#include <ejit/ejit.h>
#include "../../deps/lightening/lightening/lightening.c"
#include "../common.h"

#define VEC_TYPE jit_operand_t
#define VEC_NAME operands
#include "../vec.h"

struct reloc_helper {
	jit_reloc_t r;
	size_t to;
};

#define VEC_TYPE struct reloc_helper
#define VEC_NAME relocs
#include "../vec.h"

#define VEC_TYPE jit_addr_t
#define VEC_NAME addrs
#include "../vec.h"

static void *alloc_arena(size_t size)
{
	return mmap(NULL, size,
	            PROT_EXEC | PROT_READ | PROT_WRITE,
	            MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
}

static void free_arena(void *arena, size_t size)
{
	munmap(arena, size);
}

static size_t grploc_count(struct ejit_func *f)
{
	return f->gpr < jit_v_num() ? 0 : f->gpr - jit_v_num();
}

static size_t frploc_count(struct ejit_func *f)
{
	return f->fpr < jit_vf_num() ? 0 : f->fpr - jit_vf_num();
}

static size_t stack_size(struct ejit_func *f)
{
	return grploc_count(f) * sizeof(jit_uword_t)
	       + frploc_count(f) * sizeof(jit_float64_t);
}

static jit_off_t stack_loc(size_t l)
{
	assert(l >= jit_v_num());
	return (l - jit_v_num()) * sizeof(jit_uword_t);
}

static jit_off_t stack_loc_f(struct ejit_func *f, size_t l)
{
	assert(l >= jit_vf_num());
	return grploc_count(f) * sizeof(jit_uword_t)
	       + (l - jit_vf_num()) * sizeof(jit_float64_t);
}


static jit_gpr_t getloc(struct ejit_func *f, jit_state_t *j, size_t l, size_t i)
{
	assert(l < f->gpr);
	if (l < jit_v_num())
		return jit_v(l);

	jit_ldxi(j, jit_r(i), JIT_SP, stack_loc(l));
	return jit_r(i);
}

static jit_fpr_t getloc_f(struct ejit_func *f, jit_state_t *j, size_t l,
                          size_t i)
{
	assert(l < f->fpr);
	if (l < jit_vf_num())
		return jit_vf(l);

	jit_ldxi_f(j, jit_f(i), JIT_SP, stack_loc_f(f, l));
	return jit_f(i);
}

static jit_fpr_t getloc_d(struct ejit_func *f, jit_state_t *j, size_t l,
                          size_t i)
{
	assert(l < f->fpr);
	if (l < jit_vf_num())
		return jit_vf(l);

	/* not that stack_loc_f assumes double, so floats technically take up
	 * more space than needed but at least we don't get any alignment issues */
	jit_ldxi_d(j, jit_f(i), JIT_SP, stack_loc_f(f, l));
	return jit_f(i);
}

static jit_gpr_t getgpr(struct ejit_func *f, size_t l, size_t i)
{
	assert(l < f->gpr);
	if (l < jit_v_num())
		return jit_v(l);

	return jit_r(i);
}

static jit_fpr_t getfpr(struct ejit_func *f, size_t l, size_t i)
{
	assert(l < f->fpr);
	if (l < jit_vf_num())
		return jit_vf(l);

	return jit_f(i);
}

static void putloc(struct ejit_func *f, jit_state_t *j, size_t l, jit_gpr_t r)
{
	assert(l < f->gpr);
	if (l < jit_v_num()) {
		assert(jit_v(l).regno == r.regno);
		return;
	}

	jit_stxi(j, stack_loc(l), JIT_SP, r);
}

static void putloc_f(struct ejit_func *f, jit_state_t *j, size_t l, jit_fpr_t r)
{
	assert(l < f->fpr);
	if (l < jit_vf_num()) {
		assert(jit_v(l).regno == r.regno);
		return;
	}

	jit_stxi_f(j, stack_loc_f(f, l), JIT_SP, r);
}

static void putloc_d(struct ejit_func *f, jit_state_t *j, size_t l, jit_fpr_t r)
{
	assert(l < f->fpr);
	if (l < jit_vf_num()) {
		assert(jit_v(l).regno == r.regno);
		return;
	}

	jit_stxi_d(j, stack_loc_f(f, l), JIT_SP, r);
}

static void compile_label(jit_state_t *j, size_t ii, struct addrs *addrs)
{
	*addrs_at(addrs, ii) = jit_address(j);
}

static void compile_movi(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r = getgpr(f, i.r0, 0);
	jit_movi(j, r, i.o);
	putloc(f, j, i.r0, r);
}

static void compile_movi_f(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t r = getfpr(f, i.r0, 0);
	jit_movi_f(j, r, i.f);
	putloc_f(f, j, i.r0, r);
}

static void compile_movi_d(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t r = getfpr(f, i.r0, 0);
	jit_movi_d(j, r, i.d);
	putloc_d(f, j, i.r0, r);
}

static void compile_movr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t to = getgpr(f, i.r0, 0);
	jit_gpr_t from = getloc(f, j, i.r1, 1);
	jit_movr(j, to, from);
	putloc(f, j, i.r0, to);
}

static void compile_movr_f(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t to = getfpr(f, i.r0, 0);
	jit_fpr_t from = getloc_f(f, j, i.r1, 1);
	jit_movr_f(j, to, from);
	putloc_f(f, j, i.r0, to);
}

static void compile_movr_d(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t to = getfpr(f, i.r0, 0);
	jit_fpr_t from = getloc_d(f, j, i.r1, 1);
	jit_movr_d(j, to, from);
	putloc_d(f, j, i.r0, to);
}

static void compile_addr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t dst = getgpr(f, i.r0, 0);
	jit_gpr_t src0 = getloc(f, j, i.r1, 1);
	jit_gpr_t src1 = getloc(f, j, i.r2, 2);
	jit_addr(j, dst, src0, src1);
	putloc(f, j, i.r0, dst);
}

static void compile_addi(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t dst = getgpr(f, i.r0, 0);
	jit_gpr_t src0 = getloc(f, j, i.r1, 1);
	jit_addi(j, dst, src0, i.o);
	putloc(f, j, i.r0, dst);
}

static void compile_subr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t dst = getgpr(f, i.r0, 0);
	jit_gpr_t src0 = getloc(f, j, i.r1, 1);
	jit_gpr_t src1 = getloc(f, j, i.r2, 2);
	jit_subr(j, dst, src0, src1);
	putloc(f, j, i.r0, dst);
}

