Lines Matching +full:0 +full:x1b8
20 ; SSE: # %bb.0:
21 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
29 ; SSE-NEXT: movlhps {{.*#+}} xmm6 = xmm6[0],xmm1[0]
35 ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm5[0]
36 ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm3[0]
46 ; AVX: # %bb.0:
47 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
56 ; AVX-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[2],ymm6[2]
59 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[3],ymm1[3]
70 ; AVX2: # %bb.0:
71 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
78 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
79 ; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,2,1,3]
80 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5],ymm0[6,7]
81 ; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
83 ; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,2,1,3]
91 ; AVX2-FP: # %bb.0:
92 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
99 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
100 ; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,2,1,3]
101 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5],ymm0[6,7]
102 ; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
104 ; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,2,1,3]
112 ; AVX2-FCP: # %bb.0:
113 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
120 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
121 ; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,2,1,3]
122 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5],ymm0[6,7]
123 ; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
125 ; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,2,1,3]
133 ; AVX512: # %bb.0:
134 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
144 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,2,4,6,8,10,1,3]
152 ; AVX512-FCP: # %bb.0:
153 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
163 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,2,4,6,8,10,1,3]
171 ; AVX512DQ: # %bb.0:
172 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
182 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,2,4,6,8,10,1,3]
190 ; AVX512DQ-FCP: # %bb.0:
191 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
201 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,2,4,6,8,10,1,3]
209 ; AVX512BW: # %bb.0:
210 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
220 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,2,4,6,8,10,1,3]
228 ; AVX512BW-FCP: # %bb.0:
229 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
239 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,2,4,6,8,10,1,3]
247 ; AVX512DQ-BW: # %bb.0:
248 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
258 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,2,4,6,8,10,1,3]
266 ; AVX512DQ-BW-FCP: # %bb.0:
267 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
277 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,2,4,6,8,10,1,3]
289 %1 = shufflevector <2 x i64> %in.vec0, <2 x i64> %in.vec1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
290 %2 = shufflevector <2 x i64> %in.vec2, <2 x i64> %in.vec3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
291 %3 = shufflevector <2 x i64> %in.vec4, <2 x i64> %in.vec5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
292 %4 = shufflevector <4 x i64> %1, <4 x i64> %2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
293 %5 = shufflevector <4 x i64> %3, <4 x i64> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef>
294 %6 = shufflevector <8 x i64> %4, <8 x i64> %5, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>
295 %interleaved.vec = shufflevector <12 x i64> %6, <12 x i64> poison, <12 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 1, i32 3, i32 5, i32 7, i32 9, i32 11>
302 ; SSE: # %bb.0:
303 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
319 ; SSE-NEXT: movlhps {{.*#+}} xmm13 = xmm13[0],xmm12[0]
324 ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm7[0]
326 ; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm5[0]
330 ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm8[0]
331 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm6[0]
333 ; SSE-NEXT: movlhps {{.*#+}} xmm9 = xmm9[0],xmm11[0]
349 ; AVX: # %bb.0:
350 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
356 ; AVX-NEXT: vmovddup {{.*#+}} xmm5 = mem[0,0]
361 ; AVX-NEXT: vblendpd {{.*#+}} ymm8 = ymm3[0,1],ymm8[2,3]
362 ; AVX-NEXT: vblendpd {{.*#+}} ymm5 = ymm8[0],ymm5[1],ymm8[2,3]
368 ; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm12[4,5,6,7]
369 ; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1],ymm11[2,3],ymm10[4,5],ymm11[6,7]
370 ; AVX-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm2[0],mem[0],ymm2[2],mem[2]
372 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm11 = xmm11[0],mem[0]
373 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm11[0,1,2,3],ymm2[4,5,6,7]
377 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm11[0],ymm0[2],ymm11[3]
381 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm1[0,1],ymm3[2,3]
382 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm1[0,1,2],ymm4[3]
383 ; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm7[0],xmm6[0]
384 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm9[0],xmm8[0]
396 ; AVX2: # %bb.0:
397 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
404 ; AVX2-NEXT: vmovddup {{.*#+}} xmm6 = xmm5[0,0]
408 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm4[0,1],ymm9[0,1]
409 ; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm9[0,1],ymm6[2,3],ymm9[4,5,6,7]
415 ; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5,6,7]
416 ; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm11[0,1,2,3,4,5],ymm5[6,7]
419 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm8[0],ymm7[0],ymm8[2],ymm7[2]
420 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
421 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
426 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],mem[6,7]
430 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
441 ; AVX2-FP: # %bb.0:
442 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
449 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm6 = xmm5[0,0]
453 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm4[0,1],ymm9[0,1]
454 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm9[0,1],ymm6[2,3],ymm9[4,5,6,7]
460 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5,6,7]
461 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm11[0,1,2,3,4,5],ymm5[6,7]
464 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm8[0],ymm7[0],ymm8[2],ymm7[2]
465 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
466 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
471 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],mem[6,7]
475 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
486 ; AVX2-FCP: # %bb.0:
487 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
494 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm6 = xmm5[0,0]
498 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm4[0,1],ymm9[0,1]
499 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm9[0,1],ymm6[2,3],ymm9[4,5,6,7]
505 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5,6,7]
506 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm11[0,1,2,3,4,5],ymm5[6,7]
509 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm8[0],ymm7[0],ymm8[2],ymm7[2]
510 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
511 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
516 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],mem[6,7]
520 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
531 ; AVX512: # %bb.0:
532 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
539 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,4,8,12,0,0,1,5]
541 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,8,12,6,7]
543 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,5,0,0,10,14,2,6]
545 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,13,4,5,6,7]
548 ; AVX512-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
559 ; AVX512-FCP: # %bb.0:
560 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
567 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,4,8,12,0,0,1,5]
569 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,8,12,6,7]
571 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,5,0,0,10,14,2,6]
573 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,13,4,5,6,7]
576 ; AVX512-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
587 ; AVX512DQ: # %bb.0:
588 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
595 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,4,8,12,0,0,1,5]
597 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,8,12,6,7]
599 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,5,0,0,10,14,2,6]
601 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,13,4,5,6,7]
604 ; AVX512DQ-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
615 ; AVX512DQ-FCP: # %bb.0:
616 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
623 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,4,8,12,0,0,1,5]
