summaryrefslogtreecommitdiffstats
path: root/src/3rdparty/pcre/sljit/sljitNativeX86_64.c
diff options
context:
space:
mode:
Diffstat (limited to 'src/3rdparty/pcre/sljit/sljitNativeX86_64.c')
-rw-r--r--src/3rdparty/pcre/sljit/sljitNativeX86_64.c602
1 files changed, 285 insertions, 317 deletions
diff --git a/src/3rdparty/pcre/sljit/sljitNativeX86_64.c b/src/3rdparty/pcre/sljit/sljitNativeX86_64.c
index 40d875b841..28f04fddd8 100644
--- a/src/3rdparty/pcre/sljit/sljitNativeX86_64.c
+++ b/src/3rdparty/pcre/sljit/sljitNativeX86_64.c
@@ -26,118 +26,116 @@
/* x86 64-bit arch dependent functions. */
-static int emit_load_imm64(struct sljit_compiler *compiler, int reg, sljit_w imm)
+static sljit_si emit_load_imm64(struct sljit_compiler *compiler, sljit_si reg, sljit_sw imm)
{
- sljit_ub *buf;
-
- buf = (sljit_ub*)ensure_buf(compiler, 1 + 2 + sizeof(sljit_w));
- FAIL_IF(!buf);
- INC_SIZE(2 + sizeof(sljit_w));
- *buf++ = REX_W | ((reg_map[reg] <= 7) ? 0 : REX_B);
- *buf++ = 0xb8 + (reg_map[reg] & 0x7);
- *(sljit_w*)buf = imm;
+ sljit_ub *inst;
+
+ inst = (sljit_ub*)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++ = MOV_r_i32 + (reg_map[reg] & 0x7);
+ *(sljit_sw*)inst = imm;
return SLJIT_SUCCESS;
}
-static sljit_ub* generate_far_jump_code(struct sljit_jump *jump, sljit_ub *code_ptr, int type)
+static sljit_ub* generate_far_jump_code(struct sljit_jump *jump, sljit_ub *code_ptr, sljit_si type)
{
if (type < SLJIT_JUMP) {
+ /* Invert type. */
*code_ptr++ = get_jump_code(type ^ 0x1) - 0x10;
*code_ptr++ = 10 + 3;
}
SLJIT_COMPILE_ASSERT(reg_map[TMP_REG3] == 9, tmp3_is_9_first);
*code_ptr++ = REX_W | REX_B;
- *code_ptr++ = 0xb8 + 1;
+ *code_ptr++ = MOV_r_i32 + 1;
jump->addr = (sljit_uw)code_ptr;
if (jump->flags & JUMP_LABEL)
jump->flags |= PATCH_MD;
else
- *(sljit_w*)code_ptr = jump->u.target;
+ *(sljit_sw*)code_ptr = jump->u.target;
- code_ptr += sizeof(sljit_w);
+ code_ptr += sizeof(sljit_sw);
*code_ptr++ = REX_B;
- *code_ptr++ = 0xff;
- *code_ptr++ = (type >= SLJIT_FAST_CALL) ? 0xd1 /* call */ : 0xe1 /* jmp */;
+ *code_ptr++ = GROUP_FF;
+ *code_ptr++ = (type >= SLJIT_FAST_CALL) ? (MOD_REG | CALL_rm | 1) : (MOD_REG | JMP_rm | 1);
return code_ptr;
}
-static sljit_ub* generate_fixed_jump(sljit_ub *code_ptr, sljit_w addr, int type)
+static sljit_ub* generate_fixed_jump(sljit_ub *code_ptr, sljit_sw addr, sljit_si type)
{
- sljit_w delta = addr - ((sljit_w)code_ptr + 1 + sizeof(sljit_hw));
+ sljit_sw delta = addr - ((sljit_sw)code_ptr + 1 + sizeof(sljit_si));
if (delta <= SLJIT_W(0x7fffffff) && delta >= SLJIT_W(-0x80000000)) {
- *code_ptr++ = (type == 2) ? 0xe8 /* call */ : 0xe9 /* jmp */;
- *(sljit_w*)code_ptr = delta;
+ *code_ptr++ = (type == 2) ? CALL_i32 : JMP_i32;
+ *(sljit_sw*)code_ptr = delta;
}
else {
SLJIT_COMPILE_ASSERT(reg_map[TMP_REG3] == 9, tmp3_is_9_second);
*code_ptr++ = REX_W | REX_B;
- *code_ptr++ = 0xb8 + 1;
- *(sljit_w*)code_ptr = addr;
- code_ptr += sizeof(sljit_w);
+ *code_ptr++ = MOV_r_i32 + 1;
+ *(sljit_sw*)code_ptr = addr;
+ code_ptr += sizeof(sljit_sw);
*code_ptr++ = REX_B;
- *code_ptr++ = 0xff;
- *code_ptr++ = (type == 2) ? 0xd1 /* call */ : 0xe1 /* jmp */;
+ *code_ptr++ = GROUP_FF;
+ *code_ptr++ = (type == 2) ? (MOD_REG | CALL_rm | 1) : (MOD_REG | JMP_rm | 1);
}
return code_ptr;
}
-SLJIT_API_FUNC_ATTRIBUTE int sljit_emit_enter(struct sljit_compiler *compiler, int args, int temporaries, int saveds, int local_size)
+SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_enter(struct sljit_compiler *compiler, sljit_si args, sljit_si scratches, sljit_si saveds, sljit_si local_size)
{
- int size, pushed_size;
- sljit_ub *buf;
+ sljit_si size, pushed_size;
+ sljit_ub *inst;
CHECK_ERROR();
- check_sljit_emit_enter(compiler, args, temporaries, saveds, local_size);
+ check_sljit_emit_enter(compiler, args, scratches, saveds, local_size);
- compiler->temporaries = temporaries;
