1 files changed, 522 insertions, 152 deletions

diff --git a/src/3rdparty/pcre2/src/sljit/sljitNativeX86_64.c b/src/3rdparty/pcre2/src/sljit/sljitNativeX86_64.c
index f37df6e1bf..f313f3f038 100644
--- a/src/3rdparty/pcre2/src/sljit/sljitNativeX86_64.c
+++ b/src/3rdparty/pcre2/src/sljit/sljitNativeX86_64.c
@@ -37,9 +37,9 @@ static sljit_s32 emit_load_imm64(struct sljit_compiler *compiler, sljit_s32 reg,
 	inst = (sljit_u8*)ensure_buf(compiler, 1 + 2 + sizeof(sljit_sw));
 	FAIL_IF(!inst);
 	INC_SIZE(2 + sizeof(sljit_sw));
-	*inst++ = REX_W | ((reg_map[reg] <= 7) ? 0 : REX_B);
-	*inst++ = U8(MOV_r_i32 | (reg_map[reg] & 0x7));
-	sljit_unaligned_store_sw(inst, imm);
+	inst[0] = REX_W | ((reg_map[reg] <= 7) ? 0 : REX_B);
+	inst[1] = U8(MOV_r_i32 | reg_lmap[reg]);
+	sljit_unaligned_store_sw(inst + 2, imm);
 	return SLJIT_SUCCESS;
 }
 
@@ -72,7 +72,7 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_uw
 	sljit_uw inst_size;
 
 	/* The immediate operand must be 32 bit. */
-	SLJIT_ASSERT(!(a & SLJIT_IMM) || compiler->mode32 || IS_HALFWORD(imma));
+	SLJIT_ASSERT(a != SLJIT_IMM || compiler->mode32 || IS_HALFWORD(imma));
 	/* Both cannot be switched on. */
 	SLJIT_ASSERT((flags & (EX86_BIN_INS | EX86_SHIFT_INS)) != (EX86_BIN_INS | EX86_SHIFT_INS));
 	/* Size flags not allowed for typed instructions. */
@@ -80,55 +80,56 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_uw
 	/* Both size flags cannot be switched on. */
 	SLJIT_ASSERT((flags & (EX86_BYTE_ARG | EX86_HALF_ARG)) != (EX86_BYTE_ARG | EX86_HALF_ARG));
 	/* SSE2 and immediate is not possible. */
-	SLJIT_ASSERT(!(a & SLJIT_IMM) || !(flags & EX86_SSE2));
-	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));
+	SLJIT_ASSERT(a != SLJIT_IMM || !(flags & EX86_SSE2));
+	SLJIT_ASSERT(((flags & (EX86_PREF_F2 | EX86_PREF_F3 | EX86_PREF_66))
+			& ((flags & (EX86_PREF_F2 | EX86_PREF_F3 | EX86_PREF_66)) - 1)) == 0);
+	SLJIT_ASSERT((flags & (EX86_VEX_EXT | EX86_REX)) != EX86_VEX_EXT);
 
 	size &= 0xf;
-	inst_size = size;
+	/* The mod r/m byte is always present. */
+	inst_size = size + 1;
 
 	if (!compiler->mode32 && !(flags & EX86_NO_REXW))
 		rex |= REX_W;
 	else if (flags & EX86_REX)
 		rex |= REX;
 
-	if (flags & (EX86_PREF_F2 | EX86_PREF_F3))
-		inst_size++;
-	if (flags & EX86_PREF_66)
+	if (flags & (EX86_PREF_F2 | EX86_PREF_F3 | EX86_PREF_66))
 		inst_size++;
 
 	/* Calculate size of b. */
-	inst_size += 1; /* mod r/m byte. */
 	if (b & SLJIT_MEM) {
-		if (!(b & OFFS_REG_MASK)) {
-			if (NOT_HALFWORD(immb)) {
-				PTR_FAIL_IF(emit_load_imm64(compiler, TMP_REG2, immb));
-				immb = 0;
-				if (b & REG_MASK)
-					b |= TO_OFFS_REG(TMP_REG2);
-				else
-					b |= TMP_REG2;
-			}
-			else if (reg_lmap[b & REG_MASK] == 4)
-				b |= TO_OFFS_REG(SLJIT_SP);
+		if (!(b & OFFS_REG_MASK) && NOT_HALFWORD(immb)) {
+			PTR_FAIL_IF(emit_load_imm64(compiler, TMP_REG2, immb));
+			immb = 0;
+			if (b & REG_MASK)
+				b |= TO_OFFS_REG(TMP_REG2);
+			else
+				b |= TMP_REG2;
 		}
 
 		if (!(b & REG_MASK))
 			inst_size += 1 + sizeof(sljit_s32); /* SIB byte required to avoid RIP based addressing. */
 		else {
-			if (reg_map[b & REG_MASK] >= 8)
-				rex |= REX_B;
-
-			if (immb != 0 && (!(b & OFFS_REG_MASK) || (b & OFFS_REG_MASK) == TO_OFFS_REG(SLJIT_SP))) {
+			if (immb != 0 && !(b & OFFS_REG_MASK)) {
 				/* Immediate operand. */
 				if (immb <= 127 && immb >= -128)
 					inst_size += sizeof(sljit_s8);
 				else
 					inst_size += sizeof(sljit_s32);
+			} else if (reg_lmap[b & REG_MASK] == 5) {
+				/* Swap registers if possible. */
+				if ((b & OFFS_REG_MASK) && (immb & 0x3) == 0 && reg_lmap[OFFS_REG(b)] != 5)
+					b = SLJIT_MEM | OFFS_REG(b) | TO_OFFS_REG(b & REG_MASK);
+				else
+					inst_size += sizeof(sljit_s8);
 			}
-			else if (reg_lmap[b & REG_MASK] == 5)
-				inst_size += sizeof(sljit_s8);
+
+			if (reg_map[b & REG_MASK] >= 8)
+				rex |= REX_B;
+
+			if (reg_lmap[b & REG_MASK] == 4 && !(b & OFFS_REG_MASK))
+				b |= TO_OFFS_REG(SLJIT_SP);
 
 			if (b & OFFS_REG_MASK) {
 				inst_size += 1; /* SIB byte. */
@@ -136,26 +137,29 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_uw
 					rex |= REX_X;
 			}
 		}
-	}
-	else if (!(flags & EX86_SSE2_OP2)) {
+	} else if (!(flags & EX86_SSE2_OP2)) {
 		if (reg_map[b] >= 8)
 			rex |= REX_B;
-	}
-	else if (freg_map[b] >= 8)
+	} else if (freg_map[b] >= 8)
 		rex |= REX_B;
 
