/*
 * IR - Lightweight JIT Compilation Framework
 * (x86/x86_64 native code generator based on DynAsm)
 * Copyright (C) 2022 Zend by Perforce.
 * Authors: Dmitry Stogov <dmitry@php.net>
 */

|.if X64
|.arch x64
|.else
|.arch x86
|.endif

|.actionlist dasm_actions
|.globals ir_lb
|.section code, cold_code, rodata, jmp_table

#define IR_IS_SIGNED_32BIT(val)   ((((intptr_t)(val)) <= 0x7fffffff) && (((intptr_t)(val)) >= (-2147483647 - 1)))
#define IR_IS_UNSIGNED_32BIT(val) (((uintptr_t)(val)) <= 0xffffffff)
#define IR_IS_32BIT(type, val)    (IR_IS_TYPE_SIGNED(type) ? IR_IS_SIGNED_32BIT((val).i64) : IR_IS_UNSIGNED_32BIT((val).u64))
#define IR_IS_FP_ZERO(insn)       ((insn.type == IR_DOUBLE) ? (insn.val.u64 == 0) : (insn.val.u32 == 0))
#define IR_MAY_USE_32BIT_ADDR(addr) \
	(ctx->code_buffer && \
	IR_IS_SIGNED_32BIT((char*)(addr) - (char*)ctx->code_buffer) && \
	IR_IS_SIGNED_32BIT((char*)(addr) - ((char*)ctx->code_buffer + ctx->code_buffer_size)))

#define IR_SPILL_POS_TO_OFFSET(offset) \
	((ctx->flags & IR_USE_FRAME_POINTER) ? \
		((offset) - (ctx->stack_frame_size - ctx->stack_frame_alignment)) : \
		((offset) + ctx->call_stack_size))

|.macro ASM_REG_OP, op, type, reg
||	switch (ir_type_size[type]) {
|| 		default:
||			IR_ASSERT(0);
||		case 1:
|			op Rb(reg)
||			break;
||		case 2:
|			op Rw(reg)
|| 			break;
||		case 4:
|			op Rd(reg)
|| 			break;
|.if X64
||		case 8:
|			op Rq(reg)
|| 			break;
|.endif
||	}
|.endmacro

|.macro ASM_MEM_OP, op, type, mem
||	switch (ir_type_size[type]) {
|| 		default:
||			IR_ASSERT(0);
||		case 1:
|			op byte mem
||			break;
||		case 2:
|			op word mem
|| 			break;
||		case 4:
|			op dword mem
|| 			break;
|.if X64
||		case 8:
|			op qword mem
|| 			break;
|.endif
||	}
|.endmacro

|.macro ASM_REG_REG_OP, op, type, dst, src
||	switch (ir_type_size[type]) {
|| 		default:
||			IR_ASSERT(0);
||		case 1:
|			op Rb(dst), Rb(src)
||			break;
||		case 2:
|			op Rw(dst), Rw(src)
|| 			break;
||		case 4:
|			op Rd(dst), Rd(src)
|| 			break;
|.if X64
||		case 8:
|			op Rq(dst), Rq(src)
|| 			break;
|.endif
||	}
|.endmacro

|.macro ASM_REG_REG_OP2, op, type, dst, src
||	switch (ir_type_size[type]) {
|| 		default:
||			IR_ASSERT(0);
||		case 1:
||		case 2:
|			op Rw(dst), Rw(src)
|| 			break;
||		case 4:
|			op Rd(dst), Rd(src)
|| 			break;
|.if X64
||		case 8:
|			op Rq(dst), Rq(src)
|| 			break;
|.endif
||	}
|.endmacro

|.macro ASM_REG_TXT_OP, op, type, dst, src
||	switch (ir_type_size[type]) {
|| 		default:
||			IR_ASSERT(0);
||		case 1:
|			op Rb(dst), src
||			break;
||		case 2:
|			op Rw(dst), src
|| 			break;
||		case 4:
|			op Rd(dst), src
|| 			break;
|.if X64
||		case 8:
|			op Rq(dst), src
|| 			break;
|.endif
||	}
|.endmacro

|.macro ASM_REG_IMM_OP, op, type, dst, src
||	switch (ir_type_size[type]) {
|| 		default:
||			IR_ASSERT(0);
||		case 1:
|			op Rb(dst), (src & 0xff)
||			break;
||		case 2:
|			op Rw(dst), (src & 0xffff)
|| 			break;
||		case 4:
|			op Rd(dst), src
|| 			break;
|.if X64
||		case 8:
|			op Rq(dst), src
|| 			break;
|.endif
||	}
|.endmacro

|.macro ASM_MEM_REG_OP, op, type, dst, src
||	switch (ir_type_size[type]) {
|| 		default:
||			IR_ASSERT(0);
||		case 1:
|			op byte dst, Rb(src)
||			break;
||		case 2:
|			op word dst, Rw(src)
|| 			break;
||		case 4:
|			op dword dst, Rd(src)
|| 			break;
|.if X64
||		case 8:
|			op qword dst, Rq(src)
|| 			break;
|.endif
||	}
|.endmacro

|.macro ASM_MEM_TXT_OP, op, type, dst, src
||	switch (ir_type_size[type]) {
|| 		default:
||			IR_ASSERT(0);
||		case 1:
|			op byte dst, src
||			break;
||		case 2:
|			op word dst, src
|| 			break;
||		case 4:
|			op dword dst, src
|| 			break;
|.if X64
||		case 8:
|			op qword dst, src
|| 			break;
|.endif
||	}
|.endmacro

|.macro ASM_MEM_IMM_OP, op, type, dst, src
||	switch (ir_type_size[type]) {
|| 		default:
||			IR_ASSERT(0);
||		case 1:
|			op byte dst, (src & 0xff)
||			break;
||		case 2:
|			op word dst, (src & 0xffff)
|| 			break;
||		case 4:
|			op dword dst, src
|| 			break;
|.if X64
||		case 8:
|			op qword dst, src
|| 			break;
|.endif
||	}
|.endmacro

|.macro ASM_REG_MEM_OP, op, type, dst, src
||	switch (ir_type_size[type]) {
|| 		default:
||			IR_ASSERT(0);
||		case 1:
|			op Rb(dst), byte src
||			break;
||		case 2:
|			op Rw(dst), word src
|| 			break;
||		case 4:
|			op Rd(dst), dword src
|| 			break;
|.if X64
||		case 8:
|			op Rq(dst), qword src
|| 			break;
|.endif
||	}
|.endmacro

|.macro ASM_REG_REG_IMUL, type, dst, src
||	switch (ir_type_size[type]) {
|| 		default:
||			IR_ASSERT(0);
||		case 2:
|			imul Rw(dst), Rw(src)
|| 			break;
||		case 4:
|			imul Rd(dst), Rd(src)
|| 			break;
|.if X64
||		case 8:
|			imul Rq(dst), Rq(src)
|| 			break;
|.endif
||	}
|.endmacro

|.macro ASM_REG_IMM_IMUL, type, dst, src
||	switch (ir_type_size[type]) {
|| 		default:
||			IR_ASSERT(0);
||		case 2:
|			imul Rw(dst), src
|| 			break;
||		case 4:
|			imul Rd(dst), src
|| 			break;
|.if X64
||		case 8:
|			imul Rq(dst), src
|| 			break;
|.endif
||	}
|.endmacro

|.macro ASM_REG_MEM_IMUL, type, dst, src
||	switch (ir_type_size[type]) {
|| 		default:
||			IR_ASSERT(0);
||		case 2:
|			imul Rw(dst), word src
|| 			break;
||		case 4:
|			imul Rd(dst), dword src
|| 			break;
|.if X64
||		case 8:
|			imul Rq(dst), qword src
|| 			break;
|.endif
||	}
|.endmacro

|.macro ASM_SSE2_REG_REG_OP, fop, dop, type, dst, src
||	if (type == IR_DOUBLE) {
|		dop xmm(dst-IR_REG_FP_FIRST), xmm(src-IR_REG_FP_FIRST)
||	} else {
||		IR_ASSERT(type == IR_FLOAT);
|		fop xmm(dst-IR_REG_FP_FIRST), xmm(src-IR_REG_FP_FIRST)
||	}
|.endmacro

|.macro ASM_SSE2_REG_MEM_OP, fop, dop, type, dst, src
||	if (type == IR_DOUBLE) {
|		dop xmm(dst-IR_REG_FP_FIRST), qword src
||	} else {
||		IR_ASSERT(type == IR_FLOAT);
|		fop xmm(dst-IR_REG_FP_FIRST), dword src
||	}
|.endmacro

|.macro ASM_AVX_REG_REG_REG_OP, fop, dop, type, dst, op1, op2
||	if (type == IR_DOUBLE) {
|		dop xmm(dst-IR_REG_FP_FIRST), xmm(op1-IR_REG_FP_FIRST), xmm(op2-IR_REG_FP_FIRST)
||	} else {
||		IR_ASSERT(type == IR_FLOAT);
|		fop xmm(dst-IR_REG_FP_FIRST), xmm(op1-IR_REG_FP_FIRST), xmm(op2-IR_REG_FP_FIRST)
||	}
|.endmacro

|.macro ASM_AVX_REG_REG_MEM_OP, fop, dop, type, dst, op1, op2
||	if (type == IR_DOUBLE) {
|		dop xmm(dst-IR_REG_FP_FIRST), xmm(op1-IR_REG_FP_FIRST), qword op2
||	} else {
||		IR_ASSERT(type == IR_FLOAT);
|		fop xmm(dst-IR_REG_FP_FIRST), xmm(op1-IR_REG_FP_FIRST), dword op2
||	}
|.endmacro

|.macro ASM_FP_REG_REG_OP, fop, dop, avx_fop, avx_dop, type, dst, src
||	if (ctx->mflags & IR_X86_AVX) {
|		ASM_SSE2_REG_REG_OP avx_fop, avx_dop, type, dst, src
||	} else {
|		ASM_SSE2_REG_REG_OP fop, dop, type, dst, src
||	}
|.endmacro

|.macro ASM_FP_MEM_REG_OP, fop, dop, avx_fop, avx_dop, type, dst, src
||	if (type == IR_DOUBLE) {
||		if (ctx->mflags & IR_X86_AVX) {
|			avx_dop qword dst, xmm(src-IR_REG_FP_FIRST)
||		} else {
|			dop qword dst, xmm(src-IR_REG_FP_FIRST)
||		}
||	} else {
||		IR_ASSERT(type == IR_FLOAT);
||		if (ctx->mflags & IR_X86_AVX) {
|			avx_fop dword dst, xmm(src-IR_REG_FP_FIRST)
||		} else {
|			fop dword dst, xmm(src-IR_REG_FP_FIRST)
||		}
||	}
|.endmacro

|.macro ASM_FP_REG_MEM_OP, fop, dop, avx_fop, avx_dop, type, dst, src
||	if (ctx->mflags & IR_X86_AVX) {
|		ASM_SSE2_REG_MEM_OP avx_fop, avx_dop, type, dst, src
||	} else {
|		ASM_SSE2_REG_MEM_OP fop, dop, type, dst, src
||	}
|.endmacro

typedef struct _ir_backend_data {
    ir_reg_alloc_data  ra_data;
	uint32_t           dessa_from_block;
	dasm_State        *dasm_state;
	int                rodata_label, jmp_table_label;
	bool               double_neg_const;
	bool               float_neg_const;
	bool               double_abs_const;
	bool               float_abs_const;
	bool               double_zero_const;
} ir_backend_data;

#define IR_GP_REG_NAME(code, name64, name32, name16, name8, name8h) \
	#name64,
#define IR_GP_REG_NAME32(code, name64, name32, name16, name8, name8h) \
	#name32,
#define IR_GP_REG_NAME16(code, name64, name32, name16, name8, name8h) \
	#name16,
#define IR_GP_REG_NAME8(code, name64, name32, name16, name8, name8h) \
	#name8,
#define IR_FP_REG_NAME(code, name) \
	#name,

static const char *_ir_reg_name[IR_REG_NUM] = {
	IR_GP_REGS(IR_GP_REG_NAME)
	IR_FP_REGS(IR_FP_REG_NAME)
};

static const char *_ir_reg_name32[IR_REG_NUM] = {
	IR_GP_REGS(IR_GP_REG_NAME32)
};

static const char *_ir_reg_name16[IR_REG_NUM] = {
	IR_GP_REGS(IR_GP_REG_NAME16)
};

static const char *_ir_reg_name8[IR_REG_NUM] = {
	IR_GP_REGS(IR_GP_REG_NAME8)
};

/* Calling Convention */
#ifdef _WIN64

static const int8_t _ir_int_reg_params[IR_REG_INT_ARGS] = {
	IR_REG_INT_ARG1,
	IR_REG_INT_ARG2,
	IR_REG_INT_ARG3,
	IR_REG_INT_ARG4,
};

static const int8_t _ir_fp_reg_params[IR_REG_FP_ARGS] = {
	IR_REG_FP_ARG1,
	IR_REG_FP_ARG2,
	IR_REG_FP_ARG3,
	IR_REG_FP_ARG4,
};

#elif defined(IR_TARGET_X64)

static const int8_t _ir_int_reg_params[IR_REG_INT_ARGS] = {
	IR_REG_INT_ARG1,
	IR_REG_INT_ARG2,
	IR_REG_INT_ARG3,
	IR_REG_INT_ARG4,
	IR_REG_INT_ARG5,
	IR_REG_INT_ARG6,
};

static const int8_t _ir_fp_reg_params[IR_REG_FP_ARGS] = {
	IR_REG_FP_ARG1,
	IR_REG_FP_ARG2,
	IR_REG_FP_ARG3,
	IR_REG_FP_ARG4,
	IR_REG_FP_ARG5,
	IR_REG_FP_ARG6,
	IR_REG_FP_ARG7,
	IR_REG_FP_ARG8,
};

#else

static const int8_t *_ir_int_reg_params = NULL;
static const int8_t *_ir_fp_reg_params = NULL;
static const int8_t _ir_int_fc_reg_params[IR_REG_INT_FCARGS] = {
	IR_REG_INT_FCARG1,
	IR_REG_INT_FCARG2,
};
static const int8_t *_ir_fp_fc_reg_params = NULL;

#endif

const char *ir_reg_name(int8_t reg, ir_type type)
{
	if (reg >= IR_REG_NUM) {
		if (reg == IR_REG_SCRATCH) {
			return "SCRATCH";
		} else {
			IR_ASSERT(reg == IR_REG_ALL);
			return "ALL";
		}
	}
	IR_ASSERT(reg >= 0 && reg < IR_REG_NUM);
	if (type == IR_VOID) {
		type = (reg < IR_REG_FP_FIRST) ? IR_ADDR : IR_DOUBLE;
	}
	if (IR_IS_TYPE_FP(type) || ir_type_size[type] == 8) {
		return _ir_reg_name[reg];
	} else if (ir_type_size[type] == 4) {
		return _ir_reg_name32[reg];
	} else if (ir_type_size[type] == 2) {
		return _ir_reg_name16[reg];
	} else {
		IR_ASSERT(ir_type_size[type] == 1);
		return _ir_reg_name8[reg];
	}
}

#define IR_RULES(_)        \
	_(CMP_INT)             \
	_(CMP_FP)              \
	_(MUL_INT)             \
	_(DIV_INT)             \
	_(MOD_INT)             \
	_(TEST_INT)            \
	_(SETCC_INT)           \
	_(TESTCC_INT)          \
	_(LEA_OB)              \
	_(LEA_SI)              \
	_(LEA_SIB)             \
	_(LEA_IB)              \
	_(LEA_SI_O)            \
	_(LEA_SIB_O)           \
	_(LEA_IB_O)            \
	_(LEA_I_OB)            \
	_(LEA_OB_I)            \
	_(LEA_OB_SI)           \
	_(LEA_SI_OB)           \
	_(LEA_B_SI)            \
	_(LEA_SI_B)            \
	_(INC)                 \
	_(DEC)                 \
	_(MUL_PWR2)            \
	_(DIV_PWR2)            \
	_(MOD_PWR2)            \
	_(BOOL_NOT_INT)        \
	_(ABS_INT)             \
	_(OP_INT)              \
	_(OP_FP)               \
	_(IMUL3)               \
	_(BINOP_INT)           \
	_(BINOP_SSE2)          \
	_(BINOP_AVX)           \
	_(SHIFT)               \
	_(SHIFT_CONST)         \
	_(COPY_INT)            \
	_(COPY_FP)             \
	_(CMP_AND_BRANCH_INT)  \
	_(CMP_AND_BRANCH_FP)   \
	_(TEST_AND_BRANCH_INT) \
	_(JCC_INT)             \
	_(GUARD_CMP_INT)       \
	_(GUARD_CMP_FP)        \
	_(GUARD_TEST_INT)      \
	_(GUARD_JCC_INT)       \
	_(GUARD_OVERFLOW)      \
	_(OVERFLOW_AND_BRANCH) \
	_(MIN_MAX_INT)         \
	_(MEM_OP_INT)          \
	_(MEM_INC)             \
	_(MEM_DEC)             \
	_(MEM_MUL_PWR2)        \
	_(MEM_DIV_PWR2)        \
	_(MEM_MOD_PWR2)        \
	_(MEM_BINOP_INT)       \
	_(MEM_SHIFT)           \
	_(MEM_SHIFT_CONST)     \
	_(REG_BINOP_INT)       \
	_(VSTORE_INT)          \
	_(VSTORE_FP)           \
	_(LOAD_INT)            \
	_(LOAD_FP)             \
	_(STORE_INT)           \
	_(STORE_FP)            \
	_(IF_INT)              \
	_(RETURN_VOID)         \
	_(RETURN_INT)          \
	_(RETURN_FP)           \
	_(BIT_COUNT)           \

#define IR_RULE_ENUM(name) IR_ ## name,

enum _ir_rule {
	IR_FIRST_RULE = IR_LAST_OP,
	IR_RULES(IR_RULE_ENUM)
	IR_LAST_RULE
};

#define IR_RULE_NAME(name)  #name,
const char *ir_rule_name[IR_LAST_OP] = {
	NULL,
	IR_RULES(IR_RULE_NAME)
	NULL
};

/* register allocation */
int ir_get_target_constraints(const ir_ctx *ctx, ir_ref ref, ir_target_constraints *constraints)
{
	uint32_t rule = ir_rule(ctx, ref);
	const ir_insn *insn;
	int n = 0;
	int flags = IR_USE_MUST_BE_IN_REG | IR_OP1_MUST_BE_IN_REG | IR_OP2_MUST_BE_IN_REG | IR_OP3_MUST_BE_IN_REG;

	constraints->def_reg = IR_REG_NONE;
	constraints->hints_count = 0;
	switch (rule & IR_RULE_MASK) {
		case IR_BINOP_INT:
			insn = &ctx->ir_base[ref];
			if (rule & IR_FUSED) {
				if (ctx->ir_base[insn->op1].op == IR_RLOAD) {
					flags = IR_OP1_MUST_BE_IN_REG | IR_OP2_SHOULD_BE_IN_REG;
				} else {
					flags = IR_OP2_MUST_BE_IN_REG;
				}
			} else {
				flags = IR_DEF_REUSES_OP1_REG | IR_USE_MUST_BE_IN_REG | IR_OP1_SHOULD_BE_IN_REG | IR_OP2_SHOULD_BE_IN_REG;
			}
			if (IR_IS_CONST_REF(insn->op2) && insn->op1 != insn->op2) {
				insn = &ctx->ir_base[insn->op2];
				if (ir_type_size[insn->type] == 8 && !IR_IS_32BIT(insn->type, insn->val)) {
					constraints->tmp_regs[0] = IR_TMP_REG(2, insn->type, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
					n = 1;
				}
			}
			break;
		case IR_IMUL3:
			flags = IR_USE_MUST_BE_IN_REG | IR_OP1_SHOULD_BE_IN_REG;
			break;
		case IR_SHIFT:
			if (rule & IR_FUSED) {
				flags = IR_OP2_MUST_BE_IN_REG;
			} else {
				flags = IR_DEF_REUSES_OP1_REG | IR_DEF_CONFLICTS_WITH_INPUT_REGS | IR_USE_MUST_BE_IN_REG | IR_OP1_SHOULD_BE_IN_REG | IR_OP2_SHOULD_BE_IN_REG;
			}
			constraints->hints[1] = IR_REG_NONE;
			constraints->hints[2] = IR_REG_RCX;
			constraints->hints_count = 3;
			constraints->tmp_regs[0] = IR_SCRATCH_REG(IR_REG_RCX, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
			n = 1;
			break;
		case IR_MUL_INT:
			/* %rax - used as input and result */
			constraints->def_reg = IR_REG_RAX;
			constraints->hints[1] = IR_REG_RAX;
			constraints->hints_count = 2;
			flags = IR_USE_SHOULD_BE_IN_REG | IR_OP1_SHOULD_BE_IN_REG | IR_OP2_SHOULD_BE_IN_REG;
			constraints->tmp_regs[0] = IR_SCRATCH_REG(IR_REG_RDX, IR_USE_SUB_REF, IR_DEF_SUB_REF);
			constraints->tmp_regs[1] = IR_SCRATCH_REG(IR_REG_RAX, IR_LOAD_SUB_REF, IR_SAVE_SUB_REF);
			n = 2;
			break;
		case IR_DIV_INT:
			/* %rax - used as input and result */
			constraints->def_reg = IR_REG_RAX;
			constraints->hints[1] = IR_REG_RAX;
			constraints->hints_count = 2;
			flags = IR_USE_SHOULD_BE_IN_REG | IR_OP1_SHOULD_BE_IN_REG | IR_OP2_SHOULD_BE_IN_REG;
			constraints->tmp_regs[0] = IR_SCRATCH_REG(IR_REG_RDX, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
			constraints->tmp_regs[1] = IR_SCRATCH_REG(IR_REG_RAX, IR_LOAD_SUB_REF, IR_SAVE_SUB_REF);
			n = 2;
			goto op2_const;
		case IR_MOD_INT:
			constraints->def_reg = IR_REG_RDX;
			constraints->hints[1] = IR_REG_RAX;
			constraints->hints_count = 2;
			flags = IR_USE_SHOULD_BE_IN_REG | IR_OP1_SHOULD_BE_IN_REG | IR_OP2_SHOULD_BE_IN_REG;
			constraints->tmp_regs[0] = IR_SCRATCH_REG(IR_REG_RAX, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
			constraints->tmp_regs[1] = IR_SCRATCH_REG(IR_REG_RDX, IR_LOAD_SUB_REF, IR_SAVE_SUB_REF);
			n = 2;
			goto op2_const;
		case IR_MIN_MAX_INT:
			flags = IR_DEF_REUSES_OP1_REG | IR_USE_MUST_BE_IN_REG | IR_OP1_SHOULD_BE_IN_REG | IR_OP2_MUST_BE_IN_REG;
op2_const:
			insn = &ctx->ir_base[ref];
			if (IR_IS_CONST_REF(insn->op2) && insn->op1 != insn->op2) {
				constraints->tmp_regs[n] = IR_TMP_REG(2, insn->type, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
				n++;
			}
			break;
		case IR_CMP_INT:
		case IR_TEST_INT:
			insn = &ctx->ir_base[ref];
			flags = IR_USE_MUST_BE_IN_REG | IR_OP1_MUST_BE_IN_REG | IR_OP2_SHOULD_BE_IN_REG;
			if (IR_IS_CONST_REF(insn->op1)) {
				const ir_insn *val_insn = &ctx->ir_base[insn->op1];
				constraints->tmp_regs[0] = IR_TMP_REG(1, val_insn->type, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
				n = 1;
			} else if (ir_rule(ctx, insn->op1) & IR_FUSED) {
				flags = IR_USE_MUST_BE_IN_REG | IR_OP2_MUST_BE_IN_REG;
			}
			if (IR_IS_CONST_REF(insn->op2) && insn->op1 != insn->op2) {
				const ir_insn *val_insn = &ctx->ir_base[insn->op2];
				flags = IR_USE_MUST_BE_IN_REG | IR_OP1_SHOULD_BE_IN_REG;
				if (ir_type_size[val_insn->type] == 8 && !IR_IS_32BIT(val_insn->type, val_insn->val)) {
					constraints->tmp_regs[n] = IR_TMP_REG(2, val_insn->type, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
					n++;
				}
			}
			break;
		case IR_CMP_FP:
			insn = &ctx->ir_base[ref];
			if (!(rule & IR_FUSED)) {
				constraints->tmp_regs[0] = IR_TMP_REG(3, IR_BOOL, IR_DEF_SUB_REF, IR_SAVE_SUB_REF);
				n = 1;
			}
			flags = IR_USE_MUST_BE_IN_REG | IR_OP1_MUST_BE_IN_REG | IR_OP2_SHOULD_BE_IN_REG;
			if (IR_IS_CONST_REF(insn->op1)) {
				const ir_insn *val_insn = &ctx->ir_base[insn->op1];
				constraints->tmp_regs[n] = IR_TMP_REG(1, val_insn->type, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
				n++;
			}
			break;
		case IR_BINOP_AVX:
			flags = IR_USE_MUST_BE_IN_REG | IR_OP1_MUST_BE_IN_REG | IR_OP2_SHOULD_BE_IN_REG;
			insn = &ctx->ir_base[ref];
			if (IR_IS_CONST_REF(insn->op1)) {
				constraints->tmp_regs[0] = IR_TMP_REG(1, insn->type, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
				n = 1;
			}
			break;
		case IR_VSTORE_INT:
			flags = IR_OP3_MUST_BE_IN_REG;
			insn = &ctx->ir_base[ref];
			if (IR_IS_CONST_REF(insn->op3)) {
				insn = &ctx->ir_base[insn->op3];
				if (ir_type_size[insn->type] == 8 && !IR_IS_32BIT(insn->type, insn->val)) {
					constraints->tmp_regs[0] = IR_TMP_REG(3, insn->type, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
					n = 1;
				}
			}
			break;
		case IR_STORE_INT:
			flags = IR_OP2_MUST_BE_IN_REG | IR_OP3_MUST_BE_IN_REG;
			insn = &ctx->ir_base[ref];
			if (IR_IS_CONST_REF(insn->op2)) {
				const ir_insn *val_insn = &ctx->ir_base[insn->op2];
				IR_ASSERT(val_insn->type == IR_ADDR);
				if (ir_type_size[val_insn->type] == 8 && !IR_IS_SIGNED_32BIT(val_insn->val.i64)) {
					constraints->tmp_regs[0] = IR_TMP_REG(2, IR_ADDR, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
					n = 1;
				}
			}
			if (IR_IS_CONST_REF(insn->op3)) {
				const ir_insn *val_insn = &ctx->ir_base[insn->op3];
				if (ir_type_size[val_insn->type] == 8 && !IR_IS_32BIT(val_insn->type, val_insn->val)) {
					constraints->tmp_regs[n] = IR_TMP_REG(3, val_insn->type, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
					n++;
				}
			}
			break;
		case IR_VSTORE_FP:
			flags = IR_OP3_MUST_BE_IN_REG;
			insn = &ctx->ir_base[ref];
			if (IR_IS_CONST_REF(insn->op3)) {
				insn = &ctx->ir_base[insn->op3];
				constraints->tmp_regs[0] = IR_TMP_REG(3, insn->type, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
				n = 1;
			}
			break;
		case IR_LOAD_FP:
		case IR_MEM_BINOP_INT:
			flags = IR_USE_MUST_BE_IN_REG | IR_OP2_MUST_BE_IN_REG;
			insn = &ctx->ir_base[ref];
			if (IR_IS_CONST_REF(insn->op2)) {
				const ir_insn *val_insn = &ctx->ir_base[insn->op2];
				IR_ASSERT(val_insn->type == IR_ADDR);
				if (ir_type_size[val_insn->type] == 8 && !IR_IS_32BIT(val_insn->type, val_insn->val)) {
					constraints->tmp_regs[0] = IR_TMP_REG(2, IR_ADDR, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
					n = 1;
				}
			}
			break;
		case IR_STORE_FP:
			flags = IR_OP2_MUST_BE_IN_REG | IR_OP3_MUST_BE_IN_REG;
			insn = &ctx->ir_base[ref];
			if (IR_IS_CONST_REF(insn->op2)) {
				IR_ASSERT(ctx->ir_base[insn->op2].type == IR_ADDR);
				constraints->tmp_regs[0] = IR_TMP_REG(2, IR_ADDR, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
				n = 1;
			}
			if (IR_IS_CONST_REF(insn->op3)) {
				insn = &ctx->ir_base[insn->op3];
				constraints->tmp_regs[n] = IR_TMP_REG(3, insn->type, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
				n++;
			}
			break;
		case IR_SWITCH:
			flags = IR_OP2_MUST_BE_IN_REG;
			insn = &ctx->ir_base[ref];
			if (IR_IS_CONST_REF(insn->op2)) {
				insn = &ctx->ir_base[insn->op2];
				constraints->tmp_regs[0] = IR_TMP_REG(2, insn->type, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
				n = 1;
			}
			if (sizeof(void*) == 8) {
				constraints->tmp_regs[n] = IR_TMP_REG(3, IR_ADDR, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
				n++;
			}
			break;
		case IR_CALL:
			insn = &ctx->ir_base[ref];
			if (IR_IS_TYPE_INT(insn->type)) {
				constraints->def_reg = IR_REG_INT_RET1;
#ifdef IR_REG_FP_RET1
			} else {
				constraints->def_reg = IR_REG_FP_RET1;
#endif
			}
			constraints->tmp_regs[0] = IR_SCRATCH_REG(IR_REG_SCRATCH, IR_USE_SUB_REF, IR_DEF_SUB_REF);
			n = 1;
			IR_FALLTHROUGH;
		case IR_TAILCALL:
			insn = &ctx->ir_base[ref];
			if (insn->inputs_count > 2) {
				constraints->hints[2] = IR_REG_NONE;
				constraints->hints_count = ir_get_args_regs(ctx, insn, constraints->hints);
				if (!IR_IS_CONST_REF(insn->op2)) {
					constraints->tmp_regs[n] = IR_TMP_REG(1, IR_ADDR, IR_LOAD_SUB_REF, IR_USE_SUB_REF);
					n++;
				}
			}
			flags = IR_USE_SHOULD_BE_IN_REG | IR_OP2_SHOULD_BE_IN_REG | IR_OP3_SHOULD_BE_IN_REG;
			break;
		case IR_BINOP_SSE2:
			flags = IR_DEF_REUSES_OP1_REG | IR_USE_MUST_BE_IN_REG | IR_OP1_SHOULD_BE_IN_REG | IR_OP2_SHOULD_BE_IN_REG;
			break;
		case IR_SHIFT_CONST:
		case IR_INC:
		case IR_DEC:
		case IR_MUL_PWR2:
		case IR_DIV_PWR2:
		case IR_MOD_PWR2:
		case IR_OP_INT:
		case IR_OP_FP:
			flags = IR_DEF_REUSES_OP1_REG | IR_USE_MUST_BE_IN_REG | IR_OP1_SHOULD_BE_IN_REG;
			break;
		case IR_BIT_COUNT:
			insn = &ctx->ir_base[ref];
			if (ir_type_size[insn->type] == 1) {
				flags = IR_USE_MUST_BE_IN_REG | IR_OP1_MUST_BE_IN_REG;
			} else {
				flags = IR_USE_MUST_BE_IN_REG | IR_OP1_SHOULD_BE_IN_REG;
			}
			if (IR_IS_CONST_REF(insn->op1)) {
				constraints->tmp_regs[0] = IR_TMP_REG(1, insn->type, IR_LOAD_SUB_REF, IR_DEF_SUB_REF);
				n = 1;
			}
			break;
		case IR_CTPOP:
			flags = IR_USE_MUST_BE_IN_REG | IR_OP1_SHOULD_BE_IN_REG;
			insn = &ctx->ir_base[ref];
			constraints->tmp_regs[0] = IR_TMP_REG(2, insn->type, IR_USE_SUB_REF, IR_SAVE_SUB_REF);
			n = 1;
			if (ir_type_size[insn->type] == 8) {
				constraints->tmp_regs[1] = IR_TMP_REG(3, insn->type, IR_USE_SUB_REF, IR_SAVE_SUB_REF);
				n = 2;
			}
			break;
		case IR_COPY_INT:
		case IR_COPY_FP:
		case IR_SEXT:
		case IR_ZEXT:
			flags = IR_DEF_REUSES_OP1_REG | IR_USE_MUST_BE_IN_REG | IR_OP1_SHOULD_BE_IN_REG;
			break;
		case IR_ABS_INT:
			flags = IR_DEF_CONFLICTS_WITH_INPUT_REGS | IR_USE_MUST_BE_IN_REG | IR_OP1_MUST_BE_IN_REG;
			break;
		case IR_PARAM:
			constraints->def_reg = ir_get_param_reg(ctx, ref);
			flags = 0;
			break;
		case IR_PI:
		case IR_PHI:
			flags = IR_USE_SHOULD_BE_IN_REG;
			break;
		case IR_RLOAD:
			constraints->def_reg = ctx->ir_base[ref].op2;
			flags = IR_USE_SHOULD_BE_IN_REG;
			break;
		case IR_EXITCALL:
			flags = IR_USE_MUST_BE_IN_REG;
			constraints->def_reg = IR_REG_INT_RET1;
			break;
		case IR_IF_INT:
		case IR_GUARD:
		case IR_GUARD_NOT:
			flags = IR_OP2_SHOULD_BE_IN_REG;
			break;
		case IR_IJMP:
			flags = IR_OP2_SHOULD_BE_IN_REG;
			break;
		case IR_RSTORE:
			flags = IR_OP3_SHOULD_BE_IN_REG;
			break;
		case IR_RETURN_INT:
			flags = IR_OP2_SHOULD_BE_IN_REG;
			constraints->hints[2] = IR_REG_INT_RET1;
			constraints->hints_count = 3;
			break;
		case IR_RETURN_FP:
#ifdef IR_REG_FP_RET1
			flags = IR_OP2_SHOULD_BE_IN_REG;
			constraints->hints[2] = IR_REG_FP_RET1;
			constraints->hints_count = 3;
#endif
			break;
		case IR_SNAPSHOT:
			flags = 0;
			break;
	}
	constraints->tmps_count = n;

	return flags;
}

/* instruction selection */
static uint32_t ir_match_insn(ir_ctx *ctx, ir_ref ref);

static void ir_match_fuse_addr(ir_ctx *ctx, ir_ref addr_ref)
{
	if (!IR_IS_CONST_REF(addr_ref)) {
		uint32_t rule = ctx->rules[addr_ref];

		if (!rule) {
			ctx->rules[addr_ref] = rule = ir_match_insn(ctx, addr_ref);
		}
		if (rule == IR_LEA_OB) {
			ir_use_list *use_list = &ctx->use_lists[addr_ref];
			ir_ref j = use_list->count;

			if (j > 1) {
				/* check if address is used only in LOAD and STORE */
				ir_ref *p = &ctx->use_edges[use_list->refs];

				do {
					ir_insn *insn = &ctx->ir_base[*p];
					if (insn->op != IR_LOAD && (insn->op != IR_STORE || insn->op3 == addr_ref)) {
						return;
					}
					p++;
				} while (--j);
			}
			ctx->rules[addr_ref] = IR_FUSED | IR_SIMPLE | IR_LEA_OB;
		}
	}
}

/* A naive check if there is a STORE or CALL between this LOAD and the fusion root */
static bool ir_match_has_mem_deps(ir_ctx *ctx, ir_ref ref, ir_ref root)
{
	if (ref + 1 != root) {
		ir_ref pos = ctx->prev_ref[root];

		do {
			ir_insn *insn = &ctx->ir_base[pos];

			if (insn->op == IR_STORE) {
				// TODO: check if LOAD and STORE addresses may alias
				return 1;
			} else if (insn->op == IR_CALL) {
				return 1;
			}
			pos = ctx->prev_ref[pos];
		} while (ref != pos);
	}
	return 0;
}

static bool ir_match_fuse_load(ir_ctx *ctx, ir_ref ref, ir_ref root)
{
	if (ir_in_same_block(ctx, ref)
	 && ctx->ir_base[ref].op == IR_LOAD
	 && ctx->use_lists[ref].count == 2
	 && !ir_match_has_mem_deps(ctx, ref, root)) {
		ir_ref addr_ref = ctx->ir_base[ref].op2;
		ir_insn *addr_insn = &ctx->ir_base[addr_ref];

		if (IR_IS_CONST_REF(addr_ref)) {
			if (addr_insn->op == IR_C_ADDR &&
				(sizeof(void*) == 4 || IR_IS_SIGNED_32BIT(addr_insn->val.i64))) {
				ctx->rules[ref] = IR_FUSED | IR_SIMPLE | IR_LOAD;
				return 1;
			}
		} else {
			ctx->rules[ref] = IR_FUSED | IR_SIMPLE | IR_LOAD;
			ir_match_fuse_addr(ctx, addr_ref);
			return 1;
		}
	} else if (ir_in_same_block(ctx, ref)
	 && ctx->ir_base[ref].op == IR_VLOAD) {
		return 1;
	}
	return 0;
}

static void ir_swap_ops(ir_insn *insn)
{
	ir_ref tmp = insn->op1;
	insn->op1 = insn->op2;
	insn->op2 = tmp;
}

static void ir_match_fuse_load_commutative_int(ir_ctx *ctx, ir_insn *insn, ir_ref root)
{
	if (IR_IS_CONST_REF(insn->op2)) {
		if (ir_type_size[insn->type] > 4 && !IR_IS_32BIT(ctx->ir_base[insn->op2].type, ctx->ir_base[insn->op2].val)
		 && !IR_IS_CONST_REF(insn->op1)
		 && ir_match_fuse_load(ctx, insn->op1, root)) {
			ir_swap_ops(insn);
		}
	} else if (!ir_match_fuse_load(ctx, insn->op2, root)) {
		if (!IR_IS_CONST_REF(insn->op1)
		 && ir_match_fuse_load(ctx, insn->op1, root)) {
			ir_swap_ops(insn);
		}
	}
}

static void ir_match_fuse_load_commutative_fp(ir_ctx *ctx, ir_insn *insn, ir_ref root)
{
	if (!IR_IS_CONST_REF(insn->op2)
	 && !ir_match_fuse_load(ctx, insn->op2, root)
	 && (IR_IS_CONST_REF(insn->op1) || ir_match_fuse_load(ctx, insn->op1, root))) {
		ir_swap_ops(insn);
	}
}

static void ir_match_fuse_load_cmp_int(ir_ctx *ctx, ir_insn *insn, ir_ref root)
{
	if (IR_IS_CONST_REF(insn->op2)) {
		if (!IR_IS_CONST_REF(insn->op1)
		 && ir_match_fuse_load(ctx, insn->op1, root)
		 && ir_type_size[ctx->ir_base[insn->op2].type] > 4
		 && !IR_IS_32BIT(ctx->ir_base[insn->op2].type, ctx->ir_base[insn->op2].val)) {
			ir_swap_ops(insn);
			if (insn->op != IR_EQ && insn->op != IR_NE) {
				insn->op ^= 3;
			}
		}
	} else if (!ir_match_fuse_load(ctx, insn->op2, root)) {
		if (!IR_IS_CONST_REF(insn->op1)
		 && ir_match_fuse_load(ctx, insn->op1, root)) {
			ir_swap_ops(insn);
			if (insn->op != IR_EQ && insn->op != IR_NE) {
				insn->op ^= 3;
			}
		}
	}
}

static void ir_match_fuse_load_cmp_fp(ir_ctx *ctx, ir_insn *insn, ir_ref root, bool direct)
{
	if (direct) {
		if (insn->op == IR_LT || insn->op == IR_LE) {
			/* swap operands to avoid P flag check */
			ir_swap_ops(insn);
			insn->op ^= 3;
		}
	} else {
		if (insn->op == IR_GT || insn->op == IR_GE) {
			/* swap operands to avoid P flag check */
			ir_swap_ops(insn);
			insn->op ^= 3;
		}
	}
	if (IR_IS_CONST_REF(insn->op2)) {
	} else if (ir_match_fuse_load(ctx, insn->op2, root)) {
	} else if ((IR_IS_CONST_REF(insn->op1) && !IR_IS_FP_ZERO(ctx->ir_base[insn->op1])) || ir_match_fuse_load(ctx, insn->op1, root)) {
		ir_swap_ops(insn);
		if (insn->op != IR_EQ && insn->op != IR_NE) {
			insn->op ^= 3;
		}
	}
}

static uint32_t ir_match_insn(ir_ctx *ctx, ir_ref ref)
{
	ir_insn *op2_insn;
	ir_insn *insn = &ctx->ir_base[ref];
	uint32_t store_rule;
	ir_op load_op;

	switch (insn->op) {
		case IR_EQ:
		case IR_NE:
		case IR_LT:
		case IR_GE:
		case IR_LE:
		case IR_GT:
		case IR_ULT:
		case IR_UGE:
		case IR_ULE:
		case IR_UGT:
			if (IR_IS_TYPE_INT(ctx->ir_base[insn->op1].type)) {
				if (IR_IS_CONST_REF(insn->op2)
				 && ctx->ir_base[insn->op2].val.i64 == 0
				 && insn->op1 == ref - 1) { /* previous instruction */
					ir_insn *op1_insn = &ctx->ir_base[insn->op1];

					if (op1_insn->op == IR_ADD ||
				        op1_insn->op == IR_SUB ||
//				        op1_insn->op == IR_MUL ||
				        op1_insn->op == IR_OR  ||
				        op1_insn->op == IR_AND ||
				        op1_insn->op == IR_XOR) {

						if (ir_op_flags[op1_insn->op] & IR_OP_FLAG_COMMUTATIVE) {
							ir_match_fuse_load_commutative_int(ctx, op1_insn, ref);
						} else {
							ir_match_fuse_load(ctx, op1_insn->op2, ref);
						}
						if (op1_insn->op == IR_AND && ctx->use_lists[insn->op1].count == 1) {
							/* v = AND(_, _); CMP(v, 0) => SKIP_TEST; TEST */
							if (IR_IS_CONST_REF(op1_insn->op2)) {
								ir_match_fuse_load(ctx, op1_insn->op1, ref);
							}
							ctx->rules[insn->op1] = IR_FUSED | IR_TEST_INT;
							return IR_TESTCC_INT;
						} else {
							/* v = BINOP(_, _); CMP(v, 0) => BINOP; SETCC */
							ctx->rules[insn->op1] = IR_BINOP_INT;
							return IR_SETCC_INT;
						}
					}
				}
				ir_match_fuse_load_cmp_int(ctx, insn, ref);
				return IR_CMP_INT;
			} else {
				ir_match_fuse_load_cmp_fp(ctx, insn, ref, 1);
				return IR_CMP_FP;
			}
			break;
		case IR_ADD:
		case IR_SUB:
			if (IR_IS_TYPE_INT(insn->type)) {
				if ((ctx->flags & IR_OPT_CODEGEN) && IR_IS_CONST_REF(insn->op2)) {
					op2_insn = &ctx->ir_base[insn->op2];
					if (IR_IS_CONST_REF(insn->op1)) {
						// const
					} else if (op2_insn->val.i64 == 0) {
						return IR_COPY_INT;
					} else if ((ir_type_size[insn->type] >= 4 && insn->op == IR_ADD && IR_IS_SIGNED_32BIT(op2_insn->val.i64)) ||
							(ir_type_size[insn->type] >= 4 && insn->op == IR_SUB && IR_IS_SIGNED_32BIT(-op2_insn->val.i64))) {
						if (ir_in_same_block(ctx, insn->op1) && ctx->use_lists[insn->op1].count == 1) {
							uint32_t rule = ctx->rules[insn->op1];