+ compiler->scratches = scratches;
compiler->saveds = saveds;
compiler->flags_saved = 0;
+#if (defined SLJIT_DEBUG && SLJIT_DEBUG)
+ compiler->logical_local_size = local_size;
+#endif
size = saveds;
/* Including the return address saved by the call instruction. */
- pushed_size = (saveds + 1) * sizeof(sljit_w);
+ pushed_size = (saveds + 1) * sizeof(sljit_sw);
#ifndef _WIN64
if (saveds >= 2)
size += saveds - 1;
#else
- /* Saving the virtual stack pointer. */
- compiler->has_locals = local_size > 0;
- if (local_size > 0) {
- size += 2;
- pushed_size += sizeof(sljit_w);
- }
if (saveds >= 4)
size += saveds - 3;
- if (temporaries >= 5) {
+ if (scratches >= 5) {
size += (5 - 4) * 2;
- pushed_size += sizeof(sljit_w);
+ pushed_size += sizeof(sljit_sw);
}
#endif
size += args * 3;
if (size > 0) {
- buf = (sljit_ub*)ensure_buf(compiler, 1 + size);
- FAIL_IF(!buf);
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + size);
+ FAIL_IF(!inst);
INC_SIZE(size);
if (saveds >= 5) {
SLJIT_COMPILE_ASSERT(reg_map[SLJIT_SAVED_EREG2] >= 8, saved_ereg2_is_hireg);
- *buf++ = REX_B;
+ *inst++ = REX_B;
PUSH_REG(reg_lmap[SLJIT_SAVED_EREG2]);
}
if (saveds >= 4) {
SLJIT_COMPILE_ASSERT(reg_map[SLJIT_SAVED_EREG1] >= 8, saved_ereg1_is_hireg);
- *buf++ = REX_B;
+ *inst++ = REX_B;
PUSH_REG(reg_lmap[SLJIT_SAVED_EREG1]);
}
if (saveds >= 3) {
#ifndef _WIN64
SLJIT_COMPILE_ASSERT(reg_map[SLJIT_SAVED_REG3] >= 8, saved_reg3_is_hireg);
- *buf++ = REX_B;
+ *inst++ = REX_B;
#else
SLJIT_COMPILE_ASSERT(reg_map[SLJIT_SAVED_REG3] < 8, saved_reg3_is_loreg);
#endif
@@ -146,7 +144,7 @@ SLJIT_API_FUNC_ATTRIBUTE int sljit_emit_enter(struct sljit_compiler *compiler, i
if (saveds >= 2) {
#ifndef _WIN64
SLJIT_COMPILE_ASSERT(reg_map[SLJIT_SAVED_REG2] >= 8, saved_reg2_is_hireg);
- *buf++ = REX_B;
+ *inst++ = REX_B;
#else
SLJIT_COMPILE_ASSERT(reg_map[SLJIT_SAVED_REG2] < 8, saved_reg2_is_loreg);
#endif
@@ -157,149 +155,138 @@ SLJIT_API_FUNC_ATTRIBUTE int sljit_emit_enter(struct sljit_compiler *compiler, i
PUSH_REG(reg_lmap[SLJIT_SAVED_REG1]);
}
#ifdef _WIN64
- if (temporaries >= 5) {
+ if (scratches >= 5) {
SLJIT_COMPILE_ASSERT(reg_map[SLJIT_TEMPORARY_EREG2] >= 8, temporary_ereg2_is_hireg);
- *buf++ = REX_B;
+ *inst++ = REX_B;
PUSH_REG(reg_lmap[SLJIT_TEMPORARY_EREG2]);
}
- if (local_size > 0) {
- SLJIT_COMPILE_ASSERT(reg_map[SLJIT_LOCALS_REG] >= 8, locals_reg_is_hireg);
- *buf++ = REX_B;
- PUSH_REG(reg_lmap[SLJIT_LOCALS_REG]);
- }
#endif
#ifndef _WIN64
if (args > 0) {
- *buf++ = REX_W;
- *buf++ = 0x8b;
- *buf++ = 0xc0 | (reg_map[SLJIT_SAVED_REG1] << 3) | 0x7;
+ *inst++ = REX_W;
+ *inst++ = MOV_r_rm;
+ *inst++ = MOD_REG | (reg_map[SLJIT_SAVED_REG1] << 3) | 0x7 /* rdi */;
}
if (args > 1) {
- *buf++ = REX_W | REX_R;
- *buf++ = 0x8b;
- *buf++ = 0xc0 | (reg_lmap[SLJIT_SAVED_REG2] << 3) | 0x6;
+ *inst++ = REX_W | REX_R;
+ *inst++ = MOV_r_rm;
+ *inst++ = MOD_REG | (reg_lmap[SLJIT_SAVED_REG2] << 3) | 0x6 /* rsi */;
}
if (args > 2) {
- *buf++ = REX_W | REX_R;
- *buf++ = 0x8b;
- *buf++ = 0xc0 | (reg_lmap[SLJIT_SAVED_REG3] << 3) | 0x2;
+ *inst++ = REX_W | REX_R;
+ *inst++ = MOV_r_rm;
+ *inst++ = MOD_REG | (reg_lmap[SLJIT_SAVED_REG3] << 3) | 0x2 /* rdx */;
}
#else
if (args > 0) {
- *buf++ = REX_W;
- *buf++ = 0x8b;
- *buf++ = 0xc0 | (reg_map[SLJIT_SAVED_REG1] << 3) | 0x1;
+ *inst++ = REX_W;
+ *inst++ = MOV_r_rm;
+ *inst++ = MOD_REG | (reg_map[SLJIT_SAVED_REG1] << 3) | 0x1 /* rcx */;
}
if (args > 1) {
- *buf++ = REX_W;
- *buf++ = 0x8b;
- *buf++ = 0xc0 | (reg_map[SLJIT_SAVED_REG2] << 3) | 0x2;
+ *inst++ = REX_W;
+ *inst++ = MOV_r_rm;
+ *inst++ = MOD_REG | (reg_map[SLJIT_SAVED_REG2] << 3) | 0x2 /* rdx */;
}
if (args > 2) {
- *buf++ = REX_W | REX_B;
- *buf++ = 0x8b;
- *buf++ = 0xc0 | (reg_map[SLJIT_SAVED_REG3] << 3) | 0x0;
+ *inst++ = REX_W | REX_B;
+ *inst++ = MOV_r_rm;
+ *inst++ = MOD_REG | (reg_map[SLJIT_SAVED_REG3] << 3) | 0x0 /* r8 */;
}
#endif
}