-	if (a & SLJIT_IMM) {
+	if ((flags & EX86_VEX_EXT) && (rex & 0x3)) {
+		SLJIT_ASSERT(size == 2);
+		size++;
+		inst_size++;
+	}
+
+	if (a == SLJIT_IMM) {
 		if (flags & EX86_BIN_INS) {
 			if (imma <= 127 && imma >= -128) {
 				inst_size += 1;
 				flags |= EX86_BYTE_ARG;
 			} else
 				inst_size += 4;
-		}
-		else if (flags & EX86_SHIFT_INS) {
-			imma &= compiler->mode32 ? 0x1f : 0x3f;
+		} else if (flags & EX86_SHIFT_INS) {
+			SLJIT_ASSERT(imma <= (compiler->mode32 ? 0x1f : 0x3f));
 			if (imma != 1) {
-				inst_size ++;
+				inst_size++;
 				flags |= EX86_BYTE_ARG;
 			}
 		} else if (flags & EX86_BYTE_ARG)
@@ -164,8 +168,7 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_uw
 			inst_size += sizeof(short);
 		else
 			inst_size += sizeof(sljit_s32);
-	}
-	else {
+	} else {
 		SLJIT_ASSERT(!(flags & EX86_SHIFT_INS) || a == SLJIT_PREF_SHIFT_REG);
 		/* reg_map[SLJIT_PREF_SHIFT_REG] is less than 8. */
 		if (!(flags & EX86_SSE2_OP1)) {
@@ -182,32 +185,34 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_uw
 	inst = (sljit_u8*)ensure_buf(compiler, 1 + inst_size);
 	PTR_FAIL_IF(!inst);
 
-	/* Encoding the byte. */
+	/* Encoding prefixes. */
 	INC_SIZE(inst_size);
 	if (flags & EX86_PREF_F2)
 		*inst++ = 0xf2;
-	if (flags & EX86_PREF_F3)
+	else if (flags & EX86_PREF_F3)
 		*inst++ = 0xf3;
-	if (flags & EX86_PREF_66)
+	else if (flags & EX86_PREF_66)
 		*inst++ = 0x66;
+
+	/* Rex is always the last prefix. */
 	if (rex)
 		*inst++ = rex;
+
 	buf_ptr = inst + size;
 
 	/* Encode mod/rm byte. */
 	if (!(flags & EX86_SHIFT_INS)) {
-		if ((flags & EX86_BIN_INS) && (a & SLJIT_IMM))
+		if ((flags & EX86_BIN_INS) && a == SLJIT_IMM)
 			*inst = (flags & EX86_BYTE_ARG) ? GROUP_BINARY_83 : GROUP_BINARY_81;
 
-		if (a & SLJIT_IMM)
+		if (a == SLJIT_IMM)
 			*buf_ptr = 0;
 		else if (!(flags & EX86_SSE2_OP1))
 			*buf_ptr = U8(reg_lmap[a] << 3);
 		else
 			*buf_ptr = U8(freg_lmap[a] << 3);
-	}
-	else {
-		if (a & SLJIT_IMM) {
+	} else {
+		if (a == SLJIT_IMM) {
 			if (imma == 1)
 				*inst = GROUP_SHIFT_1;
 			else
@@ -223,7 +228,7 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_uw
 	} else if (b & REG_MASK) {
 		reg_lmap_b = reg_lmap[b & REG_MASK];
 
-		if (!(b & OFFS_REG_MASK) || (b & OFFS_REG_MASK) == TO_OFFS_REG(SLJIT_SP) || reg_lmap_b == 5) {
+		if (!(b & OFFS_REG_MASK) || (b & OFFS_REG_MASK) == TO_OFFS_REG(SLJIT_SP)) {
 			if (immb != 0 || reg_lmap_b == 5) {
 				if (immb <= 127 && immb >= -128)
 					*buf_ptr |= 0x40;
@@ -234,8 +239,9 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_uw
 			if (!(b & OFFS_REG_MASK))
 				*buf_ptr++ |= reg_lmap_b;
 			else {
-				*buf_ptr++ |= 0x04;
-				*buf_ptr++ = U8(reg_lmap_b | (reg_lmap[OFFS_REG(b)] << 3));
+				buf_ptr[0] |= 0x04;
+				buf_ptr[1] = U8(reg_lmap_b | (reg_lmap[OFFS_REG(b)] << 3));
+				buf_ptr += 2;
 			}
 
 			if (immb != 0 || reg_lmap_b == 5) {
@@ -246,20 +252,26 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_uw
 					buf_ptr += sizeof(sljit_s32);
 				}
 			}
+		} else {
+			if (reg_lmap_b == 5)
+				*buf_ptr |= 0x40;
+
+			buf_ptr[0] |= 0x04;
+			buf_ptr[1] = U8(reg_lmap_b | (reg_lmap[OFFS_REG(b)] << 3) | (immb << 6));
+			buf_ptr += 2;
+
+			if (reg_lmap_b == 5)
+				*buf_ptr++ = 0;
 		}
-		else {
-			*buf_ptr++ |= 0x04;
-			*buf_ptr++ = U8(reg_lmap_b | (reg_lmap[OFFS_REG(b)] << 3) | (immb << 6));
-		}
-	}
-	else {
-		*buf_ptr++ |= 0x04;
-		*buf_ptr++ = 0x25;
+	} else {
+		buf_ptr[0] |= 0x04;
+		buf_ptr[1] = 0x25;
+		buf_ptr += 2;
 		sljit_unaligned_store_s32(buf_ptr, (sljit_s32)immb); /* 32 bit displacement. */
 		buf_ptr += sizeof(sljit_s32);
 	}
 