625 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,8,12,6,7]
627 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,5,0,0,10,14,2,6]
629 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,13,4,5,6,7]
632 ; AVX512DQ-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
643 ; AVX512BW: # %bb.0:
644 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
651 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,4,8,12,0,0,1,5]
653 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,8,12,6,7]
655 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,5,0,0,10,14,2,6]
657 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,13,4,5,6,7]
660 ; AVX512BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
671 ; AVX512BW-FCP: # %bb.0:
672 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
679 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,4,8,12,0,0,1,5]
681 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,8,12,6,7]
683 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,5,0,0,10,14,2,6]
685 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,13,4,5,6,7]
688 ; AVX512BW-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
699 ; AVX512DQ-BW: # %bb.0:
700 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
707 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,4,8,12,0,0,1,5]
709 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,8,12,6,7]
711 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,5,0,0,10,14,2,6]
713 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,13,4,5,6,7]
716 ; AVX512DQ-BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
727 ; AVX512DQ-BW-FCP: # %bb.0:
728 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
735 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,4,8,12,0,0,1,5]
737 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,8,12,6,7]
739 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,5,0,0,10,14,2,6]
741 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,13,4,5,6,7]
744 ; AVX512DQ-BW-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
759 %1 = shufflevector <4 x i64> %in.vec0, <4 x i64> %in.vec1, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
760 %2 = shufflevector <4 x i64> %in.vec2, <4 x i64> %in.vec3, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
761 %3 = shufflevector <4 x i64> %in.vec4, <4 x i64> %in.vec5, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
762 %4 = shufflevector <8 x i64> %1, <8 x i64> %2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
763 %5 = shufflevector <8 x i64> %3, <8 x i64> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
764 %6 = shufflevector <16 x i64> %4, <16 x i64> %5, <24 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23>
765 %interleaved.vec = shufflevector <24 x i64> %6, <24 x i64> poison, <24 x i32> <i32 0, i32 4, i32 8, i32 12, i32 16, i32 20, i32 1, i32 5, i32 9, i32 13, i32 17, i32 21, i32 2, i32 6, i32 10, i32 14, i32 18, i32 22, i32 3, i32 7, i32 11, i32 15, i32 19, i32 23>
772 ; SSE: # %bb.0:
783 ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
791 ; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm1[0]
794 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
796 ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm10[0]
797 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
799 ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
801 ; SSE-NEXT: movlhps {{.*#+}} xmm10 = xmm10[0],xmm11[0]
803 ; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
805 ; SSE-NEXT: movlhps {{.*#+}} xmm11 = xmm11[0],xmm12[0]
807 ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
809 ; SSE-NEXT: movlhps {{.*#+}} xmm12 = xmm12[0],xmm13[0]
811 ; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
813 ; SSE-NEXT: movlhps {{.*#+}} xmm13 = xmm13[0],xmm0[0]
816 ; SSE-NEXT: movlhps {{.*#+}} xmm15 = xmm15[0],xmm14[0]
818 ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
820 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
822 ; SSE-NEXT: movlhps {{.*#+}} xmm14 = xmm14[0],xmm0[0]
824 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
828 ; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]
833 ; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm1[0]
838 ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]
843 ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm3[0]
845 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
853 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
855 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
861 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
863 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
868 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
870 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
872 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
875 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
883 ; AVX: # %bb.0:
887 ; AVX-NEXT: vmovddup {{.*#+}} xmm0 = mem[0,0]
889 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
894 ; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm11[0,1],ymm2[2,3]
895 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3]
896 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
903 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm9[4,5,6,7]
904 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm7[2,3],ymm2[4,5],ymm7[6,7]
905 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
910 ; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm12[4,5,6,7]
911 ; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1],ymm10[2,3],ymm7[4,5],ymm10[6,7]
912 ; AVX-NEXT: vmovddup {{.*#+}} xmm10 = mem[0,0]
916 ; AVX-NEXT: vblendpd {{.*#+}} ymm15 = ymm14[0,1],ymm15[2,3]
917 ; AVX-NEXT: vblendpd {{.*#+}} ymm10 = ymm15[0],ymm10[1],ymm15[2,3]
923 ; AVX-NEXT: vshufpd {{.*#+}} ymm13 = ymm13[0],ymm14[0],ymm13[2],ymm14[3]
930 ; AVX-NEXT: vshufpd {{.*#+}} ymm14 = ymm11[0],ymm14[0],ymm11[2],ymm14[3]
934 ; AVX-NEXT: vblendpd {{.*#+}} ymm11 = ymm11[0,1],ymm15[2,3]
935 ; AVX-NEXT: vblendpd {{.*#+}} ymm11 = ymm11[0,1,2],ymm0[3]
937 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
938 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
940 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm15[6,7]
944 ; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm15[0,1],ymm2[2,3]
945 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3]
947 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm2 = xmm2[0],mem[0]
948 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
950 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm15[6,7]
951 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm12[0],xmm4[0]
952 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm8[0],xmm6[0]
953 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
954 ; AVX-NEXT: # xmm3 = xmm3[0],mem[0]
955 ; AVX-NEXT: vmovlhps {{.*#+}} xmm5 = xmm9[0],xmm5[0]
956 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
969 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
971 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
977 ; AVX2: # %bb.0:
982 ; AVX2-NEXT: vmovddup {{.*#+}} xmm0 = xmm3[0,0]
987 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm1[0,1],ymm2[0,1]
988 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]
996 ; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm14[4,5,6,7]
997 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5],ymm2[6,7]
1000 ; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm14[4,5,6,7]
1003 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5],ymm3[6,7]
1004 ; AVX2-NEXT: vmovddup {{.*#+}} xmm5 = xmm5[0,0]
1006 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm4[0,1],ymm8[0,1]
1007 ; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm8[0,1],ymm5[2,3],ymm8[4,5,6,7]
1013 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm9[0],ymm6[2],ymm9[2]
1018 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm10[0],ymm7[0],ymm10[2],ymm7[2]
1025 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm14[2,3],ymm4[4,5,6,7]
1027 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm11[0],ymm9[0],ymm11[2],ymm9[2]
1030 ; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],ymm11[6,7]
1037 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm12[2,3],ymm1[4,5,6,7]
1038 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm8[0],ymm13[0],ymm8[2],ymm13[2]
1041 ; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm12[6,7]
1042 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
1043 ; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],mem[6,7]
1044 ; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],mem[6,7]
1061 ; AVX2-FP: # %bb.0:
1066 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm0 = xmm3[0,0]
1071 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm1[0,1],ymm2[0,1]
1072 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]
1080 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm14[4,5,6,7]
1081 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5],ymm2[6,7]
1084 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm14[4,5,6,7]
1087 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5],ymm3[6,7]
1088 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm5 = xmm5[0,0]