							if (!rule) {
								ctx->rules[insn->op1] = rule = ir_match_insn(ctx, insn->op1);
							}
							if (rule == IR_LEA_SI) {
								/* z = MUL(Y, 2|4|8) ... ADD(z, imm32) => SKIP ... LEA [Y*2|4|8+im32] */
								ctx->rules[insn->op1] = IR_FUSED | IR_SIMPLE | IR_LEA_SI;
								return IR_LEA_SI_O;
							} else if (rule == IR_LEA_SIB) {
								/* z = ADD(X, MUL(Y, 2|4|8)) ... ADD(z, imm32) => SKIP ... LEA [X+Y*2|4|8+im32] */
								ctx->rules[insn->op1] = IR_FUSED | IR_SIMPLE | IR_LEA_SIB;
								return IR_LEA_SIB_O;
							} else if (rule == IR_LEA_IB) {
								/* z = ADD(X, Y) ... ADD(z, imm32) => SKIP ... LEA [X+Y+im32] */
								ctx->rules[insn->op1] = IR_FUSED | IR_SIMPLE | IR_LEA_IB;
								return IR_LEA_IB_O;
							}
						}
						/* ADD(X, imm32) => LEA [X+imm32] */
						return IR_LEA_OB;
					} else if (op2_insn->val.i64 == 1 || op2_insn->val.i64 == -1) {
						if (insn->op == IR_ADD) {
							if (op2_insn->val.i64 == 1) {
								/* ADD(_, 1) => INC */
								return IR_INC;
						    } else {
								/* ADD(_, -1) => DEC */
								return IR_DEC;
						    }
						} else {
							if (op2_insn->val.i64 == 1) {
								/* SUB(_, 1) => DEC */
								return IR_DEC;
						    } else {
								/* SUB(_, -1) => INC */
								return IR_INC;
						    }
						}
					}
				} else if ((ctx->flags & IR_OPT_CODEGEN) && insn->op == IR_ADD && ir_type_size[insn->type] >= 4) {
					if (ir_in_same_block(ctx, insn->op1) && ctx->use_lists[insn->op1].count == 1) {
						uint32_t rule =ctx->rules[insn->op1];
						if (!rule) {
							ctx->rules[insn->op1] = rule = ir_match_insn(ctx, insn->op1);
						}
						if (rule == IR_LEA_OB) {
							ctx->rules[insn->op1] = IR_FUSED | IR_SIMPLE | IR_LEA_OB;
							if (ir_in_same_block(ctx, insn->op2) && ctx->use_lists[insn->op2].count == 1) {
								rule = ctx->rules[insn->op2];
								if (!rule) {
									ctx->rules[insn->op2] = rule = ir_match_insn(ctx, insn->op2);
								}
								if (rule == IR_LEA_SI) {
									/* x = ADD(X, imm32) ... y = MUL(Y, 2|4|8) ... ADD(x, y) => SKIP ... SKIP ... LEA */
									ctx->rules[insn->op2] = IR_FUSED | IR_SIMPLE | IR_LEA_SI;
									return IR_LEA_OB_SI;
								}
							}
							/* x = ADD(X, imm32) ... ADD(x, Y) => SKIP ... LEA */
							return IR_LEA_OB_I;
						}
					}
					if (ir_in_same_block(ctx, insn->op2) && ctx->use_lists[insn->op2].count == 1) {
						uint32_t rule = ctx->rules[insn->op2];
						if (!rule) {
							ctx->rules[insn->op2] = rule = ir_match_insn(ctx, insn->op2);
						}
						if (rule == IR_LEA_OB) {
							ctx->rules[insn->op2] = IR_FUSED | IR_SIMPLE | IR_LEA_OB;
							if (ir_in_same_block(ctx, insn->op1) && ctx->use_lists[insn->op1].count == 1) {
								rule =ctx->rules[insn->op1];
								if (!rule) {
									ctx->rules[insn->op1] = rule = ir_match_insn(ctx, insn->op1);
								}
								if (rule == IR_LEA_SI) {
									/* x = ADD(X, imm32) ... y = MUL(Y, 2|4|8) ... ADD(y, x) => SKIP ... SKIP ... LEA */
									ctx->rules[insn->op1] = IR_FUSED | IR_SIMPLE | IR_LEA_SI;
									return IR_LEA_SI_OB;
								}
							}
							/* x = ADD(X, imm32) ... ADD(Y, x) => SKIP ... LEA */
							return IR_LEA_I_OB;
						}
					}
					/* ADD(X, Y) => LEA [X + Y] */
					return IR_LEA_IB;
				}
binop_int:
				if (ir_op_flags[insn->op] & IR_OP_FLAG_COMMUTATIVE) {
					ir_match_fuse_load_commutative_int(ctx, insn, ref);
				} else {
					ir_match_fuse_load(ctx, insn->op2, ref);
				}
				return IR_BINOP_INT;
			} else {
binop_fp:
				if (ir_op_flags[insn->op] & IR_OP_FLAG_COMMUTATIVE) {
					ir_match_fuse_load_commutative_fp(ctx, insn, ref);
				} else {
					ir_match_fuse_load(ctx, insn->op2, ref);
				}
				if (ctx->mflags & IR_X86_AVX) {
					return IR_BINOP_AVX;
				} else {
					return IR_BINOP_SSE2;
				}
			}
			break;
		case IR_MUL:
			if (IR_IS_TYPE_INT(insn->type)) {
				if ((ctx->flags & IR_OPT_CODEGEN) && IR_IS_CONST_REF(insn->op2)) {
					op2_insn = &ctx->ir_base[insn->op2];
					if (IR_IS_CONST_REF(insn->op1)) {
						// const
					} else if (op2_insn->val.u64 == 0) {
						// 0
					} else if (op2_insn->val.u64 == 1) {
						return IR_COPY_INT;
					} else if (ir_type_size[insn->type] >= 4 &&
							(op2_insn->val.u64 == 2 || op2_insn->val.u64 == 4 || op2_insn->val.u64 == 8)) {
						/* MUL(X, 2|4|8) => LEA [X*2|4|8] */
						return IR_LEA_SI;
					} else if (ir_type_size[insn->type] >= 4 &&
							(op2_insn->val.u64 == 3 || op2_insn->val.u64 == 5 || op2_insn->val.u64 == 9)) {
						/* MUL(X, 3|5|9) => LEA [X+X*2|4|8] */
						return IR_LEA_SIB;
					} else if (IR_IS_POWER_OF_TWO(op2_insn->val.u64)) {
						/* MUL(X, PWR2) => SHL */
						return IR_MUL_PWR2;
					} else if (IR_IS_TYPE_SIGNED(insn->type)
					 && ir_type_size[insn->type] != 1
					 && IR_IS_SIGNED_32BIT(op2_insn->val.i64)
					 && !IR_IS_CONST_REF(insn->op1)) {
						/* MUL(_, imm32) => IMUL */
						ir_match_fuse_load(ctx, insn->op1, ref);
						return IR_IMUL3;
					}
				}
				/* Prefer IMUL over MUL because it's more flexible and uses less registers ??? */
//				if (IR_IS_TYPE_SIGNED(insn->type) && ir_type_size[insn->type] != 1) {
				if (ir_type_size[insn->type] != 1) {
					goto binop_int;
				}
				ir_match_fuse_load(ctx, insn->op2, ref);
				return IR_MUL_INT;
			} else {
				goto binop_fp;
			}
			break;
		case IR_ADD_OV:
		case IR_SUB_OV:
			IR_ASSERT(IR_IS_TYPE_INT(insn->type));
			goto binop_int;
		case IR_MUL_OV:
			IR_ASSERT(IR_IS_TYPE_INT(insn->type));
			if (IR_IS_TYPE_SIGNED(insn->type) && ir_type_size[insn->type] != 1) {
				if ((ctx->flags & IR_OPT_CODEGEN) && IR_IS_CONST_REF(insn->op2)) {
					op2_insn = &ctx->ir_base[insn->op2];
					if (IR_IS_SIGNED_32BIT(op2_insn->val.i64)
					 && !IR_IS_CONST_REF(insn->op1)) {
						/* MUL(_, imm32) => IMUL */
						ir_match_fuse_load(ctx, insn->op1, ref);
						return IR_IMUL3;
					}
				}
				goto binop_int;
			}
			ir_match_fuse_load(ctx, insn->op2, ref);
			return IR_MUL_INT;
		case IR_DIV:
			if (IR_IS_TYPE_INT(insn->type)) {
				if ((ctx->flags & IR_OPT_CODEGEN) && IR_IS_CONST_REF(insn->op2)) {
					op2_insn = &ctx->ir_base[insn->op2];
					if (IR_IS_CONST_REF(insn->op1)) {
						// const
					} else if (op2_insn->val.u64 == 1) {
						return IR_COPY_INT;
					} else if (IR_IS_POWER_OF_TWO(op2_insn->val.u64)) {
						/* DIV(X, PWR2) => SHR */
						return IR_DIV_PWR2;
					}
				}
				ir_match_fuse_load(ctx, insn->op2, ref);
				return IR_DIV_INT;
			} else {
				goto binop_fp;
			}
			break;
		case IR_MOD:
			if ((ctx->flags & IR_OPT_CODEGEN) && IR_IS_CONST_REF(insn->op2)) {
				op2_insn = &ctx->ir_base[insn->op2];
				if (IR_IS_CONST_REF(insn->op1)) {
					// const
				} else if (IR_IS_TYPE_UNSIGNED(insn->type)
				 && IR_IS_POWER_OF_TWO(op2_insn->val.u64)
				 && IR_IS_UNSIGNED_32BIT(op2_insn->val.u64 - 1)) {
					/* MOD(X, PWR2) => AND */
					return IR_MOD_PWR2;
				}
			}
			ir_match_fuse_load(ctx, insn->op2, ref);
			return IR_MOD_INT;
		case IR_BSWAP:
		case IR_NOT:
			if (insn->type == IR_BOOL) {
				IR_ASSERT(IR_IS_TYPE_INT(ctx->ir_base[insn->op1].type)); // TODO: IR_BOOL_NOT_FP
				return IR_BOOL_NOT_INT;
			} else {
				IR_ASSERT(IR_IS_TYPE_INT(insn->type));
				return IR_OP_INT;
			}
			break;
		case IR_NEG:
			if (IR_IS_TYPE_INT(insn->type)) {
				return IR_OP_INT;
			} else {
				return IR_OP_FP;
			}
		case IR_ABS:
			if (IR_IS_TYPE_INT(insn->type)) {
				return IR_ABS_INT; // movl %edi, %eax; negl %eax; cmovs %edi, %eax
			} else {
				return IR_OP_FP;
			}
		case IR_OR:
			if ((ctx->flags & IR_OPT_CODEGEN) && IR_IS_CONST_REF(insn->op2)) {
				op2_insn = &ctx->ir_base[insn->op2];
				if (IR_IS_CONST_REF(insn->op1)) {
					// const
				} else if (op2_insn->val.i64 == 0) {
					return IR_COPY_INT;
				} else if (op2_insn->val.i64 == -1) {
					// -1
				}
			}
			goto binop_int;
		case IR_AND:
			if ((ctx->flags & IR_OPT_CODEGEN) && IR_IS_CONST_REF(insn->op2)) {
				op2_insn = &ctx->ir_base[insn->op2];
				if (IR_IS_CONST_REF(insn->op1)) {
					// const
				} else if (op2_insn->val.i64 == 0) {
					// 0
				} else if (op2_insn->val.i64 == -1) {
					return IR_COPY_INT;
				}
			}
			goto binop_int;
		case IR_XOR:
			if ((ctx->flags & IR_OPT_CODEGEN) && IR_IS_CONST_REF(insn->op2)) {
				op2_insn = &ctx->ir_base[insn->op2];
				if (IR_IS_CONST_REF(insn->op1)) {
					// const
				}
			}
			goto binop_int;
		case IR_SHL:
			if (IR_IS_CONST_REF(insn->op2)) {
				if (ctx->flags & IR_OPT_CODEGEN) {
					op2_insn = &ctx->ir_base[insn->op2];
					if (IR_IS_CONST_REF(insn->op1)) {
						// const
					} else if (op2_insn->val.u64 == 0) {
						return IR_COPY_INT;
					} else if (ir_type_size[insn->type] >= 4) {
						if (op2_insn->val.u64 == 1) {
							// lea [op1*2]
						} else if (op2_insn->val.u64 == 2) {
							// lea [op1*4]
						} else if (op2_insn->val.u64 == 3) {
							// lea [op1*8]
						}
					}
				}
				return IR_SHIFT_CONST;
			}
			return IR_SHIFT;
		case IR_SHR:
		case IR_SAR:
		case IR_ROL:
		case IR_ROR:
			if (IR_IS_CONST_REF(insn->op2)) {
				if (ctx->flags & IR_OPT_CODEGEN) {
					op2_insn = &ctx->ir_base[insn->op2];
					if (IR_IS_CONST_REF(insn->op1)) {
						// const
					} else if (op2_insn->val.u64 == 0) {
						return IR_COPY_INT;
					}
				}
				return IR_SHIFT_CONST;
			}
			return IR_SHIFT;
		case IR_MIN:
		case IR_MAX:
			if (IR_IS_TYPE_INT(insn->type)) {
				return IR_MIN_MAX_INT;
			} else {
				goto binop_fp;
			}
			break;
//		case IR_COND:
		case IR_COPY:
			if (IR_IS_TYPE_INT(insn->type)) {
				return IR_COPY_INT;
			} else {
				return IR_COPY_FP;
			}
			break;
		case IR_CALL:
			ctx->flags2 |= IR_HAS_CALLS;
			IR_FALLTHROUGH;
		case IR_TAILCALL:
			if (ir_in_same_block(ctx, insn->op2)) {
				ir_match_fuse_load(ctx, insn->op2, ref);
			}
			return insn->op;
		case IR_VAR:
			return IR_SKIPPED | IR_VAR;
		case IR_PARAM:
			return ctx->use_lists[ref].count > 0 ? IR_PARAM : IR_SKIPPED | IR_PARAM;
		case IR_ALLOCA:
			/* alloca() may be used only in functions */
			if (ctx->flags & IR_FUNCTION) {
				ctx->flags |= IR_USE_FRAME_POINTER;
				ctx->flags2 |= IR_HAS_ALLOCA;
			}
			return IR_ALLOCA;
		case IR_VSTORE:
			if (IR_IS_TYPE_INT(ctx->ir_base[insn->op3].type)) {
				store_rule = IR_VSTORE_INT;
				load_op = IR_VLOAD;
store_int:
				if ((ctx->flags & IR_OPT_CODEGEN)
				 && ir_in_same_block(ctx, insn->op3)
				 && (ctx->use_lists[insn->op3].count == 1 ||
				     (ctx->use_lists[insn->op3].count == 2
				   && (ctx->ir_base[insn->op3].op == IR_ADD_OV ||
				       ctx->ir_base[insn->op3].op == IR_SUB_OV)))) {
					ir_insn *op_insn = &ctx->ir_base[insn->op3];
					uint32_t rule = ctx->rules[insn->op3];

					if (!rule) {
						ctx->rules[insn->op3] = rule = ir_match_insn(ctx, insn->op3);
					}
					if ((rule == IR_BINOP_INT && op_insn->op != IR_MUL) || rule == IR_LEA_OB || rule == IR_LEA_IB) {
						if (insn->op1 == op_insn->op1
						 && ctx->ir_base[op_insn->op1].op == load_op
						 && ctx->ir_base[op_insn->op1].op2 == insn->op2
						 && ctx->use_lists[op_insn->op1].count == 2) {
							/* l = LOAD(_, a) ... v = BINOP(l, _) ... STORE(l, a, v) => SKIP ... SKIP_MEM_BINOP ... MEM_BINOP */
							ctx->rules[insn->op3] = IR_FUSED | IR_BINOP_INT;
							ctx->rules[op_insn->op1] = IR_SKIPPED | load_op;
							if (ctx->ir_base[op_insn->op2].op == IR_LOAD) {
								ir_match_fuse_addr(ctx, ctx->ir_base[op_insn->op2].op2);
								ctx->rules[op_insn->op2] = IR_LOAD_INT;
							}
							ir_match_fuse_addr(ctx, insn->op2);
							return IR_MEM_BINOP_INT;
						} else if ((ir_op_flags[op_insn->op] & IR_OP_FLAG_COMMUTATIVE)
						 && insn->op1 == op_insn->op2
						 && ctx->ir_base[op_insn->op2].op == load_op
						 && ctx->ir_base[op_insn->op2].op2 == insn->op2
						 && ctx->use_lists[op_insn->op2].count == 2) {
							/* l = LOAD(_, a) ... v = BINOP(_, l) ... STORE(l, a, v) => SKIP ... SKIP_MEM_BINOP ... MEM_BINOP */
							ir_swap_ops(op_insn);
							ctx->rules[insn->op3] = IR_FUSED | IR_BINOP_INT;
							ctx->rules[op_insn->op1] = IR_SKIPPED | load_op;
							if (ctx->ir_base[op_insn->op2].op == IR_LOAD) {
								ir_match_fuse_addr(ctx, ctx->ir_base[op_insn->op2].op2);
								ctx->rules[op_insn->op2] = IR_LOAD_INT;
							}
							ir_match_fuse_addr(ctx, insn->op2);
							return IR_MEM_BINOP_INT;
						}
					} else if (rule == IR_INC) {
						if (insn->op1 == op_insn->op1
						 && ctx->ir_base[op_insn->op1].op == load_op
						 && ctx->ir_base[op_insn->op1].op2 == insn->op2
						 && ctx->use_lists[op_insn->op1].count == 2) {
							/* l = LOAD(_, a) ... v = INC(l) ... STORE(l, a, v) => SKIP ... SKIP ... MEM_INC */
							ctx->rules[insn->op3] = IR_SKIPPED | IR_INC;
							ctx->rules[op_insn->op1] = IR_SKIPPED | load_op;
							return IR_MEM_INC;
						}
					} else if (rule == IR_DEC) {
						if (insn->op1 == op_insn->op1
						 && ctx->ir_base[op_insn->op1].op == load_op
						 && ctx->ir_base[op_insn->op1].op2 == insn->op2
						 && ctx->use_lists[op_insn->op1].count == 2){
							/* l = LOAD(_, a) ... v = DEC(l) ... STORE(l, a, v) => SKIP ... SKIP ... MEM_DEC */
							ctx->rules[insn->op3] = IR_SKIPPED | IR_DEC;
							ctx->rules[op_insn->op1] = IR_SKIPPED | load_op;
							return IR_MEM_DEC;
						}
					} else if (rule == IR_MUL_PWR2) {
						if (insn->op1 == op_insn->op1
						 && ctx->ir_base[op_insn->op1].op == load_op
						 && ctx->ir_base[op_insn->op1].op2 == insn->op2
						 && ctx->use_lists[op_insn->op1].count == 2) {
							/* l = LOAD(_, a) ... v = MUL_PWR2(l) ... STORE(l, a, v) => SKIP ... SKIP ... MEM_MUL_PWR2 */
							ctx->rules[insn->op3] = IR_SKIPPED | IR_MUL_PWR2;
							ctx->rules[op_insn->op1] = IR_SKIPPED | load_op;
							return IR_MEM_MUL_PWR2;
						}
					} else if (rule == IR_DIV_PWR2) {
						if (insn->op1 == op_insn->op1
						 && ctx->ir_base[op_insn->op1].op == load_op
						 && ctx->ir_base[op_insn->op1].op2 == insn->op2
						 && ctx->use_lists[op_insn->op1].count == 2) {
							/* l = LOAD(_, a) ... v = DIV_PWR2(l) ... STORE(l, a, v) => SKIP ... SKIP ... MEM_DIV_PWR2 */
							ctx->rules[insn->op3] = IR_SKIPPED | IR_DIV_PWR2;
							ctx->rules[op_insn->op1] = IR_SKIPPED | load_op;
							return IR_MEM_DIV_PWR2;
						}
					} else if (rule == IR_MOD_PWR2) {
						if (insn->op1 == op_insn->op1
						 && ctx->ir_base[op_insn->op1].op == load_op
						 && ctx->ir_base[op_insn->op1].op2 == insn->op2
						 && ctx->use_lists[op_insn->op1].count == 2) {
							/* l = LOAD(_, a) ... v = MOD_PWR2(l) ... STORE(l, a, v) => SKIP ... SKIP ... MEM_MOD_PWR2 */
							ctx->rules[insn->op3] = IR_SKIPPED | IR_MOD_PWR2;
							ctx->rules[op_insn->op1] = IR_SKIPPED | load_op;
							return IR_MEM_MOD_PWR2;
						}
					} else if (rule == IR_SHIFT) {
						if (insn->op1 == op_insn->op1
						 && ctx->ir_base[op_insn->op1].op == load_op
						 && ctx->ir_base[op_insn->op1].op2 == insn->op2
						 && ctx->use_lists[op_insn->op1].count == 2) {
							/* l = LOAD(_, a) ... v = SHIFT(l, _) ... STORE(l, a, v) => SKIP ... SKIP_SHIFT ... MEM_SHIFT */
							ctx->rules[insn->op3] = IR_FUSED | IR_SHIFT;
							ctx->rules[op_insn->op1] = IR_SKIPPED | load_op;
							if (ctx->ir_base[op_insn->op2].op == IR_LOAD) {
								ir_match_fuse_addr(ctx, ctx->ir_base[op_insn->op2].op2);
								ctx->rules[op_insn->op2] = IR_LOAD_INT;
							}
							return IR_MEM_SHIFT;
						}
					} else if (rule == IR_SHIFT_CONST) {
						if (insn->op1 == op_insn->op1
						 && ctx->ir_base[op_insn->op1].op == load_op
						 && ctx->ir_base[op_insn->op1].op2 == insn->op2
						 && ctx->use_lists[op_insn->op1].count == 2) {
							/* l = LOAD(_, a) ... v = SHIFT(l, CONST) ... STORE(l, a, v) => SKIP ... SKIP ... MEM_SHIFT_CONST */
							ctx->rules[insn->op3] = IR_SKIPPED | IR_SHIFT_CONST;
							ctx->rules[op_insn->op1] = IR_SKIPPED | load_op;
							return IR_MEM_SHIFT_CONST;
						}
					} else if (rule == IR_OP_INT && op_insn->op != IR_BSWAP) {
						if (insn->op1 == op_insn->op1
						 && ctx->ir_base[op_insn->op1].op == load_op
						 && ctx->ir_base[op_insn->op1].op2 == insn->op2
						 && ctx->use_lists[op_insn->op1].count == 2) {
							/* l = LOAD(_, a) ... v = OP(l) ... STORE(l, a, v) => SKIP ... SKIP ... MEM_OP */
							ctx->rules[insn->op3] = IR_SKIPPED | IR_OP_INT;
							ctx->rules[op_insn->op1] = IR_SKIPPED | load_op;
							return IR_MEM_OP_INT;
						}
					}
				}
				return store_rule;
			} else {
				return IR_VSTORE_FP;
			}
			break;
		case IR_LOAD:
			ir_match_fuse_addr(ctx, insn->op2);
			if (IR_IS_TYPE_INT(insn->type)) {
				return IR_LOAD_INT;
			} else {
				return IR_LOAD_FP;
			}
			break;
		case IR_STORE:
			ir_match_fuse_addr(ctx, insn->op2);
			if (IR_IS_TYPE_INT(ctx->ir_base[insn->op3].type)) {
				store_rule = IR_STORE_INT;
				load_op = IR_LOAD;
				goto store_int;
			} else {
				return IR_STORE_FP;
			}
			break;
		case IR_RLOAD:
			if (IR_REGSET_IN(IR_REGSET_UNION((ir_regset)ctx->fixed_regset, IR_REGSET_FIXED), insn->op2)) {
				return IR_SKIPPED | IR_RLOAD;
			}
			return IR_RLOAD;
		case IR_RSTORE:
			if (IR_IS_TYPE_INT(ctx->ir_base[insn->op2].type)) {
				if ((ctx->flags & IR_OPT_CODEGEN)
				 && ir_in_same_block(ctx, insn->op2)
				 && ctx->use_lists[insn->op2].count == 1
				 && IR_IS_TYPE_INT(ctx->ir_base[insn->op2].type)) {
					ir_insn *op_insn = &ctx->ir_base[insn->op2];

					if (op_insn->op == IR_ADD ||
				        op_insn->op == IR_SUB ||
//				        op_insn->op == IR_MUL ||
				        op_insn->op == IR_OR  ||
				        op_insn->op == IR_AND ||
				        op_insn->op == IR_XOR) {
						if (insn->op1 == op_insn->op1
						 && ctx->ir_base[op_insn->op1].op == IR_RLOAD
						 && ctx->ir_base[op_insn->op1].op2 == insn->op3
						 && ctx->use_lists[op_insn->op1].count == 2) {
							/* l = RLOAD(r) ... v = BINOP(l, _) ... RSTORE(l, r, v) => SKIP ... SKIP_REG_BINOP ... REG_BINOP */
							ctx->rules[insn->op2] = IR_FUSED | IR_BINOP_INT;
							ctx->rules[op_insn->op1] = IR_SKIPPED | IR_RLOAD;
							return IR_REG_BINOP_INT;
						} else if ((ir_op_flags[op_insn->op] & IR_OP_FLAG_COMMUTATIVE)
						 && insn->op1 == op_insn->op2
						 && ctx->ir_base[op_insn->op2].op == IR_RLOAD
						 && ctx->ir_base[op_insn->op2].op2 == insn->op3
						 && ctx->use_lists[op_insn->op2].count == 2) {
							/* l = RLOAD(r) ... v = BINOP(x, l) ... RSTORE(l, r, v) => SKIP ... SKIP_REG_BINOP ... REG_BINOP */
							ir_swap_ops(op_insn);
							ctx->rules[insn->op2] = IR_FUSED | IR_BINOP_INT;
							ctx->rules[op_insn->op1] = IR_SKIPPED | IR_RLOAD;
							return IR_REG_BINOP_INT;
						}
					}
				}
			}
			if (ir_in_same_block(ctx, insn->op2)) {
				ir_match_fuse_load(ctx, insn->op2, ref);
			}
			return IR_RSTORE;
		case IR_START:
		case IR_BEGIN:
		case IR_IF_TRUE:
		case IR_IF_FALSE:
		case IR_CASE_VAL:
		case IR_CASE_DEFAULT:
		case IR_MERGE:
		case IR_LOOP_BEGIN:
		case IR_UNREACHABLE:
			return IR_SKIPPED | insn->op;
		case IR_RETURN:
			if (!insn->op2) {
				return IR_RETURN_VOID;
			} else if (IR_IS_TYPE_INT(ctx->ir_base[insn->op2].type)) {
				return IR_RETURN_INT;
			} else {
				return IR_RETURN_FP;
			}
		case IR_IF:
			if (ir_in_same_block(ctx, insn->op2) && ctx->use_lists[insn->op2].count == 1) {
				op2_insn = &ctx->ir_base[insn->op2];
				if (op2_insn->op >= IR_EQ && op2_insn->op <= IR_UGT) {
					if (IR_IS_TYPE_INT(ctx->ir_base[op2_insn->op1].type)) {
						if (IR_IS_CONST_REF(op2_insn->op2)
						 && ctx->ir_base[op2_insn->op2].val.i64 == 0
						 && op2_insn->op1 == insn->op2 - 1) { /* previous instruction */
							ir_insn *op1_insn = &ctx->ir_base[op2_insn->op1];

							if (op1_insn->op == IR_ADD ||
						        op1_insn->op == IR_SUB ||
//						        op1_insn->op == IR_MUL ||
						        op1_insn->op == IR_OR  ||
						        op1_insn->op == IR_AND ||
						        op1_insn->op == IR_XOR) {

								if (ir_op_flags[op1_insn->op] & IR_OP_FLAG_COMMUTATIVE) {
									ir_match_fuse_load_commutative_int(ctx, op1_insn, ref);
								} else {
									ir_match_fuse_load(ctx, op1_insn->op2, ref);
								}
								if (op1_insn->op == IR_AND && ctx->use_lists[op2_insn->op1].count == 1) {
									/* v = AND(_, _); c = CMP(v, 0) ... IF(c) => SKIP_TEST; SKIP ... TEST_AND_BRANCH */
									if (IR_IS_CONST_REF(op1_insn->op2)) {
										ir_match_fuse_load(ctx, op1_insn->op1, ref);
									}
									ctx->rules[op2_insn->op1] = IR_FUSED | IR_TEST_INT;
									ctx->rules[insn->op2] = IR_FUSED | IR_SIMPLE | IR_NOP;
									return IR_TEST_AND_BRANCH_INT;
								} else {
									/* v = BINOP(_, _); c = CMP(v, 0) ... IF(c) => BINOP; SKIP_CMP ... JCC */
									ctx->rules[op2_insn->op1] = IR_BINOP_INT;
									ctx->rules[insn->op2] = IR_FUSED | IR_CMP_INT;
									return IR_JCC_INT;
								}
							}
						}
						/* c = CMP(_, _) ... IF(c) => SKIP_CMP ... CMP_AND_BRANCH */
						ir_match_fuse_load_cmp_int(ctx, op2_insn, ref);
						ctx->rules[insn->op2] = IR_FUSED | IR_CMP_INT;
						return IR_CMP_AND_BRANCH_INT;
					} else {
						/* c = CMP(_, _) ... IF(c) => SKIP_CMP ... CMP_AND_BRANCH */
						ir_match_fuse_load_cmp_fp(ctx, op2_insn, ref, 1);
						ctx->rules[insn->op2] = IR_FUSED | IR_CMP_FP;
						return IR_CMP_AND_BRANCH_FP;
					}
				} else if (op2_insn->op == IR_AND) {
					/* c = AND(_, _) ... IF(c) => SKIP_TEST ... TEST_AND_BRANCH */
					ir_match_fuse_load_commutative_int(ctx, op2_insn, ref);
					if (IR_IS_CONST_REF(op2_insn->op2)) {
						ir_match_fuse_load(ctx, op2_insn->op1, ref);
					}
					ctx->rules[insn->op2] = IR_FUSED | IR_TEST_INT;
					return IR_TEST_AND_BRANCH_INT;
				} else if (op2_insn->op == IR_OVERFLOW) {
					/* c = OVERFLOW(_) ... IF(c) => SKIP_OVERFLOW ... OVERFLOW_AND_BRANCH */
					ctx->rules[insn->op2] = IR_FUSED | IR_SIMPLE | IR_OVERFLOW;
					return IR_OVERFLOW_AND_BRANCH;
				}
			}
			if (IR_IS_TYPE_INT(ctx->ir_base[insn->op2].type)) {
				if (insn->op2 == ref - 1 /* previous instruction */
				 && ir_in_same_block(ctx, insn->op2)) {
					op2_insn = &ctx->ir_base[insn->op2];
					if (op2_insn->op == IR_ADD ||
					    op2_insn->op == IR_SUB ||
//					    op2_insn->op == IR_MUL ||
					    op2_insn->op == IR_OR  ||
					    op2_insn->op == IR_AND ||
					    op2_insn->op == IR_XOR) {

							/* v = BINOP(_, _); IF(v) => BINOP; JCC */
						if (ir_op_flags[op2_insn->op] & IR_OP_FLAG_COMMUTATIVE) {
							ir_match_fuse_load_commutative_int(ctx, op2_insn, ref);
						} else {
							ir_match_fuse_load(ctx, op2_insn->op2, ref);
						}
						ctx->rules[insn->op2] = IR_BINOP_INT;
						return IR_JCC_INT;
					}
				} else if ((ctx->flags & IR_OPT_CODEGEN)
				 && insn->op1 == ref - 1 /* previous instruction */
				 && insn->op2 == ref - 2 /* previous instruction */
				 && ir_in_same_block(ctx, insn->op2)
				 && ctx->use_lists[insn->op2].count == 2
				 && IR_IS_TYPE_INT(ctx->ir_base[insn->op2].type)) {
					ir_insn *store_insn = &ctx->ir_base[insn->op1];

					if (store_insn->op == IR_STORE && store_insn->op3 == insn->op2) {
						ir_insn *op_insn = &ctx->ir_base[insn->op2];

						if (op_insn->op == IR_ADD ||
						    op_insn->op == IR_SUB ||
//						    op_insn->op == IR_MUL ||
						    op_insn->op == IR_OR  ||
						    op_insn->op == IR_AND ||
						    op_insn->op == IR_XOR) {
							if (ctx->ir_base[op_insn->op1].op == IR_LOAD
							 && ctx->ir_base[op_insn->op1].op2 == store_insn->op2) {
								if (ir_in_same_block(ctx, op_insn->op1)
								 && ctx->use_lists[op_insn->op1].count == 2
								 && store_insn->op1 == op_insn->op1) {
									/* v = MEM_BINOP(_, _); IF(v) => MEM_BINOP; JCC */
									ctx->rules[insn->op2] = IR_FUSED | IR_BINOP_INT;
									ctx->rules[op_insn->op1] = IR_SKIPPED | IR_LOAD;
									ir_match_fuse_addr(ctx, store_insn->op2);
									ctx->rules[insn->op1] = IR_MEM_BINOP_INT;
									return IR_JCC_INT;
								}
							} else if ((ir_op_flags[op_insn->op] & IR_OP_FLAG_COMMUTATIVE)
							 && ctx->ir_base[op_insn->op2].op == IR_LOAD
							 && ctx->ir_base[op_insn->op2].op2 == store_insn->op2) {
								if (ir_in_same_block(ctx, op_insn->op2)
								 && ctx->use_lists[op_insn->op2].count == 2
								 && store_insn->op1 == op_insn->op2) {
									/* v = MEM_BINOP(_, _); IF(v) => MEM_BINOP; JCC */
									ir_swap_ops(op_insn);
									ctx->rules[insn->op2] = IR_FUSED | IR_BINOP_INT;
									ctx->rules[op_insn->op1] = IR_SKIPPED | IR_LOAD;
									ir_match_fuse_addr(ctx, store_insn->op2);
									ctx->rules[insn->op1] = IR_MEM_BINOP_INT;
									return IR_JCC_INT;
								}
							}
						}
					}
				}
				ir_match_fuse_load(ctx, insn->op2, ref);
				return IR_IF_INT;
			} else {
				IR_ASSERT(0 && "NIY IR_IF_FP");
				break;
			}
		case IR_GUARD:
		case IR_GUARD_NOT:
			if (ir_in_same_block(ctx, insn->op2) && ctx->use_lists[insn->op2].count == 1) {
				op2_insn = &ctx->ir_base[insn->op2];
				if (op2_insn->op >= IR_EQ && op2_insn->op <= IR_UGT
					// TODO: register allocator may clobber operands of CMP before they are used in the GUARD_CMP
				 && (insn->op2 == ref - 1 ||
				     (insn->op2 == ctx->prev_ref[ref] - 1
				   && ctx->ir_base[ctx->prev_ref[ref]].op == IR_SNAPSHOT))) {
					if (IR_IS_TYPE_INT(ctx->ir_base[op2_insn->op1].type)) {
						if (IR_IS_CONST_REF(op2_insn->op2)
						 && ctx->ir_base[op2_insn->op2].val.i64 == 0) {
							if (op2_insn->op1 == insn->op2 - 1) { /* previous instruction */
								ir_insn *op1_insn = &ctx->ir_base[op2_insn->op1];

								if (op1_insn->op == IR_ADD ||
							        op1_insn->op == IR_SUB ||
//							        op1_insn->op == IR_MUL ||
							        op1_insn->op == IR_OR  ||
							        op1_insn->op == IR_AND ||
							        op1_insn->op == IR_XOR) {

									if (ir_op_flags[op1_insn->op] & IR_OP_FLAG_COMMUTATIVE) {
										ir_match_fuse_load_commutative_int(ctx, op1_insn, ref);
									} else {
										ir_match_fuse_load(ctx, op1_insn->op2, ref);
									}
									/* v = BINOP(_, _); c = CMP(v, 0) ... IF(c) => BINOP; SKIP_CMP ... GUARD_JCC */
									ctx->rules[op2_insn->op1] = IR_BINOP_INT;
									ctx->rules[insn->op2] = IR_FUSED | IR_CMP_INT;
									return IR_GUARD_JCC_INT;
								}
							} else if ((ctx->flags & IR_OPT_CODEGEN)
							 && op2_insn->op1 == insn->op2 - 2 /* before previous instruction */
							 && ir_in_same_block(ctx, op2_insn->op1)
							 && ctx->use_lists[op2_insn->op1].count == 2) {
								ir_insn *store_insn = &ctx->ir_base[insn->op2 - 1];

								if (store_insn->op == IR_STORE && store_insn->op3 == op2_insn->op1) {
									ir_insn *op_insn = &ctx->ir_base[op2_insn->op1];

									if (op_insn->op == IR_ADD ||
									    op_insn->op == IR_SUB ||
//									    op_insn->op == IR_MUL ||
									    op_insn->op == IR_OR  ||
									    op_insn->op == IR_AND ||
									    op_insn->op == IR_XOR) {
										if (ctx->ir_base[op_insn->op1].op == IR_LOAD
										 && ctx->ir_base[op_insn->op1].op2 == store_insn->op2) {
											if (ir_in_same_block(ctx, op_insn->op1)
											 && ctx->use_lists[op_insn->op1].count == 2
											 && store_insn->op1 == op_insn->op1) {
												/* v = MEM_BINOP(_, _); IF(v) => MEM_BINOP; GUARD_JCC */
												ctx->rules[op2_insn->op1] = IR_FUSED | IR_BINOP_INT;
												ctx->rules[op_insn->op1] = IR_SKIPPED | IR_LOAD;
												ir_match_fuse_addr(ctx, store_insn->op2);
												ctx->rules[insn->op2 - 1] = IR_MEM_BINOP_INT;
												ctx->rules[insn->op2] = IR_SKIPPED | IR_NOP;
												return IR_GUARD_JCC_INT;
											}
										} else if ((ir_op_flags[op_insn->op] & IR_OP_FLAG_COMMUTATIVE)
										 && ctx->ir_base[op_insn->op2].op == IR_LOAD
										 && ctx->ir_base[op_insn->op2].op2 == store_insn->op2) {
											if (ir_in_same_block(ctx, op_insn->op2)
											 && ctx->use_lists[op_insn->op2].count == 2
											 && store_insn->op1 == op_insn->op2) {
												/* v = MEM_BINOP(_, _); IF(v) => MEM_BINOP; JCC */
												ir_swap_ops(op_insn);
												ctx->rules[op2_insn->op1] = IR_FUSED | IR_BINOP_INT;
												ctx->rules[op_insn->op1] = IR_SKIPPED | IR_LOAD;
												ir_match_fuse_addr(ctx, store_insn->op2);
												ctx->rules[insn->op2 - 1] = IR_MEM_BINOP_INT;
												ctx->rules[insn->op2] = IR_SKIPPED | IR_NOP;
												return IR_GUARD_JCC_INT;
											}
										}
									}
								}
							}
						}
						/* c = CMP(_, _) ... GUARD(c) => SKIP_CMP ... GUARD_CMP */
						ir_match_fuse_load_cmp_int(ctx, op2_insn, ref);
						ctx->rules[insn->op2] = IR_FUSED | IR_CMP_INT;
						return IR_GUARD_CMP_INT;
					} else {
						/* c = CMP(_, _) ... GUARD(c) => SKIP_CMP ... GUARD_CMP */
						ir_match_fuse_load_cmp_fp(ctx, op2_insn, ref, insn->op == IR_GUARD_NOT);
						ctx->rules[insn->op2] = IR_FUSED | IR_CMP_FP;
						return IR_GUARD_CMP_FP;
					}
				} else if (op2_insn->op == IR_AND) { // TODO: OR, XOR. etc
					/* c = AND(_, _) ... GUARD(c) => SKIP_TEST ... GUARD_TEST */
					ir_match_fuse_load_commutative_int(ctx, op2_insn, ref);
					if (IR_IS_CONST_REF(op2_insn->op2)) {
						ir_match_fuse_load(ctx, op2_insn->op1, ref);
					}
					ctx->rules[insn->op2] = IR_FUSED | IR_TEST_INT;
					return IR_GUARD_TEST_INT;
				} else if (op2_insn->op == IR_OVERFLOW) {
					/* c = OVERFLOW(_) ... GUARD(c) => SKIP_OVERFLOW ... GUARD_OVERFLOW */
					ctx->rules[insn->op2] = IR_FUSED | IR_SIMPLE | IR_OVERFLOW;
					return IR_GUARD_OVERFLOW;
				}
			}
			ir_match_fuse_load(ctx, insn->op2, ref);
			return insn->op;
		case IR_IJMP:
			if (ir_in_same_block(ctx, insn->op2)) {
				ir_match_fuse_load(ctx, insn->op2, ref);
			}
			return insn->op;
		case IR_SEXT:
		case IR_ZEXT:
		case IR_BITCAST:
		case IR_INT2FP:
		case IR_FP2INT:
		case IR_FP2FP:
			ir_match_fuse_load(ctx, insn->op1, ref);
			return insn->op;
		case IR_CTLZ:
		case IR_CTTZ:
			ir_match_fuse_load(ctx, insn->op1, ref);
			return IR_BIT_COUNT;
		case IR_CTPOP:
			ir_match_fuse_load(ctx, insn->op1, ref);
			return (ctx->mflags & IR_X86_BMI1) ? IR_BIT_COUNT : IR_CTPOP;
		default:
			break;
	}

	return insn->op;
}

static void ir_match_insn2(ir_ctx *ctx, ir_ref ref, uint32_t rule)
{
	if (rule == IR_LEA_IB) {
		ir_insn *insn = &ctx->ir_base[ref];

		if (insn->op1 == insn->op2) {
			/* pass */
		} else if (ir_match_fuse_load(ctx, insn->op2, ref) ||
		    (ctx->ir_base[insn->op2].op == IR_PARAM
		  && ctx->use_lists[insn->op2].count == 1
		  && ir_get_param_reg(ctx, insn->op2) == IR_REG_NONE)) {
			ctx->rules[ref] = IR_BINOP_INT;
		} else if (ir_match_fuse_load(ctx, insn->op1, ref) ||
		    (ctx->ir_base[insn->op1].op == IR_PARAM
		  && ctx->use_lists[insn->op1].count == 1
		  && ir_get_param_reg(ctx, insn->op1) == IR_REG_NONE)) {
			/* swap for better load fusion */
			ir_swap_ops(insn);
			ctx->rules[ref] = IR_BINOP_INT;
		}
	}
}

/* code generation */
static int32_t ir_ref_spill_slot(ir_ctx *ctx, ir_ref ref, ir_reg *reg)
{
	int32_t offset;

	IR_ASSERT(ref >= 0 && ctx->vregs[ref] && ctx->live_intervals[ctx->vregs[ref]]);
	offset = ctx->live_intervals[ctx->vregs[ref]]->stack_spill_pos;
	IR_ASSERT(offset != -1);
	if (ctx->live_intervals[ctx->vregs[ref]]->flags & IR_LIVE_INTERVAL_SPILL_SPECIAL) {
		IR_ASSERT(ctx->spill_base != IR_REG_NONE);
		*reg = ctx->spill_base;
		return offset;
	}
	*reg = (ctx->flags & IR_USE_FRAME_POINTER) ? IR_REG_FRAME_POINTER : IR_REG_STACK_POINTER;
	return IR_SPILL_POS_TO_OFFSET(offset);
}

static bool ir_is_same_spill_slot(ir_ctx *ctx, ir_ref ref, ir_reg reg, int32_t offset)
{
	ir_reg fp;

	return ir_ref_spill_slot(ctx, ref, &fp) == offset && reg == fp;
}

static int32_t ir_var_spill_slot(ir_ctx *ctx, ir_ref ref, ir_reg *reg)
{
	ir_insn *var_insn = &ctx->ir_base[ref];

	IR_ASSERT(var_insn->op == IR_VAR);
	*reg = (ctx->flags & IR_USE_FRAME_POINTER) ? IR_REG_FRAME_POINTER : IR_REG_STACK_POINTER;
	return IR_SPILL_POS_TO_OFFSET(var_insn->op3);
}

static bool ir_may_avoid_spill_load(ir_ctx *ctx, ir_ref ref, ir_ref use)
{
	ir_live_interval *ival;