static void compile_subi(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t dst = getgpr(f, i.r0, 0);
	jit_gpr_t src0 = getloc(f, j, i.r1, 1);
	jit_subi(j, dst, src0, i.o);
	putloc(f, j, i.r0, dst);
}

static void compile_subr_f(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_f(f, j, i.r1, 1);
	jit_fpr_t r2 = getloc_f(f, j, i.r2, 2);
	jit_subr_f(j, r0, r1, r2);
	putloc_f(f, j, i.r0, r0);
}

static void compile_subr_d(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_d(f, j, i.r1, 1);
	jit_fpr_t r2 = getloc_d(f, j, i.r2, 2);
	jit_subr_d(j, r0, r1, r2);
	putloc_d(f, j, i.r0, r0);
}

static void compile_mulr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_mulr(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_mulr_f(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_f(f, j, i.r1, 1);
	jit_fpr_t r2 = getloc_f(f, j, i.r2, 2);
	jit_mulr_f(j, r0, r1, r2);
	putloc_f(f, j, i.r0, r0);
}

static void compile_mulr_d(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_d(f, j, i.r1, 1);
	jit_fpr_t r2 = getloc_d(f, j, i.r2, 2);
	jit_mulr_d(j, r0, r1, r2);
	putloc_d(f, j, i.r0, r0);
}

static void compile_andi(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_andi(j, r0, r1, i.o);
	putloc(f, j, i.r0, r0);
}

static void compile_andr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_andr(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_ori(struct ejit_func *f, jit_state_t *j,
                        struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_ori(j, r0, r1, i.o);
	putloc(f, j, i.r0, r0);
}

static void compile_orr(struct ejit_func *f, jit_state_t *j,
                        struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_orr(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_xori(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_xori(j, r0, r1, i.o);
	putloc(f, j, i.r0, r0);
}

static void compile_xorr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_xorr(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_divr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_divr(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_divr_u(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_divr_u(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_divr_d(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_d(f, j, i.r1, 1);
	jit_fpr_t r2 = getloc_d(f, j, i.r2, 2);
	jit_divr_d(j, r0, r1, r2);
	putloc_d(f, j, i.r0, r0);
}

static void compile_divr_f(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_f(f, j, i.r1, 1);
	jit_fpr_t r2 = getloc_f(f, j, i.r2, 2);
	jit_divr_f(j, r0, r1, r2);
	putloc_f(f, j, i.r0, r0);
}

static void compile_remr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_remr(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_remr_u(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_remr_u(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_absr_f(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_f(f, j, i.r1, 1);
	jit_absr_f(j, r0, r1);
	putloc_f(f, j, i.r0, r0);
}

static void compile_absr_d(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_d(f, j, i.r1, 1);
	jit_absr_f(j, r0, r1);
	putloc_d(f, j, i.r0, r0);
}

static void compile_addr_f(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_f(f, j, i.r1, 1);
	jit_fpr_t r2 = getloc_f(f, j, i.r2, 2);
	jit_addr_f(j, r0, r1, r2);
	putloc_f(f, j, i.r0, r0);
}

static void compile_addr_d(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_d(f, j, i.r1, 1);
	jit_fpr_t r2 = getloc_d(f, j, i.r2, 2);
	jit_addr_d(j, r0, r1, r2);
	putloc_d(f, j, i.r0, r0);
}

static void compile_lshi(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_lshi(j, r0, r1, i.o);
	putloc(f, j, i.r0, r0);
}

static void compile_lshr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_lshr(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_rshi(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_rshi(j, r0, r1, i.o);
	putloc(f, j, i.r0, r0);
}

static void compile_rshi_u(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_rshi_u(j, r0, r1, i.o);
	putloc(f, j, i.r0, r0);
}

static void compile_rshr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_rshr(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_rshr_u(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_rshr_u(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_sti8(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r0 = getloc(f, j, i.r0, 0);
	jit_sti_c(j, i.p, r0);
}

static void compile_sti16(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_gpr_t r0 = getloc(f, j, i.r0, 0);
	jit_sti_s(j, i.p, r0);
}

static void compile_sti32(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_gpr_t r0 = getloc(f, j, i.r0, 0);
	jit_sti_i(j, i.p, r0);
}

static void compile_sti64(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_gpr_t r0 = getloc(f, j, i.r0, 0);
	jit_sti_l(j, i.p, r0);
}

static void compile_stif(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_fpr_t r0 = getloc_f(f, j, i.r0, 0);
	jit_sti_f(j, i.p, r0);
}

static void compile_stid(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_fpr_t r0 = getloc_d(f, j, i.r0, 0);
	jit_sti_d(j, i.p, r0);
}

static void compile_stxi8(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_gpr_t r0 = getloc(f, j, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_stxi_c(j, i.o, r1, r0);
}

static void compile_stxi16(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getloc(f, j, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_stxi_s(j, i.o, r1, r0);
}

static void compile_stxi32(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getloc(f, j, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_stxi_i(j, i.o, r1, r0);
}

static void compile_stxi64(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getloc(f, j, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_stxi_l(j, i.o, r1, r0);
}

static void compile_stxif(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_fpr_t r0 = getloc_f(f, j, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_stxi_f(j, i.o, r1, r0);
}

static void compile_stxid(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_fpr_t r0 = getloc_d(f, j, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_stxi_d(j, i.o, r1, r0);
}

static void compile_stxr8(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_gpr_t r0 = getloc(f, j, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_stxr_c(j, r2, r1, r0);
}

static void compile_stxr16(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getloc(f, j, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_stxr_s(j, r2, r1, r0);
}

static void compile_stxr32(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getloc(f, j, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_stxr_i(j, r2, r1, r0);
}

static void compile_stxr64(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getloc(f, j, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_stxr_l(j, r2, r1, r0);
}

static void compile_stxrf(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_fpr_t r0 = getloc_f(f, j, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_stxr_f(j, r2, r1, r0);
}

static void compile_stxrd(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_fpr_t r0 = getloc_d(f, j, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_stxr_d(j, r2, r1, r0);
}

static void compile_ldiu8(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_ldi_uc(j, r0, i.p);
	putloc(f, j, i.r0, r0);
}

static void compile_ldiu16(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_ldi_us(j, r0, i.p);
	putloc(f, j, i.r0, r0);
}

static void compile_ldiu32(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_ldi_ui(j, r0, i.p);
	putloc(f, j, i.r0, r0);
}

static void compile_ldiu64(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_ldi_l(j, r0, i.p);
	putloc(f, j, i.r0, r0);
}

