Lines Matching refs:SSE

2 …known-unknown -mattr=+sse2 | FileCheck %s --check-prefix=ALL --check-prefix=SSE --check-prefix=SSE2
3 …wn-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=ALL --check-prefix=SSE --check-prefix=SSE41
10 …-unknown -mattr=+sse2 | FileCheck %s --check-prefix=ALL --check-prefix=X32-SSE --check-prefix=X32-…
83 ; X32-SSE-LABEL: var_rotate_v2i64:
84 ; X32-SSE: # BB#0:
85 ; X32-SSE-NEXT: movdqa {{.*#+}} xmm2 = [64,0,64,0]
86 ; X32-SSE-NEXT: psubq %xmm1, %xmm2
87 ; X32-SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[2,3,0,1]
88 ; X32-SSE-NEXT: movdqa %xmm0, %xmm4
89 ; X32-SSE-NEXT: psllq %xmm3, %xmm4
90 ; X32-SSE-NEXT: movq {{.*#+}} xmm1 = xmm1[0],zero
91 ; X32-SSE-NEXT: movdqa %xmm0, %xmm3
92 ; X32-SSE-NEXT: psllq %xmm1, %xmm3
93 ; X32-SSE-NEXT: movsd {{.*#+}} xmm4 = xmm3[0],xmm4[1]
94 ; X32-SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm2[2,3,0,1]
95 ; X32-SSE-NEXT: movdqa %xmm0, %xmm1
96 ; X32-SSE-NEXT: psrlq %xmm3, %xmm1
97 ; X32-SSE-NEXT: movq {{.*#+}} xmm2 = xmm2[0],zero
98 ; X32-SSE-NEXT: psrlq %xmm2, %xmm0
99 ; X32-SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
100 ; X32-SSE-NEXT: orpd %xmm4, %xmm1
101 ; X32-SSE-NEXT: movapd %xmm1, %xmm0
102 ; X32-SSE-NEXT: retl
213 ; X32-SSE-LABEL: var_rotate_v4i32:
214 ; X32-SSE: # BB#0:
215 ; X32-SSE-NEXT: movdqa {{.*#+}} xmm2 = [32,32,32,32]
216 ; X32-SSE-NEXT: psubd %xmm1, %xmm2
217 ; X32-SSE-NEXT: pslld $23, %xmm1
218 ; X32-SSE-NEXT: paddd .LCPI1_1, %xmm1
219 ; X32-SSE-NEXT: cvttps2dq %xmm1, %xmm1
220 ; X32-SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[1,1,3,3]
221 ; X32-SSE-NEXT: pmuludq %xmm0, %xmm1
222 ; X32-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
223 ; X32-SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]
224 ; X32-SSE-NEXT: pmuludq %xmm3, %xmm4
225 ; X32-SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm4[0,2,2,3]
226 ; X32-SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
227 ; X32-SSE-NEXT: movdqa %xmm2, %xmm3
228 ; X32-SSE-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero…
229 ; X32-SSE-NEXT: movdqa %xmm0, %xmm4
230 ; X32-SSE-NEXT: psrld %xmm3, %xmm4
231 ; X32-SSE-NEXT: movdqa %xmm2, %xmm3
232 ; X32-SSE-NEXT: psrlq $32, %xmm3
233 ; X32-SSE-NEXT: movdqa %xmm0, %xmm5
234 ; X32-SSE-NEXT: psrld %xmm3, %xmm5
235 ; X32-SSE-NEXT: movsd {{.*#+}} xmm4 = xmm5[0],xmm4[1]