	IR_ASSERT(ctx->vregs[ref] && ctx->live_intervals[ctx->vregs[ref]]);
	ival = ctx->live_intervals[ctx->vregs[ref]];
	while (ival) {
		ir_use_pos *use_pos = ival->use_pos;
		while (use_pos) {
			if (IR_LIVE_POS_TO_REF(use_pos->pos) == use) {
				return !use_pos->next || use_pos->next->op_num == 0;
			}
			use_pos = use_pos->next;
		}
		ival = ival->next;
	}
	return 0;
}

static void ir_emit_load_imm_int(ir_ctx *ctx, ir_type type, ir_reg reg, int64_t val)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	IR_ASSERT(IR_IS_TYPE_INT(type));
	if (val == 0) {
		|	ASM_REG_REG_OP xor, type, reg, reg
	} else if (ir_type_size[type] == 8) {
		IR_ASSERT(sizeof(void*) == 8);
|.if X64
		if (IR_IS_UNSIGNED_32BIT(val)) {
			|	mov Rd(reg), (uint32_t)val // zero extended load
		} else if (IR_IS_SIGNED_32BIT(val)) {
			|	mov Rq(reg), (int32_t)val // sign extended load
		} else {
			|	mov64 Ra(reg), val
		}
|.endif
	} else {
		|	ASM_REG_IMM_OP mov, type, reg, (int32_t)val // sign extended load
	}
}

static void ir_emit_load_mem_int(ir_ctx *ctx, ir_type type, ir_reg reg, ir_reg base_reg, int32_t offset)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	if (base_reg != IR_REG_NONE) {
		|	ASM_REG_MEM_OP mov, type, reg, [Ra(base_reg)+offset]
	} else {
		|	ASM_REG_MEM_OP mov, type, reg, [offset]
	}
}

static void ir_emit_load_imm_fp(ir_ctx *ctx, ir_type type, ir_reg reg, ir_ref src)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_insn *insn = &ctx->ir_base[src];
	int label;

	if (type == IR_FLOAT && insn->val.u32 == 0) {
		if (ctx->mflags & IR_X86_AVX) {
			|	vxorps xmm(reg-IR_REG_FP_FIRST), xmm(reg-IR_REG_FP_FIRST), xmm(reg-IR_REG_FP_FIRST)
		} else {
			|	xorps xmm(reg-IR_REG_FP_FIRST), xmm(reg-IR_REG_FP_FIRST)
		}
	} else if (type == IR_DOUBLE && insn->val.u64 == 0) {
		if (ctx->mflags & IR_X86_AVX) {
			|	vxorpd xmm(reg-IR_REG_FP_FIRST), xmm(reg-IR_REG_FP_FIRST), xmm(reg-IR_REG_FP_FIRST)
		} else {
			|	xorpd xmm(reg-IR_REG_FP_FIRST), xmm(reg-IR_REG_FP_FIRST)
		}
	} else {
		label = ctx->cfg_blocks_count - src;
		insn->const_flags |= IR_CONST_EMIT;
		|	ASM_FP_REG_MEM_OP movss, movsd, vmovss, vmovsd, type, reg, [=>label]
	}
}

static void ir_emit_load_mem_fp(ir_ctx *ctx, ir_type type, ir_reg reg, ir_reg base_reg, int32_t offset)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	if (base_reg != IR_REG_NONE) {
		|	ASM_FP_REG_MEM_OP movss, movsd, vmovss, vmovsd, type, reg, [Ra(base_reg)+offset]
	} else {
		|	ASM_FP_REG_MEM_OP movss, movsd, vmovss, vmovsd, type, reg, [Ra(base_reg)+offset]
	}
}

static void ir_emit_load(ir_ctx *ctx, ir_type type, ir_reg reg, ir_ref src)
{
	int32_t offset;
	ir_reg fp;

	if (IR_IS_CONST_REF(src)) {
		if (IR_IS_TYPE_INT(type)) {
			ir_insn *insn = &ctx->ir_base[src];

			IR_ASSERT(insn->op != IR_STR && insn->op != IR_SYM && insn->op != IR_FUNC);
			ir_emit_load_imm_int(ctx, type, reg, insn->val.i64);
		} else {
			ir_emit_load_imm_fp(ctx, type, reg, src);
		}
	} else {
		offset = ir_ref_spill_slot(ctx, src, &fp);
		if (IR_IS_TYPE_INT(type)) {
			ir_emit_load_mem_int(ctx, type, reg, fp, offset);
		} else {
			ir_emit_load_mem_fp(ctx, type, reg, fp, offset);
		}
	}
}

static void ir_emit_store_mem_int(ir_ctx *ctx, ir_type type, ir_reg base_reg, int32_t offset, ir_reg reg)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	|	ASM_MEM_REG_OP mov, type, [Ra(base_reg)+offset], reg
}

static void ir_emit_store_mem_fp(ir_ctx *ctx, ir_type type, ir_reg base_reg, int32_t offset, ir_reg reg)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	|	ASM_FP_MEM_REG_OP movss, movsd, vmovss, vmovsd, type, [Ra(base_reg)+offset], reg
}

static void ir_emit_store(ir_ctx *ctx, ir_type type, ir_ref dst, ir_reg reg)
{
	int32_t offset;
	ir_reg fp;

	IR_ASSERT(dst >= 0);
	offset = ir_ref_spill_slot(ctx, dst, &fp);
	if (IR_IS_TYPE_INT(type)) {
		ir_emit_store_mem_int(ctx, type, fp, offset, reg);
	} else {
		ir_emit_store_mem_fp(ctx, type, fp, offset, reg);
	}
}

static void ir_emit_store_imm(ir_ctx *ctx, ir_type type, ir_ref dst, int32_t imm)
{
	int32_t offset;
	ir_reg fp;
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	IR_ASSERT(dst >= 0);
	IR_ASSERT(IR_IS_TYPE_INT(type));
	offset = ir_ref_spill_slot(ctx, dst, &fp);

	|	ASM_MEM_IMM_OP mov, type, [Ra(fp)+offset], imm
}

static void ir_emit_mov(ir_ctx *ctx, ir_type type, ir_reg dst, ir_reg src)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	|	ASM_REG_REG_OP mov, type, dst, src
}

static void ir_emit_fp_mov(ir_ctx *ctx, ir_type type, ir_reg dst, ir_reg src)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	|	ASM_FP_REG_REG_OP movaps, movapd, vmovaps, vmovapd, type, dst, src
}

static int32_t ir_fuse_addr(ir_ctx *ctx, ir_ref ref, ir_reg *preg)
{
	ir_insn *addr_insn = &ctx->ir_base[ref];
	ir_reg reg;

	IR_ASSERT(addr_insn->op == IR_ADD);
	IR_ASSERT(!IR_IS_CONST_REF(addr_insn->op1) && IR_IS_CONST_REF(addr_insn->op2));
	reg = ctx->regs[ref][1];
	if (IR_REG_SPILLED(reg)) {
		reg = IR_REG_NUM(reg);
		ir_emit_load(ctx, IR_ADDR, reg, addr_insn->op1);
	}
	*preg = reg;
	return ctx->ir_base[addr_insn->op2].val.i32;
}

static int32_t ir_fuse_load(ir_ctx *ctx, ir_ref ref, ir_reg *preg)
{
	ir_insn *load_insn = &ctx->ir_base[ref];
	ir_reg reg = ctx->regs[ref][2];

	IR_ASSERT(load_insn->op == IR_LOAD);
	if (IR_IS_CONST_REF(load_insn->op2)) {
		*preg = reg;
		if (reg == IR_REG_NONE) {
			ir_insn *addr_insn = &ctx->ir_base[load_insn->op2];

			IR_ASSERT(addr_insn->op == IR_C_ADDR);
			IR_ASSERT(sizeof(void*) == 4 || IR_IS_SIGNED_32BIT(addr_insn->val.i64));
			return addr_insn->val.i32;
		} else {
			ir_emit_load(ctx, IR_ADDR, reg, load_insn->op2);
			return 0;
		}
	} else if (reg == IR_REG_NONE) {
		return ir_fuse_addr(ctx, load_insn->op2, preg);
	} else {
		if (IR_REG_SPILLED(reg)) {
			reg = IR_REG_NUM(reg);
			ir_emit_load(ctx, IR_ADDR, reg, load_insn->op2);
		}
		*preg = reg;
		return 0;
	}
}

static void ir_emit_prologue(ir_ctx *ctx)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	if (ctx->flags & IR_USE_FRAME_POINTER) {
		|	push Ra(IR_REG_RBP)
		|	mov Ra(IR_REG_RBP), Ra(IR_REG_RSP)
	}
	if (ctx->stack_frame_size + ctx->call_stack_size) {
		if (ctx->fixed_stack_red_zone) {
			IR_ASSERT(ctx->stack_frame_size + ctx->call_stack_size <= ctx->fixed_stack_red_zone);
		} else {
			|	sub Ra(IR_REG_RSP), (ctx->stack_frame_size + ctx->call_stack_size)
		}
	}
	if (ctx->used_preserved_regs) {
		int offset;
		uint32_t i;
		ir_regset used_preserved_regs = (ir_regset)ctx->used_preserved_regs;

		if (ctx->flags & IR_USE_FRAME_POINTER) {
			offset = 0;
		} else {
			offset = ctx->stack_frame_size + ctx->call_stack_size;
		}
		for (i = 0; i < IR_REG_NUM; i++) {
			if (IR_REGSET_IN(used_preserved_regs, i)) {
				if (i < IR_REG_FP_FIRST) {
					ir_reg fp = (ctx->flags & IR_USE_FRAME_POINTER) ? IR_REG_FRAME_POINTER : IR_REG_STACK_POINTER;

					offset -= sizeof(void*);
					|	mov aword [Ra(fp)+offset], Ra(i)
				} else {
					ir_reg fp = (ctx->flags & IR_USE_FRAME_POINTER) ? IR_REG_FRAME_POINTER : IR_REG_STACK_POINTER;

					offset -= sizeof(void*);
					if (ctx->mflags & IR_X86_AVX) {
						|	vmovsd qword [Ra(fp)+offset], xmm(i-IR_REG_FP_FIRST)
					} else {
						|	movsd qword [Ra(fp)+offset], xmm(i-IR_REG_FP_FIRST)
					}
				}
			}
		}
	}
}

static void ir_emit_epilogue(ir_ctx *ctx)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	if (ctx->used_preserved_regs) {
		int offset;
		uint32_t i;
		ir_regset used_preserved_regs = (ir_regset)ctx->used_preserved_regs;

		if (ctx->flags & IR_USE_FRAME_POINTER) {
			offset = 0;
		} else {
			offset = ctx->stack_frame_size + ctx->call_stack_size;
		}
		for (i = 0; i < IR_REG_NUM; i++) {
			if (IR_REGSET_IN(used_preserved_regs, i)) {
				if (i < IR_REG_FP_FIRST) {
					ir_reg fp = (ctx->flags & IR_USE_FRAME_POINTER) ? IR_REG_FRAME_POINTER : IR_REG_STACK_POINTER;

					offset -= sizeof(void*);
					|	mov Ra(i), aword [Ra(fp)+offset]
				} else {
					ir_reg fp = (ctx->flags & IR_USE_FRAME_POINTER) ? IR_REG_FRAME_POINTER : IR_REG_STACK_POINTER;

					offset -= sizeof(void*);
					if (ctx->mflags & IR_X86_AVX) {
						|	vmovsd xmm(i-IR_REG_FP_FIRST), qword [Ra(fp)+offset]
					} else {
						|	movsd xmm(i-IR_REG_FP_FIRST), qword [Ra(fp)+offset]
					}
				}
			}
		}
	}

	if (ctx->flags & IR_USE_FRAME_POINTER) {
		|	mov Ra(IR_REG_RSP), Ra(IR_REG_RBP)
		|	pop Ra(IR_REG_RBP)
	} else if (ctx->stack_frame_size + ctx->call_stack_size) {
		if (ctx->fixed_stack_red_zone) {
			IR_ASSERT(ctx->stack_frame_size + ctx->call_stack_size <= ctx->fixed_stack_red_zone);
		} else {
			|	add Ra(IR_REG_RSP), (ctx->stack_frame_size + ctx->call_stack_size)
		}
	}
}

static void ir_emit_binop_int(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = insn->type;
	ir_ref op1 = insn->op1;
	ir_ref op2 = insn->op2;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];
	ir_reg op2_reg = ctx->regs[def][2];

	IR_ASSERT(def_reg != IR_REG_NONE);

	if (op1_reg != IR_REG_NONE && IR_REG_SPILLED(op1_reg)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, op1);
	}
	if (def_reg != op1_reg) {
		if (op1_reg != IR_REG_NONE) {
			ir_emit_mov(ctx, type, def_reg, op1_reg);
		} else {
			ir_emit_load(ctx, type, def_reg, op1);
		}
		if (op1 == op2) {
			op2_reg = def_reg;
		}
	}

	if (op2_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op2_reg) || IR_IS_CONST_REF(op2)) {
			op2_reg = IR_REG_NUM(op2_reg);
			if (op1 != op2) {
				ir_emit_load(ctx, type, op2_reg, op2);
			}
		}
		switch (insn->op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_ADD:
			case IR_ADD_OV:
				|	ASM_REG_REG_OP add, type, def_reg, op2_reg
				break;
			case IR_SUB:
			case IR_SUB_OV:
				|	ASM_REG_REG_OP sub, type, def_reg, op2_reg
				break;
			case IR_MUL:
			case IR_MUL_OV:
				|	ASM_REG_REG_IMUL type, def_reg, op2_reg
				break;
			case IR_OR:
				|	ASM_REG_REG_OP or, type, def_reg, op2_reg
				break;
			case IR_AND:
				|	ASM_REG_REG_OP and, type, def_reg, op2_reg
				break;
			case IR_XOR:
				|	ASM_REG_REG_OP xor, type, def_reg, op2_reg
				break;
		}
	} else if (IR_IS_CONST_REF(op2)) {
		ir_insn *val_insn = &ctx->ir_base[op2];
		int32_t val;

		IR_ASSERT(IR_IS_32BIT(val_insn->type, val_insn->val));
		val = val_insn->val.i32;
		switch (insn->op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_ADD:
			case IR_ADD_OV:
				|	ASM_REG_IMM_OP add, type, def_reg, val
				break;
			case IR_SUB:
			case IR_SUB_OV:
				|	ASM_REG_IMM_OP sub, type, def_reg, val
				break;
			case IR_MUL:
			case IR_MUL_OV:
				|	ASM_REG_IMM_IMUL type, def_reg, val
				break;
			case IR_OR:
				|	ASM_REG_IMM_OP or, type, def_reg, val
				break;
			case IR_AND:
				|	ASM_REG_IMM_OP and, type, def_reg, val
				break;
			case IR_XOR:
				|	ASM_REG_IMM_OP xor, type, def_reg, val
				break;
		}
	} else {
		int32_t offset = 0;

		if (ir_rule(ctx, op2) & IR_FUSED) {
			offset = ir_fuse_load(ctx, op2, &op2_reg);
		} else {
			offset = ir_ref_spill_slot(ctx, op2, &op2_reg);
		}
		if (op2_reg != IR_REG_NONE) {
			switch (insn->op) {
				default:
					IR_ASSERT(0 && "NIY binary op");
				case IR_ADD:
				case IR_ADD_OV:
					|	ASM_REG_MEM_OP add, type, def_reg, [Ra(op2_reg)+offset]
					break;
				case IR_SUB:
				case IR_SUB_OV:
					|	ASM_REG_MEM_OP sub, type, def_reg, [Ra(op2_reg)+offset]
					break;
				case IR_MUL:
				case IR_MUL_OV:
					|	ASM_REG_MEM_IMUL type, def_reg, [Ra(op2_reg)+offset]
					break;
				case IR_OR:
					|	ASM_REG_MEM_OP or, type, def_reg, [Ra(op2_reg)+offset]
					break;
				case IR_AND:
					|	ASM_REG_MEM_OP and, type, def_reg, [Ra(op2_reg)+offset]
					break;
				case IR_XOR:
					|	ASM_REG_MEM_OP xor, type, def_reg, [Ra(op2_reg)+offset]
					break;
			}
		} else {
			switch (insn->op) {
				default:
					IR_ASSERT(0 && "NIY binary op");
				case IR_ADD:
				case IR_ADD_OV:
					|	ASM_REG_MEM_OP add, type, def_reg, [offset]
					break;
				case IR_SUB:
				case IR_SUB_OV:
					|	ASM_REG_MEM_OP sub, type, def_reg, [offset]
					break;
				case IR_MUL:
				case IR_MUL_OV:
					|	ASM_REG_MEM_IMUL type, def_reg, [offset]
					break;
				case IR_OR:
					|	ASM_REG_MEM_OP or, type, def_reg, [offset]
					break;
				case IR_AND:
					|	ASM_REG_MEM_OP and, type, def_reg, [offset]
					break;
				case IR_XOR:
					|	ASM_REG_MEM_OP xor, type, def_reg, [offset]
					break;
			}
		}
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_imul3(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = insn->type;
	ir_ref op1 = insn->op1;
	ir_ref op2 = insn->op2;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];
	ir_insn *val_insn = &ctx->ir_base[op2];
	int32_t val;

	IR_ASSERT(def_reg != IR_REG_NONE);
	IR_ASSERT(!IR_IS_CONST_REF(op1));
	IR_ASSERT(IR_IS_CONST_REF(op2));
	IR_ASSERT(IR_IS_32BIT(val_insn->type, val_insn->val));
	val = val_insn->val.i32;

	if (op1_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op1_reg)) {
			op1_reg = IR_REG_NUM(op1_reg);
			ir_emit_load(ctx, type, op1_reg, op1);
		}
		switch (ir_type_size[type]) {
			default:
				IR_ASSERT(0);
			case 2:
				|	imul Rw(def_reg), Rw(op1_reg), val
				break;
			case 4:
				|	imul Rd(def_reg), Rd(op1_reg), val
				break;
|.if X64
||			case 8:
|				imul Rq(def_reg), Rq(op1_reg), val
||				break;
|.endif
		}
	} else {
		int32_t offset = 0;

		if (ir_rule(ctx, op1) & IR_FUSED) {
			offset = ir_fuse_load(ctx, op1, &op1_reg);
		} else {
			offset = ir_ref_spill_slot(ctx, op1, &op1_reg);
		}
		switch (ir_type_size[type]) {
			default:
				IR_ASSERT(0);
			case 2:
				|	imul Rw(def_reg), word [Ra(op1_reg)+offset], val
				break;
			case 4:
				|	imul Rd(def_reg), dword [Ra(op1_reg)+offset], val
				break;
|.if X64
||			case 8:
|					imul Rq(def_reg), qword [Ra(op1_reg)+offset], val
||	 			break;
|.endif
		}
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_min_max_int(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = insn->type;
	ir_ref op1 = insn->op1;
	ir_ref op2 = insn->op2;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];
	ir_reg op2_reg = ctx->regs[def][2];

	IR_ASSERT(def_reg != IR_REG_NONE && op2_reg != IR_REG_NONE);

	if (op1_reg != IR_REG_NONE && IR_REG_SPILLED(op1_reg)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, op1);
	}
	if (def_reg != op1_reg) {
		if (op1_reg != IR_REG_NONE) {
			ir_emit_mov(ctx, type, def_reg, op1_reg);
		} else {
			ir_emit_load(ctx, type, def_reg, op1);
		}
	}

	if (IR_REG_SPILLED(op2_reg) || IR_IS_CONST_REF(op2)) {
		op2_reg = IR_REG_NUM(op2_reg);
		if (op1 != op2) {
			ir_emit_load(ctx, type, op2_reg, op2);
		}
	}

	if (op1 == op2) {
		return;
	}

	|	ASM_REG_REG_OP cmp, type, def_reg, op2_reg
	if (insn->op == IR_MIN) {
		if (IR_IS_TYPE_SIGNED(type)) {
			|	ASM_REG_REG_OP2 cmovg, type, def_reg, op2_reg
		} else {
			|	ASM_REG_REG_OP2 cmova, type, def_reg, op2_reg
		}
	} else {
		IR_ASSERT(insn->op == IR_MAX);
		if (IR_IS_TYPE_SIGNED(type)) {
			|	ASM_REG_REG_OP2 cmova, type, def_reg, op2_reg
		} else {
			|	ASM_REG_REG_OP2 cmovg, type, def_reg, op2_reg
		}
	}

	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_overflow(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_type type = ctx->ir_base[insn->op1].type;

	IR_ASSERT(def_reg != IR_REG_NONE);
	IR_ASSERT(IR_IS_TYPE_INT(type));
	if (IR_IS_TYPE_SIGNED(type)) {
		|	seto Rb(def_reg)
	} else {
		|	setc Rb(def_reg)
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, insn->type, def, def_reg);
	}
}

static void ir_emit_overflow_and_branch(ir_ctx *ctx, uint32_t b, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_insn *overflow_insn = &ctx->ir_base[insn->op2];
	ir_type type = ctx->ir_base[overflow_insn->op1].type;
	uint32_t true_block, false_block, next_block;
	bool reverse = 0;

	ir_get_true_false_blocks(ctx, b, &true_block, &false_block, &next_block);
	if (true_block == next_block) {
		reverse = 1;
		true_block = false_block;
		false_block = 0;
	} else if (false_block == next_block) {
		false_block = 0;
	}

	if (IR_IS_TYPE_SIGNED(type)) {
		if (reverse) {
			|	jno =>true_block
		} else {
			|	jo =>true_block
		}
	} else {
		if (reverse) {
			|	jnc =>true_block
		} else {
			|	jc =>true_block
		}
	}
	if (false_block) {
		|	jmp =>false_block
	}
}

static void ir_emit_mem_binop_int(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_insn *op_insn = &ctx->ir_base[insn->op3];
	ir_type type = op_insn->type;
	ir_ref op2 = op_insn->op2;
	ir_reg op2_reg = ctx->regs[insn->op3][2];
	ir_reg reg;
	int32_t offset = 0;

	if (insn->op == IR_STORE) {
		reg = ctx->regs[def][2];
		if (IR_IS_CONST_REF(insn->op2)) {
			if (reg == IR_REG_NONE) {
				offset = ctx->ir_base[insn->op2].val.i32;
			} else {
				ir_emit_load(ctx, IR_ADDR, reg, insn->op2);
			}
		} else if (reg == IR_REG_NONE) {
			offset = ir_fuse_addr(ctx, insn->op2, &reg);
		} else if (IR_REG_SPILLED(reg)) {
			reg = IR_REG_NUM(reg);
			ir_emit_load(ctx, IR_ADDR, reg, insn->op2);
		}
	} else {
		IR_ASSERT(insn->op == IR_VSTORE);
		offset = ir_var_spill_slot(ctx, insn->op2, &reg);
	}

	if (op2_reg == IR_REG_NONE) {
		ir_val *val = &ctx->ir_base[op2].val;

		IR_ASSERT(IR_IS_CONST_REF(op2) && (ir_type_size[type] != 8 || IR_IS_32BIT(type, ctx->ir_base[op2].val)));
		switch (op_insn->op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_ADD:
			case IR_ADD_OV:
				if (reg != IR_REG_NONE) {
					|	ASM_MEM_IMM_OP add, type, [Ra(reg)+offset], val->i32
				} else {
					|	ASM_MEM_IMM_OP add, type, [offset], val->i32
				}
				break;
			case IR_SUB:
			case IR_SUB_OV:
				if (reg != IR_REG_NONE) {
					|	ASM_MEM_IMM_OP sub, type, [Ra(reg)+offset], val->i32
				} else {
					|	ASM_MEM_IMM_OP sub, type, [offset], val->i32
				}
				break;
			case IR_OR:
				if (reg != IR_REG_NONE) {
					|	ASM_MEM_IMM_OP or, type, [Ra(reg)+offset], val->i32
				} else {
					|	ASM_MEM_IMM_OP or, type, [offset], val->i32
				}
				break;
			case IR_AND:
				if (reg != IR_REG_NONE) {
					|	ASM_MEM_IMM_OP and, type, [Ra(reg)+offset], val->i32
				} else {
					|	ASM_MEM_IMM_OP and, type, [offset], val->i32
				}
				break;
			case IR_XOR:
				if (reg != IR_REG_NONE) {
					|	ASM_MEM_IMM_OP xor, type, [Ra(reg)+offset], val->i32
				} else {
					|	ASM_MEM_IMM_OP xor, type, [offset], val->i32
				}
				break;
		}
	} else {
		if (IR_REG_SPILLED(op2_reg) || IR_IS_CONST_REF(op2)) {
			op2_reg = IR_REG_NUM(op2_reg);
			ir_emit_load(ctx, type, op2_reg, op2);
		}
		switch (op_insn->op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_ADD:
			case IR_ADD_OV:
				if (reg != IR_REG_NONE) {
					|	ASM_MEM_REG_OP add, type, [Ra(reg)+offset], op2_reg
				} else {
					|	ASM_MEM_REG_OP add, type, [offset], op2_reg
				}
				break;
			case IR_SUB:
			case IR_SUB_OV:
				if (reg != IR_REG_NONE) {
					|	ASM_MEM_REG_OP sub, type, [Ra(reg)+offset], op2_reg
				} else {
					|	ASM_MEM_REG_OP sub, type, [offset], op2_reg
				}
				break;
			case IR_OR:
				if (reg != IR_REG_NONE) {
					|	ASM_MEM_REG_OP or, type, [Ra(reg)+offset], op2_reg
				} else {
					|	ASM_MEM_REG_OP or, type, [offset], op2_reg
				}
				break;
			case IR_AND:
				if (reg != IR_REG_NONE) {
					|	ASM_MEM_REG_OP and, type, [Ra(reg)+offset], op2_reg
				} else {
					|	ASM_MEM_REG_OP and, type, [offset], op2_reg
				}
				break;
			case IR_XOR:
				if (reg != IR_REG_NONE) {
					|	ASM_MEM_REG_OP xor, type, [Ra(reg)+offset], op2_reg
				} else {
					|	ASM_MEM_REG_OP xor, type, [offset], op2_reg
				}
				break;
		}
	}
}

static void ir_emit_reg_binop_int(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_insn *op_insn = &ctx->ir_base[insn->op2];
	ir_type type = op_insn->type;
	ir_ref op2 = op_insn->op2;
	ir_reg op2_reg = ctx->regs[insn->op2][2];
	ir_reg reg;

	IR_ASSERT(insn->op == IR_RSTORE);
	reg = insn->op3;

	if (op2_reg == IR_REG_NONE) {
		ir_val *val = &ctx->ir_base[op2].val;

		IR_ASSERT(IR_IS_CONST_REF(op2) && (ir_type_size[type] != 8 || IR_IS_32BIT(type, ctx->ir_base[op2].val)));
		switch (op_insn->op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_ADD:
				|	ASM_REG_IMM_OP add, type, reg, val->i32
				break;
			case IR_SUB:
				|	ASM_REG_IMM_OP sub, type, reg, val->i32
				break;
			case IR_OR:
				|	ASM_REG_IMM_OP or, type, reg, val->i32
				break;
			case IR_AND:
				|	ASM_REG_IMM_OP and, type, reg, val->i32
				break;
			case IR_XOR:
				|	ASM_REG_IMM_OP xor, type, reg, val->i32
				break;
		}
	} else {
		if (IR_REG_SPILLED(op2_reg) || IR_IS_CONST_REF(op2)) {
			op2_reg = IR_REG_NUM(op2_reg);
			ir_emit_load(ctx, type, op2_reg, op2);
		}
		switch (op_insn->op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_ADD:
				|	ASM_REG_REG_OP add, type, reg, op2_reg
				break;
			case IR_SUB:
				|	ASM_REG_REG_OP sub, type, reg, op2_reg
				break;
			case IR_OR:
				|	ASM_REG_REG_OP or, type, reg, op2_reg
				break;
			case IR_AND:
				|	ASM_REG_REG_OP and, type, reg, op2_reg
				break;
			case IR_XOR:
				|	ASM_REG_REG_OP xor, type, reg, op2_reg
				break;
		}
	}
}

static void ir_emit_mul_div_mod_pwr2(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = insn->type;
	ir_ref op1 = insn->op1;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];

	IR_ASSERT(def_reg != IR_REG_NONE);

	if (op1_reg != IR_REG_NONE && IR_REG_SPILLED(op1_reg)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, op1);
	}
	if (def_reg != op1_reg) {
		if (op1_reg != IR_REG_NONE) {
			ir_emit_mov(ctx, type, def_reg, op1_reg);
		} else {
			ir_emit_load(ctx, type, def_reg, op1);
		}
	}
	if (insn->op == IR_MUL) {
		uint32_t shift = IR_LOG2(ctx->ir_base[insn->op2].val.u64);
		if (shift == 1) {
			|	ASM_REG_REG_OP add, insn->type, def_reg, def_reg
		} else {
			|	ASM_REG_IMM_OP shl, insn->type, def_reg, shift
		}
	} else if (insn->op == IR_DIV) {
		uint32_t shift = IR_LOG2(ctx->ir_base[insn->op2].val.u64);
		|	ASM_REG_IMM_OP shr, insn->type, def_reg, shift
	} else {
		IR_ASSERT(insn->op == IR_MOD);
		uint64_t mask = ctx->ir_base[insn->op2].val.u64 - 1;
		IR_ASSERT(IR_IS_UNSIGNED_32BIT(mask));
		|	ASM_REG_IMM_OP and, insn->type, def_reg, mask
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_mem_mul_div_mod_pwr2(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_insn *op_insn = &ctx->ir_base[insn->op3];
	ir_type type = op_insn->type;
	ir_reg reg;
	int32_t offset = 0;

	if (insn->op == IR_STORE) {
		reg = ctx->regs[def][2];
		if (IR_IS_CONST_REF(insn->op2)) {
			if (reg == IR_REG_NONE) {
				offset = ctx->ir_base[insn->op2].val.i32;
			} else {
				ir_emit_load(ctx, IR_ADDR, reg, insn->op2);
			}
		} else if (reg == IR_REG_NONE) {
			offset = ir_fuse_addr(ctx, insn->op2, &reg);
		} else if (IR_REG_SPILLED(reg)) {
			reg = IR_REG_NUM(reg);
			ir_emit_load(ctx, IR_ADDR, reg, insn->op2);
		}
	} else {
		IR_ASSERT(insn->op == IR_VSTORE);
		offset = ir_var_spill_slot(ctx, insn->op2, &reg);
	}

	if (op_insn->op == IR_MUL) {
		uint32_t shift = IR_LOG2(ctx->ir_base[op_insn->op2].val.u64);
		|	ASM_MEM_IMM_OP shl, type, [Ra(reg)+offset], shift
	} else if (op_insn->op == IR_DIV) {
		uint32_t shift = IR_LOG2(ctx->ir_base[op_insn->op2].val.u64);
		|	ASM_MEM_IMM_OP shr, type, [Ra(reg)+offset], shift
	} else {
		IR_ASSERT(op_insn->op == IR_MOD);
		uint64_t mask = ctx->ir_base[op_insn->op2].val.u64 - 1;
		IR_ASSERT(IR_IS_UNSIGNED_32BIT(mask));
		|	ASM_MEM_IMM_OP and, type, [Ra(reg)+offset], mask
	}
}

static void ir_emit_shift(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = insn->type;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];
	ir_reg op2_reg = ctx->regs[def][2];

	IR_ASSERT(def_reg != IR_REG_NONE && def_reg != IR_REG_RCX);
	if (op1_reg != IR_REG_NONE && IR_REG_SPILLED(op1_reg)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, insn->op1);
	}
	if (op2_reg != IR_REG_NONE && IR_REG_SPILLED(op2_reg)) {
		op2_reg = IR_REG_NUM(op2_reg);
		ir_emit_load(ctx, type, op2_reg, insn->op2);
	}
	if (op2_reg != IR_REG_RCX) {
		if (op1_reg == IR_REG_RCX) {
			ir_emit_mov(ctx, type, def_reg, op1_reg);
			op1_reg = def_reg;
		}
		if (op2_reg != IR_REG_NONE) {
			ir_emit_mov(ctx, type, IR_REG_RCX, op2_reg);
		} else {
			ir_emit_load(ctx, type, IR_REG_RCX, insn->op2);
		}
	}
	if (def_reg != op1_reg) {
		if (op1_reg != IR_REG_NONE) {
			ir_emit_mov(ctx, type, def_reg, op1_reg);
		} else {
			ir_emit_load(ctx, type, def_reg, insn->op1);
		}
	}
	switch (insn->op) {
		default:
			IR_ASSERT(0);
		case IR_SHL:
			|	ASM_REG_TXT_OP shl, insn->type, def_reg, cl
			break;
		case IR_SHR:
			|	ASM_REG_TXT_OP shr, insn->type, def_reg, cl
			break;
		case IR_SAR:
			|	ASM_REG_TXT_OP sar, insn->type, def_reg, cl
			break;
		case IR_ROL:
			|	ASM_REG_TXT_OP rol, insn->type, def_reg, cl
			break;
		case IR_ROR:
			|	ASM_REG_TXT_OP ror, insn->type, def_reg, cl
			break;
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_mem_shift(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_insn *op_insn = &ctx->ir_base[insn->op3];
	ir_type type = op_insn->type;
	ir_ref op2 = op_insn->op2;
	ir_reg op2_reg = ctx->regs[insn->op3][2];
	ir_reg reg;
	int32_t offset = 0;

	if (insn->op == IR_STORE) {
		reg = ctx->regs[def][2];
		if (IR_IS_CONST_REF(insn->op2)) {
			if (reg == IR_REG_NONE) {
				offset = ctx->ir_base[insn->op2].val.i32;
			} else {
				ir_emit_load(ctx, IR_ADDR, reg, insn->op2);
			}
		} else if (reg == IR_REG_NONE) {
			offset = ir_fuse_addr(ctx, insn->op2, &reg);
		} else if (IR_REG_SPILLED(reg)) {
			reg = IR_REG_NUM(reg);
			ir_emit_load(ctx, IR_ADDR, reg, insn->op2);
		}
	} else {
		IR_ASSERT(insn->op == IR_VSTORE);
		offset = ir_var_spill_slot(ctx, insn->op2, &reg);
	}

	if (op2_reg != IR_REG_NONE && IR_REG_SPILLED(op2_reg)) {
		op2_reg = IR_REG_NUM(op2_reg);
		ir_emit_load(ctx, type, op2_reg, op2);
	}
	if (op2_reg != IR_REG_RCX) {
		if (op2_reg != IR_REG_NONE) {
			ir_emit_mov(ctx, type, IR_REG_RCX, op2_reg);
		} else {
			ir_emit_load(ctx, type, IR_REG_RCX, op2);
		}
	}
	switch (op_insn->op) {
		default:
			IR_ASSERT(0);
		case IR_SHL:
			|	ASM_MEM_TXT_OP shl, type, [Ra(reg)+offset], cl
			break;
		case IR_SHR:
			|	ASM_MEM_TXT_OP shr, type, [Ra(reg)+offset], cl
			break;
		case IR_SAR:
			|	ASM_MEM_TXT_OP sar, type, [Ra(reg)+offset], cl
			break;
		case IR_ROL:
			|	ASM_MEM_TXT_OP rol, type, [Ra(reg)+offset], cl
			break;
		case IR_ROR:
			|	ASM_MEM_TXT_OP ror, type, [Ra(reg)+offset], cl
			break;
	}
}

static void ir_emit_shift_const(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	int32_t shift;
	ir_type type = insn->type;
	ir_ref op1 = insn->op1;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];

	IR_ASSERT(IR_IS_SIGNED_32BIT(ctx->ir_base[insn->op2].val.i64));
	shift = ctx->ir_base[insn->op2].val.i32;
	IR_ASSERT(def_reg != IR_REG_NONE);

	if (op1_reg != IR_REG_NONE && IR_REG_SPILLED(op1_reg)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, op1);
	}
	if (def_reg != op1_reg) {
		if (op1_reg != IR_REG_NONE) {
			ir_emit_mov(ctx, type, def_reg, op1_reg);
		} else {
			ir_emit_load(ctx, type, def_reg, op1);
		}
	}
	switch (insn->op) {
		default:
			IR_ASSERT(0);
		case IR_SHL:
			|	ASM_REG_IMM_OP shl, insn->type, def_reg, shift
			break;
		case IR_SHR:
			|	ASM_REG_IMM_OP shr, insn->type, def_reg, shift
			break;
		case IR_SAR:
			|	ASM_REG_IMM_OP sar, insn->type, def_reg, shift
			break;
		case IR_ROL:
			|	ASM_REG_IMM_OP rol, insn->type, def_reg, shift
			break;
		case IR_ROR:
			|	ASM_REG_IMM_OP ror, insn->type, def_reg, shift
			break;
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_mem_shift_const(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_insn *op_insn = &ctx->ir_base[insn->op3];
	ir_type type = op_insn->type;
	int32_t shift;
	ir_reg reg;
	int32_t offset = 0;

	IR_ASSERT(IR_IS_SIGNED_32BIT(ctx->ir_base[op_insn->op2].val.i64));
	shift = ctx->ir_base[op_insn->op2].val.i32;
	if (insn->op == IR_STORE) {
		reg = ctx->regs[def][2];
		if (IR_IS_CONST_REF(insn->op2)) {
			if (reg == IR_REG_NONE) {
				offset = ctx->ir_base[insn->op2].val.i32;
			} else {
				ir_emit_load(ctx, IR_ADDR, reg, insn->op2);
			}
		} else if (reg == IR_REG_NONE) {
			offset = ir_fuse_addr(ctx, insn->op2, &reg);
		} else if (IR_REG_SPILLED(reg)) {
			reg = IR_REG_NUM(reg);
			ir_emit_load(ctx, IR_ADDR, reg, insn->op2);
		}
	} else {
		IR_ASSERT(insn->op == IR_VSTORE);
		offset = ir_var_spill_slot(ctx, insn->op2, &reg);
	}

	switch (op_insn->op) {
		default:
			IR_ASSERT(0);
		case IR_SHL:
			|	ASM_MEM_IMM_OP shl, type, [Ra(reg)+offset], shift
			break;
		case IR_SHR:
			|	ASM_MEM_IMM_OP shr, type, [Ra(reg)+offset], shift
			break;
		case IR_SAR:
			|	ASM_MEM_IMM_OP sar, type, [Ra(reg)+offset], shift
			break;
		case IR_ROL:
			|	ASM_MEM_IMM_OP rol, type, [Ra(reg)+offset], shift
			break;
		case IR_ROR:
			|	ASM_MEM_IMM_OP ror, type, [Ra(reg)+offset], shift
			break;
	}
}

static void ir_emit_op_int(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = insn->type;
	ir_ref op1 = insn->op1;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];

	IR_ASSERT(def_reg != IR_REG_NONE);

	if (op1_reg != IR_REG_NONE && IR_REG_SPILLED(op1_reg)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, op1);
	}
	if (def_reg != op1_reg) {
		if (op1_reg != IR_REG_NONE) {
			ir_emit_mov(ctx, type, def_reg, op1_reg);
		} else {
			ir_emit_load(ctx, type, def_reg, op1);
		}
	}
	if (insn->op == IR_ADD) {
		|	ASM_REG_OP inc, insn->type, def_reg
	} else if (insn->op == IR_SUB) {
		|	ASM_REG_OP dec, insn->type, def_reg
	} else if (insn->op == IR_NOT) {
		|	ASM_REG_OP not, insn->type, def_reg
	} else if (insn->op == IR_NEG) {
		|	ASM_REG_OP neg, insn->type, def_reg
	} else {
		IR_ASSERT(insn->op == IR_BSWAP);
		switch (ir_type_size[insn->type]) {
			default:
				IR_ASSERT(0);
			case 4:
				|	bswap Rd(def_reg)
				break;
			case 8:
				IR_ASSERT(sizeof(void*) == 8);
|.if X64
				|	bswap Rq(def_reg)
|.endif
				break;
		}
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_bit_count(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = insn->type;
	ir_ref op1 = insn->op1;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];

	IR_ASSERT(def_reg != IR_REG_NONE);

	if (op1_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op1_reg) || IR_IS_CONST_REF(op1)) {
			op1_reg = IR_REG_NUM(op1_reg);
			ir_emit_load(ctx, type, op1_reg, op1);
		}
		switch (ir_type_size[insn->type]) {
			default:
				IR_ASSERT(0);
			case 2:
				if (insn->op == IR_CTLZ) {
					if (ctx->mflags & IR_X86_BMI1) {
						|	lzcnt Rw(def_reg), Rw(op1_reg)
					} else {
						|	bsr Rw(def_reg), Rw(op1_reg)
						|	xor Rw(def_reg), 0xf
					}
				} else if (insn->op == IR_CTTZ) {
					if (ctx->mflags & IR_X86_BMI1) {
						|	tzcnt Rw(def_reg), Rw(op1_reg)
					} else {
						|	bsf Rw(def_reg), Rw(op1_reg)
					}
				} else {
					IR_ASSERT(insn->op == IR_CTPOP);
					|	popcnt Rw(def_reg), Rw(op1_reg)
				}
				break;
			case 1:
				|   movzx Rd(op1_reg), Rb(op1_reg)
				if (insn->op == IR_CTLZ) {
					if (ctx->mflags & IR_X86_BMI1) {
						|	lzcnt Rd(def_reg), Rd(op1_reg)
						|	sub Rd(def_reg), 24
					} else {
						|	bsr Rd(def_reg), Rd(op1_reg)
						|	xor Rw(def_reg), 0x7
					}
					break;
				}
				IR_FALLTHROUGH;
			case 4:
				if (insn->op == IR_CTLZ) {
					if (ctx->mflags & IR_X86_BMI1) {
						|	lzcnt Rd(def_reg), Rd(op1_reg)
					} else {
						|	bsr Rd(def_reg), Rd(op1_reg)
						|	xor Rw(def_reg), 0x1f
					}
				} else if (insn->op == IR_CTTZ) {
					if (ctx->mflags & IR_X86_BMI1) {
						|	tzcnt Rd(def_reg), Rd(op1_reg)
					} else {
						|	bsf Rd(def_reg), Rd(op1_reg)
					}
				} else {
					IR_ASSERT(insn->op == IR_CTPOP);
					|	popcnt Rd(def_reg), Rd(op1_reg)
				}
				break;
|.if X64
			case 8:
				if (insn->op == IR_CTLZ) {
					if (ctx->mflags & IR_X86_BMI1) {
						|	lzcnt Rq(def_reg), Rq(op1_reg)
					} else {
						|	bsr Rq(def_reg), Rq(op1_reg)
						|	xor Rw(def_reg), 0x3f
					}
				} else if (insn->op == IR_CTTZ) {
					if (ctx->mflags & IR_X86_BMI1) {
						|	tzcnt Rq(def_reg), Rq(op1_reg)
					} else {
						|	bsf Rq(def_reg), Rq(op1_reg)
					}
				} else {
					IR_ASSERT(insn->op == IR_CTPOP);
					|	popcnt Rq(def_reg), Rq(op1_reg)
				}
				break;
|.endif
		}
	} else {
		int32_t offset = 0;

		if (ir_rule(ctx, op1) & IR_FUSED) {
			offset = ir_fuse_load(ctx, op1, &op1_reg);
		} else {
			offset = ir_ref_spill_slot(ctx, op1, &op1_reg);
		}
		if (op1_reg != IR_REG_NONE) {
			switch (ir_type_size[insn->type]) {
				default:
					IR_ASSERT(0);
				case 2:
					if (insn->op == IR_CTLZ) {
						if (ctx->mflags & IR_X86_BMI1) {
							|	lzcnt Rw(def_reg), word [Ra(op1_reg)+offset]
						} else {
							|	bsr Rw(def_reg), word [Ra(op1_reg)+offset]
							|	xor Rw(def_reg), 0xf
						}
					} else if (insn->op == IR_CTTZ) {
						if (ctx->mflags & IR_X86_BMI1) {
							|	tzcnt Rw(def_reg), word [Ra(op1_reg)+offset]
						} else {
							|	bsf Rw(def_reg), word [Ra(op1_reg)+offset]
						}
					} else {
						|	popcnt Rw(def_reg), word [Ra(op1_reg)+offset]
					}
					break;
				case 4:
					if (insn->op == IR_CTLZ) {
						if (ctx->mflags & IR_X86_BMI1) {
							|	lzcnt Rd(def_reg), dword [Ra(op1_reg)+offset]
						} else {
							|	bsr Rd(def_reg), dword [Ra(op1_reg)+offset]
							|	xor Rw(def_reg), 0x1f
						}
					} else if (insn->op == IR_CTTZ) {
						if (ctx->mflags & IR_X86_BMI1) {
							|	tzcnt Rd(def_reg), dword [Ra(op1_reg)+offset]
						} else {
							|	bsf Rd(def_reg), dword [Ra(op1_reg)+offset]
						}
					} else {
						|	popcnt Rd(def_reg), dword [Ra(op1_reg)+offset]
					}
					break;
|.if X64
				case 8:
					if (insn->op == IR_CTLZ) {
						if (ctx->mflags & IR_X86_BMI1) {
							|	lzcnt Rq(def_reg), qword [Ra(op1_reg)+offset]
						} else {
							|	bsr Rq(def_reg), qword [Ra(op1_reg)+offset]
							|	xor Rw(def_reg), 0x3f
						}
					} else if (insn->op == IR_CTTZ) {
						if (ctx->mflags & IR_X86_BMI1) {
							|	tzcnt Rq(def_reg), qword [Ra(op1_reg)+offset]
						} else {
							|	bsf Rq(def_reg), qword [Ra(op1_reg)+offset]
						}
					} else {
						|	popcnt Rq(def_reg), qword [Ra(op1_reg)+offset]
					}
					break;
|.endif
			}
		} else {
			switch (ir_type_size[insn->type]) {
				default:
					IR_ASSERT(0);
				case 2:
					if (insn->op == IR_CTLZ) {
						if (ctx->mflags & IR_X86_BMI1) {
							|	lzcnt Rw(def_reg), word [offset]
						} else {
							|	bsr Rw(def_reg), word [offset]
							|	xor Rw(def_reg), 0xf
						}
					} else if (insn->op == IR_CTTZ) {
						if (ctx->mflags & IR_X86_BMI1) {
							|	tzcnt Rw(def_reg), word [offset]
						} else {
							|	bsf Rw(def_reg), word [offset]
						}
					} else {
						|	popcnt Rw(def_reg), word [offset]
					}
					break;
				case 4:
					if (insn->op == IR_CTLZ) {
						if (ctx->mflags & IR_X86_BMI1) {
							|	lzcnt  Rd(def_reg), dword [offset]
						} else {
							|	bsr Rw(def_reg), word [offset]
							|	xor Rw(def_reg), 0x1f
						}
					} else if (insn->op == IR_CTTZ) {
						if (ctx->mflags & IR_X86_BMI1) {
							|	tzcnt  Rd(def_reg), dword [offset]
						} else {
							|	bsf  Rd(def_reg), dword [offset]
						}
					} else {
						|	popcnt Rd(def_reg), dword [offset]
					}
					break;
|.if X64
				case 8:
					if (insn->op == IR_CTLZ) {
						if (ctx->mflags & IR_X86_BMI1) {
							|	lzcnt Rq(def_reg), qword [offset]
						} else {
							|	bsr Rw(def_reg), word [offset]
							|	xor Rw(def_reg), 0x1f
						}
					} else if (insn->op == IR_CTTZ) {
						if (ctx->mflags & IR_X86_BMI1) {
							|	tzcnt Rq(def_reg), qword [offset]
						} else {
							|	bsf Rq(def_reg), qword [offset]
						}
					} else {
						|	popcnt Rq(def_reg), qword [offset]
					}
					break;
|.endif
			}
		}
	}

