+void sse2_punpcklwd( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
+{
+ DUMP_RR( dst, src );
+ emit_3ub(p, 0x66, 0x0f, 0x61);
+ emit_modrm( p, dst, src );
+}
+
+void sse2_punpckldq( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
+{
+ DUMP_RR( dst, src );
+ emit_3ub(p, 0x66, 0x0f, 0x62);
+ emit_modrm( p, dst, src );
+}
+
+void sse2_punpcklqdq( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
+{
+ DUMP_RR( dst, src );
+ emit_3ub(p, 0x66, 0x0f, 0x6c);
+ emit_modrm( p, dst, src );
+}
+
+void sse2_psllw_imm( struct x86_function *p, struct x86_reg dst, unsigned imm )
+{
+ DUMP_RI(dst, imm);
+ emit_3ub(p, 0x66, 0x0f, 0x71);
+ emit_modrm_noreg(p, 6, dst);
+ emit_1ub(p, imm);
+}
+
+void sse2_pslld_imm( struct x86_function *p, struct x86_reg dst, unsigned imm )
+{
+ DUMP_RI(dst, imm);
+ emit_3ub(p, 0x66, 0x0f, 0x72);
+ emit_modrm_noreg(p, 6, dst);
+ emit_1ub(p, imm);
+}
+
+void sse2_psllq_imm( struct x86_function *p, struct x86_reg dst, unsigned imm )
+{
+ DUMP_RI(dst, imm);
+ emit_3ub(p, 0x66, 0x0f, 0x73);
+ emit_modrm_noreg(p, 6, dst);
+ emit_1ub(p, imm);
+}
+
+void sse2_psrlw_imm( struct x86_function *p, struct x86_reg dst, unsigned imm )
+{
+ DUMP_RI(dst, imm);
+ emit_3ub(p, 0x66, 0x0f, 0x71);
+ emit_modrm_noreg(p, 2, dst);
+ emit_1ub(p, imm);
+}
+
+void sse2_psrld_imm( struct x86_function *p, struct x86_reg dst, unsigned imm )
+{
+ DUMP_RI(dst, imm);
+ emit_3ub(p, 0x66, 0x0f, 0x72);
+ emit_modrm_noreg(p, 2, dst);
+ emit_1ub(p, imm);
+}
+
+void sse2_psrlq_imm( struct x86_function *p, struct x86_reg dst, unsigned imm )
+{
+ DUMP_RI(dst, imm);
+ emit_3ub(p, 0x66, 0x0f, 0x73);
+ emit_modrm_noreg(p, 2, dst);
+ emit_1ub(p, imm);
+}
+
+void sse2_psraw_imm( struct x86_function *p, struct x86_reg dst, unsigned imm )
+{
+ DUMP_RI(dst, imm);
+ emit_3ub(p, 0x66, 0x0f, 0x71);
+ emit_modrm_noreg(p, 4, dst);
+ emit_1ub(p, imm);
+}
+
+void sse2_psrad_imm( struct x86_function *p, struct x86_reg dst, unsigned imm )
+{
+ DUMP_RI(dst, imm);
+ emit_3ub(p, 0x66, 0x0f, 0x72);
+ emit_modrm_noreg(p, 4, dst);
+ emit_1ub(p, imm);
+}
+
+void sse2_por( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
+{
+ DUMP_RR(dst, src);
+ emit_3ub(p, 0x66, 0x0f, 0xeb);
+ emit_modrm(p, dst, src);
+}