void asm_x64_cmp_i32_with_r32(asm_x64_t* as, int src_i32, int src_r32) {
if (SIGNED_FIT8(src_i32)) {
asm_x64_write_byte_2(as, OPCODE_CMP_I8_WITH_RM32, MODRM_R64(7) | MODRM_RM_REG | MODRM_RM_R64(src_r32));
asm_x64_write_byte_1(as, src_i32 & 0xff);
} else {
asm_x64_write_byte_2(as, OPCODE_CMP_I32_WITH_RM32, MODRM_R64(7) | MODRM_RM_REG | MODRM_RM_R64(src_r32));
asm_x64_write_word32(as, src_i32);
}
}
void asm_x64_sub_i32_from_r32(asm_x64_t* as, int src_i32, int dest_r32) {
if (SIGNED_FIT8(src_i32)) {
// defaults to 32 bit operation
asm_x64_write_byte_2(as, OPCODE_SUB_I8_FROM_RM64, MODRM_R64(5) | MODRM_RM_REG | MODRM_RM_R64(dest_r32));
asm_x64_write_byte_1(as, src_i32 & 0xff);
} else {
// defaults to 32 bit operation
asm_x64_write_byte_2(as, OPCODE_SUB_I32_FROM_RM64, MODRM_R64(5) | MODRM_RM_REG | MODRM_RM_R64(dest_r32));
asm_x64_write_word32(as, src_i32);
}
}
void asm_x64_add_i32_to_r32(asm_x64_t* as, int src_i32, int dest_r32)
{
assert(dest_r32 != REG_RSP); // in this case i think src_i32 must be 64 bits
if (SIGNED_FIT8(src_i32))
{
asm_x64_write_byte_2(as, OPCODE_ADD_I8_TO_RM32, MODRM_R64(0) | MODRM_RM_REG | MODRM_RM_R64(dest_r32));
asm_x64_write_byte_1(as, src_i32 & 0xff);
}
else
{
asm_x64_write_byte_2(as, OPCODE_ADD_I32_TO_RM32, MODRM_R64(0) | MODRM_RM_REG | MODRM_RM_R64(dest_r32));
asm_x64_write_word32(as, src_i32);
}
}
void asm_x64_pop_r64(asm_x64_t *as, int dest_r64) {
if (dest_r64 < 8) {
asm_x64_write_byte_1(as, OPCODE_POP_R64 | dest_r64);
} else {
asm_x64_write_byte_2(as, REX_PREFIX | REX_B, OPCODE_POP_R64 | (dest_r64 & 7));
}
}
STATIC void asm_x64_lea_disp_to_r64(asm_x64_t *as, int src_r64, int src_disp, int dest_r64) {
// use REX prefix for 64 bit operation
assert(src_r64 < 8);
assert(dest_r64 < 8);
asm_x64_write_byte_2(as, REX_PREFIX | REX_W, OPCODE_LEA_MEM_TO_R64);
asm_x64_write_r64_disp(as, dest_r64, src_r64, src_disp);
}
void asm_x64_push_r64(asm_x64_t *as, int src_r64) {
if (src_r64 < 8) {
asm_x64_write_byte_1(as, OPCODE_PUSH_R64 | src_r64);
} else {
asm_x64_write_byte_2(as, REX_PREFIX | REX_B, OPCODE_PUSH_R64 | (src_r64 & 7));
}
}
void asm_x64_mov_i64_to_r64(asm_x64_t *as, int64_t src_i64, int dest_r64) {
// cpu defaults to i32 to r64
// to mov i64 to r64 need to use REX prefix
asm_x64_write_byte_2(as,
REX_PREFIX | REX_W | (dest_r64 < 8 ? 0 : REX_B),
OPCODE_MOV_I64_TO_R64 | (dest_r64 & 7));
asm_x64_write_word64(as, src_i64);
}
void asm_x64_mov_r32_to_mem32(asm_x64_t *as, int src_r64, int dest_r64, int dest_disp) {
if (src_r64 < 8 && dest_r64 < 8) {
asm_x64_write_byte_1(as, OPCODE_MOV_R64_TO_RM64);
} else {
asm_x64_write_byte_2(as, REX_PREFIX | REX_R_FROM_R64(src_r64) | REX_B_FROM_R64(dest_r64), OPCODE_MOV_R64_TO_RM64);
}
asm_x64_write_r64_disp(as, src_r64, dest_r64, dest_disp);
}
STATIC void asm_x64_mov_i32_to_r64(asm_x64_t *as, int src_i32, int dest_r64) {
// cpu defaults to i32 to r64, with zero extension
if (dest_r64 < 8) {