- local_size = ((local_size + pushed_size + 16 - 1) & ~(16 - 1)) - pushed_size;
-#ifdef _WIN64
- local_size += 4 * sizeof(sljit_w);
+ local_size = ((local_size + FIXED_LOCALS_OFFSET + pushed_size + 16 - 1) & ~(16 - 1)) - pushed_size;
compiler->local_size = local_size;
+#ifdef _WIN64
if (local_size > 1024) {
- /* Allocate the stack for the function itself. */
- buf = (sljit_ub*)ensure_buf(compiler, 1 + 4);
- FAIL_IF(!buf);
- INC_SIZE(4);
- *buf++ = REX_W;
- *buf++ = 0x83;
- *buf++ = 0xc0 | (5 << 3) | 4;
+ /* Allocate stack for the callback, which grows the stack. */
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + 4 + (3 + sizeof(sljit_si)));
+ FAIL_IF(!inst);
+ INC_SIZE(4 + (3 + sizeof(sljit_si)));
+ *inst++ = REX_W;
+ *inst++ = GROUP_BINARY_83;
+ *inst++ = MOD_REG | SUB | 4;
/* Pushed size must be divisible by 8. */
SLJIT_ASSERT(!(pushed_size & 0x7));
if (pushed_size & 0x8) {
- *buf++ = 5 * sizeof(sljit_w);
- local_size -= 5 * sizeof(sljit_w);
+ *inst++ = 5 * sizeof(sljit_sw);
+ local_size -= 5 * sizeof(sljit_sw);
} else {
- *buf++ = 4 * sizeof(sljit_w);
- local_size -= 4 * sizeof(sljit_w);
+ *inst++ = 4 * sizeof(sljit_sw);
+ local_size -= 4 * sizeof(sljit_sw);
}
- FAIL_IF(emit_load_imm64(compiler, SLJIT_TEMPORARY_REG1, local_size));
- FAIL_IF(sljit_emit_ijump(compiler, SLJIT_CALL1, SLJIT_IMM, SLJIT_FUNC_OFFSET(sljit_touch_stack)));
- }
-#else
- compiler->local_size = local_size;
- if (local_size > 0) {
+ /* Second instruction */
+ SLJIT_COMPILE_ASSERT(reg_map[SLJIT_SCRATCH_REG1] < 8, temporary_reg1_is_loreg);
+ *inst++ = REX_W;
+ *inst++ = MOV_rm_i32;
+ *inst++ = MOD_REG | reg_lmap[SLJIT_SCRATCH_REG1];
+ *(sljit_si*)inst = local_size;
+#if (defined SLJIT_VERBOSE && SLJIT_VERBOSE) || (defined SLJIT_DEBUG && SLJIT_DEBUG)
+ compiler->skip_checks = 1;
#endif
- /* In case of Win64, local_size is always > 4 * sizeof(sljit_w) */
- if (local_size <= 127) {
- buf = (sljit_ub*)ensure_buf(compiler, 1 + 4);
- FAIL_IF(!buf);
- INC_SIZE(4);
- *buf++ = REX_W;
- *buf++ = 0x83;
- *buf++ = 0xc0 | (5 << 3) | 4;
- *buf++ = local_size;
- }
- else {
- buf = (sljit_ub*)ensure_buf(compiler, 1 + 7);
- FAIL_IF(!buf);
- INC_SIZE(7);
- *buf++ = REX_W;
- *buf++ = 0x81;
- *buf++ = 0xc0 | (5 << 3) | 4;
- *(sljit_hw*)buf = local_size;
- buf += sizeof(sljit_hw);
- }
-#ifndef _WIN64
+ FAIL_IF(sljit_emit_ijump(compiler, SLJIT_CALL1, SLJIT_IMM, SLJIT_FUNC_OFFSET(sljit_grow_stack)));
}
#endif
-
-#ifdef _WIN64
- if (compiler->has_locals) {
- buf = (sljit_ub*)ensure_buf(compiler, 1 + 5);
- FAIL_IF(!buf);
- INC_SIZE(5);
- *buf++ = REX_W | REX_R;
- *buf++ = 0x8d;
- *buf++ = 0x40 | (reg_lmap[SLJIT_LOCALS_REG] << 3) | 0x4;
- *buf++ = 0x24;
- *buf = 4 * sizeof(sljit_w);
+ SLJIT_ASSERT(local_size > 0);
+ if (local_size <= 127) {
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + 4);
+ FAIL_IF(!inst);
+ INC_SIZE(4);
+ *inst++ = REX_W;
+ *inst++ = GROUP_BINARY_83;
+ *inst++ = MOD_REG | SUB | 4;
+ *inst++ = local_size;
+ }
+ else {
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + 7);
+ FAIL_IF(!inst);
+ INC_SIZE(7);
+ *inst++ = REX_W;
+ *inst++ = GROUP_BINARY_81;
+ *inst++ = MOD_REG | SUB | 4;
+ *(sljit_si*)inst = local_size;
+ inst += sizeof(sljit_si);
}
+#ifdef _WIN64
+ /* Save xmm6 with MOVAPS instruction. */
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + 5);
+ FAIL_IF(!inst);
+ INC_SIZE(5);
+ *inst++ = GROUP_0F;
+ *(sljit_si*)inst = 0x20247429;
#endif
return SLJIT_SUCCESS;
}
-SLJIT_API_FUNC_ATTRIBUTE void sljit_set_context(struct sljit_compiler *compiler, int args, int temporaries, int saveds, int local_size)
+SLJIT_API_FUNC_ATTRIBUTE void sljit_set_context(struct sljit_compiler *compiler, sljit_si args, sljit_si scratches, sljit_si saveds, sljit_si local_size)
{
- int pushed_size;
+ sljit_si pushed_size;
CHECK_ERROR_VOID();
- check_sljit_set_context(compiler, args, temporaries, saveds, local_size);