	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_ctpop(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = insn->type;
	ir_ref op1 = insn->op1;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];
	ir_reg tmp_reg = ctx->regs[def][2];
|.if X64
||	ir_reg const_reg = ctx->regs[def][3];
|.endif

	IR_ASSERT(def_reg != IR_REG_NONE && tmp_reg != IR_REG_NONE);
	if (IR_IS_CONST_REF(op1) || op1_reg == IR_REG_NONE) {
		ir_emit_load(ctx, type, def_reg, op1);
		if (ir_type_size[insn->type] == 1) {
			|	movzx Rd(def_reg), Rb(def_reg)
		} else if (ir_type_size[insn->type] == 2) {
			|	movzx Rd(def_reg), Rw(def_reg)
		}
	} else {
		if (IR_REG_SPILLED(op1_reg)) {
			op1_reg = IR_REG_NUM(op1_reg);
			ir_emit_load(ctx, type, op1_reg, op1);
		}
		switch (ir_type_size[insn->type]) {
			default:
				IR_ASSERT(0);
			case 1:
				|	movzx Rd(def_reg), Rb(op1_reg)
				break;
			case 2:
				|	movzx Rd(def_reg), Rw(op1_reg)
				break;
			case 4:
				|	mov Rd(def_reg), Rd(op1_reg)
				break;
|.if X64
||			case 8:
				|	mov Rq(def_reg), Rq(op1_reg)
||				break;
|.endif
		}
	}
	switch (ir_type_size[insn->type]) {
		default:
			IR_ASSERT(0);
		case 1:
			|	mov Rd(tmp_reg), Rd(def_reg)
			|	shr Rd(def_reg), 1
			|	and Rd(def_reg), 0x55
			|	sub Rd(tmp_reg), Rd(def_reg)
			|	mov Rd(def_reg), Rd(tmp_reg)
			|	and Rd(def_reg), 0x33
			|	shr Rd(tmp_reg), 2
			|	and Rd(tmp_reg), 0x33
			|	add Rd(tmp_reg), Rd(def_reg)
			|	mov Rd(def_reg), Rd(tmp_reg)
			|	shr Rd(def_reg), 4
			|	add Rd(def_reg), Rd(tmp_reg)
			|	and Rd(def_reg), 0x0f
			break;
		case 2:
			|	mov Rd(tmp_reg), Rd(def_reg)
			|	shr Rd(def_reg), 1
			|	and Rd(def_reg), 0x5555
			|	sub Rd(tmp_reg), Rd(def_reg)
			|	mov Rd(def_reg), Rd(tmp_reg)
			|	and Rd(def_reg), 0x3333
			|	shr Rd(tmp_reg), 2
			|	and Rd(tmp_reg), 0x3333
			|	add Rd(tmp_reg), Rd(def_reg)
			|	mov Rd(def_reg), Rd(tmp_reg)
			|	shr Rd(def_reg), 4
			|	add Rd(def_reg), Rd(tmp_reg)
			|	and Rd(def_reg), 0x0f0f
			|	mov	Rd(tmp_reg), Rd(def_reg)
			|	shr Rd(tmp_reg), 8
			|	and Rd(def_reg), 0x0f
			|	add Rd(def_reg), Rd(tmp_reg)
			break;
		case 4:
			|	mov Rd(tmp_reg), Rd(def_reg)
			|	shr Rd(def_reg), 1
			|	and Rd(def_reg), 0x55555555
			|	sub Rd(tmp_reg), Rd(def_reg)
			|	mov Rd(def_reg), Rd(tmp_reg)
			|	and Rd(def_reg), 0x33333333
			|	shr Rd(tmp_reg), 2
			|	and Rd(tmp_reg), 0x33333333
			|	add Rd(tmp_reg), Rd(def_reg)
			|	mov Rd(def_reg), Rd(tmp_reg)
			|	shr Rd(def_reg), 4
			|	add Rd(def_reg), Rd(tmp_reg)
			|	and Rd(def_reg), 0x0f0f0f0f
			|	imul Rd(def_reg), 0x01010101
			|	shr Rd(def_reg), 24
			break;
|.if X64
||		case 8:
||			IR_ASSERT(const_reg != IR_REG_NONE);
			|	mov Rq(tmp_reg), Rq(def_reg)
			|	shr Rq(def_reg), 1
			|	mov64 Rq(const_reg), 0x5555555555555555
			|	and Rq(def_reg), Rq(const_reg)
			|	sub Rq(tmp_reg), Rq(def_reg)
			|	mov Rq(def_reg), Rq(tmp_reg)
			|	mov64 Rq(const_reg), 0x3333333333333333
			|	and Rq(def_reg), Rq(const_reg)
			|	shr Rq(tmp_reg), 2
			|	and Rq(tmp_reg), Rq(const_reg)
			|	add Rq(tmp_reg), Rq(def_reg)
			|	mov Rq(def_reg), Rq(tmp_reg)
			|	shr Rq(def_reg), 4
			|	add Rq(def_reg), Rq(tmp_reg)
			|	mov64 Rq(const_reg), 0x0f0f0f0f0f0f0f0f
			|	and Rq(def_reg), Rq(const_reg)
			|	mov64 Rq(const_reg), 0x0101010101010101
			|	imul Rq(def_reg), Rq(const_reg)
			|	shr Rq(def_reg), 56
||			break;
|.endif
	}

	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_mem_op_int(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_insn *op_insn = &ctx->ir_base[insn->op3];
	ir_type type = op_insn->type;
	ir_reg reg;
	int32_t offset = 0;

	if (insn->op == IR_STORE) {
		reg = ctx->regs[def][2];
		if (IR_IS_CONST_REF(insn->op2)) {
			if (reg == IR_REG_NONE) {
				offset = ctx->ir_base[insn->op2].val.i32;
			} else {
				ir_emit_load(ctx, IR_ADDR, reg, insn->op2);
			}
		} else if (reg == IR_REG_NONE) {
			offset = ir_fuse_addr(ctx, insn->op2, &reg);
		} else if (IR_REG_SPILLED(reg)) {
			reg = IR_REG_NUM(reg);
			ir_emit_load(ctx, IR_ADDR, reg, insn->op2);
		}
	} else {
		IR_ASSERT(insn->op == IR_VSTORE);
		offset = ir_var_spill_slot(ctx, insn->op2, &reg);
	}

	if (op_insn->op == IR_ADD) {
		|	ASM_MEM_OP inc, type, [Ra(reg)+offset]
	} else if (op_insn->op == IR_SUB) {
		|	ASM_MEM_OP dec, type, [Ra(reg)+offset]
	} else if (op_insn->op == IR_NOT) {
		|	ASM_MEM_OP not, type, [Ra(reg)+offset]
	} else {
		IR_ASSERT(op_insn->op == IR_NEG);
		|	ASM_MEM_OP neg, type, [Ra(reg)+offset]
	}
}

static void ir_emit_abs_int(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = insn->type;
	ir_ref op1 = insn->op1;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];

	IR_ASSERT(def_reg != IR_REG_NONE && op1_reg != IR_REG_NONE);

	if (op1_reg != IR_REG_NONE && IR_REG_SPILLED(op1_reg)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, op1);
	}

	IR_ASSERT(def_reg != op1_reg);

	ir_emit_mov(ctx, insn->type, def_reg, op1_reg);
	|	ASM_REG_OP neg, insn->type, def_reg
	|	ASM_REG_REG_OP2, cmovs, type, def_reg, op1_reg
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_bool_not_int(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = ctx->ir_base[insn->op1].type;
	ir_ref op1 = insn->op1;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];

	IR_ASSERT(def_reg != IR_REG_NONE);

	if (op1_reg != IR_REG_NONE && IR_REG_SPILLED(op1_reg)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, op1);
	}

	if (op1_reg != IR_REG_NONE) {
		|	ASM_REG_REG_OP test, type, op1_reg, op1_reg
	} else {
		ir_reg fp;
		int32_t offset = ir_ref_spill_slot(ctx, op1, &fp);

		|	ASM_MEM_IMM_OP cmp, type, [Ra(fp)+offset], 0
	}
	|	sete Rb(def_reg)

	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_mul_div_mod(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = insn->type;
	ir_ref op1 = insn->op1;
	ir_ref op2 = insn->op2;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];
	ir_reg op2_reg = ctx->regs[def][2];
	int32_t offset = 0;

	if (op1_reg != IR_REG_NONE && IR_REG_SPILLED(op1_reg)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, op1);
	}
	if (op1_reg != IR_REG_RAX) {
		if (op1_reg != IR_REG_NONE) {
			ir_emit_mov(ctx, type, IR_REG_RAX, op1_reg);
		} else {
			ir_emit_load(ctx, type, IR_REG_RAX, op1);
		}
	}
	if (op2_reg == IR_REG_NONE && op1 == op2) {
		op2_reg = IR_REG_RAX;
	} else if (IR_IS_CONST_REF(op2)) {
		if (insn->op == IR_MUL || insn->op == IR_MUL_OV) {
			op2_reg = IR_REG_RDX;
		} else {
			IR_ASSERT(op2_reg != IR_REG_NONE);
		}
		ir_emit_load(ctx, type, op2_reg, op2);
	}
	if (insn->op == IR_MUL || insn->op == IR_MUL_OV) {
		IR_ASSERT(!IR_IS_TYPE_SIGNED(insn->type));
		if (op2_reg != IR_REG_NONE) {
			if (IR_REG_SPILLED(op2_reg)) {
				op2_reg = IR_REG_NUM(op2_reg);
				ir_emit_load(ctx, type, op2_reg, op2);
			}
			|	ASM_REG_OP mul, type, op2_reg
		} else {
			if (ir_rule(ctx, op2) & IR_FUSED) {
				offset = ir_fuse_load(ctx, op2, &op2_reg);
			} else {
				offset = ir_ref_spill_slot(ctx, op2, &op2_reg);
			}
			|	ASM_MEM_OP mul, type, [Ra(op2_reg)+offset]
		}
	} else {
		if (IR_IS_TYPE_SIGNED(type)) {
			if (ir_type_size[type] == 8) {
				|	cqo
			} else if (ir_type_size[type] == 4) {
				|	cdq
			} else if (ir_type_size[type] == 2) {
				|	cwd
			}
			if (op2_reg != IR_REG_NONE) {
				if (IR_REG_SPILLED(op2_reg)) {
					op2_reg = IR_REG_NUM(op2_reg);
					ir_emit_load(ctx, type, op2_reg, op2);
				}
				|	ASM_REG_OP idiv, type, op2_reg
			} else {
				if (ir_rule(ctx, op2) & IR_FUSED) {
					offset = ir_fuse_load(ctx, op2, &op2_reg);
				} else {
					offset = ir_ref_spill_slot(ctx, op2, &op2_reg);
				}
				|	ASM_MEM_OP idiv, type, [Ra(op2_reg)+offset]
			}
		} else {
			|	ASM_REG_REG_OP xor, type, IR_REG_RDX, IR_REG_RDX
			if (op2_reg != IR_REG_NONE) {
				if (IR_REG_SPILLED(op2_reg)) {
					op2_reg = IR_REG_NUM(op2_reg);
					ir_emit_load(ctx, type, op2_reg, op2);
				}
				|	ASM_REG_OP div, type, op2_reg
			} else {
				if (ir_rule(ctx, op2) & IR_FUSED) {
					offset = ir_fuse_load(ctx, op2, &op2_reg);
				} else {
					offset = ir_ref_spill_slot(ctx, op2, &op2_reg);
				}
				|	ASM_MEM_OP div, type, [Ra(op2_reg)+offset]
			}
		}
	}

	if (insn->op == IR_MUL || insn->op == IR_MUL_OV || insn->op == IR_DIV) {
		if (def_reg != IR_REG_NONE) {
			if (def_reg != IR_REG_RAX) {
				ir_emit_mov(ctx, type, def_reg, IR_REG_RAX);
			}
			if (IR_REG_SPILLED(ctx->regs[def][0])) {
				ir_emit_store(ctx, type, def, def_reg);
			}
		} else {
			ir_emit_store(ctx, type, def, IR_REG_RAX);
		}
	} else {
		IR_ASSERT(insn->op == IR_MOD);
		if (ir_type_size[type] == 1) {
			if (def_reg != IR_REG_NONE) {
				|	mov al, ah
				if (def_reg != IR_REG_RAX) {
					|	mov Rb(def_reg), al
				}
				if (IR_REG_SPILLED(ctx->regs[def][0])) {
					ir_emit_store(ctx, type, def, def_reg);
				}
			} else {
				ir_reg fp;
				int32_t offset = ir_ref_spill_slot(ctx, def, &fp);

				|	mov byte [Ra(fp)+offset], ah
			}
		} else {
			if (def_reg != IR_REG_NONE) {
				if (def_reg != IR_REG_RDX) {
					ir_emit_mov(ctx, type, def_reg, IR_REG_RDX);
				}
				if (IR_REG_SPILLED(ctx->regs[def][0])) {
					ir_emit_store(ctx, type, def, def_reg);
				}
			} else {
				ir_emit_store(ctx, type, def, IR_REG_RDX);
			}
		}
	}
}

static void ir_rodata(ir_ctx *ctx)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	|.rodata
	if (!data->rodata_label) {
		int label = data->rodata_label = ctx->cfg_blocks_count + ctx->consts_count + 2;
		|=>label:
	}
}

static void ir_emit_op_fp(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = insn->type;
	ir_ref op1 = insn->op1;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];

	IR_ASSERT(def_reg != IR_REG_NONE);

	if (op1_reg != IR_REG_NONE && IR_REG_SPILLED(op1_reg)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, op1);
	}
	if (def_reg != op1_reg) {
		if (op1_reg != IR_REG_NONE) {
			ir_emit_fp_mov(ctx, type, def_reg, op1_reg);
		} else {
			ir_emit_load(ctx, type, def_reg, op1);
		}
	}
	if (insn->op == IR_NEG) {
		if (insn->type == IR_DOUBLE) {
			if (!data->double_neg_const) {
				data->double_neg_const = 1;
				ir_rodata(ctx);
				|.align 16
				|->double_neg_const:
				|.dword 0, 0x80000000, 0, 0
				|.code
			}
			if (ctx->mflags & IR_X86_AVX) {
				|	vxorpd xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), [->double_neg_const]
			} else {
				|	xorpd xmm(def_reg-IR_REG_FP_FIRST), [->double_neg_const]
			}
		} else {
			IR_ASSERT(insn->type == IR_FLOAT);
			if (!data->float_neg_const) {
				data->float_neg_const = 1;
				ir_rodata(ctx);
				|.align 16
				|->float_neg_const:
				|.dword 0x80000000, 0, 0, 0
				|.code
			}
			if (ctx->mflags & IR_X86_AVX) {
				|	vxorps xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), [->float_neg_const]
			} else {
				|	xorps xmm(def_reg-IR_REG_FP_FIRST), [->float_neg_const]
			}
		}
	} else {
		IR_ASSERT(insn->op == IR_ABS);
		if (insn->type == IR_DOUBLE) {
			if (!data->double_abs_const) {
				data->double_abs_const = 1;
				ir_rodata(ctx);
				|.align 16
				|->double_abs_const:
				|.dword 0xffffffff, 0x7fffffff, 0, 0
				|.code
			}
			if (ctx->mflags & IR_X86_AVX) {
				|	vandpd xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), [->double_abs_const]
			} else {
				|	andpd xmm(def_reg-IR_REG_FP_FIRST), [->double_abs_const]
			}
		} else {
			IR_ASSERT(insn->type == IR_FLOAT);
			if (!data->float_abs_const) {
				data->float_abs_const = 1;
				ir_rodata(ctx);
				|.align 16
				|->float_abs_const:
				|.dword 0x7fffffff, 0, 0, 0
				|.code
			}
			if (ctx->mflags & IR_X86_AVX) {
				|	vandps xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), [->float_abs_const]
			} else {
				|	andps xmm(def_reg-IR_REG_FP_FIRST), [->float_abs_const]
			}
		}
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, insn->type, def, def_reg);
	}
}

static void ir_emit_binop_sse2(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = insn->type;
	ir_ref op1 = insn->op1;
	ir_ref op2 = insn->op2;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];
	ir_reg op2_reg = ctx->regs[def][2];

	IR_ASSERT(def_reg != IR_REG_NONE);

	if (op1_reg != IR_REG_NONE && IR_REG_SPILLED(op1_reg)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, op1);
	}
	if (def_reg != op1_reg) {
		if (op1_reg != IR_REG_NONE) {
			ir_emit_fp_mov(ctx, type, def_reg, op1_reg);
		} else {
			ir_emit_load(ctx, type, def_reg, op1);
		}
		if (op1 == op2) {
			op2_reg = def_reg;
		}
	}
	if (op2_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op2_reg) || IR_IS_CONST_REF(op2)) {
			op2_reg = IR_REG_NUM(op2_reg);
			if (op1 != op2) {
				ir_emit_load(ctx, type, op2_reg, op2);
			}
		}
		switch (insn->op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_ADD:
				|	ASM_SSE2_REG_REG_OP addss, addsd, type, def_reg, op2_reg
				break;
			case IR_SUB:
				|	ASM_SSE2_REG_REG_OP subss, subsd, type, def_reg, op2_reg
				break;
			case IR_MUL:
				|	ASM_SSE2_REG_REG_OP mulss, mulsd, type, def_reg, op2_reg
				break;
			case IR_DIV:
				|	ASM_SSE2_REG_REG_OP divss, divsd, type, def_reg, op2_reg
				break;
			case IR_MIN:
				|	ASM_SSE2_REG_REG_OP minss, minsd, type, def_reg, op2_reg
				break;
			case IR_MAX:
				|	ASM_SSE2_REG_REG_OP maxss, maxsd, type, def_reg, op2_reg
				break;
		}
	} else if (IR_IS_CONST_REF(op2)) {
		ir_insn *val_insn = &ctx->ir_base[op2];
		int label = ctx->cfg_blocks_count - op2;

		val_insn->const_flags |= IR_CONST_EMIT;
		switch (insn->op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_ADD:
				|	ASM_SSE2_REG_MEM_OP addss, addsd, type, def_reg, [=>label]
				break;
			case IR_SUB:
				|	ASM_SSE2_REG_MEM_OP subss, subsd, type, def_reg, [=>label]
				break;
			case IR_MUL:
				|	ASM_SSE2_REG_MEM_OP mulss, mulsd, type, def_reg, [=>label]
				break;
			case IR_DIV:
				|	ASM_SSE2_REG_MEM_OP divss, divsd, type, def_reg, [=>label]
				break;
			case IR_MIN:
				|	ASM_SSE2_REG_MEM_OP minss, minsd, type, def_reg, [=>label]
				break;
			case IR_MAX:
				|	ASM_SSE2_REG_MEM_OP maxss, maxsd, type, def_reg, [=>label]
				break;
		}
	} else {
		int32_t offset = 0;

		if (ir_rule(ctx, op2) & IR_FUSED) {
			offset = ir_fuse_load(ctx, op2, &op2_reg);
		} else {
			offset = ir_ref_spill_slot(ctx, op2, &op2_reg);
		}
		switch (insn->op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_ADD:
				|	ASM_SSE2_REG_MEM_OP addss, addsd, type, def_reg, [Ra(op2_reg)+offset]
				break;
			case IR_SUB:
				|	ASM_SSE2_REG_MEM_OP subss, subsd, type, def_reg, [Ra(op2_reg)+offset]
				break;
			case IR_MUL:
				|	ASM_SSE2_REG_MEM_OP mulss, mulsd, type, def_reg, [Ra(op2_reg)+offset]
				break;
			case IR_DIV:
				|	ASM_SSE2_REG_MEM_OP divss, divsd, type, def_reg, [Ra(op2_reg)+offset]
				break;
			case IR_MIN:
				|	ASM_SSE2_REG_MEM_OP minss, minsd, type, def_reg, [Ra(op2_reg)+offset]
				break;
			case IR_MAX:
				|	ASM_SSE2_REG_MEM_OP maxss, maxsd, type, def_reg, [Ra(op2_reg)+offset]
				break;
		}
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, insn->type, def, def_reg);
	}
}

static void ir_emit_binop_avx(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = insn->type;
	ir_ref op1 = insn->op1;
	ir_ref op2 = insn->op2;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];
	ir_reg op2_reg = ctx->regs[def][2];

	IR_ASSERT(def_reg != IR_REG_NONE && op1_reg != IR_REG_NONE);

	if (IR_REG_SPILLED(op1_reg) || IR_IS_CONST_REF(op1)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, op1);
	}
	if (op2_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op2_reg) || IR_IS_CONST_REF(op2)) {
			op2_reg = IR_REG_NUM(op2_reg);
			if (op1 != op2) {
				ir_emit_load(ctx, type, op2_reg, op2);
			}
		}
		switch (insn->op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_ADD:
				|	ASM_AVX_REG_REG_REG_OP vaddss, vaddsd, type, def_reg, op1_reg, op2_reg
				break;
			case IR_SUB:
				|	ASM_AVX_REG_REG_REG_OP vsubss, vsubsd, type, def_reg, op1_reg, op2_reg
				break;
			case IR_MUL:
				|	ASM_AVX_REG_REG_REG_OP vmulss, vmulsd, type, def_reg, op1_reg, op2_reg
				break;
			case IR_DIV:
				|	ASM_AVX_REG_REG_REG_OP vdivss, vdivsd, type, def_reg, op1_reg, op2_reg
				break;
			case IR_MIN:
				|	ASM_AVX_REG_REG_REG_OP vminss, vminsd, type, def_reg, op1_reg, op2_reg
				break;
			case IR_MAX:
				|	ASM_AVX_REG_REG_REG_OP vmaxss, vmaxsd, type, def_reg, op1_reg, op2_reg
				break;
		}
	} else if (IR_IS_CONST_REF(op2)) {
		ir_insn *val_insn = &ctx->ir_base[op2];
		int label = ctx->cfg_blocks_count - op2;

		val_insn->const_flags |= IR_CONST_EMIT;
		switch (insn->op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_ADD:
				|	ASM_AVX_REG_REG_MEM_OP vaddss, vaddsd, type, def_reg, op1_reg, [=>label]
				break;
			case IR_SUB:
				|	ASM_AVX_REG_REG_MEM_OP vsubss, vsubsd, type, def_reg, op1_reg, [=>label]
				break;
			case IR_MUL:
				|	ASM_AVX_REG_REG_MEM_OP vmulss, vmulsd, type, def_reg, op1_reg, [=>label]
				break;
			case IR_DIV:
				|	ASM_AVX_REG_REG_MEM_OP vdivss, vdivsd, type, def_reg, op1_reg, [=>label]
				break;
			case IR_MIN:
				|	ASM_AVX_REG_REG_MEM_OP vminss, vminsd, type, def_reg, op1_reg, [=>label]
				break;
			case IR_MAX:
				|	ASM_AVX_REG_REG_MEM_OP vmaxss, vmaxsd, type, def_reg, op1_reg, [=>label]
				break;
		}
	} else {
		int32_t offset = 0;

		if (ir_rule(ctx, op2) & IR_FUSED) {
			offset = ir_fuse_load(ctx, op2, &op2_reg);
		} else {
			offset = ir_ref_spill_slot(ctx, op2, &op2_reg);
		}
		switch (insn->op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_ADD:
				|	ASM_AVX_REG_REG_MEM_OP vaddss, vaddsd, type, def_reg, op1_reg, [Ra(op2_reg)+offset]
				break;
			case IR_SUB:
				|	ASM_AVX_REG_REG_MEM_OP vsubss, vsubsd, type, def_reg, op1_reg, [Ra(op2_reg)+offset]
				break;
			case IR_MUL:
				|	ASM_AVX_REG_REG_MEM_OP vmulss, vmulsd, type, def_reg, op1_reg, [Ra(op2_reg)+offset]
				break;
			case IR_DIV:
				|	ASM_AVX_REG_REG_MEM_OP vdivss, vdivsd, type, def_reg, op1_reg, [Ra(op2_reg)+offset]
				break;
			case IR_MIN:
				|	ASM_AVX_REG_REG_MEM_OP vminss, vminsd, type, def_reg, op1_reg, [Ra(op2_reg)+offset]
				break;
			case IR_MAX:
				|	ASM_AVX_REG_REG_MEM_OP vmaxss, vmaxsd, type, def_reg, op1_reg, [Ra(op2_reg)+offset]
				break;
		}
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, insn->type, def, def_reg);
	}
}

static void ir_emit_cmp_int_common(ir_ctx *ctx, ir_type type, ir_insn *insn, ir_reg op1_reg, ir_ref op1, ir_reg op2_reg, ir_ref op2)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	if (op1_reg != IR_REG_NONE) {
		if (op2_reg != IR_REG_NONE) {
			|	ASM_REG_REG_OP cmp, type, op1_reg, op2_reg
		} else if (IR_IS_CONST_REF(op2) && ctx->ir_base[op2].val.u64 == 0) {
			|	ASM_REG_REG_OP test, type, op1_reg, op1_reg
		} else if (IR_IS_CONST_REF(op2)) {
			ir_insn *val_insn = &ctx->ir_base[op2];

			IR_ASSERT(IR_IS_32BIT(val_insn->type, val_insn->val));
			|	ASM_REG_IMM_OP cmp, type, op1_reg, val_insn->val.i32
		} else {
			int32_t offset = 0;

			if (ir_rule(ctx, op2) & IR_FUSED) {
				offset = ir_fuse_load(ctx, op2, &op2_reg);
			} else {
				offset = ir_ref_spill_slot(ctx, op2, &op2_reg);
			}
			if (op2_reg != IR_REG_NONE) {
				|	ASM_REG_MEM_OP cmp, type, op1_reg, [Ra(op2_reg)+offset]
			} else {
				|	ASM_REG_MEM_OP cmp, type, op1_reg, [offset]
			}
		}
	} else if (IR_IS_CONST_REF(insn->op1)) {
		IR_ASSERT(0);
	} else {
		int32_t offset = 0;

		if (ir_rule(ctx, insn->op1) & IR_FUSED) {
			offset = ir_fuse_load(ctx, insn->op1, &op1_reg);
		} else {
			offset = ir_ref_spill_slot(ctx, insn->op1, &op1_reg);
		}
		if (op2_reg != IR_REG_NONE) {
			if (op1_reg == IR_REG_NONE) {
				|	ASM_MEM_REG_OP cmp, type, [offset], op2_reg
			} else {
				|	ASM_MEM_REG_OP cmp, type, [Ra(op1_reg)+offset], op2_reg
			}
		} else {
			IR_ASSERT(!IR_IS_CONST_REF(op1));
			IR_ASSERT(IR_IS_CONST_REF(op2));
			IR_ASSERT(IR_IS_32BIT(ctx->ir_base[op2].type, ctx->ir_base[op2].val));
			if (op1_reg == IR_REG_NONE) {
				|	ASM_MEM_IMM_OP cmp, type, [offset], ctx->ir_base[op2].val.i32
			} else {
				|	ASM_MEM_IMM_OP cmp, type, [Ra(op1_reg)+offset], ctx->ir_base[op2].val.i32
			}
		}
	}
}

static void _ir_emit_setcc_int(ir_ctx *ctx, uint8_t op, ir_reg def_reg)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	switch (op) {
		default:
			IR_ASSERT(0 && "NIY binary op");
		case IR_EQ:
			|	sete Rb(def_reg)
			break;
		case IR_NE:
			|	setne Rb(def_reg)
			break;
		case IR_LT:
			|	setl Rb(def_reg)
			break;
		case IR_GE:
			|	setge Rb(def_reg)
			break;
		case IR_LE:
			|	setle Rb(def_reg)
			break;
		case IR_GT:
			|	setg Rb(def_reg)
			break;
		case IR_ULT:
			|	setb Rb(def_reg)
			break;
		case IR_UGE:
			|	setae Rb(def_reg)
			break;
		case IR_ULE:
			|	setbe Rb(def_reg)
			break;
		case IR_UGT:
			|	seta Rb(def_reg)
			break;
	}
}

static void ir_emit_cmp_int(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = ctx->ir_base[insn->op1].type;
	ir_op op = insn->op;
	ir_ref op1 = insn->op1;
	ir_ref op2 = insn->op2;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];
	ir_reg op2_reg = ctx->regs[def][2];

	IR_ASSERT(def_reg != IR_REG_NONE);
	if (op1_reg != IR_REG_NONE && (IR_IS_CONST_REF(op1) || IR_REG_SPILLED(op1_reg))) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, op1);
	}
	if (op2_reg != IR_REG_NONE && (IR_IS_CONST_REF(op2) || IR_REG_SPILLED(op2_reg))) {
		op2_reg = IR_REG_NUM(op2_reg);
		if (op1 != op2) {
			ir_emit_load(ctx, type, op2_reg, op2);
		}
	}
	if (IR_IS_CONST_REF(op2) && ctx->ir_base[op2].val.u64 == 0) {
		if (op == IR_ULT) {
			/* always false */
			|	xor Ra(def_reg), Ra(def_reg)
			if (IR_REG_SPILLED(ctx->regs[def][0])) {
				ir_emit_store(ctx, insn->type, def, def_reg);
			}
			return;
		} else if (op == IR_UGE) {
			/* always true */
			|	ASM_REG_IMM_OP mov, insn->type, def_reg, 1
			if (IR_REG_SPILLED(ctx->regs[def][0])) {
				ir_emit_store(ctx, insn->type, def, def_reg);
			}
			return;
		} else if (op == IR_ULE) {
			op = IR_EQ;
		} else if (op == IR_UGT) {
			op = IR_NE;
		}
	}
	ir_emit_cmp_int_common(ctx, type, insn, op1_reg, op1, op2_reg, op2);
	_ir_emit_setcc_int(ctx, op, def_reg);
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, insn->type, def, def_reg);
	}
}

static void ir_emit_test_int_common(ir_ctx *ctx, ir_ref ref, ir_op op)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_insn *binop_insn = &ctx->ir_base[ref];
	ir_type type = binop_insn->type;
	ir_ref op1 = binop_insn->op1;
	ir_ref op2 = binop_insn->op2;
	ir_reg op1_reg = ctx->regs[ref][1];
	ir_reg op2_reg = ctx->regs[ref][2];

	IR_ASSERT(binop_insn->op == IR_AND);
	if (op1_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op1_reg)) {
			op1_reg = IR_REG_NUM(op1_reg);
			ir_emit_load(ctx, type, op1_reg, op1);
		}
		if (op2_reg != IR_REG_NONE) {
			if (IR_REG_SPILLED(op2_reg) || IR_IS_CONST_REF(op2)) {
				op2_reg = IR_REG_NUM(op2_reg);
				if (op1 != op2) {
					ir_emit_load(ctx, type, op2_reg, op2);
				}
			}
			|	ASM_REG_REG_OP test, type, op1_reg, op2_reg
		} else if (IR_IS_CONST_REF(op2)) {
			ir_insn *val_insn = &ctx->ir_base[op2];
			int32_t val;

			IR_ASSERT(IR_IS_32BIT(val_insn->type, val_insn->val));
			val = val_insn->val.i32;
			if ((op == IR_EQ || op == IR_NE) && val == 0xff && (sizeof(void*) == 8 || op1_reg <= IR_REG_R3)) {
				|	test Rb(op1_reg), Rb(op1_reg)
			} else if ((op == IR_EQ || op == IR_NE) && val == 0xff00 && op1_reg <= IR_REG_R3) {
				if (op1_reg == IR_REG_RAX) {
					|	test ah, ah
				} else if (op1_reg == IR_REG_RBX) {
					|	test bh, bh
				} else if (op1_reg == IR_REG_RCX) {
					|	test ch, ch
				} else if (op1_reg == IR_REG_RDX) {
					|	test dh, dh
				} else {
					IR_ASSERT(0);
				}
			} else if ((op == IR_EQ || op == IR_NE) && val == 0xffff) {
				|	test Rw(op1_reg), Rw(op1_reg)
			} else if ((op == IR_EQ || op == IR_NE) && val == -1) {
				|	test Rd(op1_reg), Rd(op1_reg)
			} else {
				|	ASM_REG_IMM_OP test, type, op1_reg, val
			}
		} else {
			int32_t offset = 0;

			if (ir_rule(ctx, op2) & IR_FUSED) {
				offset = ir_fuse_load(ctx, op2, &op2_reg);
			} else {
				offset = ir_ref_spill_slot(ctx, op2, &op2_reg);
			}
			|	ASM_REG_MEM_OP test, type, op1_reg, [Ra(op2_reg)+offset]
		}
	} else if (IR_IS_CONST_REF(op1)) {
		IR_ASSERT(0);
	} else {
		int32_t offset = 0;

		if (ir_rule(ctx, op1) & IR_FUSED) {
			offset = ir_fuse_load(ctx, op1, &op1_reg);
		} else {
			offset = ir_ref_spill_slot(ctx, op1, &op1_reg);
		}
		if (op2_reg != IR_REG_NONE) {
			if (IR_REG_SPILLED(op2_reg) || IR_IS_CONST_REF(op2)) {
				op2_reg = IR_REG_NUM(op2_reg);
				if (op1 != op2) {
					ir_emit_load(ctx, type, op2_reg, op2);
				}
			}
			if (op1_reg == IR_REG_NONE) {
				|	ASM_MEM_REG_OP test, type, [offset], op2_reg
			} else {
				|	ASM_MEM_REG_OP test, type, [Ra(op1_reg)+offset], op2_reg
			}
		} else {
			IR_ASSERT(!IR_IS_CONST_REF(op1));
			IR_ASSERT(IR_IS_CONST_REF(op2));
			IR_ASSERT(IR_IS_32BIT(ctx->ir_base[op2].type, ctx->ir_base[op2].val));
			if (op1_reg == IR_REG_NONE) {
				|	ASM_MEM_IMM_OP test, type, [offset], ctx->ir_base[op2].val.i32
			} else {
				|	ASM_MEM_IMM_OP test, type, [Ra(op1_reg)+offset], ctx->ir_base[op2].val.i32
			}
		}
	}
}

static void ir_emit_testcc_int(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);

	IR_ASSERT(def_reg != IR_REG_NONE);
	ir_emit_test_int_common(ctx, insn->op1, insn->op);
	_ir_emit_setcc_int(ctx, insn->op, def_reg);
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, insn->type, def, def_reg);
	}
}

static void ir_emit_setcc_int(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);

	IR_ASSERT(def_reg != IR_REG_NONE);
	_ir_emit_setcc_int(ctx, insn->op, def_reg);
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, insn->type, def, def_reg);
	}
}

static ir_op ir_emit_cmp_fp_common(ir_ctx *ctx, ir_ref cmp_ref, ir_insn *cmp_insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = ctx->ir_base[cmp_insn->op1].type;
	ir_op op = cmp_insn->op;
	ir_ref op1, op2;
	ir_reg op1_reg, op2_reg;

	op1 = cmp_insn->op1;
	op2 = cmp_insn->op2;
	op1_reg = ctx->regs[cmp_ref][1];
	op2_reg = ctx->regs[cmp_ref][2];

	if (op1_reg == IR_REG_NONE && op2_reg != IR_REG_NONE && (op == IR_EQ || op == IR_NE)) {
		ir_ref tmp;
		ir_reg tmp_reg;

		tmp = op1;
		op1 = op2;
		op2 = tmp;
		tmp_reg = op1_reg;
		op1_reg = op2_reg;
		op2_reg = tmp_reg;
	}


	IR_ASSERT(op1_reg != IR_REG_NONE);
	if (IR_REG_SPILLED(op1_reg) || IR_IS_CONST_REF(op1)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, op1);
	}
	if (op2_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op2_reg) || IR_IS_CONST_REF(op2)) {
			op2_reg = IR_REG_NUM(op2_reg);
			if (op1 != op2) {
				ir_emit_load(ctx, type, op2_reg, op2);
			}
		}
		|	ASM_FP_REG_REG_OP ucomiss, ucomisd, vucomiss, vucomisd, type, op1_reg, op2_reg
	} else if (IR_IS_CONST_REF(op2)) {
		ir_insn *val_insn = &ctx->ir_base[op2];
		int label = ctx->cfg_blocks_count - op2;

		val_insn->const_flags |= IR_CONST_EMIT;
		|	ASM_FP_REG_MEM_OP ucomiss, ucomisd, vucomiss, vucomisd, type, op1_reg, [=>label]
	} else {
		int32_t offset = 0;

		if (ir_rule(ctx, op2) & IR_FUSED) {
			offset = ir_fuse_load(ctx, op2, &op2_reg);
		} else {
			offset = ir_ref_spill_slot(ctx, op2, &op2_reg);
		}
		|	ASM_FP_REG_MEM_OP ucomiss, ucomisd, vucomiss, vucomisd, type, op1_reg, [Ra(op2_reg)+offset]
	}
	return op;
}

static void ir_emit_cmp_fp(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_op op = ir_emit_cmp_fp_common(ctx, def, insn);
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg tmp_reg = ctx->regs[def][3];

	IR_ASSERT(def_reg != IR_REG_NONE);
	switch (op) {
		default:
			IR_ASSERT(0 && "NIY binary op");
		case IR_EQ:
			|	setnp Rb(def_reg)
			|	mov Rd(tmp_reg), 0
			|	cmovne Rd(def_reg), Rd(tmp_reg)
			break;
		case IR_NE:
			|	setp Rb(def_reg)
			|	mov Rd(tmp_reg), 1
			|	cmovne Rd(def_reg), Rd(tmp_reg)
			break;
		case IR_LT:
			|	setnp Rb(def_reg)
			|	mov Rd(tmp_reg), 0
			|	cmovae Rd(def_reg), Rd(tmp_reg)
			break;
		case IR_GE:
			|	setae Rb(def_reg)
			break;
		case IR_LE:
			|	setnp Rb(def_reg)
			|	mov Rd(tmp_reg), 0
			|	cmova Rd(def_reg), Rd(tmp_reg)
			break;
		case IR_GT:
			|	seta Rb(def_reg)
			break;
		case IR_ULT:
			|	setb Rb(def_reg)
			break;
		case IR_UGE:
			|	setp Rb(def_reg)
			|	mov Rd(tmp_reg), 1
			|	cmovae Rd(def_reg), Rd(tmp_reg)
			break;
		case IR_ULE:
			|	setbe Rb(def_reg)
			break;
		case IR_UGT:
			|	setp Rb(def_reg)
			|	mov Rd(tmp_reg), 1
			|	cmova Rd(def_reg), Rd(tmp_reg)
			break;
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, insn->type, def, def_reg);
	}
}

static void ir_emit_jmp_true(ir_ctx *ctx, uint32_t b, ir_ref def)
{
	uint32_t true_block, false_block, next_block;
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	ir_get_true_false_blocks(ctx, b, &true_block, &false_block, &next_block);
	if (true_block != next_block) {
		|	jmp =>true_block
	}
}

static void ir_emit_jmp_false(ir_ctx *ctx, uint32_t b, ir_ref def)
{
	uint32_t true_block, false_block, next_block;
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	ir_get_true_false_blocks(ctx, b, &true_block, &false_block, &next_block);
	if (false_block != next_block) {
		|	jmp =>false_block
	}
}

static void ir_emit_jcc(ir_ctx *ctx, uint8_t op, uint32_t b, ir_ref def, ir_insn *insn, bool int_cmp)
{
	uint32_t true_block, false_block, next_block;
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	ir_get_true_false_blocks(ctx, b, &true_block, &false_block, &next_block);
	if (true_block == next_block) {
		/* swap to avoid unconditional JMP */
		if (int_cmp || op == IR_EQ || op == IR_NE) {
			op ^= 1; // reverse
		} else {
			op ^= 5; // reverse
		}
		true_block = false_block;
		false_block = 0;
	} else if (false_block == next_block) {
		false_block = 0;
	}

	if (int_cmp) {
		switch (op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_EQ:
				|	je =>true_block
				break;
			case IR_NE:
				|	jne =>true_block
				break;
			case IR_LT:
				|	jl =>true_block
				break;
			case IR_GE:
				|	jge =>true_block
				break;
			case IR_LE:
				|	jle =>true_block
				break;
			case IR_GT:
				|	jg =>true_block
				break;
			case IR_ULT:
				|	jb =>true_block
				break;
			case IR_UGE:
				|	jae =>true_block
				break;
			case IR_ULE:
				|	jbe =>true_block
				break;
			case IR_UGT:
				|	ja =>true_block
				break;
		}
	} else {
		switch (op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_EQ:
				if (!false_block) {
					|	jp >1
					|	je =>true_block
					|1:
				} else {
					|	jp =>false_block
					|	je =>true_block
				}
				break;
			case IR_NE:
				|	jne =>true_block
				|	jp =>true_block
				break;
			case IR_LT:
				if (!false_block) {
					|	jp >1
					|	jb =>true_block
					|1:
				} else {
					|	jp =>false_block
					|	jb =>true_block
				}
				break;
			case IR_GE:
				|	jae =>true_block
				break;
			case IR_LE:
				if (!false_block) {
					|	jp >1
					|	jbe =>true_block
					|1:
				} else {
					|	jp =>false_block
					|	jbe =>true_block
				}
				break;
			case IR_GT:
				|	ja =>true_block
				break;
			case IR_ULT:
				|	jb =>true_block
				break;
			case IR_UGE:
				|	jp =>true_block
				|	jae =>true_block
				break;
			case IR_ULE:
				|	jbe =>true_block
				break;
			case IR_UGT:
				|	jp =>true_block
				|	ja =>true_block
				break;
		}
	}
	if (false_block) {
		|	jmp =>false_block
	}
}

static void ir_emit_cmp_and_branch_int(ir_ctx *ctx, uint32_t b, ir_ref def, ir_insn *insn)
{
	ir_insn *cmp_insn = &ctx->ir_base[insn->op2];
	ir_op op = cmp_insn->op;
	ir_type type = ctx->ir_base[cmp_insn->op1].type;
	ir_ref op1 = cmp_insn->op1;
	ir_ref op2 = cmp_insn->op2;
	ir_reg op1_reg = ctx->regs[insn->op2][1];
	ir_reg op2_reg = ctx->regs[insn->op2][2];

	if (op1_reg != IR_REG_NONE && (IR_IS_CONST_REF(op1) || IR_REG_SPILLED(op1_reg))) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, op1);
	}
	if (op2_reg != IR_REG_NONE && (IR_IS_CONST_REF(op2) || IR_REG_SPILLED(op2_reg))) {
		op2_reg = IR_REG_NUM(op2_reg);
		if (op1 != op2) {
			ir_emit_load(ctx, type, op2_reg, op2);
		}
	}
	if (IR_IS_CONST_REF(op2) && ctx->ir_base[op2].val.u64 == 0) {
		if (op == IR_ULT) {
			/* always false */
			ir_emit_jmp_false(ctx, b, def);
			return;
		} else if (op == IR_UGE) {
			/* always true */
			ir_emit_jmp_true(ctx, b, def);
			return;
		} else if (op == IR_ULE) {
			op = IR_EQ;
		} else if (op == IR_UGT) {
			op = IR_NE;
		}
	}

	bool same_comparison = 0;
	ir_insn *prev_insn = &ctx->ir_base[insn->op1];
	if (prev_insn->op == IR_IF_TRUE || prev_insn->op == IR_IF_FALSE) {
		if (ir_rule(ctx, prev_insn->op1) == IR_CMP_AND_BRANCH_INT) {
			prev_insn = &ctx->ir_base[prev_insn->op1];
			prev_insn = &ctx->ir_base[prev_insn->op2];
			if (prev_insn->op1 == cmp_insn->op1 && prev_insn->op2 == cmp_insn->op2) {
				same_comparison = true;
			}
		}
	}
	if (!same_comparison) {
		ir_emit_cmp_int_common(ctx, type, cmp_insn, op1_reg, op1, op2_reg, op2);
	}
	ir_emit_jcc(ctx, op, b, def, insn, 1);
}

static void ir_emit_test_and_branch_int(ir_ctx *ctx, uint32_t b, ir_ref def, ir_insn *insn)
{
	ir_ref op2 = insn->op2;
	ir_op op = ctx->ir_base[op2].op;

	if (op >= IR_EQ && op <= IR_UGT) {
		op2 = ctx->ir_base[op2].op1;
	} else {
		IR_ASSERT(op == IR_AND);
		op = IR_NE;
	}

	ir_emit_test_int_common(ctx, op2, op);
	ir_emit_jcc(ctx, op, b, def, insn, 1);
}

static void ir_emit_cmp_and_branch_fp(ir_ctx *ctx, uint32_t b, ir_ref def, ir_insn *insn)
{
	ir_op op = ir_emit_cmp_fp_common(ctx, insn->op2, &ctx->ir_base[insn->op2]);
	ir_emit_jcc(ctx, op, b, def, insn, 0);
}

static void ir_emit_if_int(ir_ctx *ctx, uint32_t b, ir_ref def, ir_insn *insn)
{
	ir_type type = ctx->ir_base[insn->op2].type;
	ir_reg op2_reg = ctx->regs[def][2];
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	if (op2_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op2_reg)) {
			op2_reg = IR_REG_NUM(op2_reg);
			ir_emit_load(ctx, type, op2_reg, insn->op2);
		}
		|	ASM_REG_REG_OP test, type, op2_reg, op2_reg
	} else if (IR_IS_CONST_REF(insn->op2)) {
		uint32_t true_block, false_block, next_block;

		ir_get_true_false_blocks(ctx, b, &true_block, &false_block, &next_block);
		if (ir_const_is_true(&ctx->ir_base[insn->op2])) {
			if (true_block != next_block) {
				|	jmp =>true_block
			}
		} else {
			if (false_block != next_block) {
				|	jmp =>false_block
			}
		}
		return;
	} else {
		int32_t offset = 0;

		if (ir_rule(ctx, insn->op2) & IR_FUSED) {
			offset = ir_fuse_load(ctx, insn->op2, &op2_reg);
		} else {
			offset = ir_ref_spill_slot(ctx, insn->op2, &op2_reg);
		}
		if (op2_reg == IR_REG_NONE) {
			|	ASM_MEM_IMM_OP cmp, type, [offset], 0
		} else {
			|	ASM_MEM_IMM_OP cmp, type, [Ra(op2_reg)+offset], 0
		}
	}
	ir_emit_jcc(ctx, IR_NE, b, def, insn, 1);
}

static void ir_emit_cond(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type = insn->type;
	ir_ref op1 = insn->op1;
	ir_ref op2 = insn->op2;
	ir_ref op3 = insn->op3;
	ir_type op1_type = ctx->ir_base[op1].type;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];
	ir_reg op2_reg = ctx->regs[def][2];
	ir_reg op3_reg = ctx->regs[def][3];