236 ; X32-SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm4[1,3,2,3]
237 ; X32-SSE-NEXT: pxor %xmm4, %xmm4
238 ; X32-SSE-NEXT: movdqa %xmm2, %xmm5
239 ; X32-SSE-NEXT: punpckhdq {{.*#+}} xmm5 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]
240 ; X32-SSE-NEXT: movdqa %xmm0, %xmm6
241 ; X32-SSE-NEXT: psrld %xmm5, %xmm6
242 ; X32-SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]
243 ; X32-SSE-NEXT: psrld %xmm2, %xmm0
244 ; X32-SSE-NEXT: movsd {{.*#+}} xmm6 = xmm0[0],xmm6[1]
245 ; X32-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,2,2,3]
246 ; X32-SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]
247 ; X32-SSE-NEXT: por %xmm1, %xmm0
248 ; X32-SSE-NEXT: retl
442 ; X32-SSE-LABEL: var_rotate_v8i16:
443 ; X32-SSE: # BB#0:
444 ; X32-SSE-NEXT: movdqa {{.*#+}} xmm3 = [16,16,16,16,16,16,16,16]
445 ; X32-SSE-NEXT: psubw %xmm1, %xmm3
446 ; X32-SSE-NEXT: psllw $12, %xmm1
447 ; X32-SSE-NEXT: movdqa %xmm1, %xmm2
448 ; X32-SSE-NEXT: psraw $15, %xmm2
449 ; X32-SSE-NEXT: movdqa %xmm0, %xmm4
450 ; X32-SSE-NEXT: psllw $8, %xmm4
451 ; X32-SSE-NEXT: pand %xmm2, %xmm4
452 ; X32-SSE-NEXT: pandn %xmm0, %xmm2
453 ; X32-SSE-NEXT: por %xmm4, %xmm2
454 ; X32-SSE-NEXT: paddw %xmm1, %xmm1
455 ; X32-SSE-NEXT: movdqa %xmm1, %xmm4
456 ; X32-SSE-NEXT: psraw $15, %xmm4
457 ; X32-SSE-NEXT: movdqa %xmm4, %xmm5
458 ; X32-SSE-NEXT: pandn %xmm2, %xmm5
459 ; X32-SSE-NEXT: psllw $4, %xmm2
460 ; X32-SSE-NEXT: pand %xmm4, %xmm2
461 ; X32-SSE-NEXT: por %xmm5, %xmm2
462 ; X32-SSE-NEXT: paddw %xmm1, %xmm1
463 ; X32-SSE-NEXT: movdqa %xmm1, %xmm4
464 ; X32-SSE-NEXT: psraw $15, %xmm4
465 ; X32-SSE-NEXT: movdqa %xmm4, %xmm5
466 ; X32-SSE-NEXT: pandn %xmm2, %xmm5
467 ; X32-SSE-NEXT: psllw $2, %xmm2
468 ; X32-SSE-NEXT: pand %xmm4, %xmm2
469 ; X32-SSE-NEXT: por %xmm5, %xmm2
470 ; X32-SSE-NEXT: paddw %xmm1, %xmm1
471 ; X32-SSE-NEXT: psraw $15, %xmm1
472 ; X32-SSE-NEXT: movdqa %xmm1, %xmm4
473 ; X32-SSE-NEXT: pandn %xmm2, %xmm4
474 ; X32-SSE-NEXT: psllw $1, %xmm2
475 ; X32-SSE-NEXT: pand %xmm1, %xmm2
476 ; X32-SSE-NEXT: psllw $12, %xmm3
477 ; X32-SSE-NEXT: movdqa %xmm3, %xmm1
478 ; X32-SSE-NEXT: psraw $15, %xmm1
479 ; X32-SSE-NEXT: movdqa %xmm1, %xmm5
480 ; X32-SSE-NEXT: pandn %xmm0, %xmm5
481 ; X32-SSE-NEXT: psrlw $8, %xmm0
482 ; X32-SSE-NEXT: pand %xmm1, %xmm0
483 ; X32-SSE-NEXT: por %xmm5, %xmm0
484 ; X32-SSE-NEXT: paddw %xmm3, %xmm3