+ check_sljit_set_context(compiler, args, scratches, saveds, local_size);
- compiler->temporaries = temporaries;
+ compiler->scratches = scratches;
compiler->saveds = saveds;
- /* Including the return address saved by the call instruction. */
- pushed_size = (saveds + 1) * sizeof(sljit_w);
-#ifdef _WIN64
- compiler->has_locals = local_size > 0;
- if (local_size > 0)
- pushed_size += sizeof(sljit_w);
- if (temporaries >= 5)
- pushed_size += sizeof(sljit_w);
+#if (defined SLJIT_DEBUG && SLJIT_DEBUG)
+ compiler->logical_local_size = local_size;
#endif
- compiler->local_size = ((local_size + pushed_size + 16 - 1) & ~(16 - 1)) - pushed_size;
+
+ /* Including the return address saved by the call instruction. */
+ pushed_size = (saveds + 1) * sizeof(sljit_sw);
#ifdef _WIN64
- compiler->local_size += 4 * sizeof(sljit_w);
+ if (scratches >= 5)
+ pushed_size += sizeof(sljit_sw);
#endif
+ compiler->local_size = ((local_size + FIXED_LOCALS_OFFSET + pushed_size + 16 - 1) & ~(16 - 1)) - pushed_size;
}
-SLJIT_API_FUNC_ATTRIBUTE int sljit_emit_return(struct sljit_compiler *compiler, int op, int src, sljit_w srcw)
+SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_return(struct sljit_compiler *compiler, sljit_si op, sljit_si src, sljit_sw srcw)
{
- int size;
- sljit_ub *buf;
+ sljit_si size;
+ sljit_ub *inst;
CHECK_ERROR();
check_sljit_emit_return(compiler, op, src, srcw);
@@ -307,25 +294,32 @@ SLJIT_API_FUNC_ATTRIBUTE int sljit_emit_return(struct sljit_compiler *compiler,
compiler->flags_saved = 0;
FAIL_IF(emit_mov_before_return(compiler, op, src, srcw));
- if (compiler->local_size > 0) {
- if (compiler->local_size <= 127) {
- buf = (sljit_ub*)ensure_buf(compiler, 1 + 4);
- FAIL_IF(!buf);
- INC_SIZE(4);
- *buf++ = REX_W;
- *buf++ = 0x83;
- *buf++ = 0xc0 | (0 << 3) | 4;
- *buf = compiler->local_size;
- }
- else {
- buf = (sljit_ub*)ensure_buf(compiler, 1 + 7);
- FAIL_IF(!buf);
- INC_SIZE(7);
- *buf++ = REX_W;
- *buf++ = 0x81;
- *buf++ = 0xc0 | (0 << 3) | 4;
- *(sljit_hw*)buf = compiler->local_size;
- }
+#ifdef _WIN64
+ /* Restore xmm6 with MOVAPS instruction. */
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + 5);
+ FAIL_IF(!inst);
+ INC_SIZE(5);
+ *inst++ = GROUP_0F;
+ *(sljit_si*)inst = 0x20247428;
+#endif
+ SLJIT_ASSERT(compiler->local_size > 0);
+ if (compiler->local_size <= 127) {
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + 4);
+ FAIL_IF(!inst);
+ INC_SIZE(4);
+ *inst++ = REX_W;
+ *inst++ = GROUP_BINARY_83;
+ *inst++ = MOD_REG | ADD | 4;
+ *inst = compiler->local_size;
+ }
+ else {
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + 7);
+ FAIL_IF(!inst);
+ INC_SIZE(7);
+ *inst++ = REX_W;
+ *inst++ = GROUP_BINARY_81;
+ *inst++ = MOD_REG | ADD | 4;
+ *(sljit_si*)inst = compiler->local_size;
}
size = 1 + compiler->saveds;
@@ -333,25 +327,19 @@ SLJIT_API_FUNC_ATTRIBUTE int sljit_emit_return(struct sljit_compiler *compiler,
if (compiler->saveds >= 2)
size += compiler->saveds - 1;
#else
- if (compiler->has_locals)
- size += 2;
if (compiler->saveds >= 4)
size += compiler->saveds - 3;
- if (compiler->temporaries >= 5)
+ if (compiler->scratches >= 5)
size += (5 - 4) * 2;
#endif
- buf = (sljit_ub*)ensure_buf(compiler, 1 + size);
- FAIL_IF(!buf);
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + size);
+ FAIL_IF(!inst);
INC_SIZE(size);
#ifdef _WIN64
- if (compiler->has_locals) {
- *buf++ = REX_B;
- POP_REG(reg_lmap[SLJIT_LOCALS_REG]);
- }
- if (compiler->temporaries >= 5) {
- *buf++ = REX_B;
+ if (compiler->scratches >= 5) {
+ *inst++ = REX_B;
POP_REG(reg_lmap[SLJIT_TEMPORARY_EREG2]);
}
#endif
@@ -359,22 +347,22 @@ SLJIT_API_FUNC_ATTRIBUTE int sljit_emit_return(struct sljit_compiler *compiler,
POP_REG(reg_map[SLJIT_SAVED_REG1]);
if (compiler->saveds >= 2) {
#ifndef _WIN64
- *buf++ = REX_B;
+ *inst++ = REX_B;
#endif
POP_REG(reg_lmap[SLJIT_SAVED_REG2]);
}
if (compiler->saveds >= 3) {
#ifndef _WIN64
- *buf++ = REX_B;