static void compile_ldif(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_ldi_f(j, r0, i.p);
	putloc_f(f, j, i.r0, r0);
}

static void compile_ldid(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_ldi_d(j, r0, i.p);
	putloc_d(f, j, i.r0, r0);
}

static void compile_ldi8(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_ldi_c(j, r0, i.p);
	putloc(f, j, i.r0, r0);
}

static void compile_ldi16(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_ldi_s(j, r0, i.p);
	putloc(f, j, i.r0, r0);
}

static void compile_ldi32(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_ldi_i(j, r0, i.p);
	putloc(f, j, i.r0, r0);
}

static void compile_ldi64(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_ldi_l(j, r0, i.p);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxiu8(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_ldxi_uc(j, r0, r1, i.o);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxiu16(struct ejit_func *f, jit_state_t *j,
                            struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_ldxi_us(j, r0, r1, i.o);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxiu32(struct ejit_func *f, jit_state_t *j,
                            struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_ldxi_ui(j, r0, r1, i.o);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxiu64(struct ejit_func *f, jit_state_t *j,
                            struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_ldxi_l(j, r0, r1, i.o);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxif(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_ldxi_f(j, r0, r1, i.o);
	putloc_f(f, j, i.r0, r0);
}

static void compile_ldxid(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_ldxi_d(j, r0, r1, i.o);
	putloc_d(f, j, i.r0, r0);
}

static void compile_ldxi8(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_ldxi_c(j, r0, r1, i.o);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxi16(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_ldxi_s(j, r0, r1, i.o);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxi32(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_ldxi_i(j, r0, r1, i.o);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxi64(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_ldxi_l(j, r0, r1, i.o);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxru8(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_ldxr_uc(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxru16(struct ejit_func *f, jit_state_t *j,
                            struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_ldxr_us(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxru32(struct ejit_func *f, jit_state_t *j,
                            struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_ldxr_ui(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxru64(struct ejit_func *f, jit_state_t *j,
                            struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_ldxr_l(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxr8(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_ldxr_c(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxr16(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_ldxr_s(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxr32(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_ldxr_i(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxr64(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_ldxr_l(j, r0, r1, r2);
	putloc(f, j, i.r0, r0);
}

static void compile_ldxrf(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_ldxr_f(j, r0, r1, r2);
	putloc_f(f, j, i.r0, r0);
}

static void compile_ldxrd(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_ldxr_d(j, r0, r1, r2);
	putloc_d(f, j, i.r0, r0);
}

static void compile_comr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_comr(j, r0, r1);
	putloc(f, j, i.r0, r0);
}

static void compile_negr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_negr(j, r0, r1);
	putloc(f, j, i.r0, r0);
}

static void compile_negr_f(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_f(f, j, i.r1, 1);
	jit_negr_f(j, r0, r1);
	putloc_f(f, j, i.r0, r0);
}

static void compile_negr_d(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_d(f, j, i.r1, 1);
	jit_negr_d(j, r0, r1);
	putloc_d(f, j, i.r0, r0);
}

static void compile_extr8(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_extr_c(j, r0, r1);
	putloc(f, j, i.r0, r0);
}

static void compile_extr16(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_extr_s(j, r0, r1);
	putloc(f, j, i.r0, r0);
}

static void compile_extr32(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_extr_i(j, r0, r1);
	putloc(f, j, i.r0, r0);
}

static void compile_extru8(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_extr_uc(j, r0, r1);
	putloc(f, j, i.r0, r0);
}

static void compile_extru16(struct ejit_func *f, jit_state_t *j,
                            struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_extr_us(j, r0, r1);
	putloc(f, j, i.r0, r0);
}

static void compile_extru32(struct ejit_func *f, jit_state_t *j,
                            struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_extr_ui(j, r0, r1);
	putloc(f, j, i.r0, r0);
}

static void compile_extrf(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_extr_f(j, r0, r1);
	putloc_f(f, j, i.r0, r0);
}

static void compile_extrd(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_extr_d(j, r0, r1);
	putloc_d(f, j, i.r0, r0);
}

static void compile_truncr_d_64(struct ejit_func *f, jit_state_t *j,
                                struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_d(f, j, i.r1, 1);
	jit_truncr_d_l(j, r0, r1);
	putloc(f, j, i.r0, r0);
}

static void compile_truncr_d_32(struct ejit_func *f, jit_state_t *j,
                                struct ejit_insn i)
{
#if __WORDSIZE == 64
	return compile_truncr_d_64(f, j, i);
#else
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_d(f, j, i.r1, 1);
	jit_truncr_d_i(j, r0, r1);
	putloc(f, j, i.r0, r0);
#endif
}


static void compile_truncr_f_64(struct ejit_func *f, jit_state_t *j,
                                struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_f(f, j, i.r1, 1);
	jit_truncr_f_l(j, r0, r1);
	putloc(f, j, i.r0, r0);
}

static void compile_truncr_f_32(struct ejit_func *f, jit_state_t *j,
                                struct ejit_insn i)
{
#if __WORDSIZE == 64
	return compile_truncr_f_64(f, j, i);
#else
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_f(f, j, i.r1, 1);
	jit_truncr_f_i(j, r0, r1);
	putloc(f, j, i.r0, r0);
#endif
}


static void compile_reg_cmp(struct ejit_func *f, jit_state_t *j,
                            struct ejit_insn i,
                            jit_reloc_t (*bcomp)(jit_state_t *, jit_gpr_t,
                                                 jit_gpr_t), long same)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	if (i.r1 == i.r2) {
		jit_movi(j, r0, same);
		putloc(f, j, i.r0, r0);
		return;
	}

	jit_gpr_t r1 = getloc(f, j, i.r1, 1);
	jit_gpr_t r2 = getloc(f, j, i.r2, 2);
	jit_reloc_t branch = bcomp(j, r1, r2);

	/* not equal */
	jit_movi(j, r0, 0);
	jit_reloc_t jump = jit_jmp(j);
	jit_patch_there(j, branch, jit_address(j));

	/* equal */
	jit_movi(j, r0, 1);
	jit_patch_there(j, jump, jit_address(j));