485 ; X32-SSE-NEXT: movdqa %xmm3, %xmm1
486 ; X32-SSE-NEXT: psraw $15, %xmm1
487 ; X32-SSE-NEXT: movdqa %xmm1, %xmm5
488 ; X32-SSE-NEXT: pandn %xmm0, %xmm5
489 ; X32-SSE-NEXT: psrlw $4, %xmm0
490 ; X32-SSE-NEXT: pand %xmm1, %xmm0
491 ; X32-SSE-NEXT: por %xmm5, %xmm0
492 ; X32-SSE-NEXT: paddw %xmm3, %xmm3
493 ; X32-SSE-NEXT: movdqa %xmm3, %xmm1
494 ; X32-SSE-NEXT: psraw $15, %xmm1
495 ; X32-SSE-NEXT: movdqa %xmm1, %xmm5
496 ; X32-SSE-NEXT: pandn %xmm0, %xmm5
497 ; X32-SSE-NEXT: psrlw $2, %xmm0
498 ; X32-SSE-NEXT: pand %xmm1, %xmm0
499 ; X32-SSE-NEXT: por %xmm5, %xmm0
500 ; X32-SSE-NEXT: paddw %xmm3, %xmm3
501 ; X32-SSE-NEXT: psraw $15, %xmm3
502 ; X32-SSE-NEXT: movdqa %xmm3, %xmm1
503 ; X32-SSE-NEXT: pandn %xmm0, %xmm1
504 ; X32-SSE-NEXT: psrlw $1, %xmm0
505 ; X32-SSE-NEXT: pand %xmm3, %xmm0
506 ; X32-SSE-NEXT: por %xmm1, %xmm0
507 ; X32-SSE-NEXT: por %xmm4, %xmm0
508 ; X32-SSE-NEXT: por %xmm2, %xmm0
509 ; X32-SSE-NEXT: retl
660 ; X32-SSE-LABEL: var_rotate_v16i8:
661 ; X32-SSE: # BB#0:
662 ; X32-SSE-NEXT: movdqa {{.*#+}} xmm4 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8]
663 ; X32-SSE-NEXT: psubb %xmm1, %xmm4
664 ; X32-SSE-NEXT: psllw $5, %xmm1
665 ; X32-SSE-NEXT: pxor %xmm3, %xmm3
666 ; X32-SSE-NEXT: pxor %xmm2, %xmm2
667 ; X32-SSE-NEXT: pcmpgtb %xmm1, %xmm2
668 ; X32-SSE-NEXT: movdqa %xmm0, %xmm5
669 ; X32-SSE-NEXT: psllw $4, %xmm5
670 ; X32-SSE-NEXT: pand .LCPI3_1, %xmm5
671 ; X32-SSE-NEXT: pand %xmm2, %xmm5
672 ; X32-SSE-NEXT: pandn %xmm0, %xmm2
673 ; X32-SSE-NEXT: por %xmm5, %xmm2
674 ; X32-SSE-NEXT: paddb %xmm1, %xmm1
675 ; X32-SSE-NEXT: pxor %xmm5, %xmm5
676 ; X32-SSE-NEXT: pcmpgtb %xmm1, %xmm5
677 ; X32-SSE-NEXT: movdqa %xmm5, %xmm6
678 ; X32-SSE-NEXT: pandn %xmm2, %xmm6
679 ; X32-SSE-NEXT: psllw $2, %xmm2
680 ; X32-SSE-NEXT: pand .LCPI3_2, %xmm2
681 ; X32-SSE-NEXT: pand %xmm5, %xmm2
682 ; X32-SSE-NEXT: por %xmm6, %xmm2
683 ; X32-SSE-NEXT: paddb %xmm1, %xmm1
684 ; X32-SSE-NEXT: pxor %xmm5, %xmm5
685 ; X32-SSE-NEXT: pcmpgtb %xmm1, %xmm5
686 ; X32-SSE-NEXT: movdqa %xmm5, %xmm1
687 ; X32-SSE-NEXT: pandn %xmm2, %xmm1
688 ; X32-SSE-NEXT: paddb %xmm2, %xmm2
689 ; X32-SSE-NEXT: pand %xmm5, %xmm2
690 ; X32-SSE-NEXT: psllw $5, %xmm4
691 ; X32-SSE-NEXT: pxor %xmm5, %xmm5
692 ; X32-SSE-NEXT: pcmpgtb %xmm4, %xmm5
693 ; X32-SSE-NEXT: movdqa %xmm5, %xmm6
694 ; X32-SSE-NEXT: pandn %xmm0, %xmm6