+ *inst++ = REX_B;
#endif
POP_REG(reg_lmap[SLJIT_SAVED_REG3]);
}
if (compiler->saveds >= 4) {
- *buf++ = REX_B;
+ *inst++ = REX_B;
POP_REG(reg_lmap[SLJIT_SAVED_EREG1]);
}
if (compiler->saveds >= 5) {
- *buf++ = REX_B;
+ *inst++ = REX_B;
POP_REG(reg_lmap[SLJIT_SAVED_EREG2]);
}
@@ -386,39 +374,32 @@ SLJIT_API_FUNC_ATTRIBUTE int sljit_emit_return(struct sljit_compiler *compiler,
/* Operators */
/* --------------------------------------------------------------------- */
-static int emit_do_imm32(struct sljit_compiler *compiler, sljit_ub rex, sljit_ub opcode, sljit_w imm)
+static sljit_si emit_do_imm32(struct sljit_compiler *compiler, sljit_ub rex, sljit_ub opcode, sljit_sw imm)
{
- sljit_ub *buf;
+ sljit_ub *inst;
+ sljit_si length = 1 + (rex ? 1 : 0) + sizeof(sljit_si);
- if (rex != 0) {
- buf = (sljit_ub*)ensure_buf(compiler, 1 + 2 + sizeof(sljit_hw));
- FAIL_IF(!buf);
- INC_SIZE(2 + sizeof(sljit_hw));
- *buf++ = rex;
- *buf++ = opcode;
- *(sljit_hw*)buf = imm;
- }
- else {
- buf = (sljit_ub*)ensure_buf(compiler, 1 + 1 + sizeof(sljit_hw));
- FAIL_IF(!buf);
- INC_SIZE(1 + sizeof(sljit_hw));
- *buf++ = opcode;
- *(sljit_hw*)buf = imm;
- }
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + length);
+ FAIL_IF(!inst);
+ INC_SIZE(length);
+ if (rex)
+ *inst++ = rex;
+ *inst++ = opcode;
+ *(sljit_si*)inst = imm;
return SLJIT_SUCCESS;
}
-static sljit_ub* emit_x86_instruction(struct sljit_compiler *compiler, int size,
+static sljit_ub* emit_x86_instruction(struct sljit_compiler *compiler, sljit_si size,
/* The register or immediate operand. */
- int a, sljit_w imma,
+ sljit_si a, sljit_sw imma,
/* The general operand (not immediate). */
- int b, sljit_w immb)
+ sljit_si b, sljit_sw immb)
{
- sljit_ub *buf;
+ sljit_ub *inst;
sljit_ub *buf_ptr;
sljit_ub rex = 0;
- int flags = size & ~0xf;
- int inst_size;
+ sljit_si flags = size & ~0xf;
+ sljit_si inst_size;
/* The immediate operand must be 32 bit. */
SLJIT_ASSERT(!(a & SLJIT_IMM) || compiler->mode32 || IS_HALFWORD(imma));
@@ -431,6 +412,9 @@ static sljit_ub* emit_x86_instruction(struct sljit_compiler *compiler, int size,
#if (defined SLJIT_SSE2 && SLJIT_SSE2)
/* 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));
#endif
size &= 0xf;
@@ -452,7 +436,7 @@ static sljit_ub* emit_x86_instruction(struct sljit_compiler *compiler, int size,
rex |= REX;
#if (defined SLJIT_SSE2 && SLJIT_SSE2)
- if (flags & EX86_PREF_F2)
+ if (flags & (EX86_PREF_F2 | EX86_PREF_F3))
inst_size++;
#endif
if (flags & EX86_PREF_66)
@@ -462,23 +446,21 @@ static sljit_ub* emit_x86_instruction(struct sljit_compiler *compiler, int size,
inst_size += 1; /* mod r/m byte. */
if (b & SLJIT_MEM) {
if ((b & 0x0f) == SLJIT_UNUSED)
- inst_size += 1 + sizeof(sljit_hw); /* SIB byte required to avoid RIP based addressing. */
+ inst_size += 1 + sizeof(sljit_si); /* SIB byte required to avoid RIP based addressing. */
else {
if (reg_map[b & 0x0f] >= 8)
rex |= REX_B;
if (immb != 0 && !(b & 0xf0)) {
/* Immediate operand. */
if (immb <= 127 && immb >= -128)
- inst_size += sizeof(sljit_b);
+ inst_size += sizeof(sljit_sb);
else
- inst_size += sizeof(sljit_hw);
+ inst_size += sizeof(sljit_si);
}
}
-#ifndef _WIN64
- if ((b & 0xf) == SLJIT_LOCALS_REG && (b & 0xf0) == 0)
+ if ((b & 0xf) == SLJIT_LOCALS_REG && !(b & 0xf0))
b |= SLJIT_LOCALS_REG << 4;
-#endif
if ((b & 0xf0) != SLJIT_UNUSED) {
inst_size += 1; /* SIB byte. */
@@ -513,7 +495,7 @@ static sljit_ub* emit_x86_instruction(struct sljit_compiler *compiler, int size,
else if (flags & EX86_HALF_ARG)
inst_size += sizeof(short);
else
- inst_size += sizeof(sljit_hw);
+ inst_size += sizeof(sljit_si);
}
else {
SLJIT_ASSERT(!(flags & EX86_SHIFT_INS) || a == SLJIT_PREF_SHIFT_REG);
@@ -530,25 +512,27 @@ static sljit_ub* emit_x86_instruction(struct sljit_compiler *compiler, int size,
if (rex)
inst_size++;
- buf = (sljit_ub*)ensure_buf(compiler, 1 + inst_size);
