static void x87_arith_op( struct x86_function *p, struct x86_reg dst, struct x86_reg arg,
unsigned char dst0ub0,
unsigned char dst0ub1,
unsigned char arg0ub0,
unsigned char arg0ub1,
unsigned char argmem_noreg)
{
assert(dst.file == file_x87);
if (arg.file == file_x87) {
if (dst.idx == 0)
emit_2ub(p, dst0ub0, dst0ub1+arg.idx);
else if (arg.idx == 0)
emit_2ub(p, arg0ub0, arg0ub1+arg.idx);
else
assert(0);
}
else if (dst.idx == 0) {
assert(arg.file == file_REG32);
emit_1ub(p, 0xd8);
emit_modrm_noreg(p, argmem_noreg, arg);
}
else
assert(0);
}
/* Always use a 32bit offset for forward jumps:
*/
unsigned char *x86_jcc_forward( struct x86_function *p,
enum x86_cc cc )
{
emit_2ub(p, 0x0f, 0x80 + cc);
emit_1i(p, 0);
return x86_get_label(p);
}
void sse_movss( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
emit_2ub(p, 0xF3, X86_TWOB);
emit_op_modrm( p, 0x10, 0x11, dst, src );
}
void sse_rsqrtps( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
emit_2ub(p, X86_TWOB, 0x52);
emit_modrm( p, dst, src );
}
void sse2_rcpps( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
emit_2ub(p, X86_TWOB, 0x53);
emit_modrm( p, dst, src );
}
void sse2_movd( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
emit_2ub(p, 0x66, X86_TWOB);
emit_op_modrm( p, 0x6e, 0x7e, dst, src );
}
void sse_prefetch1( struct x86_function *p, struct x86_reg ptr)
{
DUMP_R( ptr );
assert(ptr.mod != mod_REG);
emit_2ub(p, 0x0f, 0x18);
emit_modrm_noreg(p, 2, ptr);
}
void mmx_emms( struct x86_function *p )
{
DUMP();
assert(p->need_emms);
emit_2ub(p, 0x0f, 0x77);
p->need_emms = 0;
}
void x86_jcc( struct x86_function *p,
enum x86_cc cc,
int label )
{
int offset = label - (x86_get_label(p) + 2);
DUMP_I(cc);
if (offset < 0) {
/*assert(p->csr - p->store > -offset);*/
if (p->csr - p->store <= -offset) {
/* probably out of memory (using the error_overflow buffer) */
return;
}
}
if (offset <= 127 && offset >= -128) {
emit_1ub(p, 0x70 + cc);
emit_1b(p, (char) offset);
}
else {
offset = label - (x86_get_label(p) + 6);
emit_2ub(p, 0x0f, 0x80 + cc);
emit_1i(p, offset);
}
}
void x87_fucompp( struct x86_function *p )
{
DUMP();
emit_2ub(p, 0xda, 0xe9);
note_x87_pop(p); /* pop twice */
note_x87_pop(p); /* pop twice */
}
void x87_fucomp( struct x86_function *p, struct x86_reg arg )
{
DUMP_R( arg );
assert(arg.file == file_x87);
emit_2ub(p, 0xdd, 0xe8+arg.idx);
note_x87_pop(p);
}
void sse2_cvtdq2ps( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
DUMP_RR( dst, src );
emit_2ub(p, X86_TWOB, 0x5b);
emit_modrm( p, dst, src );
}
void sse_xorps( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
DUMP_RR( dst, src );
emit_2ub(p, X86_TWOB, 0x57);
emit_modrm( p, dst, src );
}
void x86_imul( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
DUMP_RR( dst, src );
emit_2ub(p, X86_TWOB, 0xAF);
emit_modrm(p, dst, src);
}
/* Always use a 32bit offset for forward jumps:
*/
int x86_jcc_forward( struct x86_function *p,
enum x86_cc cc )
{
DUMP_I(cc);
emit_2ub(p, 0x0f, 0x80 + cc);
emit_1i(p, 0);
return x86_get_label(p);
}
void x87_fdivrp( struct x86_function *p, struct x86_reg dst )
