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 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 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);
}
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_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 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 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_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 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 );
}
/**
* 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);
}
/**
* 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);
}