-	if (a & SLJIT_IMM) {
+	if (a == SLJIT_IMM) {
 		if (flags & EX86_BYTE_ARG)
 			*buf_ptr = U8(imma);
 		else if (flags & EX86_HALF_ARG)
@@ -268,7 +280,78 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_uw
 			sljit_unaligned_store_s32(buf_ptr, (sljit_s32)imma);
 	}
 
-	return !(flags & EX86_SHIFT_INS) ? inst : (inst + 1);
+	return inst;
+}
+
+static sljit_s32 emit_vex_instruction(struct sljit_compiler *compiler, sljit_uw op,
+	/* The first and second register operand. */
+	sljit_s32 a, sljit_s32 v,
+	/* The general operand (not immediate). */
+	sljit_s32 b, sljit_sw immb)
+{
+	sljit_u8 *inst;
+	sljit_u8 vex = 0;
+	sljit_u8 vex_m = 0;
+	sljit_uw size;
+
+	SLJIT_ASSERT(((op & (EX86_PREF_F2 | EX86_PREF_F3 | EX86_PREF_66))
+			& ((op & (EX86_PREF_F2 | EX86_PREF_F3 | EX86_PREF_66)) - 1)) == 0);
+
+	op |= EX86_REX;
+
+	if (op & VEX_OP_0F38)
+		vex_m = 0x2;
+	else if (op & VEX_OP_0F3A)
+		vex_m = 0x3;
+
+	if ((op & VEX_W) || ((op & VEX_AUTO_W) && !compiler->mode32)) {
+		if (vex_m == 0)
+			vex_m = 0x1;
+
+		vex |= 0x80;
+	}
+
+	if (op & EX86_PREF_66)
+		vex |= 0x1;
+	else if (op & EX86_PREF_F2)
+		vex |= 0x3;
+	else if (op & EX86_PREF_F3)
+		vex |= 0x2;
+
+	op &= ~(EX86_PREF_66 | EX86_PREF_F2 | EX86_PREF_F3);
+
+	if (op & VEX_256)
+		vex |= 0x4;
+
+	vex = U8(vex | ((((op & VEX_SSE2_OPV) ? freg_map[v] : reg_map[v]) ^ 0xf) << 3));
+
+	size = op & ~(sljit_uw)0xff;
+	size |= (vex_m == 0) ? (EX86_VEX_EXT | 2) : 3;
+
+	inst = emit_x86_instruction(compiler, size, a, 0, b, immb);
+	FAIL_IF(!inst);
+
+	SLJIT_ASSERT((inst[-1] & 0xf0) == REX);
+
+	/* If X or B is present in REX prefix. */
+	if (vex_m == 0 && inst[-1] & 0x3)
+		vex_m = 0x1;
+
+	if (vex_m == 0) {
+		vex |= U8(((inst[-1] >> 2) ^ 0x1) << 7);
+
+		inst[-1] = 0xc5;
+		inst[0] = vex;
+		inst[1] = U8(op);
+		return SLJIT_SUCCESS;
+	}
+
+	vex_m |= U8((inst[-1] ^ 0x7) << 5);
+	inst[-1] = 0xc4;
+	inst[0] = vex_m;
+	inst[1] = vex;
+	inst[2] = U8(op);
+	return SLJIT_SUCCESS;
 }
 
 /* --------------------------------------------------------------------- */
@@ -360,13 +443,19 @@ static sljit_u8* generate_put_label_code(struct sljit_put_label *put_label, slji
 	return code_ptr;
 }
 
+#ifdef _WIN64
+typedef struct {
+	sljit_sw regs[2];
+} sljit_sse2_reg;
+#endif /* _WIN64 */
+
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_enter(struct sljit_compiler *compiler,
 	sljit_s32 options, sljit_s32 arg_types, sljit_s32 scratches, sljit_s32 saveds,
 	sljit_s32 fscratches, sljit_s32 fsaveds, sljit_s32 local_size)
 {
 	sljit_uw size;
 	sljit_s32 word_arg_count = 0;
-	sljit_s32 saved_arg_count = 0;
+	sljit_s32 saved_arg_count = SLJIT_KEPT_SAVEDS_COUNT(options);
 	sljit_s32 saved_regs_size, tmp, i;
 #ifdef _WIN64
 	sljit_s32 saved_float_regs_size;
@@ -379,16 +468,19 @@ 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);
 
+	if (options & SLJIT_ENTER_REG_ARG)
+		arg_types = 0;
+
 	/* Emit ENDBR64 at function entry if needed.  */
 	FAIL_IF(emit_endbranch(compiler));
 
 	compiler->mode32 = 0;
 
 	/* Including the return address saved by the call instruction. */
-	saved_regs_size = GET_SAVED_REGISTERS_SIZE(scratches, saveds, 1);
+	saved_regs_size = GET_SAVED_REGISTERS_SIZE(scratches, saveds - saved_arg_count, 1);
 
 	tmp = SLJIT_S0 - saveds;
-	for (i = SLJIT_S0; i > tmp; i--) {
+	for (i = SLJIT_S0 - saved_arg_count; i > tmp; i--) {
 		size = reg_map[i] >= 8 ? 2 : 1;
 		inst = (sljit_u8*)ensure_buf(compiler, 1 + size);
 		FAIL_IF(!inst);
@@ -410,7 +502,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_enter(struct sljit_compiler *compi
 