1090 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm4[0,1],ymm8[0,1]
1091 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm8[0,1],ymm5[2,3],ymm8[4,5,6,7]
1097 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm9[0],ymm6[2],ymm9[2]
1102 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm10[0],ymm7[0],ymm10[2],ymm7[2]
1109 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm14[2,3],ymm4[4,5,6,7]
1111 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm11[0],ymm9[0],ymm11[2],ymm9[2]
1114 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],ymm11[6,7]
1121 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm12[2,3],ymm1[4,5,6,7]
1122 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm8[0],ymm13[0],ymm8[2],ymm13[2]
1125 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm12[6,7]
1126 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1127 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],mem[6,7]
1128 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],mem[6,7]
1145 ; AVX2-FCP: # %bb.0:
1150 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm0 = xmm3[0,0]
1155 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm1[0,1],ymm2[0,1]
1156 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7]
1164 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm14[4,5,6,7]
1165 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5],ymm2[6,7]
1168 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm14[4,5,6,7]
1171 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5],ymm3[6,7]
1172 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm5 = xmm5[0,0]
1174 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm4[0,1],ymm8[0,1]
1175 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm8[0,1],ymm5[2,3],ymm8[4,5,6,7]
1181 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm9[0],ymm6[2],ymm9[2]
1186 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm10[0],ymm7[0],ymm10[2],ymm7[2]
1193 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm14[2,3],ymm4[4,5,6,7]
1195 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm11[0],ymm9[0],ymm11[2],ymm9[2]
1198 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],ymm11[6,7]
1205 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm12[2,3],ymm1[4,5,6,7]
1206 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm8[0],ymm13[0],ymm8[2],ymm13[2]
1209 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm12[6,7]
1210 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1211 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],mem[6,7]
1212 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],mem[6,7]
1229 ; AVX512: # %bb.0:
1230 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
1237 ; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
1239 ; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,0,4,12]
1249 ; AVX512-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1252 ; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]
1257 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,9,0,4,5,6,7]
1259 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,9,4,5,6,7]
1262 ; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1265 ; AVX512-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]
1268 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,13,0,4,5,6,7]
1270 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,13,4,5,6,7]
1273 ; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1275 ; AVX512-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1279 ; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm5[4,5,6,7]
1280 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm6 = [14,0,2,3,4,5,15,0]
1282 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
1284 ; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,8,1,9,0,8,1,9]
1285 ; AVX512-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
1288 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],mem[0]
1290 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm6 {%k1}
1292 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,8,6,7]
1296 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm2
1297 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm6 = [10,0,2,3,4,5,11,0]
1299 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,10,2,3,4,5,6,11]
1301 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,12,6,7]
1313 ; AVX512-FCP: # %bb.0:
1314 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1321 ; AVX512-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
1323 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,0,4,12]
1333 ; AVX512-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1336 ; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]
1341 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,9,0,4,5,6,7]
1343 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,9,4,5,6,7]
1346 ; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1349 ; AVX512-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]
1352 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,13,0,4,5,6,7]
1354 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,13,4,5,6,7]
1357 ; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1359 ; AVX512-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1363 ; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm5[4,5,6,7]
1364 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [14,0,2,3,4,5,15,0]
1366 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
1368 ; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,8,1,9,0,8,1,9]
1369 ; AVX512-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
1372 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],mem[0]
1374 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm6 {%k1}
1376 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,8,6,7]
1380 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm2
1381 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [10,0,2,3,4,5,11,0]
1383 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,10,2,3,4,5,6,11]
1385 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,12,6,7]
1397 ; AVX512DQ: # %bb.0:
1398 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
1405 ; AVX512DQ-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
1407 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,0,4,12]
1417 ; AVX512DQ-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1420 ; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]
1425 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,9,0,4,5,6,7]
1427 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,9,4,5,6,7]
1430 ; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1433 ; AVX512DQ-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]
1436 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,13,0,4,5,6,7]
1438 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,13,4,5,6,7]
1441 ; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1443 ; AVX512DQ-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1447 ; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm5[4,5,6,7]
1448 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm6 = [14,0,2,3,4,5,15,0]
1450 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
1452 ; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,8,1,9,0,8,1,9]
1453 ; AVX512DQ-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
1456 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],mem[0]
1458 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm6 {%k1}
1460 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,8,6,7]
1464 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm2
1465 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm6 = [10,0,2,3,4,5,11,0]
1467 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,10,2,3,4,5,6,11]
1469 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,12,6,7]
1481 ; AVX512DQ-FCP: # %bb.0:
1482 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1489 ; AVX512DQ-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
1491 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,0,4,12]
1501 ; AVX512DQ-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1504 ; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]
1509 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,9,0,4,5,6,7]
1511 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,9,4,5,6,7]
1514 ; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1517 ; AVX512DQ-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]
1520 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,13,0,4,5,6,7]
1522 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,13,4,5,6,7]
1525 ; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1527 ; AVX512DQ-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1531 ; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm5[4,5,6,7]
1532 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [14,0,2,3,4,5,15,0]
1534 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
1536 ; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,8,1,9,0,8,1,9]
1537 ; AVX512DQ-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
1540 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],mem[0]
