 << 3);
 
 	return SLJIT_SUCCESS;
 }
@@ -1284,12 +1450,6 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op1(struct sljit_compiler *compile
 	compiler->mode32 = op_flags & SLJIT_I32_OP;
 #endif
 
-	if (dst == SLJIT_UNUSED && !HAS_FLAGS(op)) {
-		if (op <= SLJIT_MOV_P && (src & SLJIT_MEM))
-			return emit_prefetch(compiler, op, src, srcw);
-		return SLJIT_SUCCESS;
-	}
-
 	op = GET_OPCODE(op);
 
 	if (op >= SLJIT_MOV && op <= SLJIT_MOV_P) {
@@ -2186,6 +2346,33 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2(struct sljit_compiler *compile
 	return SLJIT_SUCCESS;
 }
 
+SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_src(struct sljit_compiler *compiler, sljit_s32 op,
+	sljit_s32 src, sljit_sw srcw)
+{
+	CHECK_ERROR();
+	CHECK(check_sljit_emit_op_src(compiler, op, src, srcw));
+	ADJUST_LOCAL_OFFSET(src, srcw);
+
+	CHECK_EXTRA_REGS(src, srcw, (void)0);
+
+	switch (op) {
+	case SLJIT_FAST_RETURN:
+		return emit_fast_return(compiler, src, srcw);
+	case SLJIT_SKIP_FRAMES_BEFORE_FAST_RETURN:
+		/* Don't adjust shadow stack if it isn't enabled.  */
+		if (!cpu_has_shadow_stack ())
+			return SLJIT_SUCCESS;
+		return adjust_shadow_stack(compiler, src, srcw, SLJIT_UNUSED, 0);
+	case SLJIT_PREFETCH_L1:
+	case SLJIT_PREFETCH_L2:
+	case SLJIT_PREFETCH_L3:
+	case SLJIT_PREFETCH_ONCE:
+		return emit_prefetch(compiler, op, src, srcw);
+	}
+
+	return SLJIT_SUCCESS;
+}
+
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_get_register_index(sljit_s32 reg)
 {
 	CHECK_REG_INDEX(check_sljit_get_register_index(reg));
diff --git a/sljit/sljitUtils.c b/sljit/sljitUtils.c
index 857492a..972cad9 100644
--- a/sljit/sljitUtils.c
+++ b/sljit/sljitUtils.c
@@ -152,15 +152,23 @@ SLJIT_API_FUNC_ATTRIBUTE void SLJIT_FUNC sljit_release_lock(void)
 
 #ifdef _WIN32
 #include "windows.h"
-#else
+#else /* !_WIN32 */
 /* Provides mmap function. */
 #include <sys/types.h>
 #include <sys/mman.h>
+
 #ifndef MAP_ANON
 #ifdef MAP_ANONYMOUS
 #define MAP_ANON MAP_ANONYMOUS
-#endif
-#endif
+#endif /* MAP_ANONYMOUS */
+#endif /* !MAP_ANON */
+
+#ifndef MADV_DONTNEED
+#ifdef POSIX_MADV_DONTNEED
+#define MADV_DONTNEED POSIX_MADV_DONTNEED
+#endif /* POSIX_MADV_DONTNEED */
+#endif /* !MADV_DONTNEED */
+
 /* For detecting the page size. */
 #include <unistd.h>
 
@@ -198,35 +206,83 @@ static SLJIT_INLINE sljit_s32 open_dev_zero(void)
 
 #endif /* SLJIT_SINGLE_THREADED */
 
-#endif
+#endif /* !MAP_ANON */
 
-#endif
+#endif /* _WIN32 */
 
 #endif /* SLJIT_UTIL_STACK || SLJIT_EXECUTABLE_ALLOCATOR */
 
 #if (defined SLJIT_UTIL_STACK && SLJIT_UTIL_STACK)
 
-/* Planning to make it even more clever in the future. */
-static sljit_sw sljit_page_align = 0;
+#if (defined SLJIT_UTIL_SIMPLE_STACK_ALLOCATION && SLJIT_UTIL_SIMPLE_STACK_ALLOCATION)
 
 SLJIT_API_FUNC_ATTRIBUTE struct sljit_stack* SLJIT_FUNC sljit_allocate_stack(sljit_uw start_size, sljit_uw max_size, void *allocator_data)
 {
 	struct sljit_stack *stack;
 	void *ptr;
-#ifdef _WIN32
-	SYSTEM_INFO si;
-#endif
 