 #ifdef _WIN64
 	local_size += SLJIT_LOCALS_OFFSET;
-	saved_float_regs_size = GET_SAVED_FLOAT_REGISTERS_SIZE(fscratches, fsaveds, 16);
+	saved_float_regs_size = GET_SAVED_FLOAT_REGISTERS_SIZE(fscratches, fsaveds, sse2_reg);
 
 	if (saved_float_regs_size > 0) {
 		saved_float_regs_offset = ((local_size + 0xf) & ~0xf);
@@ -519,16 +611,12 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_enter(struct sljit_compiler *compi
 
 		tmp = SLJIT_FS0 - fsaveds;
 		for (i = SLJIT_FS0; i > tmp; i--) {
-			inst = emit_x86_instruction(compiler, 2 | EX86_SSE2, i, 0, SLJIT_MEM1(SLJIT_SP), saved_float_regs_offset);
-			*inst++ = GROUP_0F;
-			*inst = MOVAPS_xm_x;
+			FAIL_IF(emit_groupf(compiler, MOVAPS_xm_x | EX86_SSE2, i, SLJIT_MEM1(SLJIT_SP), saved_float_regs_offset));
 			saved_float_regs_offset += 16;
 		}
 
 		for (i = fscratches; i >= SLJIT_FIRST_SAVED_FLOAT_REG; i--) {
-			inst = emit_x86_instruction(compiler, 2 | EX86_SSE2, i, 0, SLJIT_MEM1(SLJIT_SP), saved_float_regs_offset);
-			*inst++ = GROUP_0F;
-			*inst = MOVAPS_xm_x;
+			FAIL_IF(emit_groupf(compiler, MOVAPS_xm_x | EX86_SSE2, i, SLJIT_MEM1(SLJIT_SP), saved_float_regs_offset));
 			saved_float_regs_offset += 16;
 		}
 	}
@@ -552,7 +640,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_set_context(struct sljit_compiler *comp
 
 #ifdef _WIN64
 	local_size += SLJIT_LOCALS_OFFSET;
-	saved_float_regs_size = GET_SAVED_FLOAT_REGISTERS_SIZE(fscratches, fsaveds, 16);
+	saved_float_regs_size = GET_SAVED_FLOAT_REGISTERS_SIZE(fscratches, fsaveds, sse2_reg);
 
 	if (saved_float_regs_size > 0)
 		local_size = ((local_size + 0xf) & ~0xf) + saved_float_regs_size;
@@ -561,15 +649,15 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_set_context(struct sljit_compiler *comp
 #endif /* _WIN64 */
 
 	/* Including the return address saved by the call instruction. */
-	saved_regs_size = GET_SAVED_REGISTERS_SIZE(scratches, saveds, 1);
+	saved_regs_size = GET_SAVED_REGISTERS_SIZE(scratches, saveds - SLJIT_KEPT_SAVEDS_COUNT(options), 1);
 	compiler->local_size = ((local_size + saved_regs_size + 0xf) & ~0xf) - saved_regs_size;
 	return SLJIT_SUCCESS;
 }
 
-static sljit_s32 emit_stack_frame_release(struct sljit_compiler *compiler)
+static sljit_s32 emit_stack_frame_release(struct sljit_compiler *compiler, sljit_s32 is_return_to)
 {
 	sljit_uw size;
-	sljit_s32 i, tmp;
+	sljit_s32 local_size, i, tmp;
 	sljit_u8 *inst;
 #ifdef _WIN64
 	sljit_s32 saved_float_regs_offset;
@@ -578,7 +666,7 @@ static sljit_s32 emit_stack_frame_release(struct sljit_compiler *compiler)
 #endif /* _WIN64 */
 
 #ifdef _WIN64
-	saved_float_regs_offset = GET_SAVED_FLOAT_REGISTERS_SIZE(fscratches, fsaveds, 16);
+	saved_float_regs_offset = GET_SAVED_FLOAT_REGISTERS_SIZE(fscratches, fsaveds, sse2_reg);
 
 	if (saved_float_regs_offset > 0) {
 		compiler->mode32 = 1;
@@ -586,42 +674,29 @@ static sljit_s32 emit_stack_frame_release(struct sljit_compiler *compiler)
 
 		tmp = SLJIT_FS0 - fsaveds;
 		for (i = SLJIT_FS0; i > tmp; i--) {
-			inst = emit_x86_instruction(compiler, 2 | EX86_SSE2, i, 0, SLJIT_MEM1(SLJIT_SP), saved_float_regs_offset);
-			*inst++ = GROUP_0F;
-			*inst = MOVAPS_x_xm;
+			FAIL_IF(emit_groupf(compiler, MOVAPS_x_xm | EX86_SSE2, i, SLJIT_MEM1(SLJIT_SP), saved_float_regs_offset));
 			saved_float_regs_offset += 16;
 		}
 
 		for (i = fscratches; i >= SLJIT_FIRST_SAVED_FLOAT_REG; i--) {
-			inst = emit_x86_instruction(compiler, 2 | EX86_SSE2, i, 0, SLJIT_MEM1(SLJIT_SP), saved_float_regs_offset);
-			*inst++ = GROUP_0F;
-			*inst = MOVAPS_x_xm;
+			FAIL_IF(emit_groupf(compiler, MOVAPS_x_xm | EX86_SSE2, i, SLJIT_MEM1(SLJIT_SP), saved_float_regs_offset));
 			saved_float_regs_offset += 16;
 		}
+
+		compiler->mode32 = 0;
 	}
 #endif /* _WIN64 */
 
-	if (compiler->local_size > 0) {
-		if (compiler->local_size <= 127) {
-			inst = (sljit_u8*)ensure_buf(compiler, 1 + 4);
-			FAIL_IF(!inst);
-			INC_SIZE(4);
-			*inst++ = REX_W;
-			*inst++ = GROUP_BINARY_83;
-			*inst++ = MOD_REG | ADD | 4;
-			*inst = U8(compiler->local_size);
-		}
-		else {
-			inst = (sljit_u8*)ensure_buf(compiler, 1 + 7);
-			FAIL_IF(!inst);
-			INC_SIZE(7);
-			*inst++ = REX_W;
-			*inst++ = GROUP_BINARY_81;
-			*inst++ = MOD_REG | ADD | 4;
-			sljit_unaligned_store_s32(inst, compiler->local_size);
-		}
+	local_size = compiler->local_size;
+
+	if (is_return_to && compiler->scratches < SLJIT_FIRST_SAVED_REG && (compiler->saveds == SLJIT_KEPT_SAVEDS_COUNT(compiler->options))) {
+		local_size += SSIZE_OF(sw);
+		is_return_to = 0;
 	}
 