	/* write final result */
	putloc(f, j, i.r0, r0);
}

static void compile_reg_d_cmp(struct ejit_func *f, jit_state_t *j,
                              struct ejit_insn i,
                              jit_reloc_t (*bcomp)(jit_state_t *,
                                                   jit_fpr_t,
                                                   jit_fpr_t)
                              )
{
	/* note that we don't check for register sameness due to NaN, which
	 * compares to itself as fast so we can't say for sure if r1 == r1 will
	 * return true */
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_d(f, j, i.r1, 1);
	jit_fpr_t r2 = getloc_d(f, j, i.r2, 2);
	jit_reloc_t branch = bcomp(j, r1, r2);

	/* not equal */
	jit_movi(j, r0, 0);
	jit_reloc_t jump = jit_jmp(j);
	jit_patch_there(j, branch, jit_address(j));

	/* equal */
	jit_movi(j, r0, 1);
	jit_patch_there(j, jump, jit_address(j));

	/* write final result */
	putloc(f, j, i.r0, r0);
}

static void compile_reg_f_cmp(struct ejit_func *f, jit_state_t *j,
                              struct ejit_insn i,
                              jit_reloc_t (*bcomp)(jit_state_t *,
                                                   jit_fpr_t,
                                                   jit_fpr_t)
                              )
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_fpr_t r1 = getloc_f(f, j, i.r1, 1);
	jit_fpr_t r2 = getloc_f(f, j, i.r2, 2);
	jit_reloc_t branch = bcomp(j, r1, r2);

	/* not equal */
	jit_movi(j, r0, 0);
	jit_reloc_t jump = jit_jmp(j);
	jit_patch_there(j, branch, jit_address(j));

	/* equal */
	jit_movi(j, r0, 1);
	jit_patch_there(j, jump, jit_address(j));

	/* write final result */
	putloc(f, j, i.r0, r0);
}

static void compile_eqr(struct ejit_func *f, jit_state_t *j, struct ejit_insn i)
{
	compile_reg_cmp(f, j, i, jit_beqr, 1);
}

static void compile_eqr_d(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	compile_reg_d_cmp(f, j, i, jit_beqr_d);
}

static void compile_eqr_f(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	compile_reg_f_cmp(f, j, i, jit_beqr_f);
}

static void compile_ner(struct ejit_func *f, jit_state_t *j, struct ejit_insn i)
{
	compile_reg_cmp(f, j, i, jit_bner, 0);
}

static void compile_ner_d(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	compile_reg_d_cmp(f, j, i, jit_bner_d);
}

static void compile_ner_f(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	compile_reg_f_cmp(f, j, i, jit_bner_f);
}

static void compile_ger(struct ejit_func *f, jit_state_t *j, struct ejit_insn i)
{
	compile_reg_cmp(f, j, i, jit_bger, 1);
}

static void compile_ger_u(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	compile_reg_cmp(f, j, i, jit_bger_u, 1);
}

static void compile_ger_d(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	compile_reg_d_cmp(f, j, i, jit_bger_d);
}

static void compile_ger_f(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	compile_reg_f_cmp(f, j, i, jit_bger_f);
}

static void compile_gtr(struct ejit_func *f, jit_state_t *j, struct ejit_insn i)
{
	compile_reg_cmp(f, j, i, jit_bgtr, 0);
}

static void compile_gtr_u(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	compile_reg_cmp(f, j, i, jit_bgtr_u, 0);
}

static void compile_gtr_d(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	compile_reg_d_cmp(f, j, i, jit_bgtr_d);
}

static void compile_gtr_f(struct ejit_func *f, jit_state_t *j,
                          struct ejit_insn i)
{
	compile_reg_f_cmp(f, j, i, jit_bgtr_f);
}

static void compile_bmci(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_reloc_t r = jit_bmci(j, r1, i.o);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bmcr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_gpr_t r2 = getloc(f, j, i.r2, 1);
	jit_reloc_t r = jit_bmcr(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bmsi(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_reloc_t r = jit_bmsi(j, r1, i.o);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bmsr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_gpr_t r2 = getloc(f, j, i.r2, 1);
	jit_reloc_t r = jit_bmsr(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_beqi(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_reloc_t r = jit_beqi(j, r1, i.o);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_beqr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_gpr_t r2 = getloc(f, j, i.r2, 1);
	jit_reloc_t r = jit_beqr(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_beqr_f(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i, struct relocs *relocs)
{
	jit_fpr_t r1 = getloc_f(f, j, i.r1, 0);
	jit_fpr_t r2 = getloc_f(f, j, i.r2, 1);
	jit_reloc_t r = jit_beqr_f(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_beqr_d(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i, struct relocs *relocs)
{
	jit_fpr_t r1 = getloc_d(f, j, i.r1, 0);
	jit_fpr_t r2 = getloc_d(f, j, i.r2, 1);
	jit_reloc_t r = jit_beqr_d(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bnei(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_reloc_t r = jit_bnei(j, r1, i.o);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bner(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_gpr_t r2 = getloc(f, j, i.r2, 1);
	jit_reloc_t r = jit_bner(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bner_f(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i, struct relocs *relocs)
{
	jit_fpr_t r1 = getloc_f(f, j, i.r1, 0);
	jit_fpr_t r2 = getloc_f(f, j, i.r2, 1);
	jit_reloc_t r = jit_bner_f(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bner_d(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i, struct relocs *relocs)
{
	jit_fpr_t r1 = getloc_d(f, j, i.r1, 0);
	jit_fpr_t r2 = getloc_d(f, j, i.r2, 1);
	jit_reloc_t r = jit_bner_d(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bger(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_gpr_t r2 = getloc(f, j, i.r2, 1);
	jit_reloc_t r = jit_bger(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bger_u(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_gpr_t r2 = getloc(f, j, i.r2, 1);
	jit_reloc_t r = jit_bger_u(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bgei(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_reloc_t r = jit_bgei(j, r1, i.o);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bgei_u(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_reloc_t r = jit_bgei_u(j, r1, i.o);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bger_f(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i, struct relocs *relocs)
{
	jit_fpr_t r1 = getloc_f(f, j, i.r1, 0);
	jit_fpr_t r2 = getloc_f(f, j, i.r2, 1);
	jit_reloc_t r = jit_bger_f(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bger_d(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i, struct relocs *relocs)
{
	jit_fpr_t r1 = getloc_d(f, j, i.r1, 0);
	jit_fpr_t r2 = getloc_d(f, j, i.r2, 1);
	jit_reloc_t r = jit_bger_d(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bgtr(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_gpr_t r2 = getloc(f, j, i.r2, 1);
	jit_reloc_t r = jit_bgtr(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bgtr_u(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_gpr_t r2 = getloc(f, j, i.r2, 1);
	jit_reloc_t r = jit_bgtr_u(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bgti(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_reloc_t r = jit_bgti(j, r1, i.o);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bgti_u(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_reloc_t r = jit_bgti_u(j, r1, i.o);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bgtr_f(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i, struct relocs *relocs)
{
	jit_fpr_t r1 = getloc_f(f, j, i.r1, 0);
	jit_fpr_t r2 = getloc_f(f, j, i.r2, 1);
	jit_reloc_t r = jit_bgtr_f(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_bgtr_d(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i, struct relocs *relocs)
{
	jit_fpr_t r1 = getloc_d(f, j, i.r1, 0);
	jit_fpr_t r2 = getloc_d(f, j, i.r2, 1);
	jit_reloc_t r = jit_bgtr_d(j, r1, r2);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_blei(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_reloc_t r = jit_blei(j, r1, i.o);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_blei_u(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_reloc_t r = jit_blei_u(j, r1, i.o);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_blti(struct ejit_func *f, jit_state_t *j,
                         struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_reloc_t r = jit_blti(j, r1, i.o);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_blti_u(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i, struct relocs *relocs)
{
	jit_gpr_t r1 = getloc(f, j, i.r1, 0);
	jit_reloc_t r = jit_blti_u(j, r1, i.o);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_jmp(struct ejit_func *f, jit_state_t *j, struct ejit_insn i,
                        struct relocs *relocs)
{
	(void)(f);
	jit_reloc_t r = jit_jmp(j);
	struct reloc_helper h = {.r = r, .to = i.r0};
	relocs_append(relocs, h);
}