 	SLJIT_UNUSED_ARG(allocator_data);
+
 	if (start_size > max_size || start_size < 1)
 		return NULL;
 
+	stack = (struct sljit_stack*)SLJIT_MALLOC(sizeof(struct sljit_stack), allocator_data);
+	if (stack == NULL)
+		return NULL;
+
+	ptr = SLJIT_MALLOC(max_size, allocator_data);
+	if (ptr == NULL) {
+		SLJIT_FREE(stack, allocator_data);
+		return NULL;
+	}
+
+	stack->min_start = (sljit_u8 *)ptr;
+ 	stack->end = stack->min_start + max_size;
+ 	stack->start = stack->end - start_size;
+	stack->top = stack->end;
+	return stack;
+}
+
+SLJIT_API_FUNC_ATTRIBUTE void SLJIT_FUNC sljit_free_stack(struct sljit_stack *stack, void *allocator_data)
+{
+	SLJIT_UNUSED_ARG(allocator_data);
+	SLJIT_FREE((void*)stack->min_start, allocator_data);
+	SLJIT_FREE(stack, allocator_data);
+}
+
+SLJIT_API_FUNC_ATTRIBUTE sljit_u8 *SLJIT_FUNC sljit_stack_resize(struct sljit_stack *stack, sljit_u8 *new_start)
+{
+	if ((new_start < stack->min_start) || (new_start >= stack->end))
+		return NULL;
+	stack->start = new_start;
+	return new_start;
+}
+
+#else /* !SLJIT_UTIL_SIMPLE_STACK_ALLOCATION */
+
 #ifdef _WIN32
+
+SLJIT_INLINE static sljit_sw get_page_alignment(void) {
+	SYSTEM_INFO si;
+	static sljit_sw sljit_page_align;
 	if (!sljit_page_align) {
 		GetSystemInfo(&si);
 		sljit_page_align = si.dwPageSize - 1;
 	}
-#else
+	return sljit_page_align;
+}
+
+SLJIT_API_FUNC_ATTRIBUTE void SLJIT_FUNC sljit_free_stack(struct sljit_stack *stack, void *allocator_data)
+{
+	SLJIT_UNUSED_ARG(allocator_data);
+	VirtualFree((void*)stack->min_start, 0, MEM_RELEASE);
+	SLJIT_FREE(stack, allocator_data);
+}
+
+#else /* ! defined _WIN32 */
+
+SLJIT_INLINE static sljit_sw get_page_alignment(void) {
+	static sljit_sw sljit_page_align;
 	if (!sljit_page_align) {
 		sljit_page_align = sysconf(_SC_PAGESIZE);
 		/* Should never happen. */
@@ -234,14 +290,36 @@ SLJIT_API_FUNC_ATTRIBUTE struct sljit_stack* SLJIT_FUNC sljit_allocate_stack(slj
 			sljit_page_align = 4096;
 		sljit_page_align--;
 	}
-#endif
+	return sljit_page_align;
+}
+
+SLJIT_API_FUNC_ATTRIBUTE void SLJIT_FUNC sljit_free_stack(struct sljit_stack *stack, void *allocator_data)
+{
+	SLJIT_UNUSED_ARG(allocator_data);
+	munmap((void*)stack->min_start, stack->end - stack->min_start);
+	SLJIT_FREE(stack, allocator_data);
+}
+
+#endif /* defined _WIN32 */
+
+SLJIT_API_FUNC_ATTRIBUTE struct sljit_stack* SLJIT_FUNC sljit_allocate_stack(sljit_uw start_size, sljit_uw max_size, void *allocator_data)
+{
+	struct sljit_stack *stack;
+	void *ptr;
+	sljit_sw page_align;
+
+	SLJIT_UNUSED_ARG(allocator_data);
+
+	if (start_size > max_size || start_size < 1)
+		return NULL;
 
 	stack = (struct sljit_stack*)SLJIT_MALLOC(sizeof(struct sljit_stack), allocator_data);
-	if (!stack)
+	if (stack == NULL)
 		return NULL;
 
 	/* Align max_size. */
-	max_size = (max_size + sljit_page_align) & ~sljit_page_align;
+	page_align = get_page_alignment();
+	max_size = (max_size + page_align) & ~page_align;
 
