void sse2_rcpss( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
emit_3ub(p, 0xF3, X86_TWOB, 0x53);
emit_modrm( p, dst, src );
}
void x86_lea( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
emit_1ub(p, 0x8d);
emit_modrm( p, dst, src );
}
void sse2_packuswb( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
emit_3ub(p, 0x66, X86_TWOB, 0x67);
emit_modrm( p, dst, src );
}
void x86_test( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
emit_1ub(p, 0x85);
emit_modrm( p, dst, src );
}
void sse2_cvtps2dq( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
emit_3ub(p, 0x66, X86_TWOB, 0x5B);
emit_modrm( p, dst, src );
}
void sse_pmovmskb( struct x86_function *p,
struct x86_reg dest,
struct x86_reg src)
{
emit_3ub(p, 0x66, X86_TWOB, 0xD7);
emit_modrm(p, dest, src);
}
/* Emits the "/0".."/7" specialized versions of the modrm ("/r") bytes.
*/
static void emit_modrm_noreg( struct x86_function *p,
unsigned op,
struct x86_reg regmem )
{
struct x86_reg dummy = x86_make_reg(file_REG32, op);
emit_modrm(p, dummy, regmem);
}
void sse_xorps( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
emit_2ub(p, X86_TWOB, 0x57);
emit_modrm( p, dst, src );
}
void sse_rsqrtss( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
emit_3ub(p, 0xF3, X86_TWOB, 0x52);
emit_modrm( p, dst, src );
}
void sse2_cvttps2dq( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
DUMP_RR( dst, src );
emit_3ub( p, 0xF3, X86_TWOB, 0x5B );
emit_modrm( p, dst, src );
}
void sse2_punpcklbw( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
DUMP_RR( dst, src );
emit_3ub(p, 0x66, X86_TWOB, 0x60);
emit_modrm( p, dst, src );
}
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_rsqrtps( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
DUMP_RR( dst, src );
emit_2ub(p, X86_TWOB, 0x52);
emit_modrm( p, dst, src );
}
void sse_addss( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
DUMP_RR( dst, src );
emit_3ub(p, 0xF3, X86_TWOB, 0x58);
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);
}
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 sse2_packsswb( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
DUMP_RR( dst, src );
emit_3ub(p, 0x66, X86_TWOB, 0x63);
emit_modrm( p, dst, src );
}
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);
}
/**
* Perform a reduced swizzle:
*/
void sse2_pshufd( struct x86_function *p,
struct x86_reg dest,
struct x86_reg arg0,
unsigned char shuf)
{
emit_3ub(p, 0x66, X86_TWOB, 0x70);
emit_modrm(p, dest, arg0);
emit_1ub(p, shuf);
}
/* 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 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);
}
/* 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);
}
/**
* Perform a reduced swizzle:
*/
void sse2_pshufd( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src,
unsigned char shuf)
{
DUMP_RRI( dst, src, shuf );
emit_3ub(p, 0x66, X86_TWOB, 0x70);
emit_modrm(p, dst, src);
emit_1ub(p, shuf);
}
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 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 );
}
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 );
}
/* Many x86 instructions have two opcodes to cope with the situations
* where the destination is a register or memory reference
* respectively. This function selects the correct opcode based on
* the arguments presented.
*/
static void emit_op_modrm( struct x86_function *p,
unsigned char op_dst_is_reg,
unsigned char op_dst_is_mem,
struct x86_reg dst,
struct x86_reg src )
{
switch (dst.mod) {
case mod_REG:
emit_1ub(p, op_dst_is_reg);
emit_modrm(p, dst, src);
break;
case mod_INDIRECT:
case mod_DISP32:
case mod_DISP8:
assert(src.mod == mod_REG);
emit_1ub(p, op_dst_is_mem);
emit_modrm(p, src, dst);
break;
default:
assert(0);
break;
}
}
void sse_unpcklps( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
{
DUMP_RR( dst, src );
emit_2ub( p, X86_TWOB, 0x14 );
emit_modrm( p, dst, src );
}
void x86_call( struct x86_function *p, struct x86_reg reg)
{
emit_1ub(p, 0xff);
emit_modrm(p, reg, reg);
}