695 ; X32-SSE-NEXT: psrlw $4, %xmm0
696 ; X32-SSE-NEXT: pand .LCPI3_3, %xmm0
697 ; X32-SSE-NEXT: pand %xmm5, %xmm0
698 ; X32-SSE-NEXT: por %xmm6, %xmm0
699 ; X32-SSE-NEXT: paddb %xmm4, %xmm4
700 ; X32-SSE-NEXT: pxor %xmm5, %xmm5
701 ; X32-SSE-NEXT: pcmpgtb %xmm4, %xmm5
702 ; X32-SSE-NEXT: movdqa %xmm5, %xmm6
703 ; X32-SSE-NEXT: pandn %xmm0, %xmm6
704 ; X32-SSE-NEXT: psrlw $2, %xmm0
705 ; X32-SSE-NEXT: pand .LCPI3_4, %xmm0
706 ; X32-SSE-NEXT: pand %xmm5, %xmm0
707 ; X32-SSE-NEXT: por %xmm6, %xmm0
708 ; X32-SSE-NEXT: paddb %xmm4, %xmm4
709 ; X32-SSE-NEXT: pcmpgtb %xmm4, %xmm3
710 ; X32-SSE-NEXT: movdqa %xmm3, %xmm4
711 ; X32-SSE-NEXT: pandn %xmm0, %xmm4
712 ; X32-SSE-NEXT: psrlw $1, %xmm0
713 ; X32-SSE-NEXT: pand .LCPI3_5, %xmm0
714 ; X32-SSE-NEXT: pand %xmm3, %xmm0
715 ; X32-SSE-NEXT: por %xmm4, %xmm0
716 ; X32-SSE-NEXT: por %xmm1, %xmm0
717 ; X32-SSE-NEXT: por %xmm2, %xmm0
718 ; X32-SSE-NEXT: retl
794 ; X32-SSE-LABEL: constant_rotate_v2i64:
795 ; X32-SSE: # BB#0:
796 ; X32-SSE-NEXT: movdqa %xmm0, %xmm2
797 ; X32-SSE-NEXT: psllq $14, %xmm2
798 ; X32-SSE-NEXT: movdqa %xmm0, %xmm1
799 ; X32-SSE-NEXT: psllq $4, %xmm1
800 ; X32-SSE-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]
801 ; X32-SSE-NEXT: movdqa %xmm0, %xmm1
802 ; X32-SSE-NEXT: psrlq $50, %xmm1
803 ; X32-SSE-NEXT: psrlq $60, %xmm0
804 ; X32-SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
805 ; X32-SSE-NEXT: orpd %xmm2, %xmm1
806 ; X32-SSE-NEXT: movapd %xmm1, %xmm0
807 ; X32-SSE-NEXT: retl
892 ; X32-SSE-LABEL: constant_rotate_v4i32:
893 ; X32-SSE: # BB#0:
894 ; X32-SSE-NEXT: movdqa {{.*#+}} xmm1 = [16,32,64,128]
895 ; X32-SSE-NEXT: movdqa %xmm0, %xmm2
896 ; X32-SSE-NEXT: pmuludq %xmm1, %xmm2
897 ; X32-SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
898 ; X32-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
899 ; X32-SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
900 ; X32-SSE-NEXT: pmuludq %xmm1, %xmm3
901 ; X32-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,2,2,3]
902 ; X32-SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
903 ; X32-SSE-NEXT: movdqa %xmm0, %xmm1
904 ; X32-SSE-NEXT: psrld $25, %xmm1
905 ; X32-SSE-NEXT: movdqa %xmm0, %xmm3
906 ; X32-SSE-NEXT: psrld $27, %xmm3
907 ; X32-SSE-NEXT: movsd {{.*#+}} xmm1 = xmm3[0],xmm1[1]
908 ; X32-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,3,2,3]