1542 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm6 {%k1}
1544 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,8,6,7]
1548 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm2
1549 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [10,0,2,3,4,5,11,0]
1551 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,10,2,3,4,5,6,11]
1553 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,12,6,7]
1565 ; AVX512BW: # %bb.0:
1566 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
1573 ; AVX512BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
1575 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,0,4,12]
1585 ; AVX512BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1588 ; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]
1593 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,9,0,4,5,6,7]
1595 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,9,4,5,6,7]
1598 ; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1601 ; AVX512BW-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]
1604 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,13,0,4,5,6,7]
1606 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,13,4,5,6,7]
1609 ; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1611 ; AVX512BW-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1615 ; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm5[4,5,6,7]
1616 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm6 = [14,0,2,3,4,5,15,0]
1618 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
1620 ; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,8,1,9,0,8,1,9]
1621 ; AVX512BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
1624 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],mem[0]
1626 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm6 {%k1}
1628 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,8,6,7]
1632 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm2
1633 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm6 = [10,0,2,3,4,5,11,0]
1635 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,10,2,3,4,5,6,11]
1637 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,12,6,7]
1649 ; AVX512BW-FCP: # %bb.0:
1650 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1657 ; AVX512BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
1659 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,0,4,12]
1669 ; AVX512BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1672 ; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]
1677 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,9,0,4,5,6,7]
1679 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,9,4,5,6,7]
1682 ; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1685 ; AVX512BW-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]
1688 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,13,0,4,5,6,7]
1690 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,13,4,5,6,7]
1693 ; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1695 ; AVX512BW-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1699 ; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm5[4,5,6,7]
1700 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [14,0,2,3,4,5,15,0]
1702 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
1704 ; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,8,1,9,0,8,1,9]
1705 ; AVX512BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
1708 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],mem[0]
1710 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm6 {%k1}
1712 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,8,6,7]
1716 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm2
1717 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [10,0,2,3,4,5,11,0]
1719 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,10,2,3,4,5,6,11]
1721 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,12,6,7]
1733 ; AVX512DQ-BW: # %bb.0:
1734 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
1741 ; AVX512DQ-BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
1743 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,0,4,12]
1753 ; AVX512DQ-BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1756 ; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]
1761 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,9,0,4,5,6,7]
1763 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,9,4,5,6,7]
1766 ; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1769 ; AVX512DQ-BW-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]
1772 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,13,0,4,5,6,7]
1774 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,13,4,5,6,7]
1777 ; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1779 ; AVX512DQ-BW-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1783 ; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm5[4,5,6,7]
1784 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm6 = [14,0,2,3,4,5,15,0]
1786 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
1788 ; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,8,1,9,0,8,1,9]
1789 ; AVX512DQ-BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
1792 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],mem[0]
1794 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm6 {%k1}
1796 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,8,6,7]
1800 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm2
1801 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm6 = [10,0,2,3,4,5,11,0]
1803 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,10,2,3,4,5,6,11]
1805 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,12,6,7]
1817 ; AVX512DQ-BW-FCP: # %bb.0:
1818 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1825 ; AVX512DQ-BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
1827 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,0,4,12]
1837 ; AVX512DQ-BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1840 ; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]
1845 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,9,0,4,5,6,7]
1847 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,9,4,5,6,7]
1850 ; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1853 ; AVX512DQ-BW-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]
1856 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,13,0,4,5,6,7]
1858 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,13,4,5,6,7]
1861 ; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1863 ; AVX512DQ-BW-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1867 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm5[4,5,6,7]
1868 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [14,0,2,3,4,5,15,0]
1870 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
1872 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,8,1,9,0,8,1,9]
1873 ; AVX512DQ-BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
1876 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],mem[0]
1878 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm6 {%k1}
1880 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,8,6,7]
1884 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm2
1885 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [10,0,2,3,4,5,11,0]
1887 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,10,2,3,4,5,6,11]
1889 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,12,6,7]
1905 %1 = shufflevector <8 x i64> %in.vec0, <8 x i64> %in.vec1, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1906 %2 = shufflevector <8 x i64> %in.vec2, <8 x i64> %in.vec3, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1907 %3 = shufflevector <8 x i64> %in.vec4, <8 x i64> %in.vec5, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1908 %4 = shufflevector <16 x i64> %1, <16 x i64> %2, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
1909 %5 = shufflevector <16 x i64> %3, <16 x i64> poison, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1910 %6 = shufflevector <32 x i64> %4, <32 x i64> %5, <48 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47>
1911 %interleaved.vec = shufflevector <48 x i64> %6, <48 x i64> poison, <48 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 7, i32 15, i32 23, i32 31, i32 39, i32 47>
1918 ; SSE: # %bb.0:
1919 ; SSE-NEXT: subq $408, %rsp # imm = 0x198
1936 ; SSE-NEXT: movlhps {{.*#+}} xmm14 = xmm14[0],xmm3[0]
1937 ; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1939 ; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1941 ; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm5[0]
1942 ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1944 ; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1946 ; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm6[0]
1947 ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1949 ; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1951 ; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm1[0]
1952 ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1954 ; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1956 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