- PTR_FAIL_IF(!buf);
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + inst_size);
+ PTR_FAIL_IF(!inst);
/* Encoding the byte. */
INC_SIZE(inst_size);
#if (defined SLJIT_SSE2 && SLJIT_SSE2)
if (flags & EX86_PREF_F2)
- *buf++ = 0xf2;
+ *inst++ = 0xf2;
+ if (flags & EX86_PREF_F3)
+ *inst++ = 0xf3;
#endif
if (flags & EX86_PREF_66)
- *buf++ = 0x66;
+ *inst++ = 0x66;
if (rex)
- *buf++ = rex;
- buf_ptr = buf + size;
+ *inst++ = rex;
+ buf_ptr = inst + size;
/* Encode mod/rm byte. */
if (!(flags & EX86_SHIFT_INS)) {
if ((flags & EX86_BIN_INS) && (a & SLJIT_IMM))
- *buf = (flags & EX86_BYTE_ARG) ? 0x83 : 0x81;
+ *inst = (flags & EX86_BYTE_ARG) ? GROUP_BINARY_83 : GROUP_BINARY_81;
if ((a & SLJIT_IMM) || (a == 0))
*buf_ptr = 0;
@@ -565,24 +549,21 @@ static sljit_ub* emit_x86_instruction(struct sljit_compiler *compiler, int size,
else {
if (a & SLJIT_IMM) {
if (imma == 1)
- *buf = 0xd1;
+ *inst = GROUP_SHIFT_1;
else
- *buf = 0xc1;
+ *inst = GROUP_SHIFT_N;
} else
- *buf = 0xd3;
+ *inst = GROUP_SHIFT_CL;
*buf_ptr = 0;
}
if (!(b & SLJIT_MEM))
#if (defined SLJIT_SSE2 && SLJIT_SSE2)
- *buf_ptr++ |= 0xc0 + ((!(flags & EX86_SSE2)) ? reg_lmap[b] : b);
+ *buf_ptr++ |= MOD_REG + ((!(flags & EX86_SSE2)) ? reg_lmap[b] : b);
#else
- *buf_ptr++ |= 0xc0 + reg_lmap[b];
+ *buf_ptr++ |= MOD_REG + reg_lmap[b];
#endif
else if ((b & 0x0f) != SLJIT_UNUSED) {
-#ifdef _WIN64
- SLJIT_ASSERT((b & 0xf0) != (SLJIT_LOCALS_REG << 4));
-#endif
if ((b & 0xf0) == SLJIT_UNUSED || (b & 0xf0) == (SLJIT_LOCALS_REG << 4)) {
if (immb != 0) {
if (immb <= 127 && immb >= -128)
@@ -602,8 +583,8 @@ static sljit_ub* emit_x86_instruction(struct sljit_compiler *compiler, int size,
if (immb <= 127 && immb >= -128)
*buf_ptr++ = immb; /* 8 bit displacement. */
else {
- *(sljit_hw*)buf_ptr = immb; /* 32 bit displacement. */
- buf_ptr += sizeof(sljit_hw);
+ *(sljit_si*)buf_ptr = immb; /* 32 bit displacement. */
+ buf_ptr += sizeof(sljit_si);
}
}
}
@@ -615,8 +596,8 @@ static sljit_ub* emit_x86_instruction(struct sljit_compiler *compiler, int size,
else {
*buf_ptr++ |= 0x04;
*buf_ptr++ = 0x25;
- *(sljit_hw*)buf_ptr = immb; /* 32 bit displacement. */
- buf_ptr += sizeof(sljit_hw);
+ *(sljit_si*)buf_ptr = immb; /* 32 bit displacement. */
+ buf_ptr += sizeof(sljit_si);
}
if (a & SLJIT_IMM) {
@@ -625,154 +606,141 @@ static sljit_ub* emit_x86_instruction(struct sljit_compiler *compiler, int size,
else if (flags & EX86_HALF_ARG)
*(short*)buf_ptr = imma;
else if (!(flags & EX86_SHIFT_INS))
- *(sljit_hw*)buf_ptr = imma;
+ *(sljit_si*)buf_ptr = imma;
}
- return !(flags & EX86_SHIFT_INS) ? buf : (buf + 1);
+ return !(flags & EX86_SHIFT_INS) ? inst : (inst + 1);
}
/* --------------------------------------------------------------------- */
/* Call / return instructions */
/* --------------------------------------------------------------------- */
-static SLJIT_INLINE int call_with_args(struct sljit_compiler *compiler, int type)
+static SLJIT_INLINE sljit_si call_with_args(struct sljit_compiler *compiler, sljit_si type)
{
- sljit_ub *buf;
+ sljit_ub *inst;
#ifndef _WIN64
- SLJIT_COMPILE_ASSERT(reg_map[SLJIT_TEMPORARY_REG2] == 6 && reg_map[SLJIT_TEMPORARY_REG1] < 8 && reg_map[SLJIT_TEMPORARY_REG3] < 8, args_registers);
+ SLJIT_COMPILE_ASSERT(reg_map[SLJIT_SCRATCH_REG2] == 6 && reg_map[SLJIT_SCRATCH_REG1] < 8 && reg_map[SLJIT_SCRATCH_REG3] < 8, args_registers);
- buf = (sljit_ub*)ensure_buf(compiler, 1 + ((type < SLJIT_CALL3) ? 3 : 6));
- FAIL_IF(!buf);
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + ((type < SLJIT_CALL3) ? 3 : 6));
+ FAIL_IF(!inst);
INC_SIZE((type < SLJIT_CALL3) ? 3 : 6);
if (type >= SLJIT_CALL3) {
- *buf++ = REX_W;
- *buf++ = 0x8b;
- *buf++ = 0xc0 | (0x2 << 3) | reg_lmap[SLJIT_TEMPORARY_REG3];
+ *inst++ = REX_W;
+ *inst++ = MOV_r_rm;
+ *inst++ = MOD_REG | (0x2 /* rdx */ << 3) | reg_lmap[SLJIT_SCRATCH_REG3];
}
- *buf++ = REX_W;