	IR_ASSERT(def_reg != IR_REG_NONE);

	if (op2_reg != IR_REG_NONE && (IR_REG_SPILLED(op2_reg) || IR_IS_CONST_REF(op2))) {
		op2_reg = IR_REG_NUM(op2_reg);
		ir_emit_load(ctx, type, op2_reg, op2);
		if (op1 == op2) {
			op1_reg = op2_reg;
		}
		if (op3 == op2) {
			op3_reg = op2_reg;
		}
	}
	if (op3_reg != IR_REG_NONE && op3 != op2 && (IR_REG_SPILLED(op3_reg) || IR_IS_CONST_REF(op3))) {
		op3_reg = IR_REG_NUM(op3_reg);
		ir_emit_load(ctx, type, op3_reg, op3);
		if (op1 == op2) {
			op1_reg = op3_reg;
		}
	}
	if (op1_reg != IR_REG_NONE && op1 != op2 && op1 != op3 && (IR_REG_SPILLED(op1_reg) || IR_IS_CONST_REF(op1))) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, op1_type, op1_reg, op1);
	}

	if (IR_IS_TYPE_INT(op1_type)) {
		if (op1_reg != IR_REG_NONE) {
			|	ASM_REG_REG_OP test, op1_type, op1_reg, op1_reg
		} else {
			ir_reg fp;
			int32_t offset = ir_ref_spill_slot(ctx, op1, &fp);

			|	ASM_MEM_IMM_OP cmp, op1_type, [Ra(fp)+offset], 0
		}
		|	je >2
	} else {
		if (!data->double_zero_const) {
			data->double_zero_const = 1;
			ir_rodata(ctx);
			|.align 16
			|->double_zero_const:
			|.dword 0, 0
			|.code
		}
		|	ASM_FP_REG_MEM_OP ucomiss, ucomisd, vucomiss, vucomisd, op1_type, op1_reg, [->double_zero_const]
		|	jp >1
		|	je >2
		|1:
	}

	if (op2_reg != IR_REG_NONE) {
		if (def_reg != op2_reg) {
			if (IR_IS_TYPE_INT(type)) {
				ir_emit_mov(ctx, type, def_reg, op2_reg);
			} else {
				ir_emit_fp_mov(ctx, type, def_reg, op2_reg);
			}
		}
	} else if (IR_IS_CONST_REF(op2) || !(ir_rule(ctx, op2) & IR_FUSED)) {
		ir_emit_load(ctx, type, def_reg, op2);
	} else {
		int32_t offset = ir_fuse_load(ctx, op2, &op2_reg);

		if (IR_IS_TYPE_INT(type)) {
			ir_emit_load_mem_int(ctx, type, def_reg, op2_reg, offset);
		} else {
			ir_emit_load_mem_fp(ctx, type, def_reg, op2_reg, offset);
		}
	}
	|	jmp >3
	|2:
	if (op3_reg != IR_REG_NONE) {
		if (def_reg != op3_reg) {
			if (IR_IS_TYPE_INT(type)) {
				ir_emit_mov(ctx, type, def_reg, op3_reg);
			} else {
				ir_emit_fp_mov(ctx, type, def_reg, op3_reg);
			}
		}
	} else if (IR_IS_CONST_REF(op3) || !(ir_rule(ctx, op3) & IR_FUSED)) {
		ir_emit_load(ctx, type, def_reg, op3);
	} else {
		int32_t offset = ir_fuse_load(ctx, op3, &op3_reg);

		if (IR_IS_TYPE_INT(type)) {
			ir_emit_load_mem_int(ctx, type, def_reg, op3_reg, offset);
		} else {
			ir_emit_load_mem_fp(ctx, type, def_reg, op3_reg, offset);
		}
	}
	|3:

	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_return_void(ir_ctx *ctx)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	ir_emit_epilogue(ctx);

#ifdef IR_TARGET_X86
	if (sizeof(void*) == 4 && (ctx->flags & IR_FASTCALL_FUNC) && ctx->param_stack_size) {
		|	ret ctx->param_stack_size
		return;
	}
#endif

	|	ret
}

static void ir_emit_return_int(ir_ctx *ctx, ir_ref ref, ir_insn *insn)
{
	ir_reg op2_reg = ctx->regs[ref][2];

	if (op2_reg != IR_REG_INT_RET1) {
		ir_type type = ctx->ir_base[insn->op2].type;

		if (op2_reg != IR_REG_NONE && !IR_REG_SPILLED(op2_reg)) {
			ir_emit_mov(ctx, type, IR_REG_INT_RET1, op2_reg);
		} else {
			ir_emit_load(ctx, type, IR_REG_INT_RET1, insn->op2);
		}
	}
	ir_emit_return_void(ctx);
}

static void ir_emit_return_fp(ir_ctx *ctx, ir_ref ref, ir_insn *insn)
{
	ir_reg op2_reg = ctx->regs[ref][2];
	ir_type type = ctx->ir_base[insn->op2].type;

#ifdef IR_REG_FP_RET1
	if (op2_reg != IR_REG_FP_RET1) {
		if (op2_reg != IR_REG_NONE && !IR_REG_SPILLED(op2_reg)) {
			ir_emit_fp_mov(ctx, type, IR_REG_FP_RET1, op2_reg);
		} else {
			ir_emit_load(ctx, type, IR_REG_FP_RET1, insn->op2);
		}
	}
#else
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	if (op2_reg == IR_REG_NONE || IR_REG_SPILLED(op2_reg)) {
		ir_reg fp;
		int32_t offset = ir_ref_spill_slot(ctx, insn->op2, &fp);

		if (type == IR_DOUBLE) {
			|	fld qword [Ra(fp)+offset]
		} else {
			IR_ASSERT(type == IR_FLOAT);
			|	fld dword [Ra(fp)+offset]
		}
	} else {
		int32_t offset = ctx->ret_slot;
		ir_reg fp;

		IR_ASSERT(offset != -1);
		offset = IR_SPILL_POS_TO_OFFSET(offset);
		fp = (ctx->flags & IR_USE_FRAME_POINTER) ? IR_REG_FRAME_POINTER : IR_REG_STACK_POINTER;
		ir_emit_store_mem_fp(ctx, type, fp, offset, op2_reg);
		if (type == IR_DOUBLE) {
			|	fld qword [Ra(fp)+offset]
		} else {
			IR_ASSERT(type == IR_FLOAT);
			|	fld dword [Ra(fp)+offset]
		}
	}
#endif
	ir_emit_return_void(ctx);
}

static void ir_emit_sext(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_type dst_type = insn->type;
	ir_type src_type = ctx->ir_base[insn->op1].type;
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];

	IR_ASSERT(IR_IS_TYPE_INT(src_type));
	IR_ASSERT(IR_IS_TYPE_INT(dst_type));
	IR_ASSERT(ir_type_size[dst_type] > ir_type_size[src_type]);
	IR_ASSERT(def_reg != IR_REG_NONE);

	if (op1_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op1_reg)) {
			op1_reg = IR_REG_NUM(op1_reg);
			ir_emit_load(ctx, src_type, op1_reg, insn->op1);
		}
		if (ir_type_size[src_type] == 1) {
			if (ir_type_size[dst_type] == 2) {
				|	movsx Rw(def_reg), Rb(op1_reg)
			} else if (ir_type_size[dst_type] == 4) {
				|	movsx Rd(def_reg), Rb(op1_reg)
			} else {
				IR_ASSERT(ir_type_size[dst_type] == 8);
				IR_ASSERT(sizeof(void*) == 8);
|.if X64
				|	movsx Rq(def_reg), Rb(op1_reg)
|.endif
			}
		} else if (ir_type_size[src_type] == 2) {
			if (ir_type_size[dst_type] == 4) {
				|	movsx Rd(def_reg), Rw(op1_reg)
			} else {
				IR_ASSERT(ir_type_size[dst_type] == 8);
				IR_ASSERT(sizeof(void*) == 8);
|.if X64
				|	movsx Rq(def_reg), Rw(op1_reg)
|.endif
			}
		} else {
			IR_ASSERT(ir_type_size[src_type] == 4);
			IR_ASSERT(ir_type_size[dst_type] == 8);
			IR_ASSERT(sizeof(void*) == 8);
|.if X64
			|	movsxd Rq(def_reg), Rd(op1_reg)
|.endif
		}
	} else if (IR_IS_CONST_REF(insn->op1)) {
		IR_ASSERT(0);
	} else {
		int32_t offset = 0;

		if (ir_rule(ctx, insn->op1) & IR_FUSED) {
			offset = ir_fuse_load(ctx, insn->op1, &op1_reg);
		} else {
			offset = ir_ref_spill_slot(ctx, insn->op1, &op1_reg);
		}

		if (ir_type_size[src_type] == 1) {
			if (ir_type_size[dst_type] == 2) {
				if (op1_reg != IR_REG_NONE) {
					|	movsx Rw(def_reg), byte [Ra(op1_reg)+offset]
				} else {
					|	movsx Rw(def_reg), byte [offset]
				}
			} else if (ir_type_size[dst_type] == 4) {
				if (op1_reg != IR_REG_NONE) {
					|	movsx Rd(def_reg), byte [Ra(op1_reg)+offset]
				} else {
					|	movsx Rd(def_reg), byte [offset]
				}
			} else {
				IR_ASSERT(ir_type_size[dst_type] == 8);
				IR_ASSERT(sizeof(void*) == 8);
|.if X64
				if (op1_reg != IR_REG_NONE) {
					|	movsx Rq(def_reg), byte [Ra(op1_reg)+offset]
				} else {
					|	movsx Rq(def_reg), byte [offset]
				}
|.endif
			}
		} else if (ir_type_size[src_type] == 2) {
			if (ir_type_size[dst_type] == 4) {
				if (op1_reg != IR_REG_NONE) {
					|	movsx Rd(def_reg), word [Ra(op1_reg)+offset]
				} else {
					|	movsx Rd(def_reg), word [offset]
				}
			} else {
				IR_ASSERT(ir_type_size[dst_type] == 8);
				IR_ASSERT(sizeof(void*) == 8);
|.if X64
				if (op1_reg != IR_REG_NONE) {
					|	movsx Rq(def_reg), word [Ra(op1_reg)+offset]
				} else {
					|	movsx Rq(def_reg), word [offset]
				}
|.endif
			}
		} else {
			IR_ASSERT(ir_type_size[src_type] == 4);
			IR_ASSERT(ir_type_size[dst_type] == 8);
			IR_ASSERT(sizeof(void*) == 8);
|.if X64
			if (op1_reg != IR_REG_NONE) {
				|	movsxd Rq(def_reg), dword [Ra(op1_reg)+offset]
			} else {
				|	movsxd Rq(def_reg), dword [offset]
			}
|.endif
		}
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, dst_type, def, def_reg);
	}
}

static void ir_emit_zext(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_type dst_type = insn->type;
	ir_type src_type = ctx->ir_base[insn->op1].type;
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];

	IR_ASSERT(IR_IS_TYPE_INT(src_type));
	IR_ASSERT(IR_IS_TYPE_INT(dst_type));
	IR_ASSERT(ir_type_size[dst_type] > ir_type_size[src_type]);
	IR_ASSERT(def_reg != IR_REG_NONE);

	if (op1_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op1_reg)) {
			op1_reg = IR_REG_NUM(op1_reg);
			ir_emit_load(ctx, src_type, op1_reg, insn->op1);
		}
		if (ir_type_size[src_type] == 1) {
			if (ir_type_size[dst_type] == 2) {
				|	movzx Rw(def_reg), Rb(op1_reg)
			} else if (ir_type_size[dst_type] == 4) {
				|	movzx Rd(def_reg), Rb(op1_reg)
			} else {
				IR_ASSERT(ir_type_size[dst_type] == 8);
				IR_ASSERT(sizeof(void*) == 8);
|.if X64
				|	movzx Rq(def_reg), Rb(op1_reg)
|.endif
			}
		} else if (ir_type_size[src_type] == 2) {
			if (ir_type_size[dst_type] == 4) {
				|	movzx Rd(def_reg), Rw(op1_reg)
			} else {
				IR_ASSERT(ir_type_size[dst_type] == 8);
				IR_ASSERT(sizeof(void*) == 8);
|.if X64
				|	movzx Rq(def_reg), Rw(op1_reg)
|.endif
			}
		} else {
			IR_ASSERT(ir_type_size[src_type] == 4);
			IR_ASSERT(ir_type_size[dst_type] == 8);
			IR_ASSERT(sizeof(void*) == 8);
|.if X64
			/* Avoid zero extension to the same register. This may be not always safe ??? */
			if (op1_reg != def_reg) {
				|	mov Rd(def_reg), Rd(op1_reg)
			}
|.endif
		}
	} else if (IR_IS_CONST_REF(insn->op1)) {
		IR_ASSERT(0);
	} else {
		int32_t offset = 0;

		if (ir_rule(ctx, insn->op1) & IR_FUSED) {
			offset = ir_fuse_load(ctx, insn->op1, &op1_reg);
		} else {
			offset = ir_ref_spill_slot(ctx, insn->op1, &op1_reg);
		}

		if (ir_type_size[src_type] == 1) {
			if (ir_type_size[dst_type] == 2) {
				if (op1_reg != IR_REG_NONE) {
					|	movzx Rw(def_reg), byte [Ra(op1_reg)+offset]
				} else {
					|	movzx Rw(def_reg), byte [offset]
				}
			} else if (ir_type_size[dst_type] == 4) {
				if (op1_reg != IR_REG_NONE) {
					|	movzx Rd(def_reg), byte [Ra(op1_reg)+offset]
				} else {
					|	movzx Rd(def_reg), byte [offset]
				}
			} else {
				IR_ASSERT(ir_type_size[dst_type] == 8);
				IR_ASSERT(sizeof(void*) == 8);
|.if X64
				if (op1_reg != IR_REG_NONE) {
					|	movzx Rq(def_reg), byte [Ra(op1_reg)+offset]
				} else {
					|	movzx Rq(def_reg), byte [offset]
				}
|.endif
			}
		} else if (ir_type_size[src_type] == 2) {
			if (ir_type_size[dst_type] == 4) {
				if (op1_reg != IR_REG_NONE) {
					|	movzx Rd(def_reg), word [Ra(op1_reg)+offset]
				} else {
					|	movzx Rd(def_reg), word [offset]
				}
			} else {
				IR_ASSERT(ir_type_size[dst_type] == 8);
				IR_ASSERT(sizeof(void*) == 8);
|.if X64
				if (op1_reg != IR_REG_NONE) {
					|	movzx Rq(def_reg), word [Ra(op1_reg)+offset]
				} else {
					|	movzx Rq(def_reg), word [offset]
				}
|.endif
			}
		} else {
			IR_ASSERT(ir_type_size[src_type] == 4);
			IR_ASSERT(ir_type_size[dst_type] == 8);
|.if X64
			if (op1_reg != IR_REG_NONE) {
				|	mov Rd(def_reg), dword [Ra(op1_reg)+offset]
			} else {
				|	mov Rd(def_reg), dword [offset]
			}
|.endif
		}
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, dst_type, def, def_reg);
	}
}

static void ir_emit_trunc(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_type dst_type = insn->type;
	ir_type src_type = ctx->ir_base[insn->op1].type;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];

	IR_ASSERT(IR_IS_TYPE_INT(src_type));
	IR_ASSERT(IR_IS_TYPE_INT(dst_type));
	IR_ASSERT(ir_type_size[dst_type] < ir_type_size[src_type]);
	IR_ASSERT(def_reg != IR_REG_NONE);
	if (op1_reg != IR_REG_NONE && IR_REG_SPILLED(op1_reg)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, src_type, op1_reg, insn->op1);
	}
	if (op1_reg != IR_REG_NONE) {
		if (op1_reg != def_reg) {
			ir_emit_mov(ctx, dst_type, def_reg, op1_reg);
		}
	} else {
		ir_emit_load(ctx, dst_type, def_reg, insn->op1);
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, dst_type, def, def_reg);
	}
}

static void ir_emit_bitcast(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_type dst_type = insn->type;
	ir_type src_type = ctx->ir_base[insn->op1].type;
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];
	int32_t offset;

	IR_ASSERT(ir_type_size[dst_type] == ir_type_size[src_type]);
	IR_ASSERT(def_reg != IR_REG_NONE);
	if (IR_IS_TYPE_INT(src_type) && IR_IS_TYPE_INT(dst_type)) {
		if (ir_rule(ctx, insn->op1) & IR_FUSED) {
			offset = ir_fuse_load(ctx, insn->op1, &op1_reg);
			ir_emit_load_mem_int(ctx, dst_type, def_reg, op1_reg, offset);
		} else if (op1_reg != IR_REG_NONE) {
			if (IR_REG_SPILLED(op1_reg)) {
				op1_reg = IR_REG_NUM(op1_reg);
				ir_emit_load(ctx, src_type, op1_reg, insn->op1);
			}
			if (op1_reg != def_reg) {
				ir_emit_mov(ctx, dst_type, def_reg, op1_reg);
			}
		} else {
			ir_emit_load(ctx, dst_type, def_reg, insn->op1);
		}
	} else if (IR_IS_TYPE_FP(src_type) && IR_IS_TYPE_FP(dst_type)) {
		if (ir_rule(ctx, insn->op1) & IR_FUSED) {
			offset = ir_fuse_load(ctx, insn->op1, &op1_reg);
			ir_emit_load_mem_fp(ctx, dst_type, def_reg, op1_reg, offset);
		} else if (op1_reg != IR_REG_NONE) {
			if (IR_REG_SPILLED(op1_reg)) {
				op1_reg = IR_REG_NUM(op1_reg);
				ir_emit_load(ctx, src_type, op1_reg, insn->op1);
			}
			if (op1_reg != def_reg) {
				ir_emit_fp_mov(ctx, dst_type, def_reg, op1_reg);
			}
		} else {
			ir_emit_load(ctx, dst_type, def_reg, insn->op1);
		}
	} else if (IR_IS_TYPE_FP(src_type)) {
		IR_ASSERT(IR_IS_TYPE_INT(dst_type));
		if (op1_reg != IR_REG_NONE) {
			if (IR_REG_SPILLED(op1_reg)) {
				op1_reg = IR_REG_NUM(op1_reg);
				ir_emit_load(ctx, src_type, op1_reg, insn->op1);
			}
			if (src_type == IR_DOUBLE) {
				IR_ASSERT(sizeof(void*) == 8);
|.if X64
				if (ctx->mflags & IR_X86_AVX) {
					|	vmovd Rq(def_reg), xmm(op1_reg-IR_REG_FP_FIRST)
				} else {
					|	movd Rq(def_reg), xmm(op1_reg-IR_REG_FP_FIRST)
				}
|.endif
			} else {
				IR_ASSERT(src_type == IR_FLOAT);
				if (ctx->mflags & IR_X86_AVX) {
					|	vmovd Rd(def_reg), xmm(op1_reg-IR_REG_FP_FIRST)
				} else {
					|	movd Rd(def_reg), xmm(op1_reg-IR_REG_FP_FIRST)
				}
			}
		} else if (IR_IS_CONST_REF(insn->op1)) {
			ir_insn *_insn = &ctx->ir_base[insn->op1];
			if (src_type == IR_DOUBLE) {
				IR_ASSERT(sizeof(void*) == 8);
|.if X64
				|	mov64 Rq(def_reg), _insn->val.i64
|.endif
			} else {
				IR_ASSERT(src_type == IR_FLOAT);
				|	mov Rd(def_reg), _insn->val.i32
			}
		} else {
			int32_t offset = 0;

			if (ir_rule(ctx, insn->op1) & IR_FUSED) {
				offset = ir_fuse_load(ctx, insn->op1, &op1_reg);
			} else {
				offset = ir_ref_spill_slot(ctx, insn->op1, &op1_reg);
			}

			if (src_type == IR_DOUBLE) {
				IR_ASSERT(sizeof(void*) == 8);
|.if X64
				|	mov Rq(def_reg), qword [Ra(op1_reg)+offset]
|.endif
			} else {
				IR_ASSERT(src_type == IR_FLOAT);
				|	mov Rd(def_reg), dword [Ra(op1_reg)+offset]
			}
		}
	} else if (IR_IS_TYPE_FP(dst_type)) {
		IR_ASSERT(IR_IS_TYPE_INT(src_type));
		if (op1_reg != IR_REG_NONE) {
			if (IR_REG_SPILLED(op1_reg)) {
				op1_reg = IR_REG_NUM(op1_reg);
				ir_emit_load(ctx, src_type, op1_reg, insn->op1);
			}
			if (dst_type == IR_DOUBLE) {
				IR_ASSERT(sizeof(void*) == 8);
|.if X64
				if (ctx->mflags & IR_X86_AVX) {
					|	vmovd xmm(def_reg-IR_REG_FP_FIRST), Rq(op1_reg)
				} else {
					|	movd xmm(def_reg-IR_REG_FP_FIRST), Rq(op1_reg)
				}
|.endif
			} else {
				IR_ASSERT(dst_type == IR_FLOAT);
				if (ctx->mflags & IR_X86_AVX) {
					|	vmovd xmm(def_reg-IR_REG_FP_FIRST), Rd(op1_reg)
				} else {
					|	movd xmm(def_reg-IR_REG_FP_FIRST), Rd(op1_reg)
				}
			}
		} else if (IR_IS_CONST_REF(insn->op1)) {
			ir_insn *val_insn = &ctx->ir_base[insn->op1];
			int label = ctx->cfg_blocks_count - insn->op1;

			val_insn->const_flags |= IR_CONST_EMIT;
			|	ASM_FP_REG_MEM_OP  movss, movsd, vmovss, vmovsd, dst_type, def_reg, [=>label]
		} else {
			int32_t offset = 0;

			if (ir_rule(ctx, insn->op1) & IR_FUSED) {
				offset = ir_fuse_load(ctx, insn->op1, &op1_reg);
			} else {
				offset = ir_ref_spill_slot(ctx, insn->op1, &op1_reg);
			}

			|	ASM_FP_REG_MEM_OP movss, movsd, vmovss, vmovsd, dst_type, def_reg, [Ra(op1_reg)+offset]
		}
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, dst_type, def, def_reg);
	}
}

static void ir_emit_int2fp(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_type dst_type = insn->type;
	ir_type src_type = ctx->ir_base[insn->op1].type;
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];
	bool src64 = 0;

	IR_ASSERT(IR_IS_TYPE_INT(src_type));
	IR_ASSERT(IR_IS_TYPE_FP(dst_type));
	IR_ASSERT(def_reg != IR_REG_NONE);
	if (IR_IS_TYPE_SIGNED(src_type) ? ir_type_size[src_type] == 8 : ir_type_size[src_type] >= 4) {
		// TODO: we might need to perform sign/zero integer extension to 32/64 bit integer
		src64 = 1;
	}
	if (op1_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op1_reg) || IR_IS_CONST_REF(insn->op1)) {
			op1_reg = IR_REG_NUM(op1_reg);
			ir_emit_load(ctx, src_type, op1_reg, insn->op1);
		}
		if (!src64) {
			if (dst_type == IR_DOUBLE) {
				if (ctx->mflags & IR_X86_AVX) {
					|	vxorps xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	vcvtsi2sd xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), Rd(op1_reg)
				} else {
					|	pxor xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	cvtsi2sd xmm(def_reg-IR_REG_FP_FIRST), Rd(op1_reg)
				}
			} else {
				IR_ASSERT(dst_type == IR_FLOAT);
				if (ctx->mflags & IR_X86_AVX) {
					|	vxorps xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	vcvtsi2ss xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), Rd(op1_reg)
				} else {
					|	pxor xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	cvtsi2ss xmm(def_reg-IR_REG_FP_FIRST), Rd(op1_reg)
				}
			}
		} else {
			IR_ASSERT(sizeof(void*) == 8);
|.if X64
			if (dst_type == IR_DOUBLE) {
				if (ctx->mflags & IR_X86_AVX) {
					|	vxorps xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	vcvtsi2sd xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), Rq(op1_reg)
				} else {
					|	pxor xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	cvtsi2sd xmm(def_reg-IR_REG_FP_FIRST), Rq(op1_reg)
				}
			} else {
				IR_ASSERT(dst_type == IR_FLOAT);
				if (ctx->mflags & IR_X86_AVX) {
					|	vxorps xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	vcvtsi2ss xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), Rq(op1_reg)
				} else {
					|	pxor xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	cvtsi2ss xmm(def_reg-IR_REG_FP_FIRST), Rq(op1_reg)
				}
			}
|.endif
		}
	} else {
		int32_t offset = 0;

		if (ir_rule(ctx, insn->op1) & IR_FUSED) {
			offset = ir_fuse_load(ctx, insn->op1, &op1_reg);
		} else {
			offset = ir_ref_spill_slot(ctx, insn->op1, &op1_reg);
		}

		if (!src64) {
			if (dst_type == IR_DOUBLE) {
				if (ctx->mflags & IR_X86_AVX) {
					|	vxorps xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	vcvtsi2sd xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), dword [Ra(op1_reg)+offset]
				} else {
					|	pxor xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	cvtsi2sd xmm(def_reg-IR_REG_FP_FIRST), dword [Ra(op1_reg)+offset]
				}
			} else {
				IR_ASSERT(dst_type == IR_FLOAT);
				if (ctx->mflags & IR_X86_AVX) {
					|	vxorps xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	vcvtsi2ss xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), dword [Ra(op1_reg)+offset]
				} else {
					|	pxor xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	cvtsi2ss xmm(def_reg-IR_REG_FP_FIRST), dword [Ra(op1_reg)+offset]
				}
			}
		} else {
			IR_ASSERT(sizeof(void*) == 8);
|.if X64
			if (dst_type == IR_DOUBLE) {
				if (ctx->mflags & IR_X86_AVX) {
					|	vxorps xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	vcvtsi2sd xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), qword [Ra(op1_reg)+offset]
				} else {
					|	pxor xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	cvtsi2sd xmm(def_reg-IR_REG_FP_FIRST), qword [Ra(op1_reg)+offset]
				}
			} else {
				IR_ASSERT(dst_type == IR_FLOAT);
				if (ctx->mflags & IR_X86_AVX) {
					|	vxorps xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	vcvtsi2ss xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), qword [Ra(op1_reg)+offset]
				} else {
					|	pxor xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST)
					|	cvtsi2ss xmm(def_reg-IR_REG_FP_FIRST), qword [Ra(op1_reg)+offset]
				}
			}
|.endif
		}
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, dst_type, def, def_reg);
	}
}

static void ir_emit_fp2int(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_type dst_type = insn->type;
	ir_type src_type = ctx->ir_base[insn->op1].type;
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];
	bool dst64 = 0;

	IR_ASSERT(IR_IS_TYPE_FP(src_type));
	IR_ASSERT(IR_IS_TYPE_INT(dst_type));
	IR_ASSERT(def_reg != IR_REG_NONE);
	if (IR_IS_TYPE_SIGNED(dst_type) ? ir_type_size[dst_type] == 8 : ir_type_size[dst_type] >= 4) {
		// TODO: we might need to perform truncation from 32/64 bit integer
		dst64 = 1;
	}
	if (op1_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op1_reg)) {
			op1_reg = IR_REG_NUM(op1_reg);
			ir_emit_load(ctx, src_type, op1_reg, insn->op1);
		}
		if (!dst64) {
			if (src_type == IR_DOUBLE) {
				if (ctx->mflags & IR_X86_AVX) {
					|	vcvtsd2si Rd(def_reg), xmm(op1_reg-IR_REG_FP_FIRST)
				} else {
					|	cvtsd2si Rd(def_reg), xmm(op1_reg-IR_REG_FP_FIRST)
				}
			} else {
				IR_ASSERT(src_type == IR_FLOAT);
				if (ctx->mflags & IR_X86_AVX) {
					|	vcvtss2si Rd(def_reg), xmm(op1_reg-IR_REG_FP_FIRST)
				} else {
					|	cvtss2si Rd(def_reg), xmm(op1_reg-IR_REG_FP_FIRST)
				}
			}
		} else {
			IR_ASSERT(sizeof(void*) == 8);
|.if X64
			if (src_type == IR_DOUBLE) {
				if (ctx->mflags & IR_X86_AVX) {
					|	vcvtsd2si Rq(def_reg), xmm(op1_reg-IR_REG_FP_FIRST)
				} else {
					|	cvtsd2si Rq(def_reg), xmm(op1_reg-IR_REG_FP_FIRST)
				}
			} else {
				IR_ASSERT(src_type == IR_FLOAT);
				if (ctx->mflags & IR_X86_AVX) {
					|	vcvtss2si Rq(def_reg), xmm(op1_reg-IR_REG_FP_FIRST)
				} else {
					|	cvtss2si Rq(def_reg), xmm(op1_reg-IR_REG_FP_FIRST)
				}
			}
|.endif
		}
	} else if (IR_IS_CONST_REF(insn->op1)) {
		ir_insn *_insn = &ctx->ir_base[insn->op1];
		int label = ctx->cfg_blocks_count - insn->op1;

		_insn->const_flags |= IR_CONST_EMIT;
		if (!dst64) {
			if (src_type == IR_DOUBLE) {
				if (ctx->mflags & IR_X86_AVX) {
					|	vcvtsd2si Rd(def_reg), qword [=>label]
				} else {
					|	cvtsd2si Rd(def_reg), qword [=>label]
				}
			} else {
				IR_ASSERT(src_type == IR_FLOAT);
				if (ctx->mflags & IR_X86_AVX) {
					|	vcvtss2si Rd(def_reg), dword [=>label]
				} else {
					|	cvtss2si Rd(def_reg), dword [=>label]
				}
			}
		} else {
			IR_ASSERT(sizeof(void*) == 8);
|.if X64
			if (src_type == IR_DOUBLE) {
				if (ctx->mflags & IR_X86_AVX) {
					|	vcvtsd2si Rq(def_reg), qword [=>label]
				} else {
					|	cvtsd2si Rq(def_reg), qword [=>label]
				}
			} else {
				IR_ASSERT(src_type == IR_FLOAT);
				if (ctx->mflags & IR_X86_AVX) {
					|	vcvtss2si Rq(def_reg), dword [=>label]
				} else {
					|	cvtss2si Rq(def_reg), dword [=>label]
				}
			}
|.endif
		}
	} else {
		int32_t offset = 0;

		if (ir_rule(ctx, insn->op1) & IR_FUSED) {
			offset = ir_fuse_load(ctx, insn->op1, &op1_reg);
		} else {
			offset = ir_ref_spill_slot(ctx, insn->op1, &op1_reg);
		}

		if (!dst64) {
			if (src_type == IR_DOUBLE) {
				if (ctx->mflags & IR_X86_AVX) {
					|	vcvtsd2si Rd(def_reg), qword [Ra(op1_reg)+offset]
				} else {
					|	cvtsd2si Rd(def_reg), qword [Ra(op1_reg)+offset]
				}
			} else {
				IR_ASSERT(src_type == IR_FLOAT);
				if (ctx->mflags & IR_X86_AVX) {
					|	vcvtss2si Rd(def_reg), dword [Ra(op1_reg)+offset]
				} else {
					|	cvtss2si Rd(def_reg), dword [Ra(op1_reg)+offset]
				}
			}
		} else {
			IR_ASSERT(sizeof(void*) == 8);
|.if X64
			if (src_type == IR_DOUBLE) {
				if (ctx->mflags & IR_X86_AVX) {
					|	vcvtsd2si Rq(def_reg), qword [Ra(op1_reg)+offset]
				} else {
					|	cvtsd2si Rq(def_reg), qword [Ra(op1_reg)+offset]
				}
			} else {
				IR_ASSERT(src_type == IR_FLOAT);
				if (ctx->mflags & IR_X86_AVX) {
					|	vcvtss2si Rq(def_reg), dword [Ra(op1_reg)+offset]
				} else {
					|	cvtss2si Rq(def_reg), dword [Ra(op1_reg)+offset]
				}
			}
|.endif
		}
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, dst_type, def, def_reg);
	}
}

static void ir_emit_fp2fp(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_type dst_type = insn->type;
	ir_type src_type = ctx->ir_base[insn->op1].type;
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];

	IR_ASSERT(IR_IS_TYPE_FP(src_type));
	IR_ASSERT(IR_IS_TYPE_FP(dst_type));
	IR_ASSERT(def_reg != IR_REG_NONE);
	if (op1_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op1_reg)) {
			op1_reg = IR_REG_NUM(op1_reg);
			ir_emit_load(ctx, src_type, op1_reg, insn->op1);
		}
		if (src_type == dst_type) {
			if (op1_reg != def_reg) {
				ir_emit_fp_mov(ctx, dst_type, def_reg, op1_reg);
			}
		} else if (src_type == IR_DOUBLE) {
			if (ctx->mflags & IR_X86_AVX) {
				|	vcvtsd2ss xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), xmm(op1_reg-IR_REG_FP_FIRST)
			} else {
				|	cvtsd2ss xmm(def_reg-IR_REG_FP_FIRST), xmm(op1_reg-IR_REG_FP_FIRST)
			}
		} else {
			IR_ASSERT(src_type == IR_FLOAT);
			if (ctx->mflags & IR_X86_AVX) {
				|	vcvtss2sd xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), xmm(op1_reg-IR_REG_FP_FIRST)
			} else {
				|	cvtss2sd xmm(def_reg-IR_REG_FP_FIRST), xmm(op1_reg-IR_REG_FP_FIRST)
			}
		}
	} else if (IR_IS_CONST_REF(insn->op1)) {
		ir_insn *_insn = &ctx->ir_base[insn->op1];
		int label = ctx->cfg_blocks_count - insn->op1;

		_insn->const_flags |= IR_CONST_EMIT;
		if (src_type == IR_DOUBLE) {
			if (ctx->mflags & IR_X86_AVX) {
				|	vcvtsd2ss xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), qword [=>label]
			} else {
				|	cvtsd2ss xmm(def_reg-IR_REG_FP_FIRST), qword [=>label]
			}
		} else {
			IR_ASSERT(src_type == IR_FLOAT);
			if (ctx->mflags & IR_X86_AVX) {
				|	vcvtss2sd xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), dword [=>label]
			} else {
				|	cvtss2sd xmm(def_reg-IR_REG_FP_FIRST), dword [=>label]
			}
		}
	} else {
		int32_t offset = 0;

		if (ir_rule(ctx, insn->op1) & IR_FUSED) {
			offset = ir_fuse_load(ctx, insn->op1, &op1_reg);
		} else {
			offset = ir_ref_spill_slot(ctx, insn->op1, &op1_reg);
		}

		if (src_type == IR_DOUBLE) {
			if (ctx->mflags & IR_X86_AVX) {
				|	vcvtsd2ss xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), qword [Ra(op1_reg)+offset]
			} else {
				|	cvtsd2ss xmm(def_reg-IR_REG_FP_FIRST), qword [Ra(op1_reg)+offset]
			}
		} else {
			IR_ASSERT(src_type == IR_FLOAT);
			if (ctx->mflags & IR_X86_AVX) {
				|	vcvtss2sd xmm(def_reg-IR_REG_FP_FIRST), xmm(def_reg-IR_REG_FP_FIRST), dword [Ra(op1_reg)+offset]
			} else {
				|	cvtss2sd xmm(def_reg-IR_REG_FP_FIRST), dword [Ra(op1_reg)+offset]
			}
		}
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, dst_type, def, def_reg);
	}
}

static void ir_emit_copy_int(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_ref type = insn->type;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];

	IR_ASSERT(def_reg != IR_REG_NONE || op1_reg != IR_REG_NONE);
	if (op1_reg != IR_REG_NONE && IR_REG_SPILLED(op1_reg)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, insn->op1);
	}
	if (def_reg == op1_reg) {
		/* same reg */
	} else if (def_reg != IR_REG_NONE && op1_reg != IR_REG_NONE) {
		ir_emit_mov(ctx, type, def_reg, op1_reg);
	} else if (def_reg != IR_REG_NONE) {
		ir_emit_load(ctx, type, def_reg, insn->op1);
	} else if (op1_reg != IR_REG_NONE) {
		ir_emit_store(ctx, type, def, op1_reg);
	} else {
		IR_ASSERT(0);
	}
	if (def_reg != IR_REG_NONE && IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_copy_fp(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_type type = insn->type;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg op1_reg = ctx->regs[def][1];

	IR_ASSERT(def_reg != IR_REG_NONE || op1_reg != IR_REG_NONE);
	if (op1_reg != IR_REG_NONE && IR_REG_SPILLED(op1_reg)) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, insn->op1);
	}
	if (def_reg == op1_reg) {
		/* same reg */
	} else if (def_reg != IR_REG_NONE && op1_reg != IR_REG_NONE) {
		ir_emit_fp_mov(ctx, type, def_reg, op1_reg);
	} else if (def_reg != IR_REG_NONE) {
		ir_emit_load(ctx, type, def_reg, insn->op1);
	} else if (op1_reg != IR_REG_NONE) {
		ir_emit_store(ctx, type, def, op1_reg);
	} else {
		IR_ASSERT(0);
	}
	if (def_reg != IR_REG_NONE && IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_vaddr(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_ref type = insn->type;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	int32_t offset;
	ir_reg fp;

	IR_ASSERT(def_reg != IR_REG_NONE);
	offset = ir_var_spill_slot(ctx, insn->op1, &fp);
	|	lea Ra(def_reg), aword [Ra(fp)+offset]
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_vload(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_insn *var_insn = &ctx->ir_base[insn->op2];
	ir_ref type = insn->type;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	ir_reg fp;
	int32_t offset;

	IR_ASSERT(var_insn->op == IR_VAR);
	fp = (ctx->flags & IR_USE_FRAME_POINTER) ? IR_REG_FRAME_POINTER : IR_REG_STACK_POINTER;
	offset = IR_SPILL_POS_TO_OFFSET(var_insn->op3);
	if (def_reg == IR_REG_NONE && ir_is_same_mem_var(ctx, def, var_insn->op3)) {
		return; // fake load
	}
	IR_ASSERT(def_reg != IR_REG_NONE);

	if (IR_IS_TYPE_INT(type)) {
		ir_emit_load_mem_int(ctx, type, def_reg, fp, offset);
	} else {
		ir_emit_load_mem_fp(ctx, type, def_reg, fp, offset);
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_vstore_int(ir_ctx *ctx, ir_ref ref, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_insn *var_insn = &ctx->ir_base[insn->op2];
	ir_insn *val_insn = &ctx->ir_base[insn->op3];
	ir_ref type = val_insn->type;
	ir_reg op3_reg = ctx->regs[ref][3];
	ir_reg fp;
	int32_t offset;

	IR_ASSERT(var_insn->op == IR_VAR);
	fp = (ctx->flags & IR_USE_FRAME_POINTER) ? IR_REG_FRAME_POINTER : IR_REG_STACK_POINTER;
	offset = IR_SPILL_POS_TO_OFFSET(var_insn->op3);
	if ((op3_reg == IR_REG_NONE || IR_REG_SPILLED(op3_reg))
	 && !IR_IS_CONST_REF(insn->op3) && ir_is_same_mem_var(ctx, insn->op3, var_insn->op3)) {
		return; // fake store
	}
	if (IR_IS_CONST_REF(insn->op3) && IR_IS_32BIT(type, val_insn->val)) {
		|	ASM_MEM_IMM_OP mov, type, [Ra(fp)+offset], val_insn->val.i32
	} else {
		IR_ASSERT(op3_reg != IR_REG_NONE);
		if (IR_REG_SPILLED(op3_reg) || IR_IS_CONST_REF(insn->op3)) {
			op3_reg = IR_REG_NUM(op3_reg);
			ir_emit_load(ctx, type, op3_reg, insn->op3);
		}
		ir_emit_store_mem_int(ctx, type, fp, offset, op3_reg);
	}
}

static void ir_emit_vstore_fp(ir_ctx *ctx, ir_ref ref, ir_insn *insn)
{
	ir_insn *var_insn = &ctx->ir_base[insn->op2];
	ir_ref type = ctx->ir_base[insn->op3].type;
	ir_reg op3_reg = ctx->regs[ref][3];
	ir_reg fp;
	int32_t offset;