 #ifdef _WIN32
 	ptr = VirtualAlloc(NULL, max_size, MEM_RESERVE, PAGE_READWRITE);
@@ -258,18 +336,18 @@ SLJIT_API_FUNC_ATTRIBUTE struct sljit_stack* SLJIT_FUNC sljit_allocate_stack(slj
 		sljit_free_stack(stack, allocator_data);
 		return NULL;
 	}
-#else
+#else /* !_WIN32 */
 #ifdef MAP_ANON
 	ptr = mmap(NULL, max_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON, -1, 0);
-#else
+#else /* !MAP_ANON */
 	if (dev_zero < 0) {
-		if (open_dev_zero()) {
+		if (open_dev_zero() != 0) {
 			SLJIT_FREE(stack, allocator_data);
 			return NULL;
 		}
 	}
 	ptr = mmap(NULL, max_size, PROT_READ | PROT_WRITE, MAP_PRIVATE, dev_zero, 0);
-#endif
+#endif /* MAP_ANON */
 	if (ptr == MAP_FAILED) {
 		SLJIT_FREE(stack, allocator_data);
 		return NULL;
@@ -277,35 +355,28 @@ SLJIT_API_FUNC_ATTRIBUTE struct sljit_stack* SLJIT_FUNC sljit_allocate_stack(slj
 	stack->min_start = (sljit_u8 *)ptr;
 	stack->end = stack->min_start + max_size;
 	stack->start = stack->end - start_size;
-#endif
+#endif /* _WIN32 */
+
 	stack->top = stack->end;
 	return stack;
 }
 
-#undef PAGE_ALIGN
-
-SLJIT_API_FUNC_ATTRIBUTE void SLJIT_FUNC sljit_free_stack(struct sljit_stack *stack, void *allocator_data)
-{
-	SLJIT_UNUSED_ARG(allocator_data);
-#ifdef _WIN32
-	VirtualFree((void*)stack->min_start, 0, MEM_RELEASE);
-#else
-	munmap((void*)stack->min_start, stack->end - stack->min_start);
-#endif
-	SLJIT_FREE(stack, allocator_data);
-}
-
 SLJIT_API_FUNC_ATTRIBUTE sljit_u8 *SLJIT_FUNC sljit_stack_resize(struct sljit_stack *stack, sljit_u8 *new_start)
 {
+#if defined _WIN32 || defined(MADV_DONTNEED)
 	sljit_uw aligned_old_start;
 	sljit_uw aligned_new_start;
+	sljit_sw page_align;
+#endif
 
 	if ((new_start < stack->min_start) || (new_start >= stack->end))
 		return NULL;
 
 #ifdef _WIN32
-	aligned_new_start = (sljit_uw)new_start & ~sljit_page_align;
-	aligned_old_start = ((sljit_uw)stack->start) & ~sljit_page_align;
+	page_align = get_page_alignment();
+
+	aligned_new_start = (sljit_uw)new_start & ~page_align;
+	aligned_old_start = ((sljit_uw)stack->start) & ~page_align;
 	if (aligned_new_start != aligned_old_start) {
 		if (aligned_new_start < aligned_old_start) {
 			if (!VirtualAlloc((void*)aligned_new_start, aligned_old_start - aligned_new_start, MEM_COMMIT, PAGE_READWRITE))
@@ -316,24 +387,24 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_u8 *SLJIT_FUNC sljit_stack_resize(struct sljit_st
 				return NULL;
 		}
 	}
-#else
+#elif defined(MADV_DONTNEED)
 	if (stack->start < new_start) {
-		aligned_new_start = (sljit_uw)new_start & ~sljit_page_align;
-		aligned_old_start = ((sljit_uw)stack->start) & ~sljit_page_align;
+		page_align = get_page_alignment();
+
+		aligned_new_start = (sljit_uw)new_start & ~page_align;
+		aligned_old_start = ((sljit_uw)stack->start) & ~page_align;
 		/* If madvise is available, we release the unnecessary space. */
-#if defined(MADV_DONTNEED)
 		if (aligned_new_start > aligned_old_start)
 			madvise((void*)aligned_old_start, aligned_new_start - aligned_old_start, MADV_DONTNEED);
-#elif defined(POSIX_MADV_DONTNEED)
-		if (aligned_new_start > aligned_old_start)
-			posix_madvise((void*)aligned_old_start, aligned_new_start - aligned_old_start, POSIX_MADV_DONTNEED);
-#endif
 	}
-#endif
+#endif /* _WIN32 */
+
 	stack->start = new_start;
 	return new_start;
 }
 
+#endif /* SLJIT_UTIL_SIMPLE_STACK_ALLOCATION */
+
 #endif /* SLJIT_UTIL_STACK */
 
 #endif
-- 
2.34.1