- *buf++ = 0x8b;
- *buf++ = 0xc0 | (0x7 << 3) | reg_lmap[SLJIT_TEMPORARY_REG1];
+ *inst++ = REX_W;
+ *inst++ = MOV_r_rm;
+ *inst++ = MOD_REG | (0x7 /* rdi */ << 3) | reg_lmap[SLJIT_SCRATCH_REG1];
#else
- SLJIT_COMPILE_ASSERT(reg_map[SLJIT_TEMPORARY_REG2] == 2 && reg_map[SLJIT_TEMPORARY_REG1] < 8 && reg_map[SLJIT_TEMPORARY_REG3] < 8, args_registers);
+ SLJIT_COMPILE_ASSERT(reg_map[SLJIT_SCRATCH_REG2] == 2 && reg_map[SLJIT_SCRATCH_REG1] < 8 && reg_map[SLJIT_SCRATCH_REG3] < 8, args_registers);
- buf = (sljit_ub*)ensure_buf(compiler, 1 + ((type < SLJIT_CALL3) ? 3 : 6));
- FAIL_IF(!buf);
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + ((type < SLJIT_CALL3) ? 3 : 6));
+ FAIL_IF(!inst);
INC_SIZE((type < SLJIT_CALL3) ? 3 : 6);
if (type >= SLJIT_CALL3) {
- *buf++ = REX_W | REX_R;
- *buf++ = 0x8b;
- *buf++ = 0xc0 | (0x0 << 3) | reg_lmap[SLJIT_TEMPORARY_REG3];
+ *inst++ = REX_W | REX_R;
+ *inst++ = MOV_r_rm;
+ *inst++ = MOD_REG | (0x0 /* r8 */ << 3) | reg_lmap[SLJIT_SCRATCH_REG3];
}
- *buf++ = REX_W;
- *buf++ = 0x8b;
- *buf++ = 0xc0 | (0x1 << 3) | reg_lmap[SLJIT_TEMPORARY_REG1];
+ *inst++ = REX_W;
+ *inst++ = MOV_r_rm;
+ *inst++ = MOD_REG | (0x1 /* rcx */ << 3) | reg_lmap[SLJIT_SCRATCH_REG1];
#endif
return SLJIT_SUCCESS;
}
-SLJIT_API_FUNC_ATTRIBUTE int sljit_emit_fast_enter(struct sljit_compiler *compiler, int dst, sljit_w dstw, int args, int temporaries, int saveds, int local_size)
+SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_fast_enter(struct sljit_compiler *compiler, sljit_si dst, sljit_sw dstw)
{
- sljit_ub *buf;
+ sljit_ub *inst;
CHECK_ERROR();
- check_sljit_emit_fast_enter(compiler, dst, dstw, args, temporaries, saveds, local_size);
-
- compiler->temporaries = temporaries;
- compiler->saveds = saveds;
- compiler->local_size = (local_size + sizeof(sljit_uw) - 1) & ~(sizeof(sljit_uw) - 1);
-#ifdef _WIN64
- compiler->local_size += 4 * sizeof(sljit_w);
-#endif
+ check_sljit_emit_fast_enter(compiler, dst, dstw);
+ ADJUST_LOCAL_OFFSET(dst, dstw);
/* For UNUSED dst. Uncommon, but possible. */
if (dst == SLJIT_UNUSED)
dst = TMP_REGISTER;
- if (dst >= SLJIT_TEMPORARY_REG1 && dst <= TMP_REGISTER) {
+ if (dst <= TMP_REGISTER) {
if (reg_map[dst] < 8) {
- buf = (sljit_ub*)ensure_buf(compiler, 1 + 1);
- FAIL_IF(!buf);
-
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + 1);
+ FAIL_IF(!inst);
INC_SIZE(1);
POP_REG(reg_lmap[dst]);
+ return SLJIT_SUCCESS;
}
- else {
- buf = (sljit_ub*)ensure_buf(compiler, 1 + 2);
- FAIL_IF(!buf);
- INC_SIZE(2);
- *buf++ = REX_B;
- POP_REG(reg_lmap[dst]);
- }
- }
- else if (dst & SLJIT_MEM) {
-#if (defined SLJIT_CONFIG_X86_64 && SLJIT_CONFIG_X86_64)
- /* REX_W is not necessary (src is not immediate). */
- compiler->mode32 = 1;
-#endif
- buf = emit_x86_instruction(compiler, 1, 0, 0, dst, dstw);
- FAIL_IF(!buf);
- *buf++ = 0x8f;
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + 2);
+ FAIL_IF(!inst);
+ INC_SIZE(2);
+ *inst++ = REX_B;
+ POP_REG(reg_lmap[dst]);
+ return SLJIT_SUCCESS;
}
+
+ /* REX_W is not necessary (src is not immediate). */
+ compiler->mode32 = 1;
+ inst = emit_x86_instruction(compiler, 1, 0, 0, dst, dstw);
+ FAIL_IF(!inst);
+ *inst++ = POP_rm;
return SLJIT_SUCCESS;
}
-SLJIT_API_FUNC_ATTRIBUTE int sljit_emit_fast_return(struct sljit_compiler *compiler, int src, sljit_w srcw)
+SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_fast_return(struct sljit_compiler *compiler, sljit_si src, sljit_sw srcw)
{
- sljit_ub *buf;
+ sljit_ub *inst;
CHECK_ERROR();
check_sljit_emit_fast_return(compiler, src, srcw);
-
- CHECK_EXTRA_REGS(src, srcw, (void)0);
+ ADJUST_LOCAL_OFFSET(src, srcw);
if ((src & SLJIT_IMM) && NOT_HALFWORD(srcw)) {
FAIL_IF(emit_load_imm64(compiler, TMP_REGISTER, srcw));
src = TMP_REGISTER;
}
- if (src >= SLJIT_TEMPORARY_REG1 && src <= TMP_REGISTER) {
+ if (src <= TMP_REGISTER) {
if (reg_map[src] < 8) {