	IR_ASSERT(var_insn->op == IR_VAR);
	fp = (ctx->flags & IR_USE_FRAME_POINTER) ? IR_REG_FRAME_POINTER : IR_REG_STACK_POINTER;
	offset = IR_SPILL_POS_TO_OFFSET(var_insn->op3);
	if ((op3_reg == IR_REG_NONE || IR_REG_SPILLED(op3_reg))
	 && !IR_IS_CONST_REF(insn->op3) && ir_is_same_mem_var(ctx, insn->op3, var_insn->op3)) {
		return; // fake store
	}
	IR_ASSERT(op3_reg != IR_REG_NONE);
	if (IR_REG_SPILLED(op3_reg) || IR_IS_CONST_REF(insn->op3)) {
		op3_reg = IR_REG_NUM(op3_reg);
		ir_emit_load(ctx, type, op3_reg, insn->op3);
	}
	ir_emit_store_mem_fp(ctx, type, fp, offset, op3_reg);
}

static void ir_emit_load_int(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_ref type = insn->type;
	ir_reg op2_reg = ctx->regs[def][2];
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	int32_t offset = 0;

	if (ctx->use_lists[def].count == 1) {
		/* dead load */
		return;
	}
	IR_ASSERT(def_reg != IR_REG_NONE);
	if (IR_IS_CONST_REF(insn->op2)) {
		void *addr = (void*)ctx->ir_base[insn->op2].val.addr;

		if (sizeof(void*) == 4 || IR_IS_SIGNED_32BIT(addr)) {
			int32_t addr32 = (int32_t)(intptr_t)addr;
			|	ASM_REG_MEM_OP mov, type, def_reg, [addr32]
			if (IR_REG_SPILLED(ctx->regs[def][0])) {
				ir_emit_store(ctx, type, def, def_reg);
			}
			return;
		}
	}
	if (op2_reg == IR_REG_NONE) {
		op2_reg = def_reg;
	}
	if (!IR_IS_CONST_REF(insn->op2) && (ir_rule(ctx, insn->op2) & IR_FUSED)) {
		offset = ir_fuse_addr(ctx, insn->op2, &op2_reg);
		if (IR_REG_SPILLED(ctx->regs[def][0]) && ir_is_same_spill_slot(ctx, def, op2_reg, offset)) {
			if (!ir_may_avoid_spill_load(ctx, def, def)) {
				ir_emit_load_mem_int(ctx, type, def_reg, op2_reg, offset);
			}
			/* avoid load to the same location (valid only when register is not reused) */
			return;
		}
	} else if (IR_REG_SPILLED(op2_reg) || IR_IS_CONST_REF(insn->op2)) {
		op2_reg = IR_REG_NUM(op2_reg);
		IR_ASSERT(ctx->ir_base[insn->op2].type == IR_ADDR);
		ir_emit_load(ctx, IR_ADDR, op2_reg, insn->op2);
	}

	ir_emit_load_mem_int(ctx, type, def_reg, op2_reg, offset);
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_load_fp(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_ref type = insn->type;
	ir_reg op2_reg = ctx->regs[def][2];
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);
	int32_t offset = 0;

	if (ctx->use_lists[def].count == 1) {
		/* dead load */
		return;
	}
	IR_ASSERT(def_reg != IR_REG_NONE);
	if (IR_IS_CONST_REF(insn->op2)) {
		if (op2_reg == IR_REG_NONE) {
			int32_t addr32 = ctx->ir_base[insn->op2].val.i32;

			IR_ASSERT(sizeof(void*) == 4 || IR_IS_SIGNED_32BIT(ctx->ir_base[insn->op2].val.i64));
			|	ASM_FP_REG_MEM_OP movss, movsd, vmovss, vmovsd, type, def_reg, [addr32]
			if (IR_REG_SPILLED(ctx->regs[def][0])) {
				ir_emit_store(ctx, type, def, def_reg);
			}
			return;
		} else {
			ir_emit_load(ctx, IR_ADDR, op2_reg, insn->op2);
		}
	} else if (op2_reg == IR_REG_NONE) {
		offset = ir_fuse_addr(ctx, insn->op2, &op2_reg);
		if (IR_REG_SPILLED(ctx->regs[def][0]) && ir_is_same_spill_slot(ctx, def, op2_reg, offset)) {
			if (!ir_may_avoid_spill_load(ctx, def, def)) {
				ir_emit_load_mem_fp(ctx, type, def_reg, op2_reg, offset);
			}
			/* avoid load to the same location (valid only when register is not reused) */
			return;
		}
	} else if (IR_REG_SPILLED(op2_reg)) {
		op2_reg = IR_REG_NUM(op2_reg);
		IR_ASSERT(ctx->ir_base[insn->op2].type == IR_ADDR);
		ir_emit_load(ctx, IR_ADDR, op2_reg, insn->op2);
	}

	ir_emit_load_mem_fp(ctx, type, def_reg, op2_reg, offset);
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_store_int(ir_ctx *ctx, ir_ref ref, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_insn *val_insn = &ctx->ir_base[insn->op3];
	ir_ref type = val_insn->type;
	ir_reg op2_reg = ctx->regs[ref][2];
	ir_reg op3_reg = ctx->regs[ref][3];
	int32_t offset = 0;

	if (IR_IS_CONST_REF(insn->op2)) {

		if (op2_reg == IR_REG_NONE) {
			int32_t addr32 = ctx->ir_base[insn->op2].val.i32;

			IR_ASSERT(sizeof(void*) == 4 || IR_IS_SIGNED_32BIT(ctx->ir_base[insn->op2].val.i64));
			if (IR_IS_CONST_REF(insn->op3) && IR_IS_32BIT(type, val_insn->val)) {
				|	ASM_MEM_IMM_OP mov, type, [addr32], val_insn->val.i32
			} else {
				IR_ASSERT(op3_reg != IR_REG_NONE);
				if (IR_REG_SPILLED(op3_reg) || IR_IS_CONST_REF(insn->op3)) {
					op3_reg = IR_REG_NUM(op3_reg);
					ir_emit_load(ctx, type, op3_reg, insn->op3);
				}
				|	ASM_MEM_REG_OP mov, type, [addr32], op3_reg
			}
			return;
		} else {
			ir_emit_load(ctx, IR_ADDR, op2_reg, insn->op2);
		}
	} else if (op2_reg == IR_REG_NONE) {
		offset = ir_fuse_addr(ctx, insn->op2, &op2_reg);
		if (!IR_IS_CONST_REF(insn->op3) && IR_REG_SPILLED(op3_reg) && ir_is_same_spill_slot(ctx, insn->op3, op2_reg, offset)) {
			if (!ir_may_avoid_spill_load(ctx, insn->op3, ref)) {
				op3_reg = IR_REG_NUM(op3_reg);
				ir_emit_load(ctx, type, op3_reg, insn->op3);
			}
			/* avoid store to the same location */
			return;
		}
	} else if (IR_REG_SPILLED(op2_reg)) {
		op2_reg = IR_REG_NUM(op2_reg);
		IR_ASSERT(ctx->ir_base[insn->op2].type == IR_ADDR);
		ir_emit_load(ctx, IR_ADDR, op2_reg, insn->op2);
	}

	if (IR_IS_CONST_REF(insn->op3) && IR_IS_32BIT(type, val_insn->val)) {
		|	ASM_MEM_IMM_OP mov, type, [Ra(op2_reg)+offset], val_insn->val.i32
	} else {
		IR_ASSERT(op3_reg != IR_REG_NONE);
		if (IR_REG_SPILLED(op3_reg) || IR_IS_CONST_REF(insn->op3)) {
			op3_reg = IR_REG_NUM(op3_reg);
			ir_emit_load(ctx, type, op3_reg, insn->op3);
		}
		ir_emit_store_mem_int(ctx, type, op2_reg, offset, op3_reg);
	}
}

static void ir_emit_store_fp(ir_ctx *ctx, ir_ref ref, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_ref type = ctx->ir_base[insn->op3].type;
	ir_reg op2_reg = ctx->regs[ref][2];
	ir_reg op3_reg = ctx->regs[ref][3];
	int32_t offset = 0;

	IR_ASSERT(op3_reg != IR_REG_NONE);
	if (IR_IS_CONST_REF(insn->op2)) {
		if (op2_reg == IR_REG_NONE) {
			int32_t addr32 = ctx->ir_base[insn->op2].val.i32;

			IR_ASSERT(sizeof(void*) == 4 || IR_IS_SIGNED_32BIT(ctx->ir_base[insn->op2].val.i64));
			if (IR_REG_SPILLED(op3_reg) || IR_IS_CONST_REF(insn->op3)) {
				op3_reg = IR_REG_NUM(op3_reg);
				ir_emit_load(ctx, type, op3_reg, insn->op3);
			}
			|	ASM_FP_MEM_REG_OP movss, movsd, vmovss, vmovsd, type, [addr32], op3_reg
			return;
		} else {
			ir_emit_load(ctx, IR_ADDR, op2_reg, insn->op2);
		}
	} else if (op2_reg == IR_REG_NONE) {
		offset = ir_fuse_addr(ctx, insn->op2, &op2_reg);
		if (!IR_IS_CONST_REF(insn->op3) && IR_REG_SPILLED(op3_reg) && ir_is_same_spill_slot(ctx, insn->op3, op2_reg, offset)) {
			if (!ir_may_avoid_spill_load(ctx, insn->op3, ref)) {
				op3_reg = IR_REG_NUM(op3_reg);
				ir_emit_load(ctx, type, op3_reg, insn->op3);
			}
			/* avoid store to the same location */
			return;
		}
	} else if (IR_REG_SPILLED(op2_reg)) {
		op2_reg = IR_REG_NUM(op2_reg);
		IR_ASSERT(ctx->ir_base[insn->op2].type == IR_ADDR);
		ir_emit_load(ctx, IR_ADDR, op2_reg, insn->op2);
	}

	if (IR_REG_SPILLED(op3_reg) || IR_IS_CONST_REF(insn->op3)) {
		op3_reg = IR_REG_NUM(op3_reg);
		ir_emit_load(ctx, type, op3_reg, insn->op3);
	}
	ir_emit_store_mem_fp(ctx, type, op2_reg, offset, op3_reg);
}

static void ir_emit_rload(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_reg src_reg = insn->op2;
	ir_type type = insn->type;

	if (IR_REGSET_IN(IR_REGSET_UNION((ir_regset)ctx->fixed_regset, IR_REGSET_FIXED), src_reg)) {
		if (ctx->vregs[def]
		 && ctx->live_intervals[ctx->vregs[def]]
		 && ctx->live_intervals[ctx->vregs[def]]->stack_spill_pos != -1) {
			ir_emit_store(ctx, type, def, src_reg);
		}
	} else {
		ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);

		if (def_reg == IR_REG_NONE) {
			/* op3 is used as a flag that the value is already stored in memory.
			 * If op3 is set we don't have to store the value once again (in case of spilling)
			 */
			if (!insn->op3 || !ir_is_same_spill_slot(ctx, def, ctx->spill_base, insn->op3)) {
				ir_emit_store(ctx, type, def, src_reg);
			}
		} else {
			if (src_reg != def_reg) {
				if (IR_IS_TYPE_INT(type)) {
					ir_emit_mov(ctx, type, def_reg, src_reg);
				} else {
					IR_ASSERT(IR_IS_TYPE_FP(type));
					ir_emit_fp_mov(ctx, type, def_reg, src_reg);
				}
			}
			if (IR_REG_SPILLED(ctx->regs[def][0])
			 && (!insn->op3 || !ir_is_same_spill_slot(ctx, def,  ctx->spill_base, insn->op3))) {
				ir_emit_store(ctx, type, def, def_reg);
			}
		}
	}
}

static void ir_emit_rstore(ir_ctx *ctx, ir_ref ref, ir_insn *insn)
{
	ir_ref type = ctx->ir_base[insn->op2].type;
	ir_reg op2_reg = ctx->regs[ref][2];
	ir_reg dst_reg = insn->op3;

	if (!IR_IS_CONST_REF(insn->op2) && (ir_rule(ctx, insn->op2) & IR_FUSED)) {
		int32_t offset = ir_fuse_load(ctx, insn->op2, &op2_reg);
		if (IR_IS_TYPE_INT(type)) {
			ir_emit_load_mem_int(ctx, type, dst_reg, op2_reg, offset);
		} else {
			ir_emit_load_mem_fp(ctx, type, dst_reg, op2_reg, offset);
		}
	} else if (op2_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op2_reg)) {
			op2_reg = IR_REG_NUM(op2_reg);
			ir_emit_load(ctx, type, op2_reg, insn->op2);
		}
		if (op2_reg != dst_reg) {
			if (IR_IS_TYPE_INT(type)) {
				ir_emit_mov(ctx, type, dst_reg, op2_reg);
			} else {
				IR_ASSERT(IR_IS_TYPE_FP(type));
				ir_emit_fp_mov(ctx, type, dst_reg, op2_reg);
			}
		}
	} else {
		ir_emit_load(ctx, type, dst_reg, insn->op2);
	}
}

static void ir_emit_alloca(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);

	if (IR_IS_CONST_REF(insn->op2)) {
		ir_insn *val = &ctx->ir_base[insn->op2];
		int32_t size = val->val.i32;

		IR_ASSERT(IR_IS_TYPE_INT(val->type));
		IR_ASSERT(IR_IS_TYPE_UNSIGNED(val->type) || val->val.i64 > 0);
		IR_ASSERT(IR_IS_SIGNED_32BIT(val->val.i64));

		if (ctx->flags2 & IR_HAS_CALLS) {
			/* Stack must be 16 byte aligned */
			size = IR_ALIGNED_SIZE(size, 16);
		} else {
			size = IR_ALIGNED_SIZE(size, 8);
		}
		|	ASM_REG_IMM_OP sub, IR_ADDR, IR_REG_RSP, size
		if (!(ctx->flags & IR_USE_FRAME_POINTER)) {
			ctx->call_stack_size += size;
		}
	} else {
		int32_t alignment = (ctx->flags2 & IR_HAS_CALLS) ? 16 : 8;
		ir_reg op2_reg = ctx->regs[def][2];
		ir_type type = ctx->ir_base[insn->op2].type;

		IR_ASSERT(ctx->flags & IR_FUNCTION);
		IR_ASSERT(ctx->flags & IR_USE_FRAME_POINTER);
		IR_ASSERT(def_reg != IR_REG_NONE);
		if (op2_reg != IR_REG_NONE && IR_REG_SPILLED(op2_reg)) {
			op2_reg = IR_REG_NUM(op2_reg);
			ir_emit_load(ctx, type, op2_reg, insn->op2);
		}
		if (def_reg != op2_reg) {
			if (op2_reg != IR_REG_NONE) {
				ir_emit_mov(ctx, type, def_reg, op2_reg);
			} else {
				ir_emit_load(ctx, type, def_reg, insn->op2);
			}
		}

		|	ASM_REG_IMM_OP add, IR_ADDR, def_reg, (alignment-1)
		|	ASM_REG_IMM_OP and, IR_ADDR, def_reg, ~(alignment-1)
		|	ASM_REG_REG_OP sub, IR_ADDR, IR_REG_RSP, def_reg
	}
	if (def_reg != IR_REG_NONE) {
		|	mov Ra(def_reg), Ra(IR_REG_RSP)
		if (IR_REG_SPILLED(ctx->regs[def][0])) {
			ir_emit_store(ctx, insn->type, def, def_reg);
		}
	} else {
		ir_emit_store(ctx, IR_ADDR, def, IR_REG_STACK_POINTER);
	}
}

static void ir_emit_afree(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;

	if (IR_IS_CONST_REF(insn->op2)) {
		ir_insn *val = &ctx->ir_base[insn->op2];
		int32_t size = val->val.i32;

		IR_ASSERT(IR_IS_TYPE_INT(val->type));
		IR_ASSERT(IR_IS_TYPE_UNSIGNED(val->type) || val->val.i64 > 0);
		IR_ASSERT(IR_IS_SIGNED_32BIT(val->val.i64));

		if (ctx->flags2 & IR_HAS_CALLS) {
			/* Stack must be 16 byte aligned */
			size = IR_ALIGNED_SIZE(size, 16);
		} else {
			size = IR_ALIGNED_SIZE(size, 8);
		}
		|	ASM_REG_IMM_OP add, IR_ADDR, IR_REG_RSP, size
		if (!(ctx->flags & IR_USE_FRAME_POINTER)) {
			ctx->call_stack_size -= size;
		}
	} else {
//		int32_t alignment = (ctx->flags2 & IR_HAS_CALLS) ? 16 : 8;
		ir_reg op2_reg = ctx->regs[def][2];
		ir_type type = ctx->ir_base[insn->op2].type;

		IR_ASSERT(ctx->flags & IR_FUNCTION);
		if (op2_reg != IR_REG_NONE && IR_REG_SPILLED(op2_reg)) {
			op2_reg = IR_REG_NUM(op2_reg);
			ir_emit_load(ctx, type, op2_reg, insn->op2);
		}

		// TODO: alignment ???

		|	ASM_REG_REG_OP add, IR_ADDR, IR_REG_RSP, op2_reg
	}
}

static void ir_emit_frame_addr(ir_ctx *ctx, ir_ref def)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);

	if (ctx->flags & IR_USE_FRAME_POINTER) {
		|	mov Ra(def_reg), Ra(IR_REG_RBP)
	} else {
		|	lea Ra(def_reg), [Ra(IR_REG_RSP)+(ctx->stack_frame_size + ctx->call_stack_size)]
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, IR_ADDR, def, def_reg);
	}
}

static void ir_emit_switch(ir_ctx *ctx, uint32_t b, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type;
	ir_block *bb;
	ir_insn *use_insn, *val;
	uint32_t n, *p, use_block;
	int i;
	int label, default_label = 0;
	int count = 0;
	ir_val min, max;
	int64_t offset;
	ir_reg op2_reg = ctx->regs[def][2];
|.if X64
||	ir_reg tmp_reg = ctx->regs[def][3];
|.endif

	type = ctx->ir_base[insn->op2].type;
	if (IR_IS_TYPE_SIGNED(type)) {
		min.u64 = 0x7fffffffffffffff;
		max.u64 = 0x8000000000000000;
	} else {
		min.u64 = 0xffffffffffffffff;
		max.u64 = 0x0;
	}

	bb = &ctx->cfg_blocks[b];
	p = &ctx->cfg_edges[bb->successors];
	for (n = bb->successors_count; n != 0; p++, n--) {
		use_block = *p;
		use_insn = &ctx->ir_base[ctx->cfg_blocks[use_block].start];
		if (use_insn->op == IR_CASE_VAL) {
			val = &ctx->ir_base[use_insn->op2];
			if (IR_IS_TYPE_SIGNED(type)) {
				IR_ASSERT(IR_IS_TYPE_SIGNED(val->type));
				min.i64 = IR_MIN(min.i64, val->val.i64);
				max.i64 = IR_MAX(max.i64, val->val.i64);
			} else {
				IR_ASSERT(!IR_IS_TYPE_SIGNED(val->type));
				min.u64 = (int64_t)IR_MIN(min.u64, val->val.u64);
				max.u64 = (int64_t)IR_MAX(max.u64, val->val.u64);
			}
			count++;
		} else {
			IR_ASSERT(use_insn->op == IR_CASE_DEFAULT);
			default_label = ir_skip_empty_target_blocks(ctx, use_block);
		}
	}

	IR_ASSERT(op2_reg != IR_REG_NONE);
|.if X64
||	IR_ASSERT(tmp_reg != IR_REG_NONE || sizeof(void*) != 8);
|.endif
	if (IR_REG_SPILLED(op2_reg)) {
		op2_reg = IR_REG_NUM(op2_reg);
		ir_emit_load(ctx, type, op2_reg, insn->op2);
	} else if (IR_IS_CONST_REF(insn->op2)) {
		ir_emit_load(ctx, type, op2_reg, insn->op2);
	}

	/* Generate a table jmp or a seqence of calls */
	if ((max.i64-min.i64) < count * 8) {
		int *labels = ir_mem_malloc(sizeof(int) * (size_t)(max.i64 - min.i64 + 1));

		for (i = 0; i <= (max.i64 - min.i64); i++) {
			labels[i] = default_label;
		}
		p = &ctx->cfg_edges[bb->successors];
		for (n = bb->successors_count; n != 0; p++, n--) {
			use_block = *p;
			use_insn = &ctx->ir_base[ctx->cfg_blocks[use_block].start];
			if (use_insn->op == IR_CASE_VAL) {
				val = &ctx->ir_base[use_insn->op2];
				label = ir_skip_empty_target_blocks(ctx, use_block);
				labels[val->val.i64 - min.i64] = label;
			}
		}

		if (IR_IS_32BIT(type, max)) {
			|	ASM_REG_IMM_OP cmp, type, op2_reg, max.i32
		} else {
			IR_ASSERT(ir_type_size[type] == 8);
			IR_ASSERT(sizeof(void*) == 8);
|.if X64
			|	mov64 Rq(tmp_reg), max.i64
			|	cmp Rq(op2_reg), Rq(tmp_reg)
|.endif
		}
		if (IR_IS_TYPE_SIGNED(type)) {
			|	jg =>default_label
		} else {
			|	ja =>default_label
		}

		if (IR_IS_32BIT(type, min)) {
			offset = -min.i64 * sizeof(void*);
			if (IR_IS_SIGNED_32BIT(offset)) {
				|	ASM_REG_IMM_OP cmp, type, op2_reg, min.i32
			} else {
				|	ASM_REG_REG_OP sub, type, op2_reg, (int32_t)offset // TODO: reg clobbering
				offset = 0;
			}
		} else {
			IR_ASSERT(sizeof(void*) == 8);
|.if X64
			|	mov64 Rq(tmp_reg), min.i64
			|	ASM_REG_REG_OP sub, type, op2_reg, tmp_reg // TODO: reg clobbering
			offset = 0;
|.endif
		}
		if (IR_IS_TYPE_SIGNED(type)) {
			|	jl =>default_label
		} else {
			|	jb =>default_label
		}
		if (sizeof(void*) == 8) {
|.if X64
			switch (ir_type_size[type]) {
				default:
					IR_ASSERT(0);
				case 1:
					if (IR_IS_TYPE_SIGNED(type)) {
						|	movsx Ra(op2_reg), Rb(op2_reg)
					} else {
						|	movzx Ra(op2_reg), Rb(op2_reg)
					}
					break;
				case 2:
					if (IR_IS_TYPE_SIGNED(type)) {
						|	movsx Ra(op2_reg), Rw(op2_reg)
					} else {
						|	movzx Ra(op2_reg), Rw(op2_reg)
					}
					break;
				case 4:
					if (IR_IS_TYPE_SIGNED(type)) {
						|	movsxd Ra(op2_reg), Rd(op2_reg)
					} else {
						|	mov Rd(op2_reg), Rd(op2_reg)
					}
					break;
				case 8:
					break;
			}
			|	lea Ra(tmp_reg), aword [>1]
			|	jmp aword [Ra(tmp_reg)+Ra(op2_reg)*8+(int32_t)offset]
|.endif
		} else {
|.if not X64
			switch (ir_type_size[type]) {
				default:
					IR_ASSERT(0 && "Unsupported type size");
				case 1:
					if (IR_IS_TYPE_SIGNED(type)) {
						|	movsx Ra(op2_reg), Rb(op2_reg)
					} else {
						|	movzx Ra(op2_reg), Rb(op2_reg)
					}
					break;
				case 2:
					if (IR_IS_TYPE_SIGNED(type)) {
						|	movsx Ra(op2_reg), Rw(op2_reg)
					} else {
						|	movzx Ra(op2_reg), Rw(op2_reg)
					}
					break;
				case 4:
					break;
			}
			|//	jmp aword [Ra(op2_reg)*4+(int32_t)offset+>1]
			|	lea Ra(op2_reg), aword [Ra(op2_reg)*4+(int32_t)offset] // TODO: reg clobbering
			|	jmp aword [Ra(op2_reg)+>1]
|.endif
		}
		|.jmp_table
		if (!data->jmp_table_label) {
			data->jmp_table_label = ctx->cfg_blocks_count + ctx->consts_count + 3;
			|=>data->jmp_table_label:
		}
		|.align aword
		|1:
		for (i = 0; i <= (max.i64 - min.i64); i++) {
			int b = labels[i];
			ir_block *bb = &ctx->cfg_blocks[b];
			ir_insn *insn = &ctx->ir_base[bb->end];

			if (insn->op == IR_IJMP && IR_IS_CONST_REF(insn->op2)) {
				ir_ref prev = ctx->prev_ref[bb->end];
				if (prev != bb->start && ctx->ir_base[prev].op == IR_SNAPSHOT) {
					prev = ctx->prev_ref[prev];
				}
				if (prev == bb->start) {
					void *addr = ir_jmp_addr(ctx, insn, &ctx->ir_base[insn->op2]);

					|	.aword &addr
					if (ctx->ir_base[bb->start].op != IR_CASE_DEFAULT) {
						bb->flags |= IR_BB_EMPTY;
					}
					continue;
				}
			}
			|	.aword =>b
		}
		|.code
		ir_mem_free(labels);
	} else {
		p = &ctx->cfg_edges[bb->successors];
		for (n = bb->successors_count; n != 0; p++, n--) {
			use_block = *p;
			use_insn = &ctx->ir_base[ctx->cfg_blocks[use_block].start];
			if (use_insn->op == IR_CASE_VAL) {
				val = &ctx->ir_base[use_insn->op2];
				label = ir_skip_empty_target_blocks(ctx, use_block);
				if (IR_IS_32BIT(type, val->val)) {
					|	ASM_REG_IMM_OP cmp, type, op2_reg, val->val.i32
				} else {
					IR_ASSERT(sizeof(void*) == 8);
|.if X64
					|	mov64 Ra(tmp_reg), val->val.i64
					|	ASM_REG_REG_OP cmp, type, op2_reg, tmp_reg
|.endif
				}
				|	je =>label
			}
		}
		if (default_label) {
			|	jmp =>default_label
		}
	}
}

static int ir_parallel_copy(ir_ctx *ctx, ir_copy *copies, int count, ir_reg tmp_reg, ir_reg tmp_fp_reg)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	int i;
	int8_t *pred, *loc, *types;
	ir_reg to, from_reg, c;
	ir_type type;
	ir_regset todo, ready;
	ir_reg last_reg = IR_REG_NONE, last_fp_reg = IR_REG_NONE;

	loc = ir_mem_malloc(IR_REG_NUM * 3 * sizeof(int8_t));
	pred = loc + IR_REG_NUM;
	types = pred + IR_REG_NUM;
	memset(loc, IR_REG_NONE, IR_REG_NUM * 2 * sizeof(int8_t));
	todo = IR_REGSET_EMPTY;
	ready = IR_REGSET_EMPTY;

	for (i = 0; i < count; i++) {
		from_reg = copies[i].from;
		to = copies[i].to;
		if (from_reg != to) {
			loc[from_reg] = from_reg;
			pred[to] = from_reg;
			types[from_reg] = copies[i].type;
			/* temporary register may be the same as some of destinations */
			if (to == tmp_reg) {
				IR_ASSERT(last_reg == IR_REG_NONE);
				last_reg = to;
			} else if (to == tmp_fp_reg) {
				IR_ASSERT(last_fp_reg == IR_REG_NONE);
				last_fp_reg = to;
			} else {
				IR_ASSERT(!IR_REGSET_IN(todo, to));
				IR_REGSET_INCL(todo, to);
			}
		}
	}

	IR_REGSET_FOREACH(todo, i) {
		if (loc[i] == IR_REG_NONE) {
			IR_REGSET_INCL(ready, i);
		}
	} IR_REGSET_FOREACH_END();

	while (1) {
		while (ready != IR_REGSET_EMPTY) {
			to = ir_regset_pop_first(&ready);
			from_reg = pred[to];
			c = loc[from_reg];
			type = types[from_reg];
			if (IR_IS_TYPE_INT(type)) {
				if (ir_type_size[type] > 2) {
					ir_emit_mov(ctx, type, to, c);
				} else if (ir_type_size[type] == 2) {
					if (IR_IS_TYPE_SIGNED(type)) {
						|	movsx Rd(to), Rw(c)
						type = IR_I32;
					} else {
						|	movzx Rd(to), Rw(c)
						type = IR_U32;
					}
				} else /* if (ir_type_size[type] == 1) */ {
					if (IR_IS_TYPE_SIGNED(type)) {
						|	movsx Rd(to), Rb(c)
						type = IR_I32;
					} else {
						|	movzx Rd(to), Rb(c)
						type = IR_U32;
					}
				}
			} else {
				ir_emit_fp_mov(ctx, type, to, c);
			}
			IR_REGSET_EXCL(todo, to);
			loc[from_reg] = to;
			if (from_reg == c && pred[from_reg] != IR_REG_NONE) {
				IR_REGSET_INCL(ready, from_reg);
			}
		}

		if (todo == IR_REGSET_EMPTY) {
			break;
		}
		to = ir_regset_pop_first(&todo);
		from_reg = pred[to];
		IR_ASSERT(to != loc[from_reg]);
		type = types[from_reg];
		if (IR_IS_TYPE_INT(type)) {
			IR_ASSERT(tmp_reg != IR_REG_NONE);
			IR_ASSERT(tmp_reg >= IR_REG_GP_FIRST && tmp_reg <= IR_REG_GP_LAST);
			ir_emit_mov(ctx, type, tmp_reg, to);
			loc[to] = tmp_reg;
		} else {
			IR_ASSERT(tmp_fp_reg != IR_REG_NONE);
			IR_ASSERT(tmp_fp_reg >= IR_REG_FP_FIRST && tmp_fp_reg <= IR_REG_FP_LAST);
			ir_emit_fp_mov(ctx, type, tmp_fp_reg, to);
			loc[to] = tmp_fp_reg;
		}
		IR_REGSET_INCL(ready, to);
	}

	if (last_reg != IR_REG_NONE) {
		to = last_reg;
		from_reg = pred[to];
		c = loc[from_reg];
		if (to != c) {
			type = types[from_reg];
			IR_ASSERT(IR_IS_TYPE_INT(type));
			if (ir_type_size[type] > 2) {
				ir_emit_mov(ctx, type, to, c);
			} else if (ir_type_size[type] == 2) {
				if (IR_IS_TYPE_SIGNED(type)) {
					|	movsx Rd(to), Rw(c)
					type = IR_I32;
				} else {
					|	movzx Rd(to), Rw(c)
					type = IR_U32;
				}
			} else /* if (ir_type_size[type] == 1) */ {
				if (IR_IS_TYPE_SIGNED(type)) {
					|	movsx Rd(to), Rb(c)
					type = IR_I32;
				} else {
					|	movzx Rd(to), Rb(c)
					type = IR_U32;
				}
			}
		}
	}

	if (last_fp_reg != IR_REG_NONE) {
		to = last_fp_reg;
		from_reg = pred[to];
		c = loc[from_reg];
		if (to != c) {
			type = types[from_reg];
			IR_ASSERT(!IR_IS_TYPE_INT(type));
			ir_emit_fp_mov(ctx, type, to, c);
		}
	}

	ir_mem_free(loc);

	return 1;
}

static int32_t ir_call_used_stack(ir_ctx *ctx, ir_insn *insn)
{
	int j, n;
	ir_type type;
	int int_param = 0;
	int fp_param = 0;
	int int_reg_params_count = IR_REG_INT_ARGS;
	int fp_reg_params_count = IR_REG_FP_ARGS;
	int32_t used_stack = 0;

#ifdef IR_HAVE_FASTCALL
	if (sizeof(void*) == 4 && ir_is_fastcall(ctx, insn)) {
		int_reg_params_count = IR_REG_INT_FCARGS;
		fp_reg_params_count = IR_REG_FP_FCARGS;
	}
#endif

	n = insn->inputs_count;
	for (j = 3; j <= n; j++) {
		type = ctx->ir_base[ir_insn_op(insn, j)].type;
		if (IR_IS_TYPE_INT(type)) {
			if (int_param >= int_reg_params_count) {
				used_stack += IR_MAX(sizeof(void*), ir_type_size[type]);
			}
			int_param++;
#ifdef _WIN64
			/* WIN64 calling convention use common couter for int and fp registers */
			fp_param++;
#endif
		} else {
			IR_ASSERT(IR_IS_TYPE_FP(type));
			if (fp_param >= fp_reg_params_count) {
				used_stack += IR_MAX(sizeof(void*), ir_type_size[type]);
			}
			fp_param++;
#ifdef _WIN64
			/* WIN64 calling convention use common couter for int and fp registers */
			int_param++;
#endif
		}
	}

	/* Reserved "home space" or "shadow store" for register arguments (used in Windows64 ABI) */
	used_stack += IR_SHADOW_ARGS;

	return used_stack;
}

static int32_t ir_emit_arguments(ir_ctx *ctx, ir_ref def, ir_insn *insn, ir_reg tmp_reg)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	int j, n;
	ir_ref arg;
	ir_insn *arg_insn;
	uint8_t type;
	ir_reg src_reg, dst_reg;
	int int_param = 0;
	int fp_param = 0;
	int count = 0;
	int int_reg_params_count = IR_REG_INT_ARGS;
	int fp_reg_params_count = IR_REG_FP_ARGS;
	const int8_t *int_reg_params = _ir_int_reg_params;
	const int8_t *fp_reg_params = _ir_fp_reg_params;
	int32_t used_stack, stack_offset = IR_SHADOW_ARGS;
	ir_copy *copies;
	bool do_pass3 = 0;
	/* For temporaries we may use any scratch registers except for registers used for parameters */
	ir_reg tmp_fp_reg = IR_REG_FP_LAST; /* Temporary register for FP loads and swap */

	n = insn->inputs_count;
	if (n < 3) {
		return 0;
	}

	if (tmp_reg == IR_REG_NONE) {
		tmp_reg = IR_REG_RAX;
	}

#ifdef IR_HAVE_FASTCALL
	if (sizeof(void*) == 4 && ir_is_fastcall(ctx, insn)) {
		int_reg_params_count = IR_REG_INT_FCARGS;
		fp_reg_params_count = IR_REG_FP_FCARGS;
		int_reg_params = _ir_int_fc_reg_params;
		fp_reg_params = _ir_fp_fc_reg_params;
	}
#endif

	if (insn->op == IR_CALL
	 && (ctx->flags & IR_PREALLOCATED_STACK)
#ifdef IR_HAVE_FASTCALL
	 && !ir_is_fastcall(ctx, insn) /* fast call functions restore stack pointer */
#endif
	) {
		// TODO: support for preallocated stack
		used_stack = 0;
	} else {
		used_stack = ir_call_used_stack(ctx, insn);
		if (IR_SHADOW_ARGS
		 && insn->op == IR_TAILCALL
		 && used_stack == IR_SHADOW_ARGS) {
			used_stack = 0;
		}
		if (ctx->fixed_call_stack_size
		 && used_stack <= ctx->fixed_call_stack_size
#ifdef IR_HAVE_FASTCALL
		 && !ir_is_fastcall(ctx, insn) /* fast call functions restore stack pointer */
#endif
		) {
			used_stack = 0;
		} else {
			/* Stack must be 16 byte aligned */
			int32_t aligned_stack = IR_ALIGNED_SIZE(used_stack, 16);
			ctx->call_stack_size += aligned_stack;
			if (aligned_stack) {
				|	sub Ra(IR_REG_RSP), aligned_stack
			}
		}
	}

	/* 1. move all register arguments that should be passed through stack
	 *    and collect arguments that should be passed through registers */
	copies = ir_mem_malloc((n - 2) * sizeof(ir_copy));
	for (j = 3; j <= n; j++) {
		arg = ir_insn_op(insn, j);
		src_reg = ir_get_alocated_reg(ctx, def, j);
		arg_insn = &ctx->ir_base[arg];
		type = arg_insn->type;
		if (IR_IS_TYPE_INT(type)) {
			if (int_param < int_reg_params_count) {
				dst_reg = int_reg_params[int_param];
			} else {
				dst_reg = IR_REG_NONE; /* pass argument through stack */
			}
			int_param++;
#ifdef _WIN64
			/* WIN64 calling convention use common couter for int and fp registers */
			fp_param++;
#endif
		} else {
			IR_ASSERT(IR_IS_TYPE_FP(type));
			if (fp_param < fp_reg_params_count) {
				dst_reg = fp_reg_params[fp_param];
			} else {
				dst_reg = IR_REG_NONE; /* pass argument through stack */
			}
			fp_param++;
#ifdef _WIN64
			/* WIN64 calling convention use common couter for int and fp registers */
			int_param++;
#endif
		}
		if (dst_reg != IR_REG_NONE) {
			if (IR_IS_CONST_REF(arg) || src_reg == IR_REG_NONE) {
				/* delay CONST->REG and MEM->REG moves to third pass */
				do_pass3 = 1;
			} else {
				if (IR_REG_SPILLED(src_reg)) {
					src_reg = IR_REG_NUM(src_reg);
					ir_emit_load(ctx, type, src_reg, arg);
				}
				if (src_reg != dst_reg) {
					/* delay REG->REG moves to second pass */
					copies[count].type = type;
					copies[count].from = src_reg;
					copies[count].to = dst_reg;
					count++;
				}
			}
		} else {
			/* Pass register arguments to stack (REG->MEM moves) */
			if (!IR_IS_CONST_REF(arg) && src_reg != IR_REG_NONE && !IR_REG_SPILLED(src_reg)) {
				if (IR_IS_TYPE_INT(type)) {
					ir_emit_store_mem_int(ctx, type, IR_REG_STACK_POINTER, stack_offset, src_reg);
				} else {
					ir_emit_store_mem_fp(ctx, type, IR_REG_STACK_POINTER, stack_offset, src_reg);
				}
			} else {
				do_pass3 = 1;
			}
			stack_offset += IR_MAX(sizeof(void*), ir_type_size[type]);
		}
	}

	/* 2. move all arguments that should be passed from one register to another (REG->REG movs) */
	if (count) {
		ir_parallel_copy(ctx, copies, count, tmp_reg, tmp_fp_reg);
	}
	ir_mem_free(copies);

	/* 3. move the remaining memory and immediate values */
	if (do_pass3) {
		stack_offset = IR_SHADOW_ARGS;
		int_param = 0;
		fp_param = 0;
		for (j = 3; j <= n; j++) {
			arg = ir_insn_op(insn, j);
			src_reg = ir_get_alocated_reg(ctx, def, j);
			arg_insn = &ctx->ir_base[arg];
			type = arg_insn->type;
			if (IR_IS_TYPE_INT(type)) {
				if (int_param < int_reg_params_count) {
					dst_reg = int_reg_params[int_param];
				} else {
					dst_reg = IR_REG_NONE; /* argument already passed through stack */
				}
				int_param++;
#ifdef _WIN64
				/* WIN64 calling convention use common couter for int and fp registers */
				fp_param++;
#endif
			} else {
				IR_ASSERT(IR_IS_TYPE_FP(type));
				if (fp_param < fp_reg_params_count) {
					dst_reg = fp_reg_params[fp_param];
				} else {
					dst_reg = IR_REG_NONE; /* argument already passed through stack */
				}
				fp_param++;
#ifdef _WIN64
				/* WIN64 calling convention use common couter for int and fp registers */
				int_param++;
#endif
			}
			if (dst_reg != IR_REG_NONE) {
				if (IR_IS_CONST_REF(arg) || src_reg == IR_REG_NONE) {
					if (IR_IS_TYPE_INT(type)) {
						if (IR_IS_CONST_REF(arg)) {
							if (type == IR_ADDR) {
								ir_insn *val_insn = &ctx->ir_base[arg];

								if (val_insn->op == IR_STR) {
									int label = ctx->cfg_blocks_count - arg;

									val_insn->const_flags |= IR_CONST_EMIT;
									|	lea Ra(dst_reg), aword [=>label]
									continue;
								} else if (val_insn->op == IR_SYM || val_insn->op == IR_FUNC) {
									void *addr = (ctx->loader && ctx->loader->resolve_sym_name) ?
										ctx->loader->resolve_sym_name(ctx->loader, ir_get_str(ctx, val_insn->val.i32)) :
										ir_resolve_sym_name(ir_get_str(ctx, val_insn->val.i32));
									if (sizeof(void*) == 4 || IR_IS_SIGNED_32BIT(addr)) {
										|	mov Ra(dst_reg), ((ptrdiff_t)addr)
									} else {
|.if X64
										|	mov64 Rq(dst_reg), ((ptrdiff_t)addr)
|.endif
									}
									continue;
								}
								IR_ASSERT(val_insn->op == IR_ADDR || val_insn->op == IR_FUNC_ADDR);
							}
							if (type == IR_I8 || type == IR_I16) {
								type = IR_I32;
							} else if (type == IR_U8 || type == IR_U16) {
								type = IR_U32;
							}
							ir_emit_load(ctx, type, dst_reg, arg);
						} else {
							ir_reg fp;
							int32_t offset = ir_ref_spill_slot(ctx, arg, &fp);

							if (ir_type_size[type] > 2) {
								ir_emit_load_mem_int(ctx, type, dst_reg, fp, offset);
							} else if (ir_type_size[type] == 2) {
								if (type == IR_I16) {
									if (fp != IR_REG_NONE) {
										|	movsx Rd(dst_reg), word [Ra(fp)+offset]
									} else {
										|	movsx Rd(dst_reg), word [offset]
									}
								} else {
									if (fp != IR_REG_NONE) {
										|	movzx Rd(dst_reg), word [Ra(fp)+offset]
									} else {
										|	movzx Rd(dst_reg), word [offset]
									}
								}
							} else {
								IR_ASSERT(ir_type_size[type] == 1);
								if (type == IR_I8) {
									if (fp != IR_REG_NONE) {
										|	movsx Rd(dst_reg), byte [Ra(fp)+offset]
									} else {
										|	movsx Rd(dst_reg), byte [offset]
									}
								} else {
									if (fp != IR_REG_NONE) {
										|	movzx Rd(dst_reg), byte [Ra(fp)+offset]
									} else {
										|	movzx Rd(dst_reg), byte [offset]
									}
								}
							}
						}
					} else {
						ir_emit_load(ctx, type, dst_reg, arg);
					}
				}
			} else {
				if (IR_IS_TYPE_INT(type)) {
					if (IR_IS_CONST_REF(arg)) {
						ir_insn *val_insn = &ctx->ir_base[arg];

						if (val_insn->op == IR_STR) {
							int label = ctx->cfg_blocks_count - arg;

							val_insn->const_flags |= IR_CONST_EMIT;
							IR_ASSERT(tmp_reg != IR_REG_NONE);
|.if X64
							|	lea Ra(tmp_reg), aword [=>label]
							|	mov [Ra(IR_REG_RSP)+stack_offset], Ra(tmp_reg)
|.else
							|	mov [Ra(IR_REG_RSP)+stack_offset], =>label
|.endif
						} else if (val_insn->op == IR_FUNC || val_insn->op == IR_SYM) {
							void *addr = (ctx->loader && ctx->loader->resolve_sym_name) ?
								ctx->loader->resolve_sym_name(ctx->loader, ir_get_str(ctx, val_insn->val.i32)) :
								ir_resolve_sym_name(ir_get_str(ctx, val_insn->val.i32));
							if (sizeof(void*) == 4) {
								|	mov aword [Ra(IR_REG_RSP)+stack_offset], ((ptrdiff_t)addr)
|.if X64
||							} else if (IR_IS_SIGNED_32BIT(addr)) {
								|	mov Ra(tmp_reg), ((ptrdiff_t)addr)
								|	mov [Ra(IR_REG_RSP)+stack_offset], Ra(tmp_reg)
||							} else {
								|	mov64 Rq(tmp_reg), ((ptrdiff_t)addr)
								|	mov [Ra(IR_REG_RSP)+stack_offset], Ra(tmp_reg)
|.endif
							}
						} else if (IR_IS_SIGNED_32BIT(val_insn->val.i64)) {
							if (ir_type_size[type] <= 4) {
								|	mov dword [Ra(IR_REG_RSP)+stack_offset], val_insn->val.i32
							} else {
								IR_ASSERT(sizeof(void*) == 8);
|.if X64
								|	mov qword [rsp+stack_offset], val_insn->val.i32
|.endif
							}
						} else {
							IR_ASSERT(sizeof(void*) == 8);
|.if X64
							IR_ASSERT(tmp_reg != IR_REG_NONE);
							|	mov64 Ra(tmp_reg), val_insn->val.i64
							|	mov [rsp+stack_offset], Ra(tmp_reg)
|.endif
						}
					} else if (src_reg == IR_REG_NONE) {
						IR_ASSERT(tmp_reg != IR_REG_NONE);
						ir_emit_load(ctx, type, tmp_reg, arg);
						ir_emit_store_mem_int(ctx, type, IR_REG_STACK_POINTER, stack_offset, tmp_reg);
					} else if (IR_REG_SPILLED(src_reg)) {
						src_reg = IR_REG_NUM(src_reg);
						ir_emit_load(ctx, type, src_reg, arg);
						ir_emit_store_mem_int(ctx, type, IR_REG_STACK_POINTER, stack_offset, src_reg);
					}
				} else {
					if (IR_IS_CONST_REF(arg)) {
						ir_val *val = &ctx->ir_base[arg].val;
						if (ir_type_size[type] == 4) {
							|	mov dword [Ra(IR_REG_RSP)+stack_offset], val->i32
						} else if (sizeof(void*) == 8) {
|.if X64
							if (val->i64 == 0) {
								|	mov qword [rsp+stack_offset], val->i32
							} else {
								IR_ASSERT(tmp_reg != IR_REG_NONE);
								|	mov64 Rq(tmp_reg), val->i64
								|	mov qword [rsp+stack_offset], Ra(tmp_reg)
							}
|.endif
						} else {
							ir_emit_load(ctx, type, tmp_fp_reg, arg);
							ir_emit_store_mem_fp(ctx, IR_DOUBLE, IR_REG_STACK_POINTER, stack_offset, tmp_fp_reg);
						}
					} else if (src_reg == IR_REG_NONE) {
						IR_ASSERT(tmp_fp_reg != IR_REG_NONE);
						ir_emit_load(ctx, type, tmp_fp_reg, arg);
						ir_emit_store_mem_fp(ctx, IR_DOUBLE, IR_REG_STACK_POINTER, stack_offset, tmp_fp_reg);
					} else if (IR_REG_SPILLED(src_reg)) {
						src_reg = IR_REG_NUM(src_reg);
						ir_emit_load(ctx, type, src_reg, arg);
						ir_emit_store_mem_fp(ctx, type, IR_REG_STACK_POINTER, stack_offset, src_reg);
					}
				}
				stack_offset += IR_MAX(sizeof(void*), ir_type_size[type]);
			}
		}
	}