+	if (local_size > 0)
+		BINARY_IMM32(ADD, local_size, SLJIT_SP, 0);
+
 	tmp = compiler->scratches;
 	for (i = SLJIT_FIRST_SAVED_REG; i <= tmp; i++) {
 		size = reg_map[i] >= 8 ? 2 : 1;
@@ -633,8 +708,8 @@ static sljit_s32 emit_stack_frame_release(struct sljit_compiler *compiler)
 		POP_REG(reg_lmap[i]);
 	}
 
-	tmp = compiler->saveds < SLJIT_NUMBER_OF_SAVED_REGISTERS ? (SLJIT_S0 + 1 - compiler->saveds) : SLJIT_FIRST_SAVED_REG;
-	for (i = tmp; i <= SLJIT_S0; i++) {
+	tmp = SLJIT_S0 - SLJIT_KEPT_SAVEDS_COUNT(compiler->options);
+	for (i = SLJIT_S0 + 1 - compiler->saveds; i <= tmp; i++) {
 		size = reg_map[i] >= 8 ? 2 : 1;
 		inst = (sljit_u8*)ensure_buf(compiler, 1 + size);
 		FAIL_IF(!inst);
@@ -644,23 +719,43 @@ static sljit_s32 emit_stack_frame_release(struct sljit_compiler *compiler)
 		POP_REG(reg_lmap[i]);
 	}
 
+	if (is_return_to)
+		BINARY_IMM32(ADD, sizeof(sljit_sw), SLJIT_SP, 0);
+
 	return SLJIT_SUCCESS;
 }
 
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_return_void(struct sljit_compiler *compiler)
 {
-	sljit_u8 *inst;
-
 	CHECK_ERROR();
 	CHECK(check_sljit_emit_return_void(compiler));
 
-	FAIL_IF(emit_stack_frame_release(compiler));
+	compiler->mode32 = 0;
 
-	inst = (sljit_u8*)ensure_buf(compiler, 1 + 1);
-	FAIL_IF(!inst);
-	INC_SIZE(1);
-	RET();
-	return SLJIT_SUCCESS;
+	FAIL_IF(emit_stack_frame_release(compiler, 0));
+	return emit_byte(compiler, RET_near);
+}
+
+SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_return_to(struct sljit_compiler *compiler,
+	sljit_s32 src, sljit_sw srcw)
+{
+	CHECK_ERROR();
+	CHECK(check_sljit_emit_return_to(compiler, src, srcw));
+
+	compiler->mode32 = 0;
+
+	if ((src & SLJIT_MEM) || (src >= SLJIT_FIRST_SAVED_REG && src <= (SLJIT_S0 - SLJIT_KEPT_SAVEDS_COUNT(compiler->options)))) {
+		ADJUST_LOCAL_OFFSET(src, srcw);
+
+		EMIT_MOV(compiler, TMP_REG2, 0, src, srcw);
+		src = TMP_REG2;
+		srcw = 0;
+	}
+
+	FAIL_IF(emit_stack_frame_release(compiler, 1));
+
+	SLJIT_SKIP_CHECKS(compiler);
+	return sljit_emit_ijump(compiler, SLJIT_JUMP, src, srcw);
 }
 
 /* --------------------------------------------------------------------- */
@@ -786,17 +881,15 @@ SLJIT_API_FUNC_ATTRIBUTE struct sljit_jump* sljit_emit_call(struct sljit_compile
 
 	compiler->mode32 = 0;
 
-	PTR_FAIL_IF(call_with_args(compiler, arg_types, NULL));
+	if ((type & 0xff) != SLJIT_CALL_REG_ARG)
+		PTR_FAIL_IF(call_with_args(compiler, arg_types, NULL));
 
 	if (type & SLJIT_CALL_RETURN) {
-		PTR_FAIL_IF(emit_stack_frame_release(compiler));
+		PTR_FAIL_IF(emit_stack_frame_release(compiler, 0));
 		type = SLJIT_JUMP | (type & SLJIT_REWRITABLE_JUMP);
 	}
 
-#if (defined SLJIT_VERBOSE && SLJIT_VERBOSE) \
-		|| (defined SLJIT_ARGUMENT_CHECKS && SLJIT_ARGUMENT_CHECKS)
-	compiler->skip_checks = 1;
-#endif
+	SLJIT_SKIP_CHECKS(compiler);
 	return sljit_emit_jump(compiler, type);
 }
 
@@ -816,41 +909,31 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_icall(struct sljit_compiler *compi
 	}
 
 	if (type & SLJIT_CALL_RETURN) {
-		if (src >= SLJIT_FIRST_SAVED_REG && src <= SLJIT_S0) {
+		if (src >= SLJIT_FIRST_SAVED_REG && src <= (SLJIT_S0 - SLJIT_KEPT_SAVEDS_COUNT(compiler->options))) {
 			EMIT_MOV(compiler, TMP_REG2, 0, src, srcw);
 			src = TMP_REG2;
 		}
 
-		FAIL_IF(emit_stack_frame_release(compiler));
-		type = SLJIT_JUMP;
+		FAIL_IF(emit_stack_frame_release(compiler, 0));
 	}
 
-	FAIL_IF(call_with_args(compiler, arg_types, &src));
+	if ((type & 0xff) != SLJIT_CALL_REG_ARG)
+		FAIL_IF(call_with_args(compiler, arg_types, &src));
 