static void compile_retval(struct ejit_func *f, jit_state_t *j,
                           struct ejit_insn i)
{
	jit_gpr_t r0 = getgpr(f, i.r0, 0);
	jit_retval(j, r0);
	putloc(f, j, i.r0, r0);
}

static void compile_retval_f(struct ejit_func *f, jit_state_t *j,
                             struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_retval_d(j, r0);
	/* convert double to float */
	jit_extr_d_f(j, r0, r0);
	putloc_f(f, j, i.r0, r0);
}

static void compile_retval_d(struct ejit_func *f, jit_state_t *j,
                             struct ejit_insn i)
{
	jit_fpr_t r0 = getfpr(f, i.r0, 0);
	jit_retval_d(j, r0);
	putloc_d(f, j, i.r0, r0);
}

static enum jit_operand_abi jit_abi_from(enum ejit_type t)
{
	switch (t) {
	case EJIT_INT8: return JIT_OPERAND_ABI_INT8;
	case EJIT_INT16: return JIT_OPERAND_ABI_INT16;
	case EJIT_INT32: return JIT_OPERAND_ABI_INT32;
	case EJIT_INT64: return JIT_OPERAND_ABI_INT64;
	case EJIT_UINT8: return JIT_OPERAND_ABI_UINT8;
	case EJIT_UINT16: return JIT_OPERAND_ABI_UINT16;
	case EJIT_UINT32: return JIT_OPERAND_ABI_UINT32;
	case EJIT_UINT64: return JIT_OPERAND_ABI_UINT64;
	case EJIT_POINTER: return JIT_OPERAND_ABI_POINTER;
	case EJIT_FLOAT: return JIT_OPERAND_ABI_FLOAT;
	case EJIT_DOUBLE: return JIT_OPERAND_ABI_DOUBLE;
	default:
	}

	abort();
}

static size_t arg_offsetof(enum ejit_type t)
{
	switch (t) {
	case EJIT_INT8: return offsetof(struct ejit_arg, i8);
	case EJIT_INT16: return offsetof(struct ejit_arg, i16);
	case EJIT_INT32: return offsetof(struct ejit_arg, i32);
	case EJIT_INT64: return offsetof(struct ejit_arg, i64);
	case EJIT_UINT8: return offsetof(struct ejit_arg, u8);
	case EJIT_UINT16: return offsetof(struct ejit_arg, u16);
	case EJIT_UINT32: return offsetof(struct ejit_arg, u32);
	case EJIT_UINT64: return offsetof(struct ejit_arg, u64);
	case EJIT_POINTER: return offsetof(struct ejit_arg, p);
	case EJIT_FLOAT: return offsetof(struct ejit_arg, f);
	case EJIT_DOUBLE: return offsetof(struct ejit_arg, d);
	default:
	};

	abort();
}

static jit_off_t arg_offset(struct ejit_insn i)
{
	/* index of ejit_arg in stack and offset of whatever type we're dealing
	 * with */
	return (sizeof(struct ejit_arg) * i.r0) + arg_offsetof(i.r1);
}

static jit_off_t type_offset(struct ejit_insn i)
{
	return (sizeof(struct ejit_arg) * i.r0) + offsetof(struct ejit_arg,
	                                                   type);
}

static void fixup_operands(struct operands *operands, size_t fixup)
{
	foreach_vec(i, *operands) {
		jit_operand_t op = *operands_at(operands, i);
		if (op.kind != JIT_OPERAND_KIND_MEM)
			continue;

		op.loc.mem.offset += fixup;
		*operands_at(operands, i) = op;
	}
}

static void compile_imm_call(jit_state_t *j, struct operands *src, struct operands *dst,
                             void *addr, size_t argc, jit_operand_t args[argc])
{
	/* each move is type + arg, so twofold */
	size_t movec = operands_len(src) / 2;
	size_t fixup = jit_align_stack(j, movec * sizeof(struct ejit_arg));
	fixup_operands(src, fixup);
	/* note, do not fix up destination! */
	/* remember to move all operands */
	jit_move_operands(j, dst->buf, src->buf, movec * 2);
	jit_calli(j, addr, argc, args);

	jit_shrink_stack(j, fixup);
}

/* adds a header that converts from our external interface (params on stack) to
 * an internal format that's effectively just the underlying ABI, makes calls
 * between functions we know are compiled a bit faster. */
static void compile_trampoline(struct ejit_func *f, jit_state_t *j)
{
	/* very important, argc we don't really do anything with but JIT_R1
	 * contains the argument stack! */

	size_t frame = jit_enter_jit_abi(j, 0, 0, 0);
	jit_load_args_2(j,
	                jit_operand_gpr(JIT_OPERAND_ABI_WORD, JIT_R0),
	                jit_operand_gpr(JIT_OPERAND_ABI_POINTER, JIT_R1));

	struct operands args = operands_create(0);