#ifdef _WIN64
	/* WIN64 calling convention requires duplcation of parameters passed in FP register into GP ones */
	if (ir_is_vararg(ctx, insn)) {
		n = IR_MIN(n, IR_MAX_REG_ARGS + 2);
		for (j = 3; j <= n; j++) {
			arg = ir_insn_op(insn, j);
			arg_insn = &ctx->ir_base[arg];
			type = arg_insn->type;
			if (IR_IS_TYPE_FP(type)) {
				src_reg = fp_reg_params[j-3];
				dst_reg = int_reg_params[j-3];
|.if X64
				if (ctx->mflags & IR_X86_AVX) {
					|	vmovd Rq(dst_reg), xmm(src_reg-IR_REG_FP_FIRST)
				} else {
					|	movd Rq(dst_reg), xmm(src_reg-IR_REG_FP_FIRST)
				}
|.endif
			}
		}
	}
#endif
#ifdef IR_REG_VARARG_FP_REGS
	/* set hidden argument to specify the number of vector registers used */
	if (ir_is_vararg(ctx, insn)) {
		fp_param = IR_MIN(fp_param, fp_reg_params_count);
		|	mov Rd(IR_REG_VARARG_FP_REGS), fp_param
	}
#endif

	return used_stack;
}

static void ir_emit_call(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_reg def_reg;
	int32_t used_stack = ir_emit_arguments(ctx, def, insn, ctx->regs[def][1]);

	if (IR_IS_CONST_REF(insn->op2)) {
		ir_insn *addr_insn = &ctx->ir_base[insn->op2];
		void *addr;

		IR_ASSERT(addr_insn->type == IR_ADDR);
		if (addr_insn->op == IR_FUNC) {
			addr = (ctx->loader && ctx->loader->resolve_sym_name) ?
				ctx->loader->resolve_sym_name(ctx->loader, ir_get_str(ctx, addr_insn->val.i32)) :
				ir_resolve_sym_name(ir_get_str(ctx, addr_insn->val.i32));
		} else {
			IR_ASSERT(addr_insn->op == IR_ADDR || addr_insn->op == IR_FUNC_ADDR);
			addr = (void*)addr_insn->val.addr;
		}
		if (sizeof(void*) == 4 || IR_MAY_USE_32BIT_ADDR(addr)) {
			|	call aword &addr
		} else {
|.if X64
||			ir_reg tmp_reg = IR_REG_RAX;

#ifdef IR_REG_VARARG_FP_REGS
||			if (ir_is_vararg(ctx, insn)) {
||				tmp_reg = IR_REG_R11;
||			}
#endif
||			if (IR_IS_SIGNED_32BIT(addr)) {
				|	mov Rq(tmp_reg), ((ptrdiff_t)addr)    // 0x48 0xc7 0xc0 <imm-32-bit>
||			} else {
				|	mov64 Rq(tmp_reg), ((ptrdiff_t)addr)  // 0x48 0xb8 <imm-64-bit>
||			}
			|	call Rq(tmp_reg)
|.endif
		}
    } else {
		ir_reg op2_reg = ctx->regs[def][2];

		if (op2_reg != IR_REG_NONE) {
			if (IR_REG_SPILLED(op2_reg)) {
				op2_reg = IR_REG_NUM(op2_reg);
				ir_emit_load(ctx, IR_ADDR, op2_reg, insn->op2);
			}
			|	call Ra(op2_reg)
		} else {
			int32_t offset;

			if (ir_rule(ctx, insn->op2) & IR_FUSED) {
				offset = ir_fuse_load(ctx, insn->op2, &op2_reg);
			} else {
				offset = ir_ref_spill_slot(ctx, insn->op2, &op2_reg);
			}

			if (op2_reg != IR_REG_NONE) {
				|	call aword [Ra(op2_reg)+offset]
			} else {
				|	call aword [offset]
			}
		}
    }

	if (used_stack) {
		int32_t aligned_stack = IR_ALIGNED_SIZE(used_stack, 16);

		ctx->call_stack_size -= aligned_stack;
		if (ir_is_fastcall(ctx, insn)) {
			aligned_stack -= used_stack;
			if (aligned_stack) {
				|	add Ra(IR_REG_RSP), aligned_stack
			}
		} else {
			|	add Ra(IR_REG_RSP), aligned_stack
		}
	}

	if (insn->type != IR_VOID) {
		if (IR_IS_TYPE_INT(insn->type)) {
			def_reg = IR_REG_NUM(ctx->regs[def][0]);
			if (def_reg != IR_REG_NONE) {
				if (def_reg != IR_REG_INT_RET1) {
					ir_emit_mov(ctx, insn->type, def_reg, IR_REG_INT_RET1);
				}
				if (IR_REG_SPILLED(ctx->regs[def][0])) {
					ir_emit_store(ctx, insn->type, def, def_reg);
				}
			} else if (ctx->use_lists[def].count > 1) {
				ir_emit_store(ctx, insn->type, def, IR_REG_INT_RET1);
			}
		} else {
			IR_ASSERT(IR_IS_TYPE_FP(insn->type));
			def_reg = IR_REG_NUM(ctx->regs[def][0]);
#ifdef IR_REG_FP_RET1
			if (def_reg != IR_REG_NONE) {
				if (def_reg != IR_REG_FP_RET1) {
					ir_emit_fp_mov(ctx, insn->type, def_reg, IR_REG_FP_RET1);
				}
				if (IR_REG_SPILLED(ctx->regs[def][0])) {
					ir_emit_store(ctx, insn->type, def, def_reg);
				}
			} else if (ctx->use_lists[def].count > 1) {
				ir_emit_store(ctx, insn->type, def, IR_REG_FP_RET1);
			}
#else
			IR_ASSERT(0); // TODO: float/double return value
#endif
		}
	}
}

static void ir_emit_tailcall(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	int32_t used_stack = ir_emit_arguments(ctx, def, insn, ctx->regs[def][1]);
	(void) used_stack;

	if (used_stack != 0) {
		ir_emit_call(ctx, def, insn);
		ir_emit_return_void(ctx);
		return;
	}

	ir_emit_epilogue(ctx);

	if (IR_IS_CONST_REF(insn->op2)) {
		ir_insn *addr_insn = &ctx->ir_base[insn->op2];
		void *addr;

		IR_ASSERT(addr_insn->type == IR_ADDR);
		if (addr_insn->op == IR_FUNC) {
			addr = (ctx->loader && ctx->loader->resolve_sym_name) ?
				ctx->loader->resolve_sym_name(ctx->loader, ir_get_str(ctx, addr_insn->val.i32)) :
				ir_resolve_sym_name(ir_get_str(ctx, addr_insn->val.i32));
		} else {
			IR_ASSERT(addr_insn->op == IR_ADDR || addr_insn->op == IR_FUNC_ADDR);
			addr = (void*)addr_insn->val.addr;
		}
		if (sizeof(void*) == 4 || IR_MAY_USE_32BIT_ADDR(addr)) {
			|	jmp aword &addr
		} else {
|.if X64
||			ir_reg tmp_reg = IR_REG_RAX;

#ifdef IR_REG_VARARG_FP_REGS
||			if (ir_is_vararg(ctx, insn)) {
||				tmp_reg = IR_REG_R11;
||			}
#endif
||			if (IR_IS_SIGNED_32BIT(addr)) {
				|	mov Rq(tmp_reg), ((ptrdiff_t)addr)    // 0x48 0xc7 0xc0 <imm-32-bit>
||			} else {
				|	mov64 Rq(tmp_reg), ((ptrdiff_t)addr)  // 0x48 0xb8 <imm-64-bit>
||			}
			|	jmp Rq(tmp_reg)
|.endif
		}
    } else {
		ir_reg op2_reg = ctx->regs[def][2];

		if (op2_reg != IR_REG_NONE) {
			if (IR_REG_SPILLED(op2_reg)) {
				op2_reg = IR_REG_NUM(op2_reg);
				ir_emit_load(ctx, IR_ADDR, op2_reg, insn->op2);
			}
			|	jmp Ra(op2_reg)
		} else {
			int32_t offset;

			if (ir_rule(ctx, insn->op2) & IR_FUSED) {
				offset = ir_fuse_load(ctx, insn->op2, &op2_reg);
			} else {
				offset = ir_ref_spill_slot(ctx, insn->op2, &op2_reg);
			}

			if (op2_reg != IR_REG_NONE) {
				|	jmp aword [Ra(op2_reg)+offset]
			} else {
				|	jmp aword [offset]
			}
		}
    }
}

static void ir_emit_ijmp(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_reg op2_reg = ctx->regs[def][2];

	if (IR_IS_CONST_REF(insn->op2)) {
		void *addr = ir_jmp_addr(ctx, insn, &ctx->ir_base[insn->op2]);

		if (sizeof(void*) == 4 || IR_MAY_USE_32BIT_ADDR(addr)) {
			|	jmp aword &addr
		} else {
|.if X64
			if (IR_IS_SIGNED_32BIT(addr)) {
				|	mov rax, ((ptrdiff_t)addr)    // 0x48 0xc7 0xc0 <imm-32-bit>
			} else {
				|	mov64 rax, ((ptrdiff_t)addr)  // 0x48 0xb8 <imm-64-bit>
			}
			|	jmp rax
|.endif
		}
	} else if (ir_rule(ctx, insn->op2) & IR_FUSED) {
	    int32_t offset;

		offset = ir_fuse_load(ctx, insn->op2, &op2_reg);
		if (op2_reg == IR_REG_NONE) {
			|	jmp aword [offset]
		} else {
			|	jmp aword [Ra(op2_reg)+offset]
		}
	} else if (op2_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op2_reg)) {
			op2_reg = IR_REG_NUM(op2_reg);
			ir_emit_load(ctx, IR_ADDR, op2_reg, insn->op2);
		}
		|	jmp Ra(op2_reg)
	} else {
		ir_reg fp;
		int32_t offset = ir_ref_spill_slot(ctx, insn->op2, &fp);

		|	jmp aword [Ra(fp)+offset]
	}
}

static bool ir_emit_guard_jcc(ir_ctx *ctx, uint32_t b, ir_ref def, uint8_t op, void *addr, bool int_cmp)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_insn *next_insn = &ctx->ir_base[def + 1];

	if (next_insn->op == IR_END || next_insn->op == IR_LOOP_END) {
		ir_block *bb = &ctx->cfg_blocks[b];
		uint32_t target;

		if (!(bb->flags & IR_BB_DESSA_MOVES)) {
			target = ctx->cfg_edges[bb->successors];
			if (UNEXPECTED(bb->successors_count == 2)) {
				if (ctx->cfg_blocks[target].flags & IR_BB_ENTRY) {
					target = ctx->cfg_edges[bb->successors + 1];
				} else {
					IR_ASSERT(ctx->cfg_blocks[ctx->cfg_edges[bb->successors + 1]].flags & IR_BB_ENTRY);
				}
			} else {
				IR_ASSERT(bb->successors_count == 1);
			}
			target = ir_skip_empty_target_blocks(ctx, target);
			if (b == ctx->cfg_blocks_count || target != ir_skip_empty_next_blocks(ctx, b + 1)) {
				if (int_cmp) {
					switch (op) {
						default:
							IR_ASSERT(0 && "NIY binary op");
						case IR_EQ:
							|	jne =>target
							break;
						case IR_NE:
							|	je =>target
							break;
						case IR_LT:
							|	jge =>target
							break;
						case IR_GE:
							|	jl =>target
							break;
						case IR_LE:
							|	jg =>target
							break;
						case IR_GT:
							|	jle =>target
							break;
						case IR_ULT:
							|	jae =>target
							break;
						case IR_UGE:
							|	jb =>target
							break;
						case IR_ULE:
							|	ja =>target
							break;
						case IR_UGT:
							|	jbe =>target
							break;
					}
				} else {
					switch (op) {
						default:
							IR_ASSERT(0 && "NIY binary op");
						case IR_EQ:
							|	jne =>target
							|	jp =>target
							break;
						case IR_NE:
							|	jp &addr
							|	je =>target
							break;
						case IR_LT:
							|	jae =>target
							break;
						case IR_GE:
							|	jp &addr
							|	jb =>target
							break;
						case IR_LE:
							|	ja =>target
							break;
						case IR_GT:
							|	jp &addr
							|	jbe =>target
							break;
					}
				}
				|	jmp &addr
				return 1;
			}
		}
	} else if (next_insn->op == IR_IJMP && IR_IS_CONST_REF(next_insn->op2)) {
		void *target_addr = ir_jmp_addr(ctx, next_insn, &ctx->ir_base[next_insn->op2]);

		if (sizeof(void*) == 4 || IR_MAY_USE_32BIT_ADDR(target_addr)) {
			if (int_cmp) {
				switch (op) {
					default:
						IR_ASSERT(0 && "NIY binary op");
					case IR_EQ:
						|	jne &target_addr
						break;
					case IR_NE:
						|	je &target_addr
						break;
					case IR_LT:
						|	jge &target_addr
						break;
					case IR_GE:
						|	jl &target_addr
						break;
					case IR_LE:
						|	jg &target_addr
						break;
					case IR_GT:
						|	jle &target_addr
						break;
					case IR_ULT:
						|	jae &target_addr
						break;
					case IR_UGE:
						|	jb &target_addr
						break;
					case IR_ULE:
						|	ja &target_addr
						break;
					case IR_UGT:
						|	jbe &target_addr
						break;
				}
			} else {
				switch (op) {
					default:
						IR_ASSERT(0 && "NIY binary op");
					case IR_EQ:
						|	jne &target_addr
						|	jp &target_addr
						break;
					case IR_NE:
						|	jp &addr
						|	je &target_addr
						break;
					case IR_LT:
						|	jae &target_addr
						break;
					case IR_GE:
						|	jp &addr
						|	jb &target_addr
						break;
					case IR_LE:
						|	ja &target_addr
						break;
					case IR_GT:
						|	jp &addr
						|	jbe &target_addr
						break;
				}
			}
			|	jmp &addr
			return 1;
		}
	}

	if (int_cmp) {
		switch (op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_EQ:
				|	je &addr
				break;
			case IR_NE:
				|	jne &addr
				break;
			case IR_LT:
				|	jl &addr
				break;
			case IR_GE:
				|	jge &addr
				break;
			case IR_LE:
				|	jle &addr
				break;
			case IR_GT:
				|	jg &addr
				break;
			case IR_ULT:
				|	jb &addr
				break;
			case IR_UGE:
				|	jae &addr
				break;
			case IR_ULE:
				|	jbe &addr
				break;
			case IR_UGT:
				|	ja &addr
				break;
		}
	} else {
		switch (op) {
			default:
				IR_ASSERT(0 && "NIY binary op");
			case IR_EQ:
				|	jp >1
				|	je &addr
				|1:
				break;
			case IR_NE:
				|	jne &addr
				|	jp &addr
				break;
			case IR_LT:
				|	jp >1
				|	jb &addr
				|1:
				break;
			case IR_GE:
				|	jae &addr
				break;
			case IR_LE:
				|	jp >1
				|	jbe &addr
				|1:
				break;
			case IR_GT:
				|	ja &addr
				break;
//			case IR_ULT: fprintf(stderr, "\tjb .LL%d\n", true_block); break;
//			case IR_UGE: fprintf(stderr, "\tjae .LL%d\n", true_block); break;
//			case IR_ULE: fprintf(stderr, "\tjbe .LL%d\n", true_block); break;
//			case IR_UGT: fprintf(stderr, "\tja .LL%d\n", true_block); break;
		}
	}
	return 0;
}

static bool ir_emit_guard(ir_ctx *ctx, uint32_t b, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_reg op2_reg = ctx->regs[def][2];
	ir_type type = ctx->ir_base[insn->op2].type;
	void *addr;

	IR_ASSERT(IR_IS_TYPE_INT(type));
	if (IR_IS_CONST_REF(insn->op2)) {
		bool is_true = ir_ref_is_true(ctx, insn->op2);

		if ((insn->op == IR_GUARD && !is_true) || (insn->op == IR_GUARD_NOT && is_true)) {
			addr = ir_jmp_addr(ctx, insn, &ctx->ir_base[insn->op3]);
			if (sizeof(void*) == 4 || IR_MAY_USE_32BIT_ADDR(addr)) {
				|	jmp aword &addr
			} else {
|.if X64
				if (IR_IS_SIGNED_32BIT(addr)) {
					|	mov rax, ((ptrdiff_t)addr)    // 0x48 0xc7 0xc0 <imm-32-bit>
				} else {
					|	mov64 rax, ((ptrdiff_t)addr)  // 0x48 0xb8 <imm-64-bit>
				}
				|	jmp aword [rax]
|.endif
			}
		}
		return 0;
	}

	if (op2_reg != IR_REG_NONE) {
		if (IR_REG_SPILLED(op2_reg)) {
			op2_reg = IR_REG_NUM(op2_reg);
			ir_emit_load(ctx, type, op2_reg, insn->op2);
		}
		|	ASM_REG_REG_OP test, type, op2_reg, op2_reg
	} else {
		int32_t offset = 0;

		if (ir_rule(ctx, insn->op2) & IR_FUSED) {
			offset = ir_fuse_load(ctx, insn->op2, &op2_reg);
		} else {
			offset = ir_ref_spill_slot(ctx, insn->op2, &op2_reg);
		}
		if (op2_reg == IR_REG_NONE) {
			|	ASM_MEM_IMM_OP cmp, type, [offset], 0
		} else {
			|	ASM_MEM_IMM_OP cmp, type, [Ra(op2_reg)+offset], 0
		}
	}

	addr = ir_jmp_addr(ctx, insn, &ctx->ir_base[insn->op3]);
	if (sizeof(void*) == 4 || IR_MAY_USE_32BIT_ADDR(addr)) {
		ir_op op;

		if (insn->op == IR_GUARD) {
			op = IR_EQ;
		} else {
			op = IR_NE;
		}
		return ir_emit_guard_jcc(ctx, b, def, op, addr, 1);
	} else {
|.if X64
		if (insn->op == IR_GUARD) {
			|	je >1
		} else {
			|	jne >1
		}
		|.cold_code
		|1:
		if (IR_IS_SIGNED_32BIT(addr)) {
			|	mov rax, ((ptrdiff_t)addr)    // 0x48 0xc7 0xc0 <imm-32-bit>
		} else {
			|	mov64 rax, ((ptrdiff_t)addr)  // 0x48 0xb8 <imm-64-bit>
		}
		|	jmp aword [rax]
		|.code
|.endif
		return 0;
	}
}

static bool ir_emit_guard_cmp_int(ir_ctx *ctx, uint32_t b, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_insn *cmp_insn = &ctx->ir_base[insn->op2];
	ir_op op = cmp_insn->op;
	ir_type type = ctx->ir_base[cmp_insn->op1].type;
	ir_ref op1 = cmp_insn->op1;
	ir_ref op2 = cmp_insn->op2;
	ir_reg op1_reg = ctx->regs[insn->op2][1];
	ir_reg op2_reg = ctx->regs[insn->op2][2];
	void *addr;

	if (op1_reg != IR_REG_NONE && (IR_IS_CONST_REF(op1) || IR_REG_SPILLED(op1_reg))) {
		op1_reg = IR_REG_NUM(op1_reg);
		ir_emit_load(ctx, type, op1_reg, op1);
	}
	if (op2_reg != IR_REG_NONE && (IR_IS_CONST_REF(op2) || IR_REG_SPILLED(op2_reg))) {
		op2_reg = IR_REG_NUM(op2_reg);
		if (op1 != op2) {
			ir_emit_load(ctx, type, op2_reg, op2);
		}
	}

	addr = ir_jmp_addr(ctx, insn, &ctx->ir_base[insn->op3]);
	if (IR_IS_CONST_REF(op2) && ctx->ir_base[op2].val.u64 == 0) {
		if (op == IR_ULT) {
			/* always false */
			if (sizeof(void*) == 4 || IR_MAY_USE_32BIT_ADDR(addr)) {
				|	jmp aword &addr
			} else {
|.if X64
				if (IR_IS_SIGNED_32BIT(addr)) {
					|	mov rax, ((ptrdiff_t)addr)    // 0x48 0xc7 0xc0 <imm-32-bit>
				} else {
					|	mov64 rax, ((ptrdiff_t)addr)  // 0x48 0xb8 <imm-64-bit>
				}
				|	jmp aword [rax]
|.endif
			}
			return 0;
		} else if (op == IR_UGE) {
			/* always true */
			return 0;
		} else if (op == IR_ULE) {
			op = IR_EQ;
		} else if (op == IR_UGT) {
			op = IR_NE;
		}
	}
	ir_emit_cmp_int_common(ctx, type, cmp_insn, op1_reg, op1, op2_reg, op2);

	if (insn->op == IR_GUARD) {
		op ^= 1; // reverse
	}

	return ir_emit_guard_jcc(ctx, b, def, op, addr, 1);
}

static bool ir_emit_guard_cmp_fp(ir_ctx *ctx, uint32_t b, ir_ref def, ir_insn *insn)
{
	ir_op op = ir_emit_cmp_fp_common(ctx, insn->op2, &ctx->ir_base[insn->op2]);
	void *addr = ir_jmp_addr(ctx, insn, &ctx->ir_base[insn->op3]);

	if (insn->op == IR_GUARD) {
		op ^= 1; // reverse
	}
	return ir_emit_guard_jcc(ctx, b, def, op, addr, 0);
}

static bool ir_emit_guard_test_int(ir_ctx *ctx, uint32_t b, ir_ref def, ir_insn *insn)
{
	void *addr = ir_jmp_addr(ctx, insn, &ctx->ir_base[insn->op3]);
	ir_op op = (insn->op == IR_GUARD) ? IR_EQ : IR_NE;

	ir_emit_test_int_common(ctx, insn->op2, op);
	return ir_emit_guard_jcc(ctx, b, def, op, addr, 1);
}

static bool ir_emit_guard_jcc_int(ir_ctx *ctx, uint32_t b, ir_ref def, ir_insn *insn)
{
	void *addr = ir_jmp_addr(ctx, insn, &ctx->ir_base[insn->op3]);
	ir_op op = ctx->ir_base[insn->op2].op;

	if (insn->op == IR_GUARD) {
		op ^= 1; // reverse
	}
	return ir_emit_guard_jcc(ctx, b, def, op, addr, 1);
}

static bool ir_emit_guard_overflow(ir_ctx *ctx, uint32_t b, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_type type;
	void *addr = ir_jmp_addr(ctx, insn, &ctx->ir_base[insn->op3]);

	type = ctx->ir_base[ctx->ir_base[insn->op2].op1].type;

	IR_ASSERT(IR_IS_TYPE_INT(type));
	if (IR_IS_TYPE_SIGNED(type)) {
		if (insn->op == IR_GUARD) {
			|	jno &addr
		} else {
			|	jo &addr
		}
	} else {
		if (insn->op == IR_GUARD) {
			|	jnc &addr
		} else {
			|	jc &addr
		}
	}
	return 0;
}

static void ir_emit_lea(ir_ctx *ctx, ir_ref def, ir_type type, ir_reg base_reg, ir_reg index_reg, uint8_t scale, int32_t offset)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);

	IR_ASSERT(def_reg != IR_REG_NONE);
	if (index_reg == IR_REG_NONE) {
		IR_ASSERT(base_reg != IR_REG_NONE);
		if (!offset) {
			if (ir_type_size[type] == 4) {
				|	lea Rd(def_reg), dword [Rd(base_reg)]
			} else {
				|	lea Ra(def_reg), aword [Ra(base_reg)]
			}
		} else {
			if (ir_type_size[type] == 4) {
				|	lea Rd(def_reg), dword [Rd(base_reg)+offset]
			} else {
				|	lea Ra(def_reg), aword [Ra(base_reg)+offset]
			}
		}
	} else {
		if (base_reg == IR_REG_NONE) {
			if (!offset) {
				switch (scale) {
					default:
						IR_ASSERT(0);
					case 2:
						if (ir_type_size[type] == 4) {
							|	lea Rd(def_reg), dword [Rd(index_reg)*2]
						} else {
							|	lea Ra(def_reg), aword [Ra(index_reg)*2]
						}
						break;
					case 4:
						if (ir_type_size[type] == 4) {
							|	lea Rd(def_reg), dword [Rd(index_reg)*4]
						} else {
							|	lea Ra(def_reg), aword [Ra(index_reg)*4]
						}
						break;
					case 8:
						if (ir_type_size[type] == 4) {
							|	lea Rd(def_reg), dword [Rd(index_reg)*8]
						} else {
							|	lea Ra(def_reg), aword [Ra(index_reg)*8]
						}
						break;
				}
			} else {
				switch (scale) {
					default:
						IR_ASSERT(0);
					case 2:
						if (ir_type_size[type] == 4) {
							|	lea Rd(def_reg), dword [Rd(index_reg)*2+offset]
						} else {
							|	lea Ra(def_reg), aword [Ra(index_reg)*2+offset]
						}
						break;
					case 4:
						if (ir_type_size[type] == 4) {
							|	lea Rd(def_reg), dword [Rd(index_reg)*4+offset]
						} else {
							|	lea Ra(def_reg), aword [Ra(index_reg)*4+offset]
						}
						break;
					case 8:
						if (ir_type_size[type] == 4) {
							|	lea Rd(def_reg), dword [Rd(index_reg)*8+offset]
						} else {
							|	lea Ra(def_reg), aword [Ra(index_reg)*8+offset]
						}
						break;
				}
			}
		} else {
			if (!offset) {
				switch (scale) {
					default:
						IR_ASSERT(0);
					case 1:
						if (ir_type_size[type] == sizeof(void*)) {
							if (def_reg == base_reg) {
								|	add Ra(def_reg), Ra(index_reg)
							} else if (def_reg == index_reg) {
								|	add Ra(def_reg), Ra(base_reg)
							} else {
								|	lea Ra(def_reg), aword [Ra(base_reg)+Ra(index_reg)]
							}
						} else {
							IR_ASSERT(sizeof(void*) == 8 && ir_type_size[type] == 4);
							if (def_reg == base_reg) {
								|	add Rd(def_reg), Rd(index_reg)
							} else if (def_reg == index_reg) {
								|	add Rd(def_reg), Rd(base_reg)
							} else {
								|	lea Rd(def_reg), dword [Rd(base_reg)+Rd(index_reg)]
							}
						}
						break;
					case 2:
						if (ir_type_size[type] == 4) {
							|	lea Rd(def_reg), dword [Rd(base_reg)+Rd(index_reg)*2]
						} else {
							|	lea Ra(def_reg), aword [Ra(base_reg)+Ra(index_reg)*2]
						}
						break;
					case 4:
						if (ir_type_size[type] == 4) {
							|	lea Rd(def_reg), dword [Rd(base_reg)+Rd(index_reg)*4]
						} else {
							|	lea Ra(def_reg), aword [Ra(base_reg)+Ra(index_reg)*4]
						}
						break;
					case 8:
						if (ir_type_size[type] == 4) {
							|	lea Rd(def_reg), dword [Rd(base_reg)+Rd(index_reg)*8]
						} else {
							|	lea Ra(def_reg), aword [Ra(base_reg)+Ra(index_reg)*8]
						}
						break;
				}
			} else {
				switch (scale) {
					default:
						IR_ASSERT(0);
					case 1:
						if (ir_type_size[type] == 4) {
							|	lea Rd(def_reg), dword [Rd(base_reg)+Rd(index_reg)+offset]
						} else {
							|	lea Ra(def_reg), aword [Ra(base_reg)+Ra(index_reg)+offset]
						}
						break;
					case 2:
						if (ir_type_size[type] == 4) {
							|	lea Rd(def_reg), dword [Rd(base_reg)+Rd(index_reg)*2+offset]
						} else {
							|	lea Ra(def_reg), aword [Ra(base_reg)+Ra(index_reg)*2+offset]
						}
						break;
					case 4:
						if (ir_type_size[type] == 4) {
							|	lea Rd(def_reg), dword [Rd(base_reg)+Rd(index_reg)*4+offset]
						} else {
							|	lea Ra(def_reg), aword [Ra(base_reg)+Ra(index_reg)*4+offset]
						}
						break;
					case 8:
						if (ir_type_size[type] == 4) {
							|	lea Rd(def_reg), dword [Rd(base_reg)+Rd(index_reg)*8+offset]
						} else {
							|	lea Ra(def_reg), aword [Ra(base_reg)+Ra(index_reg)*8+offset]
						}
						break;
				}
			}
		}
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, type, def, def_reg);
	}
}

static void ir_emit_tls(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_reg reg = IR_REG_NUM(ctx->regs[def][0]);

	if (ctx->use_lists[def].count == 1) {
		/* dead load */
		return;
	}

|.if X64WIN
|	gs
|	mov Ra(reg), aword [0x58]
|	mov Ra(reg), aword [Ra(reg)+insn->op2]
|	mov Ra(reg), aword [Ra(reg)+insn->op3]
|.elif WIN
|	fs
|	mov Ra(reg), aword [0x2c]
|	mov Ra(reg), aword [Ra(reg)+insn->op2]
|	mov Ra(reg), aword [Ra(reg)+insn->op3]
|.elif X64APPLE
|	gs
||	if (insn->op3 == IR_NULL) {
|		mov Ra(reg), aword [insn->op2]
||	} else {
|		mov Ra(reg), aword [insn->op2]
|		mov Ra(reg), aword [Ra(reg)+insn->op3]
||	}
|.elif X64
|	fs
||	if (insn->op3 == IR_NULL) {
|		mov Ra(reg), aword [insn->op2]
||	} else {
|		mov Ra(reg), [0x8]
|		mov Ra(reg), aword [Ra(reg)+insn->op2]
|		mov Ra(reg), aword [Ra(reg)+insn->op3]
||	}
|.else
|	gs
||	if (insn->op3 == IR_NULL) {
|		mov Ra(reg), aword [insn->op2]
||	} else {
|		mov Ra(reg), [0x4]
|		mov Ra(reg), aword [Ra(reg)+insn->op2]
|		mov Ra(reg), aword [Ra(reg)+insn->op3]
||	}
|	.endif
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, IR_ADDR, def, reg);
	}
}

static void ir_emit_exitcall(ir_ctx *ctx, ir_ref def, ir_insn *insn)
{
	ir_backend_data *data = ctx->data;
	dasm_State **Dst = &data->dasm_state;
	ir_reg def_reg = IR_REG_NUM(ctx->regs[def][0]);

	IR_ASSERT(def_reg != IR_REG_NONE);

	|.if X64
	|	sub rsp, 16*8+16*8+8 /* CPU regs + SSE regs */
	|	mov aword [rsp+0*8], rax
	|	mov aword [rsp+1*8], rcx
	|	mov aword [rsp+2*8], rdx
	|	mov aword [rsp+3*8], rbx
	|	mov aword [rsp+5*8], rbp
	|	mov aword [rsp+6*8], rsi
	|	mov aword [rsp+7*8], rdi
	|	mov aword [rsp+8*8], r8
	|	mov aword [rsp+9*8], r9
	|	mov aword [rsp+10*8], r10
	|	mov aword [rsp+11*8], r11
	|	mov aword [rsp+12*8], r12
	|	mov aword [rsp+13*8], r13
	|	mov aword [rsp+14*8], r14
	|	mov aword [rsp+15*8], r15
	|	movsd qword [rsp+16*8+0*8], xmm0
	|	movsd qword [rsp+16*8+1*8], xmm1
	|	movsd qword [rsp+16*8+2*8], xmm2
	|	movsd qword [rsp+16*8+3*8], xmm3
	|	movsd qword [rsp+16*8+4*8], xmm4
	|	movsd qword [rsp+16*8+5*8], xmm5
	|	movsd qword [rsp+16*8+6*8], xmm6
	|	movsd qword [rsp+16*8+7*8], xmm7
	|	movsd qword [rsp+16*8+8*8], xmm8
	|	movsd qword [rsp+16*8+9*8], xmm9
	|	movsd qword [rsp+16*8+10*8], xmm10
	|	movsd qword [rsp+16*8+11*8], xmm11
	|	movsd qword [rsp+16*8+12*8], xmm12
	|	movsd qword [rsp+16*8+13*8], xmm13
	|	movsd qword [rsp+16*8+14*8], xmm14
	|	movsd qword [rsp+16*8+15*8], xmm15
	|
	|	mov Ra(IR_REG_INT_ARG2), rsp
	|	lea Ra(IR_REG_INT_ARG1), [rsp+16*8+16*8+16]
	|	mov aword [rsp+4*8], Ra(IR_REG_INT_ARG1)
	|	mov Ra(IR_REG_INT_ARG1), [rsp+16*8+16*8+8]
	|.if X64WIN
	|	sub rsp, 32 /* shadow space */
	|.endif
	|.else
	|	sub esp, 8*4+8*8+12 /* CPU regs + SSE regs */
	|	mov aword [esp+0*4], eax
	|	mov aword [esp+1*4], ecx
	|	mov aword [esp+2*4], edx
	|	mov aword [esp+3*4], ebx
	|	mov aword [esp+5*4], ebp
	|	mov aword [esp+6*4], esi
	|	mov aword [esp+7*4], edi
	|	movsd qword [esp+8*4+0*8], xmm0
	|	movsd qword [esp+8*4+1*8], xmm1
	|	movsd qword [esp+8*4+2*8], xmm2
	|	movsd qword [esp+8*4+3*8], xmm3
	|	movsd qword [esp+8*4+4*8], xmm4
	|	movsd qword [esp+8*4+5*8], xmm5
	|	movsd qword [esp+8*4+6*8], xmm6
	|	movsd qword [esp+8*4+7*8], xmm7
	|
	|	mov Ra(IR_REG_INT_FCARG2), esp
	|	lea Ra(IR_REG_INT_FCARG1), [esp+8*4+8*8+16]
	|	mov aword [esp+4*4], Ra(IR_REG_INT_FCARG1)
	|	mov Ra(IR_REG_INT_FCARG1), [esp+8*4+8*8+12]
	|.endif

	if (IR_IS_CONST_REF(insn->op2)) {
		void *addr = ir_jmp_addr(ctx, insn, &ctx->ir_base[insn->op2]);

		if (sizeof(void*) == 4 || IR_MAY_USE_32BIT_ADDR(addr)) {
			|	call aword &addr
		} else {
|.if X64
			if (IR_IS_SIGNED_32BIT(addr)) {
				|	mov rax, ((ptrdiff_t)addr)    // 0x48 0xc7 0xc0 <imm-32-bit>
			} else {
				|	mov64 rax, ((ptrdiff_t)addr)  // 0x48 0xb8 <imm-64-bit>
			}
			|	call rax
|.endif
		}
	} else {
		IR_ASSERT(0);
	}

	//  restore SP
	|.if X64WIN
	|	add rsp, 32+16*8+16*8+16 /* shadow space + CPU regs + SSE regs */
	|.elif X64
	|	add rsp, 16*8+16*8+16 /* CPU regs + SSE regs */
	|.else
	|	add esp, 8*4+8*8+16 /* CPU regs + SSE regs */
	|.endif

	if (def_reg != IR_REG_INT_RET1) {
		ir_emit_mov(ctx, insn->type, def_reg, IR_REG_INT_RET1);
	}
	if (IR_REG_SPILLED(ctx->regs[def][0])) {
		ir_emit_store(ctx, insn->type, def, def_reg);
	}
}

static void ir_emit_param_move(ir_ctx *ctx, uint8_t type, ir_reg from_reg, ir_reg to_reg, ir_ref to, int32_t offset)
{
	ir_reg fp = (ctx->flags & IR_USE_FRAME_POINTER) ? IR_REG_FRAME_POINTER : IR_REG_STACK_POINTER;

	IR_ASSERT(from_reg != IR_REG_NONE || to_reg != IR_REG_NONE);

	if (IR_IS_TYPE_INT(type)) {
		if (from_reg != IR_REG_NONE) {
			if (to_reg != IR_REG_NONE) {
				ir_emit_mov(ctx, type, to_reg, from_reg);
			} else {
				ir_emit_store(ctx, type, to, from_reg);
			}
		} else {
			ir_emit_load_mem_int(ctx, type, to_reg, fp, offset);
		}
	} else {
		if (from_reg != IR_REG_NONE) {
			if (to_reg != IR_REG_NONE) {
				ir_emit_fp_mov(ctx, type, to_reg, from_reg);
			} else {
				ir_emit_store(ctx, type, to, from_reg);
			}
		} else {
			ir_emit_load_mem_fp(ctx, type, to_reg, fp, offset);
		}
	}
}

static void ir_emit_load_params(ir_ctx *ctx)
{
	ir_use_list *use_list = &ctx->use_lists[1];
	ir_insn *insn;
	ir_ref i, n, *p, use;
	int int_param_num = 0;
	int fp_param_num = 0;
	ir_reg src_reg;
	ir_reg dst_reg;
	// TODO: Calling convention specific
	int int_reg_params_count = IR_REG_INT_ARGS;
	int fp_reg_params_count = IR_REG_FP_ARGS;
	const int8_t *int_reg_params = _ir_int_reg_params;
	const int8_t *fp_reg_params = _ir_fp_reg_params;
	int32_t stack_offset = 0;

#ifdef IR_TARGET_X86
	if (sizeof(void*) == 4 && (ctx->flags & IR_FASTCALL_FUNC)) {
		int_reg_params_count = IR_REG_INT_FCARGS;
		fp_reg_params_count = IR_REG_FP_FCARGS;
		int_reg_params = _ir_int_fc_reg_params;
		fp_reg_params = _ir_fp_fc_reg_params;
	}
#endif

	if (ctx->flags & IR_USE_FRAME_POINTER) {
		stack_offset = sizeof(void*) * 2; /* skip old frame pointer and return address */
	} else {
		stack_offset = sizeof(void*) + ctx->stack_frame_size + ctx->call_stack_size; /* skip return address */
	}
	n = use_list->count;
	for (i = 0, p = &ctx->use_edges[use_list->refs]; i < n; i++, p++) {
		use = *p;
		insn = &ctx->ir_base[use];
		if (insn->op == IR_PARAM) {
			if (IR_IS_TYPE_INT(insn->type)) {
				if (int_param_num < int_reg_params_count) {
					src_reg = int_reg_params[int_param_num];
				} else {
					src_reg = IR_REG_NONE;
				}
				int_param_num++;
#ifdef _WIN64
				/* WIN64 calling convention use common couter for int and fp registers */
				fp_param_num++;
#endif
			} else {
				if (fp_param_num < fp_reg_params_count) {
					src_reg = fp_reg_params[fp_param_num];
				} else {
					src_reg = IR_REG_NONE;
				}
				fp_param_num++;
#ifdef _WIN64
				/* WIN64 calling convention use common couter for int and fp registers */
				int_param_num++;
#endif
			}
			if (ctx->vregs[use]) {
				dst_reg = IR_REG_NUM(ctx->regs[use][0]);
				IR_ASSERT(src_reg != IR_REG_NONE || dst_reg != IR_REG_NONE ||
					stack_offset == ctx->live_intervals[ctx->vregs[use]]->stack_spill_pos +
						((ctx->flags & IR_USE_FRAME_POINTER) ? -ctx->stack_frame_size : ctx->call_stack_size));
				if (src_reg != dst_reg) {
					ir_emit_param_move(ctx, insn->type, src_reg, dst_reg, use, stack_offset);
				}
				if (dst_reg != IR_REG_NONE && IR_REG_SPILLED(ctx->regs[use][0])) {
					ir_emit_store(ctx, insn->type, use, dst_reg);
				}
			}
			if (src_reg == IR_REG_NONE) {
				if (sizeof(void*) == 8) {
					stack_offset += sizeof(void*);
				} else {
					stack_offset += IR_MAX(sizeof(void*), ir_type_size[insn->type]);
				}
			}
		}
	}
}

static ir_reg ir_get_free_reg(ir_type type, ir_regset available)
{
	if (IR_IS_TYPE_INT(type)) {
		available = IR_REGSET_INTERSECTION(available, IR_REGSET_GP);
	} else {
		IR_ASSERT(IR_IS_TYPE_FP(type));
		available = IR_REGSET_INTERSECTION(available, IR_REGSET_FP);
	}
	IR_ASSERT(!IR_REGSET_IS_EMPTY(available));
	return IR_REGSET_FIRST(available);
}

static int ir_fix_dessa_tmps(ir_ctx *ctx, uint8_t type, ir_ref from, ir_ref to)
{
	ir_backend_data *data = ctx->data;
	ir_ref ref = ctx->cfg_blocks[data->dessa_from_block].end;

	if (to == 0) {
		if (IR_IS_TYPE_INT(type)) {
			if (ctx->regs[ref][0] == IR_REG_NONE) {
				ctx->regs[ref][0] = IR_REG_RAX;
			}
		} else {
			IR_ASSERT(IR_IS_TYPE_FP(type));
			if (ctx->regs[ref][1] == IR_REG_NONE) {
				ctx->regs[ref][1] = IR_REG_XMM0;
			}
		}
	} else if (from != 0) {
		if (IR_IS_TYPE_INT(type)) {
			if (ctx->regs[ref][0] == IR_REG_NONE) {
				ctx->regs[ref][0] = IR_REG_RAX;
			}
		} else {
			IR_ASSERT(IR_IS_TYPE_FP(type));
			if (ctx->regs[ref][1] == IR_REG_NONE) {
				ctx->regs[ref][1] = IR_REG_XMM0;
			}
		}
	}
	return 1;
}

static void ir_fix_param_spills(ir_ctx *ctx)
{
	ir_use_list *use_list = &ctx->use_lists[1];
	ir_insn *insn;
	ir_ref i, n, *p, use;
	int int_param_num = 0;
	int fp_param_num = 0;
	ir_reg src_reg;
	// TODO: Calling convention specific
	int int_reg_params_count = IR_REG_INT_ARGS;
	int fp_reg_params_count = IR_REG_FP_ARGS;
	const int8_t *int_reg_params = _ir_int_reg_params;
	const int8_t *fp_reg_params = _ir_fp_reg_params;
	int32_t stack_start = 0;
	int32_t stack_offset = 0;

#ifdef IR_TARGET_X86
	if (sizeof(void*) == 4 && (ctx->flags & IR_FASTCALL_FUNC)) {
		int_reg_params_count = IR_REG_INT_FCARGS;
		fp_reg_params_count = IR_REG_FP_FCARGS;
		int_reg_params = _ir_int_fc_reg_params;
		fp_reg_params = _ir_fp_fc_reg_params;
	}
#endif

	if (ctx->flags & IR_USE_FRAME_POINTER) {
		/* skip old frame pointer and return address */
		stack_start = sizeof(void*) * 2 + (ctx->stack_frame_size - ctx->stack_frame_alignment);
	} else {
		 /* skip return address */
		stack_start = sizeof(void*) + ctx->stack_frame_size;
	}
	n = use_list->count;
	for (i = 0, p = &ctx->use_edges[use_list->refs]; i < n; i++, p++) {
		use = *p;
		insn = &ctx->ir_base[use];
		if (insn->op == IR_PARAM) {
			if (IR_IS_TYPE_INT(insn->type)) {
				if (int_param_num < int_reg_params_count) {
					src_reg = int_reg_params[int_param_num];
				} else {
					src_reg = IR_REG_NONE;
				}
				int_param_num++;
#ifdef _WIN64
				/* WIN64 calling convention use common couter for int and fp registers */
				fp_param_num++;
#endif
			} else {
				if (fp_param_num < fp_reg_params_count) {
					src_reg = fp_reg_params[fp_param_num];
				} else {
					src_reg = IR_REG_NONE;
				}
				fp_param_num++;
#ifdef _WIN64
				/* WIN64 calling convention use common couter for int and fp registers */
				int_param_num++;
#endif
			}
			if (src_reg == IR_REG_NONE) {
				if (ctx->vregs[use]) {
					ir_live_interval *ival = ctx->live_intervals[ctx->vregs[use]];
					if ((ival->flags & IR_LIVE_INTERVAL_MEM_PARAM)
					 && ival->stack_spill_pos == -1
					 && (ival->next || ival->reg == IR_REG_NONE)) {
						ival->stack_spill_pos = stack_start + stack_offset;
						ctx->regs[use][0] = IR_REG_NONE;
					}
				}
				if (sizeof(void*) == 8) {
					stack_offset += sizeof(void*);
				} else {
					stack_offset += IR_MAX(sizeof(void*), ir_type_size[insn->type]);
				}
			}
		}
	}