1957 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1959 ; SSE-NEXT: movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1961 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm4[0]
1962 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1964 ; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1966 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
1967 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1969 ; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1972 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
1973 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1975 ; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1979 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
1980 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1982 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1986 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
1987 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1989 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1993 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
1994 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1996 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2000 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2001 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2003 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2007 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2010 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2014 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2015 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2017 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2021 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2022 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2024 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2028 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2029 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2034 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2035 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2040 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2041 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2046 ; SSE-NEXT: movlhps {{.*#+}} xmm13 = xmm13[0],xmm0[0]
2051 ; SSE-NEXT: movlhps {{.*#+}} xmm11 = xmm11[0],xmm0[0]
2056 ; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]
2061 ; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm1[0]
2066 ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]
2071 ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm3[0]
2073 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
2089 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2091 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2093 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2095 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2097 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2099 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2101 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2103 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2107 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2109 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2111 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2113 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2115 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2117 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2119 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2121 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2123 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2125 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2127 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2129 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2131 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2133 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2135 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2137 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2139 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2141 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2143 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2145 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2147 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2149 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2151 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2153 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2155 ; SSE-NEXT: addq $408, %rsp # imm = 0x198
2159 ; AVX: # %bb.0:
2160 ; AVX-NEXT: subq $440, %rsp # imm = 0x1B8
2163 ; AVX-NEXT: vmovddup {{.*#+}} xmm1 = mem[0,0]
2165 ; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2167 ; AVX-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2169 ; AVX-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2174 ; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm11[0,1],ymm2[2,3]
2175 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3]
2176 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2178 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2180 ; AVX-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2182 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2186 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]
2187 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5],ymm2[6,7]
2188 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2189 ; AVX-NEXT: vmovddup {{.*#+}} xmm1 = mem[0,0]
2191 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2194 ; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm0[0,1],ymm2[2,3]
2195 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3]
2196 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2198 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2201 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm2[4,5,6,7]
2203 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5],ymm1[6,7]
2204 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2209 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm5[0,1],ymm1[2,3]
2210 ; AVX-NEXT: vmovddup {{.*#+}} xmm2 = mem[0,0]
2211 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3]
2212 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2217 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm2[4,5,6,7]
2219 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5],ymm1[6,7]
2220 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2226 ; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm1[0,1],ymm2[2,3]
2227 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = mem[0,0]
2228 ; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2,3]
2229 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2234 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm2[0,1,2,3],ymm4[4,5,6,7]
2236 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1],ymm2[2,3],ymm4[4,5],ymm2[6,7]
2237 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2244 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm11[0],ymm2[2],ymm11[3]
2245 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2252 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm11[0],ymm0[2],ymm11[3]
2253 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2260 ; AVX-NEXT: vshufpd {{.*#+}} ymm5 = ymm5[0],ymm11[0],ymm5[2],ymm11[3]
2261 ; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2268 ; AVX-NEXT: vshufpd {{.*#+}} ymm11 = ymm5[0],ymm11[0],ymm5[2],ymm11[3]
2270 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm5 = xmm5[0],mem[0]
2271 ; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],mem[4,5,6,7]
2273 ; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm12[6,7]
2274 ; AVX-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2278 ; AVX-NEXT: vblendpd {{.*#+}} ymm5 = ymm5[0,1],ymm12[2,3]
2279 ; AVX-NEXT: vblendpd {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3]
2280 ; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2282 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm4 = xmm4[0],mem[0]
2283 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
2285 ; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm4[0,1,2,3,4,5],ymm5[6,7]
2289 ; AVX-NEXT: vblendpd {{.*#+}} ymm4 = ymm4[0,1],ymm14[2,3]
2290 ; AVX-NEXT: vblendpd {{.*#+}} ymm4 = ymm4[0,1,2],ymm2[3]
2292 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm2 = xmm2[0],mem[0]
2293 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
2295 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm14[6,7]
2299 ; AVX-NEXT: vblendpd {{.*#+}} ymm13 = ymm14[0,1],ymm13[2,3]
2300 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm13[0,1,2],ymm0[3]
2302 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm13 = xmm13[0],mem[0]
2303 ; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],mem[4,5,6,7]
2305 ; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]
2309 ; AVX-NEXT: vblendpd {{.*#+}} ymm12 = ymm14[0,1],ymm12[2,3]
2310 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm12[0,1,2],ymm1[3]
2311 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm12 # 16-byte Folded Reload
2312 ; AVX-NEXT: # xmm12 = xmm15[0],mem[0]
2313 ; AVX-NEXT: vmovlhps {{.*#+}} xmm9 = xmm9[0],xmm10[0]
2314 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