	foreach_vec(ii, f->insns) {
		struct ejit_insn i = *insns_at(&f->insns, ii);
		switch (i.op) {
		case PARAM: {
			jit_operand_t p = jit_operand_mem(jit_abi_from(i.r1), JIT_R1, arg_offset(i));
			operands_append(&args, p);
			break;
		}

		case PARAM_F: {
			jit_operand_t p = jit_operand_mem(jit_abi_from(i.r1), JIT_R1, arg_offset(i));
			operands_append(&args, p);
			break;
		}

		case START: {
			/* callr should be smart enough to avoid JIT_R0 if some
			 * other register wants to write to it */
			jit_reloc_t r = jit_mov_addr(j, JIT_R0);
			jit_callr(j, JIT_R0, operands_len(&args), args.buf);
			jit_leave_jit_abi(j, 0, 0, frame);
			jit_ret(j); /* should just forward the return value */

			f->direct_call = jit_address(j);
			jit_patch_here(j, r);

			operands_destroy(&args);
			return;
		}

		default: abort();
		}
	}
}

static size_t compile_fn_body(struct ejit_func *f, jit_state_t *j, void *arena,
                              size_t size)
{
	jit_begin(j, arena, size);
	compile_trampoline(f, j);

	size_t gprs = f->gpr >= jit_v_num() ? jit_v_num() : f->gpr;
	size_t fprs = f->fpr >= jit_vf_num() ? jit_vf_num() : f->fpr;
	size_t frame = jit_enter_jit_abi(j, gprs, fprs, 0);
	size_t stack = jit_align_stack(j, stack_size(f));

	struct operands src = operands_create();
	struct operands dst = operands_create();
	struct operands direct = operands_create();

	struct relocs relocs = relocs_create(sizeof(struct reloc_helper));
	struct addrs addrs = addrs_create();
	addrs_reserve(&addrs, insns_len(&f->insns));

	size_t label = 0;
	foreach_vec(ii, f->insns) {
		/* if we've hit a label, add it to our vector of label addresses */
		if (label < labels_len(&f->labels)) {
			if (*labels_at(&f->labels, label) == ii) {
				compile_label(j, ii, &addrs);
				label++;
			}
		}

		struct ejit_insn i = *insns_at(&f->insns, ii);
		switch (i.op) {
		case MOVR: compile_movr(f, j, i); break;
		case MOVR_F: compile_movr_f(f, j, i); break;
		case MOVR_D: compile_movr_d(f, j, i); break;

		case MOVI: compile_movi(f, j, i); break;
		case MOVI_F: compile_movi_f(f, j, i); break;
		case MOVI_D: compile_movi_d(f, j, i); break;

		case ADDR: compile_addr(f, j, i); break;
		case ADDI: compile_addi(f, j, i); break;
		case ADDR_F: compile_addr_f(f, j, i); break;
		case ADDR_D: compile_addr_d(f, j, i); break;

		case SUBR: compile_subr(f, j, i); break;
		case SUBI: compile_subi(f, j, i); break;
		case SUBR_F: compile_subr_f(f, j, i); break;
		case SUBR_D: compile_subr_d(f, j, i); break;

		case MULR: compile_mulr(f, j, i); break;
		case MULR_F: compile_mulr_f(f, j, i); break;
		case MULR_D: compile_mulr_d(f, j, i); break;

		case ANDI: compile_andi(f, j, i); break;
		case ANDR: compile_andr(f, j, i); break;

		case ORI: compile_ori(f, j, i); break;
		case ORR: compile_orr(f, j, i); break;

		case XORI: compile_xori(f, j, i); break;
		case XORR: compile_xorr(f, j, i); break;

		case DIVR: compile_divr(f, j, i); break;
		case DIVR_U: compile_divr_u(f, j, i); break;
		case DIVR_F: compile_divr_f(f, j, i); break;
		case DIVR_D: compile_divr_d(f, j, i); break;

		case REMR: compile_remr(f, j, i); break;
		case REMR_U: compile_remr_u(f, j, i); break;

		case ABSR_F: compile_absr_f(f, j, i); break;
		case ABSR_D: compile_absr_d(f, j, i); break;

		case LSHI: compile_lshi(f, j, i); break;
		case LSHR: compile_lshr(f, j, i); break;
		case RSHI: compile_rshi(f, j, i); break;
		case RSHI_U: compile_rshi_u(f, j, i); break;
		case RSHR: compile_rshr(f, j, i); break;
		case RSHR_U: compile_rshr_u(f, j, i); break;

		case STI8: compile_sti8(f, j, i); break;
		case STI16: compile_sti16(f, j, i); break;
		case STI32: compile_sti32(f, j, i); break;
		case STI64: compile_sti64(f, j, i); break;
		case STIF: compile_stif(f, j, i); break;
		case STID: compile_stid(f, j, i); break;

		case STXI8: compile_stxi8(f, j, i); break;
		case STXI16: compile_stxi16(f, j, i); break;
		case STXI32: compile_stxi32(f, j, i); break;
		case STXI64: compile_stxi64(f, j, i); break;
		case STXIF: compile_stxif(f, j, i); break;
		case STXID: compile_stxid(f, j, i); break;

		case STXR8: compile_stxr8(f, j, i); break;
		case STXR16: compile_stxr16(f, j, i); break;
		case STXR32: compile_stxr32(f, j, i); break;
		case STXR64: compile_stxr64(f, j, i); break;
		case STXRF: compile_stxrf(f, j, i); break;
		case STXRD: compile_stxrd(f, j, i); break;

		case LDI8: compile_ldi8(f, j, i); break;
		case LDI16: compile_ldi16(f, j, i); break;
		case LDI32: compile_ldi32(f, j, i); break;
		case LDI64: compile_ldi64(f, j, i); break;
		case LDIU8: compile_ldiu8(f, j, i); break;
		case LDIU16: compile_ldiu16(f, j, i); break;
		case LDIU32: compile_ldiu32(f, j, i); break;
		case LDIU64: compile_ldiu64(f, j, i); break;
		case LDIF: compile_ldif(f, j, i); break;
		case LDID: compile_ldid(f, j, i); break;

		case LDXI8: compile_ldxi8(f, j, i); break;
		case LDXI16: compile_ldxi16(f, j, i); break;
		case LDXI32: compile_ldxi32(f, j, i); break;
		case LDXI64: compile_ldxi64(f, j, i); break;
		case LDXIU8: compile_ldxiu8(f, j, i); break;
		case LDXIU16: compile_ldxiu16(f, j, i); break;
		case LDXIU32: compile_ldxiu32(f, j, i); break;
		case LDXIU64: compile_ldxiu64(f, j, i); break;
		case LDXIF: compile_ldxif(f, j, i); break;
		case LDXID: compile_ldxid(f, j, i); break;