909 ; X32-SSE-NEXT: movdqa %xmm0, %xmm3
910 ; X32-SSE-NEXT: psrld $26, %xmm3
911 ; X32-SSE-NEXT: psrld $28, %xmm0
912 ; X32-SSE-NEXT: movsd {{.*#+}} xmm3 = xmm0[0],xmm3[1]
913 ; X32-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[0,2,2,3]
914 ; X32-SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
915 ; X32-SSE-NEXT: por %xmm2, %xmm0
916 ; X32-SSE-NEXT: retl
1020 ; X32-SSE-LABEL: constant_rotate_v8i16:
1021 ; X32-SSE: # BB#0:
1022 ; X32-SSE-NEXT: movdqa {{.*#+}} xmm2 = [1,2,4,8,16,32,64,128]
1023 ; X32-SSE-NEXT: pmullw %xmm0, %xmm2
1024 ; X32-SSE-NEXT: movdqa {{.*#+}} xmm1 = [0,65535,65535,65535,65535,65535,65535,65535]
1025 ; X32-SSE-NEXT: movdqa %xmm1, %xmm3
1026 ; X32-SSE-NEXT: pandn %xmm0, %xmm3
1027 ; X32-SSE-NEXT: psrlw $8, %xmm0
1028 ; X32-SSE-NEXT: pand %xmm1, %xmm0
1029 ; X32-SSE-NEXT: por %xmm3, %xmm0
1030 ; X32-SSE-NEXT: movdqa {{.*#+}} xmm1 = [0,65535,65535,65535,65535,0,0,0]
1031 ; X32-SSE-NEXT: movdqa %xmm1, %xmm3
1032 ; X32-SSE-NEXT: pandn %xmm0, %xmm3
1033 ; X32-SSE-NEXT: psrlw $4, %xmm0
1034 ; X32-SSE-NEXT: pand %xmm1, %xmm0
1035 ; X32-SSE-NEXT: por %xmm3, %xmm0
1036 ; X32-SSE-NEXT: movdqa {{.*#+}} xmm1 = [0,65535,65535,0,0,65535,65535,0]
1037 ; X32-SSE-NEXT: movdqa %xmm1, %xmm3
1038 ; X32-SSE-NEXT: pandn %xmm0, %xmm3
1039 ; X32-SSE-NEXT: psrlw $2, %xmm0
1040 ; X32-SSE-NEXT: pand %xmm1, %xmm0
1041 ; X32-SSE-NEXT: por %xmm3, %xmm0
1042 ; X32-SSE-NEXT: movdqa {{.*#+}} xmm3 = [65535,0,65535,0,65535,0,65535,0]
1043 ; X32-SSE-NEXT: movdqa %xmm0, %xmm1
1044 ; X32-SSE-NEXT: pand %xmm3, %xmm1
1045 ; X32-SSE-NEXT: psrlw $1, %xmm0
1046 ; X32-SSE-NEXT: pandn %xmm0, %xmm3
1047 ; X32-SSE-NEXT: por %xmm2, %xmm3
1048 ; X32-SSE-NEXT: por %xmm3, %xmm1
1049 ; X32-SSE-NEXT: movdqa %xmm1, %xmm0
1050 ; X32-SSE-NEXT: retl
1196 ; X32-SSE-LABEL: constant_rotate_v16i8:
1197 ; X32-SSE: # BB#0:
1198 ; X32-SSE-NEXT: movdqa {{.*#+}} xmm3 = [0,1,2,3,4,5,6,7,8,7,6,5,4,3,2,1]
1199 ; X32-SSE-NEXT: psllw $5, %xmm3
1200 ; X32-SSE-NEXT: pxor %xmm2, %xmm2
1201 ; X32-SSE-NEXT: pxor %xmm1, %xmm1
1202 ; X32-SSE-NEXT: pcmpgtb %xmm3, %xmm1
1203 ; X32-SSE-NEXT: movdqa %xmm0, %xmm4
1204 ; X32-SSE-NEXT: psllw $4, %xmm4
1205 ; X32-SSE-NEXT: pand .LCPI7_1, %xmm4
1206 ; X32-SSE-NEXT: pand %xmm1, %xmm4
1207 ; X32-SSE-NEXT: pandn %xmm0, %xmm1
1208 ; X32-SSE-NEXT: por %xmm4, %xmm1
1209 ; X32-SSE-NEXT: paddb %xmm3, %xmm3