{
DUMP_R( dst );
assert(dst.file == file_x87);
assert(dst.idx >= 1);
emit_2ub(p, 0xde, 0xf0+dst.idx);
note_x87_pop(p);
}
void sse_movlhps( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
assert(dst.mod == mod_REG && src.mod == mod_REG);
emit_2ub(p, X86_TWOB, 0x16);
emit_modrm( p, dst, src );
}
void x87_fst( struct x86_function *p, struct x86_reg dst )
{
if (dst.file == file_x87)
emit_2ub(p, 0xdd, 0xd0 + dst.idx);
else {
emit_1ub(p, 0xd9);
emit_modrm_noreg(p, 2, dst);
}
}
void sse_movhlps( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
DUMP_RR( dst, src );
assert(dst.mod == mod_REG && src.mod == mod_REG);
emit_2ub(p, X86_TWOB, 0x12);
emit_modrm( p, dst, src );
}
void sse_cmpps( struct x86_function *p,
struct x86_reg dest,
struct x86_reg arg0,
unsigned char cc)
{
emit_2ub(p, X86_TWOB, 0xC2);
emit_modrm(p, dest, arg0);
emit_1ub(p, cc);
}
/* Shufps can also be used to implement a reduced swizzle when dest ==
* arg0.
*/
void sse_shufps( struct x86_function *p,
struct x86_reg dest,
struct x86_reg arg0,
unsigned char shuf)
{
emit_2ub(p, X86_TWOB, 0xC6);
emit_modrm(p, dest, arg0);
emit_1ub(p, shuf);
}
void x87_fld( struct x86_function *p, struct x86_reg arg )
{
if (arg.file == file_x87)
emit_2ub(p, 0xd9, 0xc0 + arg.idx);
else {
emit_1ub(p, 0xd9);
emit_modrm_noreg(p, 0, arg);
}
}
void x87_fcomp( struct x86_function *p, struct x86_reg dst )
{
if (dst.file == file_x87)
emit_2ub(p, 0xd8, 0xd8 + dst.idx);
else {
emit_1ub(p, 0xd8);
emit_modrm_noreg(p, 3, dst);
}
}
void x87_fcom( struct x86_function *p, struct x86_reg dst )
{
DUMP_R( dst );
if (dst.file == file_x87)
emit_2ub(p, 0xd8, 0xd0 + dst.idx);
else {
emit_1ub(p, 0xd8);
emit_modrm_noreg(p, 2, dst);
}
}
/* Shufps can also be used to implement a reduced swizzle when dest ==
* arg0.
*/
void sse_shufps( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src,
unsigned char shuf)
{
DUMP_RRI( dst, src, shuf );
emit_2ub(p, X86_TWOB, 0xC6);
emit_modrm(p, dst, src);
emit_1ub(p, shuf);
}
void sse_cmpps( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src,
enum sse_cc cc)
{
DUMP_RRI( dst, src, cc );
emit_2ub(p, X86_TWOB, 0xC2);
emit_modrm(p, dst, src);
emit_1ub(p, cc);
}
void x87_fstp( struct x86_function *p, struct x86_reg dst )
{
DUMP_R( dst );
if (dst.file == file_x87)
emit_2ub(p, 0xdd, 0xd8 + dst.idx);
else {
emit_1ub(p, 0xd9);
emit_modrm_noreg(p, 3, dst);
}
note_x87_pop(p);
}
void sse_movntps( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src)
{
DUMP_RR( dst, src );
assert(dst.mod != mod_REG);
assert(src.mod == mod_REG);
emit_2ub(p, 0x0f, 0x2b);
emit_modrm(p, src, dst);
}
void sse_cvtps2pi( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
assert(dst.file == file_MMX &&
(src.file == file_XMM || src.mod != mod_REG));
p->need_emms = 1;
emit_2ub(p, X86_TWOB, 0x2d);
emit_modrm( p, dst, src );
}
void mmx_packuswb( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
assert(dst.file == file_MMX &&
(src.file == file_MMX || src.mod != mod_REG));
p->need_emms = 1;
emit_2ub(p, X86_TWOB, 0x67);
emit_modrm( p, dst, src );
}