		case LDXR8: compile_ldxr8(f, j, i); break;
		case LDXR16: compile_ldxr16(f, j, i); break;
		case LDXR32: compile_ldxr32(f, j, i); break;
		case LDXR64: compile_ldxr64(f, j, i); break;
		case LDXRU8: compile_ldxru8(f, j, i); break;
		case LDXRU16: compile_ldxru16(f, j, i); break;
		case LDXRU32: compile_ldxru32(f, j, i); break;
		case LDXRU64: compile_ldxru64(f, j, i); break;
		case LDXRF: compile_ldxrf(f, j, i); break;
		case LDXRD: compile_ldxrd(f, j, i); break;

		case COMR: compile_comr(f, j, i); break;

		case NEGR: compile_negr(f, j, i); break;
		case NEGR_F: compile_negr_f(f, j, i); break;
		case NEGR_D: compile_negr_d(f, j, i); break;

		case EXTR8: compile_extr8(f, j, i); break;
		case EXTR16: compile_extr16(f, j, i); break;
		case EXTR32: compile_extr32(f, j, i); break;
		case EXTRU8: compile_extru8(f, j, i); break;
		case EXTRU16: compile_extru16(f, j, i); break;
		case EXTRU32: compile_extru32(f, j, i); break;
		case EXTRF: compile_extrf(f, j, i); break;
		case EXTRD: compile_extrd(f, j, i); break;

		case TRUNCR_D_32: compile_truncr_d_32(f, j, i); break;
		case TRUNCR_D_64: compile_truncr_d_64(f, j, i); break;
		case TRUNCR_F_32: compile_truncr_f_32(f, j, i); break;
		case TRUNCR_F_64: compile_truncr_f_64(f, j, i); break;

		case EQR: compile_eqr(f, j, i); break;
		case EQR_F: compile_eqr_f(f, j, i); break;
		case EQR_D: compile_eqr_d(f, j, i); break;

		case NER: compile_ner(f, j, i); break;
		case NER_F: compile_ner_f(f, j, i); break;
		case NER_D: compile_ner_d(f, j, i); break;

		case GER: compile_ger(f, j, i); break;
		case GER_U: compile_ger_u(f, j, i); break;
		case GER_F: compile_ger_f(f, j, i); break;
		case GER_D: compile_ger_d(f, j, i); break;

		case GTR: compile_gtr(f, j, i); break;
		case GTR_U: compile_gtr_u(f, j, i); break;
		case GTR_F: compile_gtr_f(f, j, i); break;
		case GTR_D: compile_gtr_d(f, j, i); break;

		case BMCI: compile_bmci(f, j, i, &relocs); break;
		case BMCR: compile_bmcr(f, j, i, &relocs); break;

		case BMSI: compile_bmsi(f, j, i, &relocs); break;
		case BMSR: compile_bmsr(f, j, i, &relocs); break;

		case BEQR: compile_beqr(f, j, i, &relocs); break;
		case BEQI: compile_beqi(f, j, i, &relocs); break;
		case BEQR_F: compile_beqr_f(f, j, i, &relocs); break;
		case BEQR_D: compile_beqr_d(f, j, i, &relocs); break;

		case BNER: compile_bner(f, j, i, &relocs); break;
		case BNEI: compile_bnei(f, j, i, &relocs); break;
		case BNER_F: compile_bner_f(f, j, i, &relocs); break;
		case BNER_D: compile_bner_d(f, j, i, &relocs); break;

		case BGER: compile_bger(f, j, i, &relocs); break;
		case BGER_U: compile_bger_u(f, j, i, &relocs); break;
		case BGEI: compile_bgei(f, j, i, &relocs); break;
		case BGEI_U: compile_bgei_u(f, j, i, &relocs); break;
		case BGER_F: compile_bger_f(f, j, i, &relocs); break;
		case BGER_D: compile_bger_d(f, j, i, &relocs); break;

		case BGTR: compile_bgtr(f, j, i, &relocs); break;
		case BGTR_U: compile_bgtr_u(f, j, i, &relocs); break;
		case BGTI: compile_bgti(f, j, i, &relocs); break;
		case BGTI_U: compile_bgti_u(f, j, i, &relocs); break;
		case BGTR_F: compile_bgtr_f(f, j, i, &relocs); break;
		case BGTR_D: compile_bgtr_d(f, j, i, &relocs); break;

		case BLEI: compile_blei(f, j, i, &relocs); break;
		case BLEI_U: compile_blei_u(f, j, i, &relocs); break;

		case BLTI: compile_blti(f, j, i, &relocs); break;
		case BLTI_U: compile_blti_u(f, j, i, &relocs); break;

		case JMP: compile_jmp(f, j, i, &relocs); break;

		case ARG: {
			jit_operand_t type =
				jit_operand_imm(JIT_OPERAND_ABI_WORD, i.r1);
			jit_operand_t arg;
			if (i.r2 < jit_v_num()) {
				/* regular register */
				arg = jit_operand_gpr(jit_abi_from(i.r1),
				                      jit_v(i.r2));
			}
			else {
				/* stack location, note that we'll fix up the SP
				 * offset before doing the actual call */
				arg = jit_operand_mem(jit_abi_from(i.r1),
				                      JIT_SP, stack_loc(i.r2));
			}

			operands_append(&src, type);
			operands_append(&src, arg);
			operands_append(&direct, arg);

			jit_operand_t to[2] = {
				jit_operand_mem(JIT_OPERAND_ABI_WORD, JIT_SP,
				                type_offset(i)),
				jit_operand_mem(jit_abi_from(i.r1), JIT_SP,
				                arg_offset(i))
			};

			operands_append(&dst, to[0]);
			operands_append(&dst, to[1]);
			break;
		}

		case ARG_F: {
			jit_operand_t type =
				jit_operand_imm(JIT_OPERAND_ABI_WORD, i.r1);
			jit_operand_t arg;
			if (i.r2 < jit_vf_num()) {
				/* regular register */
				arg = jit_operand_fpr(jit_abi_from(i.r1),
				                      jit_vf(i.r2));
			}
			else {
				/* stack location, note that we'll fix up the SP
				 * offset before doing the actual call */
				arg = jit_operand_mem(jit_abi_from(i.r1),
				                      JIT_SP,
				                      stack_loc_f(f, i.r2));
			}