-#if (defined SLJIT_VERBOSE && SLJIT_VERBOSE) \
-		|| (defined SLJIT_ARGUMENT_CHECKS && SLJIT_ARGUMENT_CHECKS)
-	compiler->skip_checks = 1;
-#endif
+	if (type & SLJIT_CALL_RETURN)
+		type = SLJIT_JUMP;
 
+	SLJIT_SKIP_CHECKS(compiler);
 	return sljit_emit_ijump(compiler, type, src, srcw);
 }
 
-SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_enter(struct sljit_compiler *compiler, sljit_s32 dst, sljit_sw dstw)
+static sljit_s32 emit_fast_enter(struct sljit_compiler *compiler, sljit_s32 dst, sljit_sw dstw)
 {
 	sljit_u8 *inst;
 
-	CHECK_ERROR();
-	CHECK(check_sljit_emit_fast_enter(compiler, dst, dstw));
-	ADJUST_LOCAL_OFFSET(dst, dstw);
-
 	if (FAST_IS_REG(dst)) {
-		if (reg_map[dst] < 8) {
-			inst = (sljit_u8*)ensure_buf(compiler, 1 + 1);
-			FAIL_IF(!inst);
-			INC_SIZE(1);
-			POP_REG(reg_lmap[dst]);
-			return SLJIT_SUCCESS;
-		}
+		if (reg_map[dst] < 8)
+			return emit_byte(compiler, U8(POP_r + reg_lmap[dst]));
 
 		inst = (sljit_u8*)ensure_buf(compiler, 1 + 2);
 		FAIL_IF(!inst);
@@ -864,7 +947,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_enter(struct sljit_compiler *
 	compiler->mode32 = 1;
 	inst = emit_x86_instruction(compiler, 1, 0, 0, dst, dstw);
 	FAIL_IF(!inst);
-	*inst++ = POP_rm;
+	*inst = POP_rm;
 	return SLJIT_SUCCESS;
 }
 
@@ -894,8 +977,8 @@ static sljit_s32 emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src
 		compiler->mode32 = 1;
 		inst = emit_x86_instruction(compiler, 1, 0, 0, src, srcw);
 		FAIL_IF(!inst);
-		*inst++ = GROUP_FF;
-		*inst |= PUSH_rm;
+		inst[0] = GROUP_FF;
+		inst[1] |= PUSH_rm;
 
 		inst = (sljit_u8*)ensure_buf(compiler, 1 + 1);
 		FAIL_IF(!inst);
@@ -906,10 +989,100 @@ static sljit_s32 emit_fast_return(struct sljit_compiler *compiler, sljit_s32 src
 	return SLJIT_SUCCESS;
 }
 
+static sljit_s32 sljit_emit_get_return_address(struct sljit_compiler *compiler,
+	sljit_s32 dst, sljit_sw dstw)
+{
+	sljit_s32 saved_regs_size;
+
+	compiler->mode32 = 0;
+	saved_regs_size = GET_SAVED_REGISTERS_SIZE(compiler->scratches, compiler->saveds - SLJIT_KEPT_SAVEDS_COUNT(compiler->options), 0);
+	return emit_mov(compiler, dst, dstw, SLJIT_MEM1(SLJIT_SP), compiler->local_size + saved_regs_size);
+}
+
 /* --------------------------------------------------------------------- */
-/*  Extend input                                                         */
+/*  Other operations                                                     */
 /* --------------------------------------------------------------------- */
 
+SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_mem(struct sljit_compiler *compiler, sljit_s32 type,
+	sljit_s32 reg,
+	sljit_s32 mem, sljit_sw memw)
+{
+	sljit_u8* inst;
+	sljit_s32 i, next, reg_idx;
+	sljit_u8 regs[2];
+
+	CHECK_ERROR();
+	CHECK(check_sljit_emit_mem(compiler, type, reg, mem, memw));
+
+	if (!(reg & REG_PAIR_MASK))
+		return sljit_emit_mem_unaligned(compiler, type, reg, mem, memw);
+
+	ADJUST_LOCAL_OFFSET(mem, memw);
+
+	compiler->mode32 = 0;
+
+	if ((mem & REG_MASK) == 0) {
+		EMIT_MOV(compiler, TMP_REG1, 0, SLJIT_IMM, memw);
+
+		mem = SLJIT_MEM1(TMP_REG1);
+		memw = 0;
+	} else if (!(mem & OFFS_REG_MASK) && ((memw < HALFWORD_MIN) || (memw > HALFWORD_MAX - SSIZE_OF(sw)))) {
+		EMIT_MOV(compiler, TMP_REG1, 0, SLJIT_IMM, memw);
+
+		mem = SLJIT_MEM2(mem & REG_MASK, TMP_REG1);
+		memw = 0;
+	}
+
+	regs[0] = U8(REG_PAIR_FIRST(reg));
+	regs[1] = U8(REG_PAIR_SECOND(reg));
+
+	next = SSIZE_OF(sw);
+
+	if (!(type & SLJIT_MEM_STORE) && (regs[0] == (mem & REG_MASK) || regs[0] == OFFS_REG(mem))) {
+		if (regs[1] == (mem & REG_MASK) || regs[1] == OFFS_REG(mem)) {
+			/* Base and offset cannot be TMP_REG1. */
+			EMIT_MOV(compiler, TMP_REG1, 0, OFFS_REG(mem), 0);
+
+			if (regs[1] == OFFS_REG(mem))
+				next = -SSIZE_OF(sw);
+
+			mem = (mem & ~OFFS_REG_MASK) | TO_OFFS_REG(TMP_REG1);
+		} else {
+			next = -SSIZE_OF(sw);
+
+			if (!(mem & OFFS_REG_MASK))
+				memw += SSIZE_OF(sw);
+		}
+	}
+
+	for (i = 0; i < 2; i++) {
+		reg_idx = next > 0 ? i : (i ^ 0x1);
+		reg = regs[reg_idx];
+
+		if ((mem & OFFS_REG_MASK) && (reg_idx == 1)) {
+			inst = (sljit_u8*)ensure_buf(compiler, (sljit_uw)(1 + 5));
+			FAIL_IF(!inst);
+
+			INC_SIZE(5);
+
+			inst[0] = U8(REX_W | ((reg_map[reg] >= 8) ? REX_R : 0) | ((reg_map[mem & REG_MASK] >= 8) ? REX_B : 0) | ((reg_map[OFFS_REG(mem)] >= 8) ? REX_X : 0));
+			inst[1] = (type & SLJIT_MEM_STORE) ? MOV_rm_r : MOV_r_rm;
+			inst[2] = 0x44 | U8(reg_lmap[reg] << 3);
+			inst[3] = U8(memw << 6) | U8(reg_lmap[OFFS_REG(mem)] << 3) | reg_lmap[mem & REG_MASK];
+			inst[4] = sizeof(sljit_sw);
+		} else if (type & SLJIT_MEM_STORE) {
+			EMIT_MOV(compiler, mem, memw, reg, 0);
+		} else {
+			EMIT_MOV(compiler, reg, 0, mem, memw);
+		}
+
+		if (!(mem & OFFS_REG_MASK))
+			memw += next;
+	}
+
+	return SLJIT_SUCCESS;
+}
+
 static sljit_s32 emit_mov_int(struct sljit_compiler *compiler, sljit_s32 sign,
 	sljit_s32 dst, sljit_sw dstw,
 	sljit_s32 src, sljit_sw srcw)
@@ -919,15 +1092,15 @@ static sljit_s32 emit_mov_int(struct sljit_compiler *compiler, sljit_s32 sign,
 