2315 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm10 # 16-byte Folded Reload
2316 ; AVX-NEXT: # xmm10 = xmm10[0],mem[0]
2317 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
2318 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
2319 ; AVX-NEXT: # xmm14 = xmm14[0],mem[0]
2320 ; AVX-NEXT: vmovlhps {{.*#+}} xmm7 = xmm7[0],xmm8[0]
2321 ; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm3[0],xmm6[0]
2323 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload
2324 ; AVX-NEXT: # xmm6 = xmm6[0],mem[0]
2325 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
2326 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
2327 ; AVX-NEXT: # xmm8 = xmm8[0],mem[0]
2328 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
2338 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
2340 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
2342 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
2346 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
2348 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
2352 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2354 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2358 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2360 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2362 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2364 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2366 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2368 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2370 ; AVX-NEXT: addq $440, %rsp # imm = 0x1B8
2375 ; AVX2: # %bb.0:
2376 ; AVX2-NEXT: subq $360, %rsp # imm = 0x168
2378 ; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2380 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2383 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
2385 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2387 ; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2390 ; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2392 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm5[0,1],ymm3[0,1]
2393 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1],ymm2[2,3],ymm3[4,5,6,7]
2394 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2402 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
2403 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
2404 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2405 ; AVX2-NEXT: vmovddup {{.*#+}} xmm1 = xmm0[0,0]
2407 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm4[0,1],ymm2[0,1]
2408 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]
2409 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2412 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
2416 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
2417 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2421 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm2[0,1],ymm0[0,1]
2423 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
2424 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
2425 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2430 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm12[4,5,6,7]
2432 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
2433 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2435 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2439 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm10 = ymm0[0,1],ymm10[0,1]
2441 ; AVX2-NEXT: vmovddup {{.*#+}} xmm12 = xmm0[0,0]
2442 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm10[0,1],ymm12[2,3],ymm10[4,5,6,7]
2443 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2448 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm14[4,5,6,7]
2450 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
2451 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2452 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2455 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
2456 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2457 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2459 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
2461 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
2462 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2465 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
2468 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
2472 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
2475 ; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm4[0,1,2,3,4,5],ymm5[6,7]
2477 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
2480 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7]
2487 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm3[0],ymm5[0],ymm3[2],ymm5[2]
2490 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm7[6,7]
2492 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload
2495 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm5[2,3],ymm3[4,5,6,7]
2502 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm5 = ymm7[0],ymm9[0],ymm7[2],ymm9[2]
2505 ; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm11[6,7]
2510 ; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1],ymm9[2,3],ymm7[4,5,6,7]
2517 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
2520 ; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm13[6,7]
2522 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload
2525 ; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1],ymm11[2,3],ymm10[4,5,6,7]
2529 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
2530 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],mem[6,7]
2531 ; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],mem[6,7]
2532 ; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],mem[6,7]
2533 ; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],mem[6,7]
2545 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2550 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2552 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2554 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2556 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2558 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2560 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2562 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2564 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2566 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2568 ; AVX2-NEXT: addq $360, %rsp # imm = 0x168
2573 ; AVX2-FP: # %bb.0:
2574 ; AVX2-FP-NEXT: subq $360, %rsp # imm = 0x168
2576 ; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2578 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2581 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
2583 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2585 ; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2588 ; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2590 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm5[0,1],ymm3[0,1]
2591 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1],ymm2[2,3],ymm3[4,5,6,7]
2592 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2600 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
2601 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
2602 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2603 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm1 = xmm0[0,0]
2605 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm4[0,1],ymm2[0,1]
2606 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]
2607 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2610 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
2614 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
2615 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2619 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm2[0,1],ymm0[0,1]
2621 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
2622 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
2623 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2628 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm12[4,5,6,7]
2630 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
2631 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2633 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2637 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm10 = ymm0[0,1],ymm10[0,1]
2639 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm12 = xmm0[0,0]
2640 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm10[0,1],ymm12[2,3],ymm10[4,5,6,7]
2641 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2646 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm14[4,5,6,7]
2648 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
2649 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2650 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2653 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
2654 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2655 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2657 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
2659 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
2660 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2663 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