			operands_append(&src, type);
			operands_append(&src, arg);
			operands_append(&direct, arg);

			jit_operand_t to[2] = {
				jit_operand_mem(JIT_OPERAND_ABI_WORD, JIT_SP,
				                type_offset(i)),
				jit_operand_mem(jit_abi_from(i.r1), JIT_SP,
				                arg_offset(i))
			};

			operands_append(&dst, to[0]);
			operands_append(&dst, to[1]);
			break;
		}

		case ESCAPEI: {
			jit_operand_t args[2] = {
				jit_operand_imm(JIT_OPERAND_ABI_WORD,
				                operands_len(&src) / 2),
				jit_operand_gpr(JIT_OPERAND_ABI_POINTER, JIT_SP)
			};
			compile_imm_call(j, &src, &dst, (void *)i.o, 2, args);
			operands_reset(&src);
			operands_reset(&dst);
			operands_reset(&direct);
			break;
		}

		case ESCAPEI_F: {
			jit_operand_t args[2] = {
				jit_operand_imm(JIT_OPERAND_ABI_WORD,
				                operands_len(&src) / 2),
				jit_operand_gpr(JIT_OPERAND_ABI_POINTER, JIT_SP)
			};
			compile_imm_call(j, &src, &dst, (void *)i.o, 2, args);
			operands_reset(&src);
			operands_reset(&dst);
			operands_reset(&direct);
			break;
		}

		case CALLI: {
			struct ejit_func *f = (struct ejit_func *)i.o;
			if (f && f->direct_call) {
				jit_calli(j, f->direct_call, operands_len(&direct), direct.buf);
				operands_reset(&src);
				operands_reset(&dst);
				operands_reset(&direct);
				break;
			}

			jit_operand_t args[3] = {
				jit_operand_imm(JIT_OPERAND_ABI_POINTER, i.o),
				jit_operand_imm(JIT_OPERAND_ABI_WORD,
				                operands_len(&src) / 2),
				jit_operand_gpr(JIT_OPERAND_ABI_POINTER, JIT_SP)
			};
			compile_imm_call(j, &src, &dst, ejit_run_func, 3, args);
			operands_reset(&src);
			operands_reset(&dst);
			operands_reset(&direct);
			break;
		}

		case RETVAL: compile_retval(f, j, i); break;
		case RETVAL_F: compile_retval_f(f, j, i); break;
		case RETVAL_D: compile_retval_d(f, j, i); break;

		case RETR: {
			jit_gpr_t r = getloc(f, j, i.r1, 0);
			/* R0 won't get overwritten by jit_leave_jit_abi */
			jit_movr(j, JIT_R0, r);
			jit_shrink_stack(j, stack);
			jit_leave_jit_abi(j, gprs, fprs, frame);
			jit_retr(j, JIT_R0);
			break;
		}

		case RETR_F: {
			jit_fpr_t r = getloc_f(f, j, i.r1, 0);
			/* convert float to double so the return types match */
			jit_extr_f_d(j, JIT_F0, r);
			jit_shrink_stack(j, stack);
			jit_leave_jit_abi(j, gprs, fprs, frame);
			jit_retr_f(j, JIT_F0);
			break;
		}

		case RETR_D: {
			jit_fpr_t r = getloc_d(f, j, i.r1, 0);
			jit_movr_d(j, JIT_F0, r);
			jit_shrink_stack(j, stack);
			jit_leave_jit_abi(j, gprs, fprs, frame);
			jit_retr_d(j, JIT_F0);
			break;
		}

		case RETI: {
			jit_shrink_stack(j, stack);
			jit_leave_jit_abi(j, gprs, fprs, frame);
			jit_reti(j, i.o);
			break;
		}

		case END: {
			/* 'void' return */
			jit_shrink_stack(j, stack);
			jit_leave_jit_abi(j, gprs, fprs, frame);
			jit_reti(j, 0);
			break;
		}

		case PARAM_F: {
			jit_operand_t to;
			if (i.r2 < jit_vf_num()) {
				/* regular register */
				to = jit_operand_fpr(jit_abi_from(i.r1),
				                     jit_vf(i.r2));
			}
			else {
				/* stack location */
				to = jit_operand_mem(jit_abi_from(i.r1), JIT_SP,
				                     stack_loc_f(f, i.r2));
			}

			operands_append(&dst, to);
			break;
		}

		case PARAM: {
			jit_operand_t to;
			if (i.r2 < jit_v_num()) {
				/* regular register */
				to = jit_operand_gpr(jit_abi_from(i.r1),
				                     jit_v(i.r2));
			}
			else {
				/* stack location */
				to = jit_operand_mem(jit_abi_from(i.r1), JIT_SP,
				                     stack_loc(i.r2));
			}

			operands_append(&dst, to);
			break;
		}

		case START: {
			/* parameters should be done by now */
			jit_load_args(j, operands_len(&dst), dst.buf);
			/* reuse for arguments */
			operands_reset(&dst);
			break;
		}

		default: abort();
		}
	}

	foreach_vec(ri, relocs) {
		struct reloc_helper h = *relocs_at(&relocs, ri);
		jit_addr_t a = *addrs_at(&addrs, h.to);
		jit_reloc_t r = h.r;

		assert(a);
		jit_patch_there(j, r, a);
	}

	operands_destroy(&src);
	operands_destroy(&dst);
	operands_destroy(&direct);
	relocs_destroy(&relocs);
	addrs_destroy(&addrs);

	if (jit_end(j, &size))
		return 0;

	return size;
}

bool ejit_compile(struct ejit_func *f, bool use_64)
{
	(void)use_64;
#if __WORDSIZE == 32
	/* can't compile 64bit code on 32bit systems, give up early */
	if (use_64)
		return false;
#endif
	if (!init_jit())
		return false;

	/* the main overhead of compilation seems to be the syscall to mmap a
	 * new arena, I might look into allocating a big buffer at once and
	 * caching it to be reused later, might allow us to compile many small
	 * functions faster */
	jit_state_t *j = jit_new_state(NULL, NULL);
	assert(j);

	void *arena = NULL;
	size_t size = 4096;

	while (1) {
		arena = alloc_arena(size);
		assert(arena);

		size_t required_size = compile_fn_body(f, j, arena, size);
		if (required_size == 0)
			break;

		free_arena(arena, size);
		size = required_size + 4096;
	}

	jit_destroy_state(j);
	f->arena = arena;
	f->size = size;
	return true;
}