- buf = (sljit_ub*)ensure_buf(compiler, 1 + 1 + 1);
- FAIL_IF(!buf);
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + 1 + 1);
+ FAIL_IF(!inst);
INC_SIZE(1 + 1);
PUSH_REG(reg_lmap[src]);
}
else {
- buf = (sljit_ub*)ensure_buf(compiler, 1 + 2 + 1);
- FAIL_IF(!buf);
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + 2 + 1);
+ FAIL_IF(!inst);
INC_SIZE(2 + 1);
- *buf++ = REX_B;
+ *inst++ = REX_B;
PUSH_REG(reg_lmap[src]);
}
}
else if (src & SLJIT_MEM) {
-#if (defined SLJIT_CONFIG_X86_64 && SLJIT_CONFIG_X86_64)
/* REX_W is not necessary (src is not immediate). */
compiler->mode32 = 1;
-#endif
- buf = emit_x86_instruction(compiler, 1, 0, 0, src, srcw);
- FAIL_IF(!buf);
- *buf++ = 0xff;
- *buf |= 6 << 3;
+ inst = emit_x86_instruction(compiler, 1, 0, 0, src, srcw);
+ FAIL_IF(!inst);
+ *inst++ = GROUP_FF;
+ *inst |= PUSH_rm;
- buf = (sljit_ub*)ensure_buf(compiler, 1 + 1);
- FAIL_IF(!buf);
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + 1);
+ FAIL_IF(!inst);
INC_SIZE(1);
}
else {
SLJIT_ASSERT(IS_HALFWORD(srcw));
/* SLJIT_IMM. */
- buf = (sljit_ub*)ensure_buf(compiler, 1 + 5 + 1);
- FAIL_IF(!buf);
+ inst = (sljit_ub*)ensure_buf(compiler, 1 + 5 + 1);
+ FAIL_IF(!inst);
INC_SIZE(5 + 1);
- *buf++ = 0x68;
- *(sljit_hw*)buf = srcw;
- buf += sizeof(sljit_hw);
+ *inst++ = PUSH_i32;
+ *(sljit_si*)inst = srcw;
+ inst += sizeof(sljit_si);
}
RET();
@@ -784,12 +752,12 @@ SLJIT_API_FUNC_ATTRIBUTE int sljit_emit_fast_return(struct sljit_compiler *compi
/* Extend input */
/* --------------------------------------------------------------------- */
-static int emit_mov_int(struct sljit_compiler *compiler, int sign,
- int dst, sljit_w dstw,
- int src, sljit_w srcw)
+static sljit_si emit_mov_int(struct sljit_compiler *compiler, sljit_si sign,
+ sljit_si dst, sljit_sw dstw,
+ sljit_si src, sljit_sw srcw)
{
- sljit_ub* code;
- int dst_r;
+ sljit_ub* inst;
+ sljit_si dst_r;
compiler->mode32 = 0;
@@ -797,32 +765,32 @@ static int emit_mov_int(struct sljit_compiler *compiler, int sign,
return SLJIT_SUCCESS; /* Empty instruction. */
if (src & SLJIT_IMM) {
- if (dst >= SLJIT_TEMPORARY_REG1 && dst <= SLJIT_NO_REGISTERS) {
+ if (dst <= TMP_REGISTER) {
if (sign || ((sljit_uw)srcw <= 0x7fffffff)) {
- code = emit_x86_instruction(compiler, 1, SLJIT_IMM, (sljit_w)(sljit_i)srcw, dst, dstw);
- FAIL_IF(!code);
- *code = 0xc7;
+ inst = emit_x86_instruction(compiler, 1, SLJIT_IMM, (sljit_sw)(sljit_si)srcw, dst, dstw);
+ FAIL_IF(!inst);
+ *inst = MOV_rm_i32;
return SLJIT_SUCCESS;
}
return emit_load_imm64(compiler, dst, srcw);
}
compiler->mode32 = 1;
- code = emit_x86_instruction(compiler, 1, SLJIT_IMM, (sljit_w)(sljit_i)srcw, dst, dstw);
- FAIL_IF(!code);
- *code = 0xc7;
+ inst = emit_x86_instruction(compiler, 1, SLJIT_IMM, (sljit_sw)(sljit_si)srcw, dst, dstw);
+ FAIL_IF(!inst);
+ *inst = MOV_rm_i32;
compiler->mode32 = 0;
return SLJIT_SUCCESS;
}
- dst_r = (dst >= SLJIT_TEMPORARY_REG1 && dst <= SLJIT_SAVED_REG3) ? dst : TMP_REGISTER;
+ dst_r = (dst <= TMP_REGISTER) ? dst : TMP_REGISTER;
- if ((dst & SLJIT_MEM) && (src >= SLJIT_TEMPORARY_REG1 && src <= SLJIT_SAVED_REG3))
+ if ((dst & SLJIT_MEM) && (src <= TMP_REGISTER))
dst_r = src;
else {
if (sign) {
- code = emit_x86_instruction(compiler, 1, dst_r, 0, src, srcw);
- FAIL_IF(!code);
- *code++ = 0x63;
+ inst = emit_x86_instruction(compiler, 1, dst_r, 0, src, srcw);
+ FAIL_IF(!inst);
+ *inst++ = MOVSXD_r_rm;
} else {
compiler->mode32 = 1;
FAIL_IF(emit_mov(compiler, dst_r, 0, src, srcw));
@@ -832,9 +800,9 @@ static int emit_mov_int(struct sljit_compiler *compiler, int sign,
if (dst & SLJIT_MEM) {
compiler->mode32 = 1;
- code = emit_x86_instruction(compiler, 1, dst_r, 0, dst, dstw);
- FAIL_IF(!code);
- *code = 0x89;
+ inst = emit_x86_instruction(compiler, 1, dst_r, 0, dst, dstw);
+ FAIL_IF(!inst);
+ *inst = MOV_rm_r;
compiler->mode32 = 0;
}
generated by cgit v1.2.3 (git 2.25.1) at 2024-05-02 22:20:36 +0000