1210 ; X32-SSE-NEXT: pxor %xmm4, %xmm4
1211 ; X32-SSE-NEXT: pcmpgtb %xmm3, %xmm4
1212 ; X32-SSE-NEXT: movdqa %xmm4, %xmm5
1213 ; X32-SSE-NEXT: pandn %xmm1, %xmm5
1214 ; X32-SSE-NEXT: psllw $2, %xmm1
1215 ; X32-SSE-NEXT: pand .LCPI7_2, %xmm1
1216 ; X32-SSE-NEXT: pand %xmm4, %xmm1
1217 ; X32-SSE-NEXT: por %xmm5, %xmm1
1218 ; X32-SSE-NEXT: paddb %xmm3, %xmm3
1219 ; X32-SSE-NEXT: pxor %xmm4, %xmm4
1220 ; X32-SSE-NEXT: pcmpgtb %xmm3, %xmm4
1221 ; X32-SSE-NEXT: movdqa %xmm4, %xmm3
1222 ; X32-SSE-NEXT: pandn %xmm1, %xmm3
1223 ; X32-SSE-NEXT: paddb %xmm1, %xmm1
1224 ; X32-SSE-NEXT: pand %xmm4, %xmm1
1225 ; X32-SSE-NEXT: movdqa {{.*#+}} xmm4 = [8,7,6,5,4,3,2,1,0,1,2,3,4,5,6,7]
1226 ; X32-SSE-NEXT: psllw $5, %xmm4
1227 ; X32-SSE-NEXT: pxor %xmm5, %xmm5
1228 ; X32-SSE-NEXT: pcmpgtb %xmm4, %xmm5
1229 ; X32-SSE-NEXT: movdqa %xmm5, %xmm6
1230 ; X32-SSE-NEXT: pandn %xmm0, %xmm6
1231 ; X32-SSE-NEXT: psrlw $4, %xmm0
1232 ; X32-SSE-NEXT: pand .LCPI7_4, %xmm0
1233 ; X32-SSE-NEXT: pand %xmm5, %xmm0
1234 ; X32-SSE-NEXT: por %xmm6, %xmm0
1235 ; X32-SSE-NEXT: paddb %xmm4, %xmm4
1236 ; X32-SSE-NEXT: pxor %xmm5, %xmm5
1237 ; X32-SSE-NEXT: pcmpgtb %xmm4, %xmm5
1238 ; X32-SSE-NEXT: movdqa %xmm5, %xmm6
1239 ; X32-SSE-NEXT: pandn %xmm0, %xmm6
1240 ; X32-SSE-NEXT: psrlw $2, %xmm0
1241 ; X32-SSE-NEXT: pand .LCPI7_5, %xmm0
1242 ; X32-SSE-NEXT: pand %xmm5, %xmm0
1243 ; X32-SSE-NEXT: por %xmm6, %xmm0
1244 ; X32-SSE-NEXT: paddb %xmm4, %xmm4
1245 ; X32-SSE-NEXT: pcmpgtb %xmm4, %xmm2
1246 ; X32-SSE-NEXT: movdqa %xmm2, %xmm4
1247 ; X32-SSE-NEXT: pandn %xmm0, %xmm4
1248 ; X32-SSE-NEXT: psrlw $1, %xmm0
1249 ; X32-SSE-NEXT: pand .LCPI7_6, %xmm0
1250 ; X32-SSE-NEXT: pand %xmm2, %xmm0
1251 ; X32-SSE-NEXT: por %xmm4, %xmm0
1252 ; X32-SSE-NEXT: por %xmm3, %xmm0
1253 ; X32-SSE-NEXT: por %xmm1, %xmm0
1254 ; X32-SSE-NEXT: retl
1266 ; SSE-LABEL: splatconstant_rotate_v2i64:
1267 ; SSE: # BB#0:
1268 ; SSE-NEXT: movdqa %xmm0, %xmm1
1269 ; SSE-NEXT: psllq $14, %xmm1
1270 ; SSE-NEXT: psrlq $50, %xmm0
1271 ; SSE-NEXT: por %xmm1, %xmm0
1272 ; SSE-NEXT: retq
1286 ; X32-SSE-LABEL: splatconstant_rotate_v2i64:
1287 ; X32-SSE: # BB#0:
1288 ; X32-SSE-NEXT: movdqa %xmm0, %xmm1
1289 ; X32-SSE-NEXT: psllq $14, %xmm1
1290 ; X32-SSE-NEXT: psrlq $50, %xmm0
1291 ; X32-SSE-NEXT: por %xmm1, %xmm0
1292 ; X32-SSE-NEXT: retl
1300 ; SSE-LABEL: splatconstant_rotate_v4i32:
1301 ; SSE: # BB#0:
1302 ; SSE-NEXT: movdqa %xmm0, %xmm1
1303 ; SSE-NEXT: pslld $4, %xmm1
1304 ; SSE-NEXT: psrld $28, %xmm0
1305 ; SSE-NEXT: por %xmm1, %xmm0
1306 ; SSE-NEXT: retq
1320 ; X32-SSE-LABEL: splatconstant_rotate_v4i32:
1321 ; X32-SSE: # BB#0:
1322 ; X32-SSE-NEXT: movdqa %xmm0, %xmm1
1323 ; X32-SSE-NEXT: pslld $4, %xmm1
1324 ; X32-SSE-NEXT: psrld $28, %xmm0
1325 ; X32-SSE-NEXT: por %xmm1, %xmm0
1326 ; X32-SSE-NEXT: retl
1334 ; SSE-LABEL: splatconstant_rotate_v8i16:
1335 ; SSE: # BB#0:
1336 ; SSE-NEXT: movdqa %xmm0, %xmm1
1337 ; SSE-NEXT: psllw $7, %xmm1
1338 ; SSE-NEXT: psrlw $9, %xmm0
1339 ; SSE-NEXT: por %xmm1, %xmm0
1340 ; SSE-NEXT: retq
1354 ; X32-SSE-LABEL: splatconstant_rotate_v8i16:
1355 ; X32-SSE: # BB#0:
1356 ; X32-SSE-NEXT: movdqa %xmm0, %xmm1
1357 ; X32-SSE-NEXT: psllw $7, %xmm1
1358 ; X32-SSE-NEXT: psrlw $9, %xmm0
1359 ; X32-SSE-NEXT: por %xmm1, %xmm0
1360 ; X32-SSE-NEXT: retl
1368 ; SSE-LABEL: splatconstant_rotate_v16i8:
1369 ; SSE: # BB#0:
1370 ; SSE-NEXT: movdqa %xmm0, %xmm1
1371 ; SSE-NEXT: psllw $4, %xmm1
1372 ; SSE-NEXT: pand {{.*}}(%rip), %xmm1
1373 ; SSE-NEXT: psrlw $4, %xmm0
1374 ; SSE-NEXT: pand {{.*}}(%rip), %xmm0
1375 ; SSE-NEXT: por %xmm1, %xmm0
1376 ; SSE-NEXT: retq
1392 ; X32-SSE-LABEL: splatconstant_rotate_v16i8:
1393 ; X32-SSE: # BB#0:
1394 ; X32-SSE-NEXT: movdqa %xmm0, %xmm1
1395 ; X32-SSE-NEXT: psllw $4, %xmm1
1396 ; X32-SSE-NEXT: pand .LCPI11_0, %xmm1
1397 ; X32-SSE-NEXT: psrlw $4, %xmm0
1398 ; X32-SSE-NEXT: pand .LCPI11_1, %xmm0
1399 ; X32-SSE-NEXT: por %xmm1, %xmm0
1400 ; X32-SSE-NEXT: retl
1412 ; SSE-LABEL: splatconstant_rotate_mask_v2i64:
1413 ; SSE: # BB#0:
1414 ; SSE-NEXT: movdqa %xmm0, %xmm1
1415 ; SSE-NEXT: psllq $15, %xmm1
1416 ; SSE-NEXT: psrlq $49, %xmm0
1417 ; SSE-NEXT: pand {{.*}}(%rip), %xmm0
1418 ; SSE-NEXT: pand {{.*}}(%rip), %xmm1
1419 ; SSE-NEXT: por %xmm0, %xmm1
1420 ; SSE-NEXT: movdqa %xmm1, %xmm0
1421 ; SSE-NEXT: retq
1438 ; X32-SSE-LABEL: splatconstant_rotate_mask_v2i64:
1439 ; X32-SSE: # BB#0:
1440 ; X32-SSE-NEXT: movdqa %xmm0, %xmm1
1441 ; X32-SSE-NEXT: psllq $15, %xmm1
1442 ; X32-SSE-NEXT: psrlq $49, %xmm0
1443 ; X32-SSE-NEXT: pand .LCPI12_0, %xmm0
1444 ; X32-SSE-NEXT: pand .LCPI12_1, %xmm1
1445 ; X32-SSE-NEXT: por %xmm0, %xmm1
1446 ; X32-SSE-NEXT: movdqa %xmm1, %xmm0
1447 ; X32-SSE-NEXT: retl
1457 ; SSE-LABEL: splatconstant_rotate_mask_v4i32:
1458 ; SSE: # BB#0:
1459 ; SSE-NEXT: movdqa %xmm0, %xmm1
1460 ; SSE-NEXT: pslld $4, %xmm1
1461 ; SSE-NEXT: psrld $28, %xmm0
1462 ; SSE-NEXT: pand {{.*}}(%rip), %xmm0
1463 ; SSE-NEXT: pand {{.*}}(%rip), %xmm1
1464 ; SSE-NEXT: por %xmm0, %xmm1
1465 ; SSE-NEXT: movdqa %xmm1, %xmm0
1466 ; SSE-NEXT: retq
1483 ; X32-SSE-LABEL: splatconstant_rotate_mask_v4i32:
1484 ; X32-SSE: # BB#0:
1485 ; X32-SSE-NEXT: movdqa %xmm0, %xmm1
1486 ; X32-SSE-NEXT: pslld $4, %xmm1
1487 ; X32-SSE-NEXT: psrld $28, %xmm0
1488 ; X32-SSE-NEXT: pand .LCPI13_0, %xmm0
1489 ; X32-SSE-NEXT: pand .LCPI13_1, %xmm1
1490 ; X32-SSE-NEXT: por %xmm0, %xmm1
1491 ; X32-SSE-NEXT: movdqa %xmm1, %xmm0
1492 ; X32-SSE-NEXT: retl
1502 ; SSE-LABEL: splatconstant_rotate_mask_v8i16:
1503 ; SSE: # BB#0:
1504 ; SSE-NEXT: movdqa %xmm0, %xmm1
1505 ; SSE-NEXT: psllw $5, %xmm1
1506 ; SSE-NEXT: psrlw $11, %xmm0
1507 ; SSE-NEXT: pand {{.*}}(%rip), %xmm0
1508 ; SSE-NEXT: pand {{.*}}(%rip), %xmm1
1509 ; SSE-NEXT: por %xmm0, %xmm1
1510 ; SSE-NEXT: movdqa %xmm1, %xmm0
1511 ; SSE-NEXT: retq
1528 ; X32-SSE-LABEL: splatconstant_rotate_mask_v8i16:
1529 ; X32-SSE: # BB#0:
1530 ; X32-SSE-NEXT: movdqa %xmm0, %xmm1
1531 ; X32-SSE-NEXT: psllw $5, %xmm1
1532 ; X32-SSE-NEXT: psrlw $11, %xmm0
1533 ; X32-SSE-NEXT: pand .LCPI14_0, %xmm0
1534 ; X32-SSE-NEXT: pand .LCPI14_1, %xmm1
1535 ; X32-SSE-NEXT: por %xmm0, %xmm1
1536 ; X32-SSE-NEXT: movdqa %xmm1, %xmm0
1537 ; X32-SSE-NEXT: retl
1547 ; SSE-LABEL: splatconstant_rotate_mask_v16i8:
1548 ; SSE: # BB#0:
1549 ; SSE-NEXT: movdqa %xmm0, %xmm1
1550 ; SSE-NEXT: psllw $4, %xmm1
1551 ; SSE-NEXT: pand {{.*}}(%rip), %xmm1
1552 ; SSE-NEXT: psrlw $4, %xmm0
1553 ; SSE-NEXT: pand {{.*}}(%rip), %xmm0
1554 ; SSE-NEXT: pand {{.*}}(%rip), %xmm0
1555 ; SSE-NEXT: pand {{.*}}(%rip), %xmm1
1556 ; SSE-NEXT: por %xmm0, %xmm1
1557 ; SSE-NEXT: movdqa %xmm1, %xmm0
1558 ; SSE-NEXT: retq
1577 ; X32-SSE-LABEL: splatconstant_rotate_mask_v16i8:
1578 ; X32-SSE: # BB#0:
1579 ; X32-SSE-NEXT: movdqa %xmm0, %xmm1
1580 ; X32-SSE-NEXT: psllw $4, %xmm1
1581 ; X32-SSE-NEXT: pand .LCPI15_0, %xmm1
1582 ; X32-SSE-NEXT: psrlw $4, %xmm0
1583 ; X32-SSE-NEXT: pand .LCPI15_1, %xmm0
1584 ; X32-SSE-NEXT: pand .LCPI15_2, %xmm0
1585 ; X32-SSE-NEXT: pand .LCPI15_3, %xmm1
1586 ; X32-SSE-NEXT: por %xmm0, %xmm1
1587 ; X32-SSE-NEXT: movdqa %xmm1, %xmm0
1588 ; X32-SSE-NEXT: retl