 	compiler->mode32 = 0;
 
-	if (src & SLJIT_IMM) {
+	if (src == SLJIT_IMM) {
 		if (FAST_IS_REG(dst)) {
-			if (sign || ((sljit_uw)srcw <= 0x7fffffff)) {
-				inst = emit_x86_instruction(compiler, 1, SLJIT_IMM, (sljit_sw)(sljit_s32)srcw, dst, dstw);
-				FAIL_IF(!inst);
-				*inst = MOV_rm_i32;
-				return SLJIT_SUCCESS;
-			}
-			return emit_load_imm64(compiler, dst, srcw);
+			if (!sign || ((sljit_u32)srcw <= 0x7fffffff))
+				return emit_do_imm32(compiler, reg_map[dst] <= 7 ? 0 : REX_B, U8(MOV_r_i32 | reg_lmap[dst]), srcw);
+
+			inst = emit_x86_instruction(compiler, 1, SLJIT_IMM, (sljit_sw)(sljit_s32)srcw, dst, dstw);
+			FAIL_IF(!inst);
+			*inst = MOV_rm_i32;
+			return SLJIT_SUCCESS;
 		}
 		compiler->mode32 = 1;
 		inst = emit_x86_instruction(compiler, 1, SLJIT_IMM, (sljit_sw)(sljit_s32)srcw, dst, dstw);
@@ -945,10 +1118,10 @@ static sljit_s32 emit_mov_int(struct sljit_compiler *compiler, sljit_s32 sign,
 		if (sign) {
 			inst = emit_x86_instruction(compiler, 1, dst_r, 0, src, srcw);
 			FAIL_IF(!inst);
-			*inst++ = MOVSXD_r_rm;
+			*inst = MOVSXD_r_rm;
 		} else {
 			compiler->mode32 = 1;
-			FAIL_IF(emit_mov(compiler, dst_r, 0, src, srcw));
+			EMIT_MOV(compiler, dst_r, 0, src, srcw);
 			compiler->mode32 = 0;
 		}
 	}
@@ -964,6 +1137,203 @@ static sljit_s32 emit_mov_int(struct sljit_compiler *compiler, sljit_s32 sign,
 	return SLJIT_SUCCESS;
 }
 