#ifdef IR_TARGET_X86
	if (sizeof(void*) == 4 && (ctx->flags & IR_FASTCALL_FUNC)) {
		ctx->param_stack_size = stack_offset;
	}
#endif
}

static void ir_allocate_unique_spill_slots(ir_ctx *ctx)
{
	uint32_t b;
	ir_block *bb;
	ir_insn *insn;
	ir_ref i, n, j, *p;
	uint32_t *rule, insn_flags;
	ir_backend_data *data = ctx->data;
	ir_regset available = 0;
	ir_target_constraints constraints;
	uint32_t def_flags;
	ir_reg reg;

#ifndef IR_REG_FP_RET1
	if (ctx->ret_type == IR_FLOAT || ctx->ret_type == IR_DOUBLE) {
		ctx->ret_slot = ir_allocate_spill_slot(ctx, ctx->ret_type, &data->ra_data);
	} else {
		ctx->ret_slot = -1;
	}
#endif

	ctx->regs = ir_mem_malloc(sizeof(ir_regs) * ctx->insns_count);
	memset(ctx->regs, IR_REG_NONE, sizeof(ir_regs) * ctx->insns_count);

	/* vregs + tmp + fixed + SRATCH + ALL */
	ctx->live_intervals = ir_mem_calloc(ctx->vregs_count + 1 + IR_REG_NUM + 2, sizeof(ir_live_interval*));

    if (!ctx->arena) {
		ctx->arena = ir_arena_create(16 * 1024);
	}

	for (b = 1, bb = ctx->cfg_blocks + b; b <= ctx->cfg_blocks_count; b++, bb++) {
		IR_ASSERT(!(bb->flags & IR_BB_UNREACHABLE));
		for (i = bb->start, insn = ctx->ir_base + i, rule = ctx->rules + i; i <= bb->end;) {
			switch (ctx->rules ? *rule : insn->op) {
				case IR_START:
				case IR_BEGIN:
				case IR_END:
				case IR_IF_TRUE:
				case IR_IF_FALSE:
				case IR_CASE_VAL:
				case IR_CASE_DEFAULT:
				case IR_MERGE:
				case IR_LOOP_BEGIN:
				case IR_LOOP_END:
					break;
				default:
					def_flags = ir_get_target_constraints(ctx, i, &constraints);
					if (ctx->rules
					 && *rule != IR_CMP_AND_BRANCH_INT
					 && *rule != IR_CMP_AND_BRANCH_FP
					 && *rule != IR_TEST_AND_BRANCH_INT
					 && *rule != IR_GUARD_CMP_INT
					 && *rule != IR_GUARD_CMP_FP) {
						available = IR_REGSET_SCRATCH;
					}
					if (ctx->vregs[i]) {
						reg = constraints.def_reg;
						if (reg != IR_REG_NONE && IR_REGSET_IN(available, reg)) {
							IR_REGSET_EXCL(available, reg);
							ctx->regs[i][0] = reg | IR_REG_SPILL_STORE;
						} else if (def_flags & IR_USE_MUST_BE_IN_REG) {
							if (insn->op == IR_VLOAD
							 && ctx->live_intervals[ctx->vregs[i]]
							 && ctx->live_intervals[ctx->vregs[i]]->stack_spill_pos != -1) {
								/* pass */
							} else if (insn->op != IR_PARAM) {
								reg = ir_get_free_reg(insn->type, available);
								IR_REGSET_EXCL(available, reg);
								ctx->regs[i][0] = reg | IR_REG_SPILL_STORE;
							}
						}
						if (!ctx->live_intervals[ctx->vregs[i]]) {
							ir_live_interval *ival = ir_arena_alloc(&ctx->arena, sizeof(ir_live_interval));
							memset(ival, 0, sizeof(ir_live_interval));
							ctx->live_intervals[ctx->vregs[i]] = ival;
							ival->type = insn->type;
							ival->reg = IR_REG_NONE;
							ival->vreg = ctx->vregs[i];
							ival->stack_spill_pos = -1;
							if (insn->op == IR_PARAM && reg == IR_REG_NONE) {
								ival->flags |= IR_LIVE_INTERVAL_MEM_PARAM;
							} else {
								ival->stack_spill_pos = ir_allocate_spill_slot(ctx, ival->type, &data->ra_data);
							}
						} else if (insn->op == IR_PARAM) {
							IR_ASSERT(0 && "unexpected PARAM");
							return;
						}
					} else if (insn->op == IR_VAR) {
						ir_use_list *use_list = &ctx->use_lists[i];
						ir_ref n = use_list->count;

						if (n > 0) {
							int32_t stack_spill_pos = insn->op3 = ir_allocate_spill_slot(ctx, insn->type, &data->ra_data);
							ir_ref i, *p, use;
							ir_insn *use_insn;

							for (i = 0, p = &ctx->use_edges[use_list->refs]; i < n; i++, p++) {
								use = *p;
								use_insn = &ctx->ir_base[use];
								if (use_insn->op == IR_VLOAD) {
									if (ctx->vregs[use]
									 && !ctx->live_intervals[ctx->vregs[use]]) {
										ir_live_interval *ival = ir_arena_alloc(&ctx->arena, sizeof(ir_live_interval));
										memset(ival, 0, sizeof(ir_live_interval));
										ctx->live_intervals[ctx->vregs[use]] = ival;
										ival->type = insn->type;
										ival->reg = IR_REG_NONE;
										ival->vreg = ctx->vregs[use];
										ival->stack_spill_pos = stack_spill_pos;
									}
								} else if (use_insn->op == IR_VSTORE) {
									if (!IR_IS_CONST_REF(use_insn->op3)
									 && ctx->vregs[use_insn->op3]
									 && !ctx->live_intervals[ctx->vregs[use_insn->op3]]) {
										ir_live_interval *ival = ir_arena_alloc(&ctx->arena, sizeof(ir_live_interval));
										memset(ival, 0, sizeof(ir_live_interval));
										ctx->live_intervals[ctx->vregs[use_insn->op3]] = ival;
										ival->type = insn->type;
										ival->reg = IR_REG_NONE;
										ival->vreg = ctx->vregs[use_insn->op3];
										ival->stack_spill_pos = stack_spill_pos;
									}
								}
							}
						}
					}

					insn_flags = ir_op_flags[insn->op];
					n = constraints.tmps_count;
					if (n) {
						do {
							n--;
							if (constraints.tmp_regs[n].type) {
								ir_reg reg = ir_get_free_reg(constraints.tmp_regs[n].type, available);
								IR_REGSET_EXCL(available, reg);
								ctx->regs[i][constraints.tmp_regs[n].num] = reg;
							} else if (constraints.tmp_regs[n].reg == IR_REG_SCRATCH) {
								available = IR_REGSET_DIFFERENCE(available, IR_REGSET_SCRATCH);
							} else {
								IR_REGSET_EXCL(available, constraints.tmp_regs[n].reg);
							}
						} while (n);
					}
					n = insn->inputs_count;
					for (j = 1, p = insn->ops + 1; j <= n; j++, p++) {
						ir_ref input = *p;
						if (IR_OPND_KIND(insn_flags, j) == IR_OPND_DATA && input > 0 && ctx->vregs[input]) {
							if ((def_flags & IR_DEF_REUSES_OP1_REG) && j == 1) {
								ir_reg reg = IR_REG_NUM(ctx->regs[i][0]);
								ctx->regs[i][1] = reg | IR_REG_SPILL_LOAD;
							} else {
								uint8_t use_flags = IR_USE_FLAGS(def_flags, j);
								ir_reg reg = (j < constraints.hints_count) ? constraints.hints[j] : IR_REG_NONE;

								if (reg != IR_REG_NONE && IR_REGSET_IN(available, reg)) {
									IR_REGSET_EXCL(available, reg);
									ctx->regs[i][j] = reg | IR_REG_SPILL_LOAD;
								} else if (j > 1 && input == insn->op1 && ctx->regs[i][1] != IR_REG_NONE) {
									ctx->regs[i][j] = ctx->regs[i][1];
								} else if (use_flags & IR_USE_MUST_BE_IN_REG) {
									reg = ir_get_free_reg(ctx->ir_base[input].type, available);
									IR_REGSET_EXCL(available, reg);
									ctx->regs[i][j] = reg | IR_REG_SPILL_LOAD;
								}
							}
						}
					}
					break;
			}
			n = ir_insn_len(insn);
			i += n;
			insn += n;
			rule += n;
		}
		if (bb->flags & IR_BB_DESSA_MOVES) {
			data->dessa_from_block = b;
			ir_gen_dessa_moves(ctx, b, ir_fix_dessa_tmps);
		}
	}

	ctx->used_preserved_regs = ctx->fixed_save_regset;
	ctx->flags |= IR_NO_STACK_COMBINE;
	ir_fix_stack_frame(ctx);
}

static void ir_preallocate_call_stack(ir_ctx *ctx)
{
	int call_stack_size, peak_call_stack_size = 0;
	ir_ref i, n;
	ir_insn *insn;

	for (i = 1, insn = ctx->ir_base + 1; i < ctx->insns_count;) {
		if (insn->op == IR_CALL) {
			call_stack_size = ir_call_used_stack(ctx, insn);
			if (call_stack_size > peak_call_stack_size
#ifdef IR_HAVE_FASTCALL
			 && !ir_is_fastcall(ctx, insn) /* fast call functions restore stack pointer */
#endif
			) {
				peak_call_stack_size = call_stack_size;
			}
		}
		n = ir_insn_len(insn);
		i += n;
		insn += n;
	}
	if (peak_call_stack_size) {
		ctx->call_stack_size = peak_call_stack_size;
		ctx->flags |= IR_PREALLOCATED_STACK;
	}
}

void ir_fix_stack_frame(ir_ctx *ctx)
{
	uint32_t additional_size = 0;

	if (ctx->used_preserved_regs) {
		ir_regset used_preserved_regs = (ir_regset)ctx->used_preserved_regs;
		ir_reg reg;
		(void) reg;

		IR_REGSET_FOREACH(used_preserved_regs, reg) {
			additional_size += sizeof(void*);
		} IR_REGSET_FOREACH_END();
	}

	ctx->stack_frame_size = IR_ALIGNED_SIZE(ctx->stack_frame_size, sizeof(void*));
	ctx->stack_frame_size += additional_size;
	ctx->stack_frame_alignment = 0;
	ctx->call_stack_size = 0;

	if (ctx->flags2 & IR_HAS_CALLS) {
		/* Stack must be 16 byte aligned */
		if (!(ctx->flags & IR_FUNCTION)) {
			while (IR_ALIGNED_SIZE(ctx->stack_frame_size, 16) != ctx->stack_frame_size) {
				ctx->stack_frame_size += sizeof(void*);
				ctx->stack_frame_alignment += sizeof(void*);
			}
		} else if (ctx->flags & IR_USE_FRAME_POINTER) {
			while (IR_ALIGNED_SIZE(ctx->stack_frame_size + sizeof(void*) * 2, 16) != ctx->stack_frame_size + sizeof(void*) * 2) {
				ctx->stack_frame_size += sizeof(void*);
				ctx->stack_frame_alignment += sizeof(void*);
			}
		} else {
			if (!(ctx->flags & IR_NO_STACK_COMBINE)) {
				ir_preallocate_call_stack(ctx);
			}
			while (IR_ALIGNED_SIZE(ctx->stack_frame_size + ctx->call_stack_size + sizeof(void*), 16) !=
					ctx->stack_frame_size + ctx->call_stack_size + sizeof(void*)) {
				ctx->stack_frame_size += sizeof(void*);
				ctx->stack_frame_alignment += sizeof(void*);
			}
		}
	}

	ir_fix_param_spills(ctx);
}

static void* dasm_labels[ir_lb_MAX];

void *ir_emit_code(ir_ctx *ctx, size_t *size_ptr)
{
	uint32_t b, n, target;
	ir_block *bb;
	ir_ref i;
	ir_insn *insn;
	uint32_t *rule;
	ir_backend_data data;
	dasm_State **Dst;
	int ret;
	void *entry;
	size_t size;

	data.ra_data.unused_slot_4 = 0;
	data.ra_data.unused_slot_2 = 0;
	data.ra_data.unused_slot_1 = 0;
	data.ra_data.handled = NULL;
	data.rodata_label = 0;
	data.jmp_table_label = 0;
	data.double_neg_const = 0;
	data.float_neg_const = 0;
	data.double_abs_const = 0;
	data.float_abs_const = 0;
	data.double_zero_const = 0;
	ctx->data = &data;

	if (!ctx->live_intervals) {
		ctx->stack_frame_size = 0;
		ctx->stack_frame_alignment = 0;
		ctx->call_stack_size = 0;
		ctx->used_preserved_regs = 0;
		ir_allocate_unique_spill_slots(ctx);
	}

	if (ctx->fixed_stack_frame_size != -1) {
		if (ctx->fixed_stack_red_zone) {
			IR_ASSERT(ctx->fixed_stack_red_zone == ctx->fixed_stack_frame_size + ctx->fixed_call_stack_size);
		}
		if (ctx->stack_frame_size > ctx->fixed_stack_frame_size) {
			// TODO: report error to caller
#ifdef IR_DEBUG_MESSAGES
			fprintf(stderr, "IR Compilation Aborted: ctx->stack_frame_size > ctx->fixed_stack_frame_size at %s:%d\n",
				__FILE__, __LINE__);
#endif
			ctx->data = NULL;
			ctx->status = IR_ERROR_FIXED_STACK_FRAME_OVERFLOW;
			return NULL;
		}
		ctx->stack_frame_size = ctx->fixed_stack_frame_size;
		ctx->call_stack_size = ctx->fixed_call_stack_size;
		ctx->stack_frame_alignment = 0;
	}

	Dst = &data.dasm_state;
	data.dasm_state = NULL;
	dasm_init(&data.dasm_state, DASM_MAXSECTION);
	dasm_setupglobal(&data.dasm_state, dasm_labels, ir_lb_MAX);
	dasm_setup(&data.dasm_state, dasm_actions);
	/* labels for each block + for each constant + rodata label + jmp_table label + for each entry */
	dasm_growpc(&data.dasm_state, ctx->cfg_blocks_count + 1 + ctx->consts_count + 1 + 1 + 1 + ctx->entries_count);

	if ((ctx->flags & IR_GEN_ENDBR) && (ctx->flags & IR_START_BR_TARGET)) {
		|.if X64
		|	endbr64
		|.else
		|	endbr32
		|.endif
	}

	if (!(ctx->flags & IR_SKIP_PROLOGUE)) {
		ir_emit_prologue(ctx);
	}
	if (ctx->flags & IR_FUNCTION) {
		ir_emit_load_params(ctx);
	}

	for (b = 1, bb = ctx->cfg_blocks + b; b <= ctx->cfg_blocks_count; b++, bb++) {
		IR_ASSERT(!(bb->flags & IR_BB_UNREACHABLE));
		if ((bb->flags & (IR_BB_START|IR_BB_ENTRY|IR_BB_EMPTY)) == IR_BB_EMPTY) {
			continue;
		}
		|=>b:

		i = bb->start;
		insn = ctx->ir_base + i;
		if (bb->flags & IR_BB_ENTRY) {
			uint32_t label = ctx->cfg_blocks_count + ctx->consts_count + 4 + insn->op3;

			|=>label:
			if ((ctx->flags & IR_GEN_ENDBR) && (ctx->flags & IR_ENTRY_BR_TARGET)) {
				|.if X64
				|	endbr64
				|.else
				|	endbr32
				|.endif
			}
			ir_emit_prologue(ctx);
			ctx->entries[insn->op3] = i;
		}

		/* skip first instruction */
		n = ir_insn_len(insn);
		i += n;
		insn += n;
		rule = ctx->rules + i;

		while (i <= bb->end) {
			if (!((*rule) & (IR_FUSED|IR_SKIPPED)))
			switch (*rule) {
				case IR_VAR:
				case IR_PARAM:
				case IR_PI:
				case IR_PHI:
				case IR_SNAPSHOT:
					break;
				case IR_LEA_OB:
					{
						ir_reg op1_reg = ctx->regs[i][1];
						int32_t offset = ctx->ir_base[insn->op2].val.i32;

						if (insn->op == IR_ADD) {
							offset = ctx->ir_base[insn->op2].val.i32;
						} else {
							IR_ASSERT(insn->op == IR_SUB);
							int64_t long_offset = ctx->ir_base[insn->op2].val.i64;
							long_offset = -long_offset;
							IR_ASSERT(IR_IS_SIGNED_32BIT(long_offset));
							offset = (int32_t)long_offset;
						}
						if (IR_REG_SPILLED(op1_reg)) {
							op1_reg = IR_REG_NUM(op1_reg);
							ir_emit_load(ctx, insn->type, op1_reg, insn->op1);
						}
						ir_emit_lea(ctx, i, insn->type, op1_reg, IR_REG_NONE, 1, offset);
					}
					break;
				case IR_LEA_SI:
					{
						ir_reg op1_reg = ctx->regs[i][1];
						int32_t scale = ctx->ir_base[insn->op2].val.i32;

						IR_ASSERT(op1_reg != IR_REG_NONE);
						if (IR_REG_SPILLED(op1_reg)) {
							op1_reg = IR_REG_NUM(op1_reg);
							ir_emit_load(ctx, insn->type, op1_reg, insn->op1);
						}
						ir_emit_lea(ctx, i, insn->type, IR_REG_NONE, op1_reg, scale, 0);
					}
					break;
				case IR_LEA_SIB:
					{
						ir_reg op1_reg = ctx->regs[i][1];
						int32_t scale = ctx->ir_base[insn->op2].val.i32;

						IR_ASSERT(op1_reg != IR_REG_NONE);
						if (IR_REG_SPILLED(op1_reg)) {
							op1_reg = IR_REG_NUM(op1_reg);
							ir_emit_load(ctx, insn->type, op1_reg, insn->op1);
						}
						ir_emit_lea(ctx, i, insn->type, op1_reg, op1_reg, scale - 1, 0);
					}
					break;
				case IR_LEA_IB:
					{
						ir_reg op1_reg = ctx->regs[i][1];
						ir_reg op2_reg = ctx->regs[i][2];

						IR_ASSERT(op1_reg != IR_REG_NONE && op2_reg != IR_REG_NONE);
						if (IR_REG_SPILLED(op1_reg)) {
							op1_reg = IR_REG_NUM(op1_reg);
							ir_emit_load(ctx, insn->type, op1_reg, insn->op1);
						}
						if (IR_REG_SPILLED(op2_reg)) {
							op2_reg = IR_REG_NUM(op2_reg);
							ir_emit_load(ctx, insn->type, op2_reg, insn->op2);
						}
						ir_emit_lea(ctx, i, insn->type, op1_reg, op2_reg, 1, 0);
					}
					break;
				case IR_LEA_OB_I:
					{
						ir_insn *op1_insn = &ctx->ir_base[insn->op1];
						ir_reg op1_reg = ctx->regs[insn->op1][1];
						ir_reg op2_reg = ctx->regs[i][2];
						int32_t offset = ctx->ir_base[op1_insn->op2].val.i32;

						IR_ASSERT(op1_reg != IR_REG_NONE && op2_reg != IR_REG_NONE);
						if (IR_REG_SPILLED(op1_reg)) {
							op1_reg = IR_REG_NUM(op1_reg);
							ir_emit_load(ctx, insn->type, op1_reg, op1_insn->op1);
						}
						if (IR_REG_SPILLED(op2_reg)) {
							op2_reg = IR_REG_NUM(op2_reg);
							ir_emit_load(ctx, insn->type, op2_reg, insn->op2);
						}
						if (op1_insn->op == IR_SUB) {
							offset = -offset;
						}
						ir_emit_lea(ctx, i, insn->type, op1_reg, op2_reg, 1, offset);
					}
					break;
				case IR_LEA_I_OB:
					{
						ir_insn *op2_insn = &ctx->ir_base[insn->op2];
						ir_reg op1_reg = ctx->regs[i][1];
						ir_reg op2_reg = ctx->regs[insn->op2][1];
						int32_t offset = ctx->ir_base[op2_insn->op2].val.i32;

						IR_ASSERT(op1_reg != IR_REG_NONE && op2_reg != IR_REG_NONE);
						if (IR_REG_SPILLED(op1_reg)) {
							op1_reg = IR_REG_NUM(op1_reg);
							ir_emit_load(ctx, insn->type, op1_reg, insn->op1);
						}
						if (IR_REG_SPILLED(op2_reg)) {
							op2_reg = IR_REG_NUM(op2_reg);
							ir_emit_load(ctx, insn->type, op2_reg, op2_insn->op1);
						}
						if (op2_insn->op == IR_SUB) {
							offset = -offset;
						}
						ir_emit_lea(ctx, i, insn->type, op1_reg, op2_reg, 1, offset);
					}
					break;
				case IR_LEA_SI_O:
					{
						ir_insn *op1_insn = &ctx->ir_base[insn->op1];
						ir_reg op1_reg = ctx->regs[insn->op1][1];
						int32_t scale = ctx->ir_base[op1_insn->op2].val.i32;
						int32_t offset = ctx->ir_base[insn->op2].val.i32;

						IR_ASSERT(op1_reg != IR_REG_NONE);
						if (IR_REG_SPILLED(op1_reg)) {
							op1_reg = IR_REG_NUM(op1_reg);
							ir_emit_load(ctx, insn->type, op1_reg, op1_insn->op1);
						}
						if (insn->op == IR_SUB) {
							offset = -offset;
						}
						ir_emit_lea(ctx, i, insn->type, IR_REG_NONE, op1_reg, scale, offset);
					}
					break;
				case IR_LEA_SIB_O:
					{
						ir_insn *op1_insn = &ctx->ir_base[insn->op1];
						ir_reg op1_reg = ctx->regs[insn->op1][1];
						int32_t scale = ctx->ir_base[op1_insn->op2].val.i32;
						int32_t offset = ctx->ir_base[insn->op2].val.i32;

						IR_ASSERT(op1_reg != IR_REG_NONE);
						if (IR_REG_SPILLED(op1_reg)) {
							op1_reg = IR_REG_NUM(op1_reg);
							ir_emit_load(ctx, insn->type, op1_reg, op1_insn->op1);
						}
						if (insn->op == IR_SUB) {
							offset = -offset;
						}
						ir_emit_lea(ctx, i, insn->type, op1_reg, op1_reg, scale - 1, offset);
					}
					break;
				case IR_LEA_IB_O:
					{
						ir_insn *op1_insn = &ctx->ir_base[insn->op1];
						ir_reg op1_reg = ctx->regs[insn->op1][1];
						ir_reg op2_reg = ctx->regs[insn->op1][2];
						int32_t offset = ctx->ir_base[insn->op2].val.i32;

						IR_ASSERT(op1_reg != IR_REG_NONE && op2_reg != IR_REG_NONE);
						if (IR_REG_SPILLED(op1_reg)) {
							op1_reg = IR_REG_NUM(op1_reg);
							ir_emit_load(ctx, insn->type, op1_reg, op1_insn->op1);
						}
						if (IR_REG_SPILLED(op2_reg)) {
							op2_reg = IR_REG_NUM(op2_reg);
							ir_emit_load(ctx, insn->type, op2_reg, op1_insn->op2);
						}
						if (insn->op == IR_SUB) {
							offset = -offset;
						}
						ir_emit_lea(ctx, i, insn->type, op1_reg, op2_reg, 1, offset);
					}
					break;
				case IR_LEA_OB_SI:
					{
						ir_insn *op1_insn = &ctx->ir_base[insn->op1];
						ir_insn *op2_insn = &ctx->ir_base[insn->op2];
						ir_reg op1_reg = ctx->regs[insn->op1][1];
						ir_reg op2_reg = ctx->regs[insn->op2][1];
						int32_t offset = ctx->ir_base[op1_insn->op2].val.i32;
						int32_t scale = ctx->ir_base[op2_insn->op2].val.i32;

						IR_ASSERT(op1_reg != IR_REG_NONE && op2_reg != IR_REG_NONE);
						if (IR_REG_SPILLED(op1_reg)) {
							op1_reg = IR_REG_NUM(op1_reg);
							ir_emit_load(ctx, insn->type, op1_reg, op1_insn->op1);
						}
						if (IR_REG_SPILLED(op2_reg)) {
							op2_reg = IR_REG_NUM(op2_reg);
							ir_emit_load(ctx, insn->type, op2_reg, op2_insn->op1);
						}
						if (op1_insn->op == IR_SUB) {
							offset = -offset;
						}
						ir_emit_lea(ctx, i, insn->type, op1_reg, op2_reg, scale, offset);
					}
					break;
				case IR_LEA_SI_OB:
					{
						ir_insn *op1_insn = &ctx->ir_base[insn->op1];
						ir_insn *op2_insn = &ctx->ir_base[insn->op2];
						ir_reg op1_reg = ctx->regs[insn->op1][1];
						ir_reg op2_reg = ctx->regs[insn->op2][1];
						int32_t scale = ctx->ir_base[op1_insn->op2].val.i32;
						int32_t offset = ctx->ir_base[op2_insn->op2].val.i32;

						IR_ASSERT(op1_reg != IR_REG_NONE && op2_reg != IR_REG_NONE);
						if (IR_REG_SPILLED(op1_reg)) {
							op1_reg = IR_REG_NUM(op1_reg);
							ir_emit_load(ctx, insn->type, op1_reg, op1_insn->op1);
						}
						if (IR_REG_SPILLED(op2_reg)) {
							op2_reg = IR_REG_NUM(op2_reg);
							ir_emit_load(ctx, insn->type, op2_reg, op2_insn->op1);
						}
						if (op1_insn->op == IR_SUB) {
							offset = -offset;
						}
						ir_emit_lea(ctx, i, insn->type, op2_reg, op1_reg, scale, offset);
					}
					break;
				case IR_LEA_B_SI:
					{
						ir_insn *op2_insn = &ctx->ir_base[insn->op2];
						ir_reg op1_reg = ctx->regs[i][1];
						ir_reg op2_reg = ctx->regs[insn->op2][1];
						int32_t scale = ctx->ir_base[op2_insn->op2].val.i32;

						IR_ASSERT(op1_reg != IR_REG_NONE && op2_reg != IR_REG_NONE);
						if (IR_REG_SPILLED(op1_reg)) {
							op1_reg = IR_REG_NUM(op1_reg);
							ir_emit_load(ctx, insn->type, op1_reg, insn->op1);
						}
						if (IR_REG_SPILLED(op2_reg)) {
							op2_reg = IR_REG_NUM(op2_reg);
							ir_emit_load(ctx, insn->type, op2_reg, op2_insn->op1);
						}
						ir_emit_lea(ctx, i, insn->type, op1_reg, op2_reg, scale, 0);
					}
					break;
				case IR_LEA_SI_B:
					{
						ir_insn *op1_insn = &ctx->ir_base[insn->op1];
						ir_reg op1_reg = ctx->regs[insn->op1][1];
						ir_reg op2_reg = ctx->regs[i][2];
						int32_t scale = ctx->ir_base[op1_insn->op2].val.i32;

						IR_ASSERT(op1_reg != IR_REG_NONE && op2_reg != IR_REG_NONE);
						if (IR_REG_SPILLED(op1_reg)) {
							op1_reg = IR_REG_NUM(op1_reg);
							ir_emit_load(ctx, insn->type, op1_reg, op1_insn->op1);
						}
						if (IR_REG_SPILLED(op2_reg)) {
							op2_reg = IR_REG_NUM(op2_reg);
							ir_emit_load(ctx, insn->type, op2_reg, insn->op2);
						}
						ir_emit_lea(ctx, i, insn->type, op2_reg, op1_reg, scale, 0);
					}
					break;
				case IR_MUL_PWR2:
				case IR_DIV_PWR2:
				case IR_MOD_PWR2:
					ir_emit_mul_div_mod_pwr2(ctx, i, insn);
					break;
				case IR_SHIFT:
					ir_emit_shift(ctx, i, insn);
					break;
				case IR_SHIFT_CONST:
					ir_emit_shift_const(ctx, i, insn);
					break;
				case IR_BIT_COUNT:
					ir_emit_bit_count(ctx, i, insn);
					break;
				case IR_CTPOP:
					ir_emit_ctpop(ctx, i, insn);
					break;
				case IR_INC:
				case IR_DEC:
				case IR_OP_INT:
					ir_emit_op_int(ctx, i, insn);
					break;
				case IR_ABS_INT:
					ir_emit_abs_int(ctx, i, insn);
					break;
				case IR_BOOL_NOT_INT:
					ir_emit_bool_not_int(ctx, i, insn);
					break;
				case IR_OP_FP:
					ir_emit_op_fp(ctx, i, insn);
					break;
				case IR_IMUL3:
					ir_emit_imul3(ctx, i, insn);
					break;
				case IR_BINOP_INT:
					ir_emit_binop_int(ctx, i, insn);
					break;
				case IR_BINOP_SSE2:
					ir_emit_binop_sse2(ctx, i, insn);
					break;
				case IR_BINOP_AVX:
					ir_emit_binop_avx(ctx, i, insn);
					break;
				case IR_MUL_INT:
				case IR_DIV_INT:
				case IR_MOD_INT:
					ir_emit_mul_div_mod(ctx, i, insn);
					break;
				case IR_CMP_INT:
					ir_emit_cmp_int(ctx, i, insn);
					break;
				case IR_TESTCC_INT:
					ir_emit_testcc_int(ctx, i, insn);
					break;
				case IR_SETCC_INT:
					ir_emit_setcc_int(ctx, i, insn);
					break;
				case IR_CMP_FP:
					ir_emit_cmp_fp(ctx, i, insn);
					break;
				case IR_SEXT:
					ir_emit_sext(ctx, i, insn);
					break;
				case IR_ZEXT:
					ir_emit_zext(ctx, i, insn);
					break;
				case IR_TRUNC:
					ir_emit_trunc(ctx, i, insn);
					break;
				case IR_BITCAST:
					ir_emit_bitcast(ctx, i, insn);
					break;
				case IR_INT2FP:
					ir_emit_int2fp(ctx, i, insn);
					break;
				case IR_FP2INT:
					ir_emit_fp2int(ctx, i, insn);
					break;
				case IR_FP2FP:
					ir_emit_fp2fp(ctx, i, insn);
					break;
				case IR_COPY_INT:
					ir_emit_copy_int(ctx, i, insn);
					break;
				case IR_COPY_FP:
					ir_emit_copy_fp(ctx, i, insn);
					break;
				case IR_CMP_AND_BRANCH_INT:
					ir_emit_cmp_and_branch_int(ctx, b, i, insn);
					break;
				case IR_CMP_AND_BRANCH_FP:
					ir_emit_cmp_and_branch_fp(ctx, b, i, insn);
					break;
				case IR_TEST_AND_BRANCH_INT:
					ir_emit_test_and_branch_int(ctx, b, i, insn);
					break;
				case IR_JCC_INT:
					{
						ir_op op = ctx->ir_base[insn->op2].op;

						if (op == IR_ADD ||
						    op == IR_SUB ||
//						    op == IR_MUL ||
						    op == IR_OR  ||
						    op == IR_AND ||
						    op == IR_XOR) {
							op = IR_NE;
						} else {
							IR_ASSERT(op >= IR_EQ && op <= IR_UGT);
						}
						ir_emit_jcc(ctx, op, b, i, insn, 1);
					}
					break;
				case IR_GUARD_CMP_INT:
					if (ir_emit_guard_cmp_int(ctx, b, i, insn)) {
						goto next_block;
					}
					break;
				case IR_GUARD_CMP_FP:
					if (ir_emit_guard_cmp_fp(ctx, b, i, insn)) {
						goto next_block;
					}
					break;
				case IR_GUARD_TEST_INT:
					if (ir_emit_guard_test_int(ctx, b, i, insn)) {
						goto next_block;
					}
					break;
				case IR_GUARD_JCC_INT:
					if (ir_emit_guard_jcc_int(ctx, b, i, insn)) {
						goto next_block;
					}
					break;
				case IR_IF_INT:
					ir_emit_if_int(ctx, b, i, insn);
					break;
				case IR_COND:
					ir_emit_cond(ctx, i, insn);
					break;
				case IR_SWITCH:
					ir_emit_switch(ctx, b, i, insn);
					break;
				case IR_MIN_MAX_INT:
					ir_emit_min_max_int(ctx, i, insn);
					break;
				case IR_OVERFLOW:
					ir_emit_overflow(ctx, i, insn);
					break;
				case IR_OVERFLOW_AND_BRANCH:
					ir_emit_overflow_and_branch(ctx, b, i, insn);
					break;
				case IR_END:
				case IR_LOOP_END:
					if (bb->flags & IR_BB_OSR_ENTRY_LOADS) {
						ir_emit_osr_entry_loads(ctx, b, bb);
					}
					if (bb->flags & IR_BB_DESSA_MOVES) {
						ir_emit_dessa_moves(ctx, b, bb);
					}
					do {
						ir_ref succ = ctx->cfg_edges[bb->successors];

						if (UNEXPECTED(bb->successors_count == 2)) {
							if (ctx->cfg_blocks[succ].flags & IR_BB_ENTRY) {
								succ = ctx->cfg_edges[bb->successors + 1];
							} else {
								IR_ASSERT(ctx->cfg_blocks[ctx->cfg_edges[bb->successors + 1]].flags & IR_BB_ENTRY);
							}
						} else {
							IR_ASSERT(bb->successors_count == 1);
						}
						target = ir_skip_empty_target_blocks(ctx, succ);
						if (b == ctx->cfg_blocks_count || target != ir_skip_empty_next_blocks(ctx, b + 1)) {
							|	jmp =>target
						}
					} while (0);
					break;
				case IR_RETURN_VOID:
					ir_emit_return_void(ctx);
					break;
				case IR_RETURN_INT:
					ir_emit_return_int(ctx, i, insn);
					break;
				case IR_RETURN_FP:
					ir_emit_return_fp(ctx, i, insn);
					break;
				case IR_CALL:
					ir_emit_call(ctx, i, insn);
					break;
				case IR_TAILCALL:
					ir_emit_tailcall(ctx, i, insn);
					break;
				case IR_IJMP:
					ir_emit_ijmp(ctx, i, insn);
					break;
				case IR_MEM_OP_INT:
				case IR_MEM_INC:
				case IR_MEM_DEC:
					ir_emit_mem_op_int(ctx, i, insn);
					break;
				case IR_MEM_BINOP_INT:
					ir_emit_mem_binop_int(ctx, i, insn);
					break;
				case IR_MEM_MUL_PWR2:
				case IR_MEM_DIV_PWR2:
				case IR_MEM_MOD_PWR2:
					ir_emit_mem_mul_div_mod_pwr2(ctx, i, insn);
					break;
				case IR_MEM_SHIFT:
					ir_emit_mem_shift(ctx, i, insn);
					break;
				case IR_MEM_SHIFT_CONST:
					ir_emit_mem_shift_const(ctx, i, insn);
					break;
				case IR_REG_BINOP_INT:
					ir_emit_reg_binop_int(ctx, i, insn);
					break;
				case IR_VADDR:
					ir_emit_vaddr(ctx, i, insn);
					break;
				case IR_VLOAD:
					ir_emit_vload(ctx, i, insn);
					break;
				case IR_VSTORE_INT:
					ir_emit_vstore_int(ctx, i, insn);
					break;
				case IR_VSTORE_FP:
					ir_emit_vstore_fp(ctx, i, insn);
					break;
				case IR_RLOAD:
					ir_emit_rload(ctx, i, insn);
					break;
				case IR_RSTORE:
					ir_emit_rstore(ctx, i, insn);
					break;
				case IR_LOAD_INT:
					ir_emit_load_int(ctx, i, insn);
					break;
				case IR_LOAD_FP:
					ir_emit_load_fp(ctx, i, insn);
					break;
				case IR_STORE_INT:
					ir_emit_store_int(ctx, i, insn);
					break;
				case IR_STORE_FP:
					ir_emit_store_fp(ctx, i, insn);
					break;
				case IR_ALLOCA:
					ir_emit_alloca(ctx, i, insn);
					break;
				case IR_AFREE:
					ir_emit_afree(ctx, i, insn);
					break;
				case IR_FRAME_ADDR:
					ir_emit_frame_addr(ctx, i);
					break;
				case IR_EXITCALL:
					ir_emit_exitcall(ctx, i, insn);
					break;
				case IR_GUARD:
				case IR_GUARD_NOT:
					if (ir_emit_guard(ctx, b, i, insn)) {
						goto next_block;
					}
					break;
				case IR_GUARD_OVERFLOW:
					if (ir_emit_guard_overflow(ctx, b, i, insn)) {
						goto next_block;
					}
					break;
				case IR_TLS:
					ir_emit_tls(ctx, i, insn);
					break;
				case IR_TRAP:
					|	int3
					break;
				default:
					IR_ASSERT(0 && "NIY rule/instruction");
					dasm_free(&data.dasm_state);
					ctx->data = NULL;
					ctx->status = IR_ERROR_UNSUPPORTED_CODE_RULE;
					return NULL;
			}
			n = ir_insn_len(insn);
			i += n;
			insn += n;
			rule += n;
		}
next_block:;
	}

	if (data.rodata_label) {
		|.rodata
	}
	for (i = IR_UNUSED + 1, insn = ctx->ir_base - i; i < ctx->consts_count; i++, insn--) {
		if (insn->const_flags & IR_CONST_EMIT) {
			if (IR_IS_TYPE_FP(insn->type)) {
				int label = ctx->cfg_blocks_count + i;

				if (!data.rodata_label) {
					data.rodata_label = ctx->cfg_blocks_count + ctx->consts_count + 2;

					|.rodata
					|=>data.rodata_label:
				}
				if (insn->type == IR_DOUBLE) {
					|.align 8
					|=>label:
					|.dword insn->val.u32, insn->val.u32_hi
				} else {
					IR_ASSERT(insn->type == IR_FLOAT);
					|.align 4
					|=>label:
					|.dword insn->val.u32
				}
			} else if (insn->op == IR_STR) {
				int label = ctx->cfg_blocks_count + i;
				const char *str = ir_get_str(ctx, insn->val.i32);
				int i = 0;

				if (!data.rodata_label) {
					data.rodata_label = ctx->cfg_blocks_count + ctx->consts_count + 2;

					|.rodata
					|=>data.rodata_label:
				}
				|.align 8
				|=>label:
				while (str[i]) {
					char c = str[i];

					if (c == '\\') {
						if (str[i+1] == '\\') {
							i++;
							c = '\\';
						} else if (str[i+1] == '\'') {
							i++;
							c = '\'';
						} else if (str[i+1] == '"') {
							i++;
							c = '"';
						} else if (str[i+1] == 'a') {
							i++;
							c = '\a';
						} else if (str[i+1] == 'b') {
							i++;
							c = '\b';
						} else if (str[i+1] == 'e') {
							i++;
							c = 27; /* '\e'; */
						} else if (str[i+1] == 'f') {
							i++;
							c = '\f';
						} else if (str[i+1] == 'n') {
							i++;
							c = '\n';
						} else if (str[i+1] == 'r') {
							i++;
							c = '\r';
						} else if (str[i+1] == 't') {
							i++;
							c = '\t';
						} else if (str[i+1] == 'v') {
							i++;
							c = '\v';
						} else if (str[i+1] == '?') {
							i++;
							c = 0x3f;
						}
					}
					|.byte c
					i++;
				}
				|.byte 0

			} else {
				IR_ASSERT(0);
			}
		}
	}
	if (data.rodata_label) {
		|.code
	}

	if (ctx->status) {
		dasm_free(&data.dasm_state);
		ctx->data = NULL;
		return NULL;
	}

	ret = dasm_link(&data.dasm_state, size_ptr);
	if (ret != DASM_S_OK) {
		IR_ASSERT(0);
		dasm_free(&data.dasm_state);
		ctx->data = NULL;
		ctx->status = IR_ERROR_LINK;
		return NULL;
	}
	size = *size_ptr;

	if (ctx->code_buffer != NULL) {
		if (IR_ALIGNED_SIZE(size, 16) > ctx->code_buffer_size) {
			dasm_free(&data.dasm_state);
			ctx->data = NULL;
			ctx->status = IR_ERROR_CODE_MEM_OVERFLOW;
			return NULL;
		}
		entry = ctx->code_buffer;
		IR_ASSERT((uintptr_t)entry % 16 == 0);
	} else {
		entry = ir_mem_mmap(size);
		if (!entry) {
			dasm_free(&data.dasm_state);
			ctx->data = NULL;
			ctx->status = IR_ERROR_CODE_MEM_OVERFLOW;
			return NULL;
		}
		ir_mem_unprotect(entry, size);
	}

	ret = dasm_encode(&data.dasm_state, entry);
	if (ret != DASM_S_OK) {
		IR_ASSERT(0);
		dasm_free(&data.dasm_state);
		if (ctx->code_buffer == NULL) {
			ir_mem_unmap(entry, size);
		}
		ctx->data = NULL;
		ctx->status = IR_ERROR_ENCODE;
		return NULL;
	}

	if (data.jmp_table_label) {
		uint32_t offset = dasm_getpclabel(&data.dasm_state, data.jmp_table_label);
		ctx->jmp_table_offset = offset;
	} else {
		ctx->jmp_table_offset = 0;
	}
	if (data.rodata_label) {
		uint32_t offset = dasm_getpclabel(&data.dasm_state, data.rodata_label);
		ctx->rodata_offset = offset;
	} else {
		ctx->rodata_offset = 0;
	}

	if (ctx->entries_count) {
		/* For all entries */
		i = ctx->entries_count;
		do {
			ir_insn *insn = &ctx->ir_base[ctx->entries[--i]];
			uint32_t offset = dasm_getpclabel(&data.dasm_state, ctx->cfg_blocks_count + ctx->consts_count + 4 + insn->op3);
			insn->op3 = offset;
		} while (i != 0);
	}

	dasm_free(&data.dasm_state);

	ir_mem_flush(entry, size);

	if (ctx->code_buffer == NULL) {
		ir_mem_protect(entry, size);
	}

	ctx->data = NULL;
	return entry;
}

const void *ir_emit_exitgroup(uint32_t first_exit_point, uint32_t exit_points_per_group, const void *exit_addr, void *code_buffer, size_t code_buffer_size, size_t *size_ptr)
{
	void *entry;
	size_t size;
	uint32_t i;
	dasm_State **Dst, *dasm_state;
	int ret;

	IR_ASSERT(code_buffer);
	IR_ASSERT(IR_IS_SIGNED_32BIT((char*)exit_addr - (char*)code_buffer));
	IR_ASSERT(IR_IS_SIGNED_32BIT((char*)exit_addr - ((char*)code_buffer + code_buffer_size)));

	Dst = &dasm_state;
	dasm_state = NULL;
	dasm_init(&dasm_state, DASM_MAXSECTION);
	dasm_setupglobal(&dasm_state, dasm_labels, ir_lb_MAX);
	dasm_setup(&dasm_state, dasm_actions);

	for (i = 0; i < exit_points_per_group - 1; i++) {
		|	push byte i
		|	.byte 0xeb, (4*(exit_points_per_group-i)-6) // jmp >1
	}
	|	push byte i
	|// 1:
	|	add aword [r4], first_exit_point
	|	jmp aword &exit_addr

	ret = dasm_link(&dasm_state, &size);
	if (ret != DASM_S_OK) {
		IR_ASSERT(0);
		dasm_free(&dasm_state);
		return NULL;
	}

	if (code_buffer != NULL) {
		if (IR_ALIGNED_SIZE(size, 16) > code_buffer_size) {
			dasm_free(&dasm_state);
			return NULL;
		}
		entry = code_buffer;
		IR_ASSERT((uintptr_t)entry % 16 == 0);
	} else {
		entry = ir_mem_mmap(size);
		ir_mem_unprotect(entry, size);
	}

	ret = dasm_encode(&dasm_state, entry);
	if (ret != DASM_S_OK) {
		IR_ASSERT(0);
		dasm_free(&dasm_state);
		if (code_buffer == NULL) {
			ir_mem_unmap(entry, size);
		}
		return NULL;
	}

	dasm_free(&dasm_state);

	ir_mem_flush(entry, size);

	if (code_buffer == NULL) {
		ir_mem_protect(entry, size);
	}

	*size_ptr = size;
	return entry;
}