asm_x64_write_byte_1(as, OPCODE_MOV_I64_TO_R64 | dest_r64);
} else {
asm_x64_write_byte_2(as, REX_PREFIX | REX_B, OPCODE_MOV_I64_TO_R64 | (dest_r64 & 7));
}
asm_x64_write_word32(as, src_i32);
}
void asm_x64_mov_mem32_to_r64zx(asm_x64_t *as, int src_r64, int src_disp, int dest_r64) {
assert(src_r64 < 8);
if (dest_r64 < 8) {
asm_x64_write_byte_1(as, OPCODE_MOV_RM64_TO_R64);
} else {
asm_x64_write_byte_2(as, REX_PREFIX | REX_R, OPCODE_MOV_RM64_TO_R64);
}
asm_x64_write_r64_disp(as, dest_r64, src_r64, src_disp);
}
void asm_x64_jcc_label(asm_x64_t *as, int jcc_type, int label) {
int dest = get_label_dest(as, label);
int rel = dest - as->code_offset;
if (dest >= 0 && rel < 0) {
// is a backwards jump, so we know the size of the jump on the first pass
// calculate rel assuming 8 bit relative jump
rel -= 2;
if (SIGNED_FIT8(rel)) {
asm_x64_write_byte_2(as, OPCODE_JCC_REL8 | jcc_type, rel & 0xff);
} else {
rel += 2;
goto large_jump;
}
} else {
// is a forwards jump, so need to assume it's large
large_jump:
rel -= 6;
asm_x64_write_byte_2(as, OPCODE_JCC_REL32_A, OPCODE_JCC_REL32_B | jcc_type);
asm_x64_write_word32(as, rel);
}
}
static void asm_x64_write_r64_disp(asm_x64_t* as, int r64, int disp_r64, int disp_offset) {
assert(disp_r64 != REG_RSP);
if (disp_offset == 0 && disp_r64 != REG_RBP) {
asm_x64_write_byte_1(as, MODRM_R64(r64) | MODRM_RM_DISP0 | MODRM_RM_R64(disp_r64));
} else if (SIGNED_FIT8(disp_offset)) {
asm_x64_write_byte_2(as, MODRM_R64(r64) | MODRM_RM_DISP8 | MODRM_RM_R64(disp_r64), IMM32_L0(disp_offset));
} else {
asm_x64_write_byte_1(as, MODRM_R64(r64) | MODRM_RM_DISP32 | MODRM_RM_R64(disp_r64));
asm_x64_write_word32(as, disp_offset);
}
}
void asm_x64_call_ind(asm_x64_t* as, void *ptr, int temp_r64) {
#ifdef __LP64__
asm_x64_mov_i64_to_r64_optimised(as, (int64_t)ptr, temp_r64);
#else
// If we get here, sizeof(int) == sizeof(void*).
asm_x64_mov_i64_to_r64_optimised(as, (int64_t)(unsigned int)ptr, temp_r64);
#endif
asm_x64_write_byte_2(as, OPCODE_CALL_RM32, MODRM_R64(2) | MODRM_RM_REG | MODRM_RM_R64(temp_r64));
// this reduces code size by 2 bytes per call, but doesn't seem to speed it up at all
// doesn't work anymore because calls are 64 bits away
/*
asm_x64_write_byte_1(as, OPCODE_CALL_REL32);
asm_x64_write_word32(as, ptr - (void*)(as->code_base + as->code_offset + 4));
*/
}
STATIC void asm_x64_write_r64_disp(asm_x64_t *as, int r64, int disp_r64, int disp_offset) {
assert(disp_r64 != ASM_X64_REG_RSP);
if (disp_r64 == ASM_X64_REG_R12) {
// special case for r12; not fully implemented
assert(SIGNED_FIT8(disp_offset));
asm_x64_write_byte_3(as, MODRM_R64(r64) | MODRM_RM_DISP8 | MODRM_RM_R64(disp_r64), 0x24, IMM32_L0(disp_offset));
return;
}
if (disp_offset == 0 && disp_r64 != ASM_X64_REG_RBP && disp_r64 != ASM_X64_REG_R13) {
asm_x64_write_byte_1(as, MODRM_R64(r64) | MODRM_RM_DISP0 | MODRM_RM_R64(disp_r64));
} else if (SIGNED_FIT8(disp_offset)) {