+static SLJIT_INLINE sljit_s32 sljit_emit_fop1_conv_f64_from_uw(struct sljit_compiler *compiler, sljit_s32 op,
+	sljit_s32 dst, sljit_sw dstw,
+	sljit_s32 src, sljit_sw srcw)
+{
+	sljit_s32 dst_r = FAST_IS_REG(dst) ? dst : TMP_FREG;
+	sljit_u8 *inst, *jump_inst1, *jump_inst2;
+	sljit_uw size1, size2;
+
+	compiler->mode32 = 0;
+
+	if (GET_OPCODE(op) == SLJIT_CONV_F64_FROM_U32) {
+		if (src != SLJIT_IMM) {
+			compiler->mode32 = 1;
+			EMIT_MOV(compiler, TMP_REG1, 0, src, srcw);
+			compiler->mode32 = 0;
+		} else
+			FAIL_IF(emit_do_imm32(compiler, reg_map[TMP_REG1] <= 7 ? 0 : REX_B, U8(MOV_r_i32 | reg_lmap[TMP_REG1]), srcw));
+
+		FAIL_IF(emit_groupf(compiler, CVTSI2SD_x_rm | EX86_SELECT_F2_F3(op) | EX86_SSE2_OP1, dst_r, TMP_REG1, 0));
+
+		compiler->mode32 = 1;
+
+		if (dst_r == TMP_FREG)
+			return emit_sse2_store(compiler, op & SLJIT_32, dst, dstw, TMP_FREG);
+		return SLJIT_SUCCESS;
+	}
+
+	if (!FAST_IS_REG(src)) {
+		EMIT_MOV(compiler, TMP_REG1, 0, src, srcw);
+		src = TMP_REG1;
+	}
+
+	BINARY_IMM32(CMP, 0, src, 0);
+
+	inst = (sljit_u8*)ensure_buf(compiler, 1 + 2);
+	FAIL_IF(!inst);
+	INC_SIZE(2);
+	inst[0] = JL_i8;
+	jump_inst1 = inst;
+
+	size1 = compiler->size;
+
+	compiler->mode32 = 0;
+	FAIL_IF(emit_groupf(compiler, CVTSI2SD_x_rm | EX86_SELECT_F2_F3(op) | EX86_SSE2_OP1, dst_r, src, 0));
+
+	inst = (sljit_u8*)ensure_buf(compiler, 1 + 2);
+	FAIL_IF(!inst);
+	INC_SIZE(2);
+	inst[0] = JMP_i8;
+	jump_inst2 = inst;
+
+	size2 = compiler->size;
+
+	jump_inst1[1] = U8(size2 - size1);
+
+	if (src != TMP_REG1)
+		EMIT_MOV(compiler, TMP_REG1, 0, src, 0);
+
+	EMIT_MOV(compiler, TMP_REG2, 0, src, 0);
+
+	inst = emit_x86_instruction(compiler, 1 | EX86_SHIFT_INS, SLJIT_IMM, 1, TMP_REG1, 0);
+	FAIL_IF(!inst);
+	inst[1] |= SHR;
+
+	compiler->mode32 = 1;
+	BINARY_IMM32(AND, 1, TMP_REG2, 0);
+
+	compiler->mode32 = 0;
+	inst = emit_x86_instruction(compiler, 1, TMP_REG1, 0, TMP_REG2, 0);
+	FAIL_IF(!inst);
+	inst[0] = OR_r_rm;
+
+	FAIL_IF(emit_groupf(compiler, CVTSI2SD_x_rm | EX86_SELECT_F2_F3(op) | EX86_SSE2_OP1, dst_r, TMP_REG1, 0));
+	compiler->mode32 = 1;
+	FAIL_IF(emit_groupf(compiler, ADDSD_x_xm | EX86_SELECT_F2_F3(op) | EX86_SSE2, dst_r, dst_r, 0));
+
+	jump_inst2[1] = U8(compiler->size - size2);
+
+	if (dst_r == TMP_FREG)
+		return emit_sse2_store(compiler, op & SLJIT_32, dst, dstw, TMP_FREG);
+	return SLJIT_SUCCESS;
+}
+
+static sljit_s32 sljit_emit_fset(struct sljit_compiler *compiler,
+	sljit_s32 freg, sljit_u8 rex, sljit_s32 is_zero)
+{
+	sljit_u8 *inst;
+	sljit_u32 size;
+
+	if (is_zero) {
+		rex = freg_map[freg] >= 8 ? (REX_R | REX_B) : 0;
+	} else {
+		if (freg_map[freg] >= 8)
+			rex |= REX_R;
+		if (reg_map[TMP_REG1] >= 8)
+			rex |= REX_B;
+	}
+
+	size = (rex != 0) ? 5 : 4;
+
+	inst = (sljit_u8*)ensure_buf(compiler, 1 + size);
+	FAIL_IF(!inst);
+	INC_SIZE(size);
+
+	*inst++ = GROUP_66;
+	if (rex != 0)
+		*inst++ = rex;
+	inst[0] = GROUP_0F;
+
+	if (is_zero) {
+		inst[1] = PXOR_x_xm;
+		inst[2] = U8(freg_lmap[freg] | (freg_lmap[freg] << 3) | MOD_REG);
+	} else {
+		inst[1] = MOVD_x_rm;
+		inst[2] = U8(reg_lmap[TMP_REG1] | (freg_lmap[freg] << 3) | MOD_REG);
+	}
+
+	return SLJIT_SUCCESS;
+}
+
+SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fset32(struct sljit_compiler *compiler,
+	sljit_s32 freg, sljit_f32 value)
+{
+	union {
+		sljit_s32 imm;
+		sljit_f32 value;
+	} u;
+
+	CHECK_ERROR();
+	CHECK(check_sljit_emit_fset32(compiler, freg, value));
+
+	u.value = value;
+
+	if (u.imm != 0) {
+		compiler->mode32 = 1;
+		EMIT_MOV(compiler, TMP_REG1, 0, SLJIT_IMM, u.imm);
+	}
+
+	return sljit_emit_fset(compiler, freg, 0, u.imm == 0);
+}
+
+SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fset64(struct sljit_compiler *compiler,
+	sljit_s32 freg, sljit_f64 value)
+{
+	union {
+		sljit_sw imm;
+		sljit_f64 value;
+	} u;
+
+	CHECK_ERROR();
+	CHECK(check_sljit_emit_fset64(compiler, freg, value));
+
+	u.value = value;
+
+	if (u.imm != 0) {
+		compiler->mode32 = 0;
+		EMIT_MOV(compiler, TMP_REG1, 0, SLJIT_IMM, u.imm);
+	}
+
+	return sljit_emit_fset(compiler, freg, REX_W, u.imm == 0);
+}
+
+SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fcopy(struct sljit_compiler *compiler, sljit_s32 op,
+	sljit_s32 freg, sljit_s32 reg)
+{
+	sljit_u8 *inst;
+	sljit_u32 size;
+	sljit_u8 rex = 0;
+
+	CHECK_ERROR();
+	CHECK(check_sljit_emit_fcopy(compiler, op, freg, reg));
+
+	if (!(op & SLJIT_32))
+		rex = REX_W;
+
+	if (freg_map[freg] >= 8)
+		rex |= REX_R;
+
+	if (reg_map[reg] >= 8)
+		rex |= REX_B;
+
+	size = (rex != 0) ? 5 : 4;
+
+	inst = (sljit_u8*)ensure_buf(compiler, 1 + size);
+	FAIL_IF(!inst);
+	INC_SIZE(size);
+
+	*inst++ = GROUP_66;
+	if (rex != 0)
+		*inst++ = rex;
+	inst[0] = GROUP_0F;
+	inst[1] = GET_OPCODE(op) == SLJIT_COPY_TO_F64 ? MOVD_x_rm : MOVD_rm_x;
+	inst[2] = U8(reg_lmap[reg] | (freg_lmap[freg] << 3) | MOD_REG);
+
+	return SLJIT_SUCCESS;
+}
+
 static sljit_s32 skip_frames_before_return(struct sljit_compiler *compiler)
 {
 	sljit_s32 tmp, size;