asm_x64_write_byte_2(as, MODRM_R64(r64) | MODRM_RM_DISP8 | MODRM_RM_R64(disp_r64), IMM32_L0(disp_offset));
} else {
asm_x64_write_byte_1(as, MODRM_R64(r64) | MODRM_RM_DISP32 | MODRM_RM_R64(disp_r64));
asm_x64_write_word32(as, disp_offset);
}
}
void asm_x64_jmp_label(asm_x64_t *as, mp_uint_t label) {
mp_uint_t dest = get_label_dest(as, label);
mp_int_t rel = dest - as->base.code_offset;
if (dest != (mp_uint_t)-1 && rel < 0) {
// is a backwards jump, so we know the size of the jump on the first pass
// calculate rel assuming 8 bit relative jump
rel -= 2;
if (SIGNED_FIT8(rel)) {
asm_x64_write_byte_2(as, OPCODE_JMP_REL8, rel & 0xff);
} else {
rel += 2;
goto large_jump;
}
} else {
// is a forwards jump, so need to assume it's large
large_jump:
rel -= 5;
asm_x64_write_byte_1(as, OPCODE_JMP_REL32);
asm_x64_write_word32(as, rel);
}
}
void asm_x64_mov_r64_to_mem64(asm_x64_t *as, int src_r64, int dest_r64, int dest_disp) {
// use REX prefix for 64 bit operation
asm_x64_write_byte_2(as, REX_PREFIX | REX_W | REX_R_FROM_R64(src_r64) | REX_B_FROM_R64(dest_r64), OPCODE_MOV_R64_TO_RM64);
asm_x64_write_r64_disp(as, src_r64, dest_r64, dest_disp);
}
void asm_x64_test_r8_with_r8(asm_x64_t* as, int src_r64_a, int src_r64_b) {
// TODO implement for other registers
assert(src_r64_a == REG_RAX);
assert(src_r64_b == REG_RAX);
asm_x64_write_byte_2(as, OPCODE_TEST_R8_WITH_RM8, MODRM_R64(src_r64_a) | MODRM_RM_REG | MODRM_RM_R64(src_r64_b));
}
void asm_x64_test_r8_with_r8(asm_x64_t *as, int src_r64_a, int src_r64_b) {
assert(src_r64_a < 8);
assert(src_r64_b < 8);
asm_x64_write_byte_2(as, OPCODE_TEST_R8_WITH_RM8, MODRM_R64(src_r64_a) | MODRM_RM_REG | MODRM_RM_R64(src_r64_b));
}
void asm_x64_sar_r32_by_imm(asm_x64_t* as, int r32, int imm) {
asm_x64_write_byte_2(as, OPCODE_SAR_RM32_BY_I8, MODRM_R64(7) | MODRM_RM_REG | MODRM_RM_R64(r32));
asm_x64_write_byte_1(as, imm);
}
void asm_x64_jmp_reg(asm_x64_t *as, int src_r64) {
assert(src_r64 < 8);
asm_x64_write_byte_2(as, OPCODE_JMP_RM64, MODRM_R64(4) | MODRM_RM_REG | MODRM_RM_R64(src_r64));
}
void asm_x64_sub_r32_from_r32(asm_x64_t* as, int src_r32, int dest_r32) {
// defaults to 32 bit operation
asm_x64_write_byte_2(as, OPCODE_SUB_R64_FROM_RM64, MODRM_R64(src_r32) | MODRM_RM_REG | MODRM_RM_R64(dest_r32));
}
void asm_x64_mov_i8_to_r8(asm_x64_t *as, int src_i8, int dest_r64) {
asm_x64_write_byte_2(as, OPCODE_MOV_I8_TO_R8 | dest_r64, src_i8);
}
void asm_x64_mov_disp_to_r64(asm_x64_t* as, int src_r64, int src_disp, int dest_r64) {
// use REX prefix for 64 bit operation
asm_x64_write_byte_2(as, REX_PREFIX | REX_W, OPCODE_MOV_RM64_TO_R64);
asm_x64_write_r64_disp(as, dest_r64, src_r64, src_disp);
}
void asm_x64_mov_r32_to_r32(asm_x64_t* as, int src_r32, int dest_r32) {
// defaults to 32 bit operation
asm_x64_write_byte_2(as, OPCODE_MOV_R64_TO_RM64, MODRM_R64(src_r32) | MODRM_RM_REG | MODRM_RM_R64(dest_r32));
}