2666 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
2670 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
2673 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm4[0,1,2,3,4,5],ymm5[6,7]
2675 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
2678 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7]
2685 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm3[0],ymm5[0],ymm3[2],ymm5[2]
2688 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm7[6,7]
2690 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload
2693 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm5[2,3],ymm3[4,5,6,7]
2700 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm5 = ymm7[0],ymm9[0],ymm7[2],ymm9[2]
2703 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm11[6,7]
2708 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1],ymm9[2,3],ymm7[4,5,6,7]
2715 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
2718 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm13[6,7]
2720 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload
2723 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1],ymm11[2,3],ymm10[4,5,6,7]
2727 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
2728 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],mem[6,7]
2729 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],mem[6,7]
2730 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],mem[6,7]
2731 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],mem[6,7]
2743 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2748 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2750 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2752 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2754 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2756 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2758 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2760 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2762 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2764 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2766 ; AVX2-FP-NEXT: addq $360, %rsp # imm = 0x168
2771 ; AVX2-FCP: # %bb.0:
2772 ; AVX2-FCP-NEXT: subq $360, %rsp # imm = 0x168
2774 ; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2776 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2779 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
2781 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2783 ; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2786 ; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2788 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm5[0,1],ymm3[0,1]
2789 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1],ymm2[2,3],ymm3[4,5,6,7]
2790 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2798 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
2799 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
2800 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2801 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm1 = xmm0[0,0]
2803 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm4[0,1],ymm2[0,1]
2804 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]
2805 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2808 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
2812 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
2813 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2817 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm2[0,1],ymm0[0,1]
2819 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
2820 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
2821 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2826 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm12[4,5,6,7]
2828 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
2829 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2831 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2835 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm10 = ymm0[0,1],ymm10[0,1]
2837 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm12 = xmm0[0,0]
2838 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm10[0,1],ymm12[2,3],ymm10[4,5,6,7]
2839 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2844 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm14[4,5,6,7]
2846 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
2847 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2848 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2851 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
2852 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2853 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2855 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
2857 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
2858 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2861 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
2864 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
2868 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
2871 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm4[0,1,2,3,4,5],ymm5[6,7]
2873 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
2876 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7]
2883 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm3[0],ymm5[0],ymm3[2],ymm5[2]
2886 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm7[6,7]
2888 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload
2891 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm5[2,3],ymm3[4,5,6,7]
2898 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm5 = ymm7[0],ymm9[0],ymm7[2],ymm9[2]
2901 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm11[6,7]
2906 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1],ymm9[2,3],ymm7[4,5,6,7]
2913 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
2916 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm13[6,7]
2918 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload
2921 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1],ymm11[2,3],ymm10[4,5,6,7]
2925 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
2926 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],mem[6,7]
2927 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],mem[6,7]
2928 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],mem[6,7]
2929 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],mem[6,7]
2941 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2946 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2948 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2950 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2952 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2954 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2956 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2958 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2960 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2962 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
2964 ; AVX2-FCP-NEXT: addq $360, %rsp # imm = 0x168
2969 ; AVX512: # %bb.0:
2970 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
2982 ; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
2985 ; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm9 = [0,0,4,12]
3001 ; AVX512-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3006 ; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm13[0,1,2,3],zmm9[4,5,6,7]
3007 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm20 = [14,0,2,3,4,5,15,0]
3009 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,14,2,3,4,5,6,15]
3012 ; AVX512-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3016 ; AVX512-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]
3022 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm23 = [0,1,13,0,4,5,6,7]
3024 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,13,4,5,6,7]
3027 ; AVX512-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3031 ; AVX512-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3]
3035 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,1,9,0,4,5,6,7]
3037 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm27 = [0,1,2,9,4,5,6,7]
3042 ; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm19[0,1,2,3],zmm28[4,5,6,7]
3055 ; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [0,8,1,9,0,8,1,9]
3056 ; AVX512-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]
3060 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm15 = xmm15[0],mem[0]
3062 ; AVX512-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm12 {%k1}
3064 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,8,6,7]
3067 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm11 = xmm21[0],mem[0]
3069 ; AVX512-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm8 {%k1}
3072 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm11 = [0,1,2,3,4,12,6,7]
3075 ; AVX512-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3080 ; AVX512-NEXT: vinserti64x4 $0, %ymm20, %zmm5, %zmm5
3081 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm20 = [10,0,2,3,4,5,11,0]
3083 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,10,2,3,4,5,6,11]
3088 ; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm2, %zmm2
3107 ; AVX512-FCP: # %bb.0:
3108 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
3120 ; AVX512-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
3123 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm9 = [0,0,4,12]
3139 ; AVX512-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3144 ; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm13[0,1,2,3],zmm9[4,5,6,7]
3145 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [14,0,2,3,4,5,15,0]
3147 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,14,2,3,4,5,6,15]
3150 ; AVX512-FCP-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3154 ; AVX512-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]
3160 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [0,1,13,0,4,5,6,7]
3162 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,13,4,5,6,7]
3165 ; AVX512-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3169 ; AVX512-FCP-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3]
3173 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,1,9,0,4,5,6,7]
3175 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [0,1,2,9,4,5,6,7]
3180 ; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm19[0,1,2,3],zmm28[4,5,6,7]
3193 ; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [0,8,1,9,0,8,1,9]
3194 ; AVX512-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]
3198 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm15 = xmm15[0],mem[0]
3200 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm12 {%k1}
3202 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,8,6,7]
3205 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm11 = xmm21[0],mem[0]
3207 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm8 {%k1}
3210 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm11 = [0,1,2,3,4,12,6,7]
3213 ; AVX512-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3218 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm20, %zmm5, %zmm5
3219 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [10,0,2,3,4,5,11,0]
3221 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,10,2,3,4,5,6,11]
3226 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm2, %zmm2
3245 ; AVX512DQ: # %bb.0:
3246 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
3258 ; AVX512DQ-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
3261 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm9 = [0,0,4,12]
3277 ; AVX512DQ-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3282 ; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm13[0,1,2,3],zmm9[4,5,6,7]
3283 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm20 = [14,0,2,3,4,5,15,0]
3285 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,14,2,3,4,5,6,15]
3288 ; AVX512DQ-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3292 ; AVX512DQ-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]
3298 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm23 = [0,1,13,0,4,5,6,7]
3300 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,13,4,5,6,7]
3303 ; AVX512DQ-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3307 ; AVX512DQ-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3]
3311 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,1,9,0,4,5,6,7]
3313 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm27 = [0,1,2,9,4,5,6,7]
3318 ; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm19[0,1,2,3],zmm28[4,5,6,7]
3331 ; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [0,8,1,9,0,8,1,9]
3332 ; AVX512DQ-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]
3336 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm15 = xmm15[0],mem[0]
3338 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm12 {%k1}
3340 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,8,6,7]
3343 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm11 = xmm21[0],mem[0]
3345 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm8 {%k1}
3348 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm11 = [0,1,2,3,4,12,6,7]
3351 ; AVX512DQ-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3356 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm20, %zmm5, %zmm5
3357 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm20 = [10,0,2,3,4,5,11,0]
3359 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,10,2,3,4,5,6,11]
3364 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm2, %zmm2
3383 ; AVX512DQ-FCP: # %bb.0:
3384 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
3396 ; AVX512DQ-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
3399 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm9 = [0,0,4,12]
3415 ; AVX512DQ-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3420 ; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm13[0,1,2,3],zmm9[4,5,6,7]
3421 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [14,0,2,3,4,5,15,0]
3423 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,14,2,3,4,5,6,15]
3426 ; AVX512DQ-FCP-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3430 ; AVX512DQ-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]
3436 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [0,1,13,0,4,5,6,7]
3438 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,13,4,5,6,7]
3441 ; AVX512DQ-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3445 ; AVX512DQ-FCP-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3]
3449 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,1,9,0,4,5,6,7]
3451 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [0,1,2,9,4,5,6,7]
3456 ; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm19[0,1,2,3],zmm28[4,5,6,7]
3469 ; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [0,8,1,9,0,8,1,9]
3470 ; AVX512DQ-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]
3474 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm15 = xmm15[0],mem[0]
3476 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm12 {%k1}
3478 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,8,6,7]
3481 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm11 = xmm21[0],mem[0]
3483 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm8 {%k1}
3486 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm11 = [0,1,2,3,4,12,6,7]
3489 ; AVX512DQ-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3494 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm20, %zmm5, %zmm5
3495 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [10,0,2,3,4,5,11,0]
3497 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,10,2,3,4,5,6,11]
3502 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm2, %zmm2
3521 ; AVX512BW: # %bb.0:
3522 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
3534 ; AVX512BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
3537 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm9 = [0,0,4,12]
3553 ; AVX512BW-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3558 ; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm13[0,1,2,3],zmm9[4,5,6,7]
3559 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm20 = [14,0,2,3,4,5,15,0]
3561 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,14,2,3,4,5,6,15]
3564 ; AVX512BW-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3568 ; AVX512BW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]
3574 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm23 = [0,1,13,0,4,5,6,7]
3576 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,13,4,5,6,7]
3579 ; AVX512BW-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3583 ; AVX512BW-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3]
3587 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,1,9,0,4,5,6,7]
3589 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm27 = [0,1,2,9,4,5,6,7]
3594 ; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm19[0,1,2,3],zmm28[4,5,6,7]
3607 ; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [0,8,1,9,0,8,1,9]
3608 ; AVX512BW-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]
3612 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm15 = xmm15[0],mem[0]
3614 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm12 {%k1}
3616 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,8,6,7]
3619 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm11 = xmm21[0],mem[0]
3621 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm8 {%k1}
3624 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm11 = [0,1,2,3,4,12,6,7]
3627 ; AVX512BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3632 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm20, %zmm5, %zmm5
3633 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm20 = [10,0,2,3,4,5,11,0]
3635 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,10,2,3,4,5,6,11]
3640 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm2, %zmm2
3659 ; AVX512BW-FCP: # %bb.0:
3660 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
3672 ; AVX512BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
3675 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm9 = [0,0,4,12]
3691 ; AVX512BW-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3696 ; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm13[0,1,2,3],zmm9[4,5,6,7]
3697 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [14,0,2,3,4,5,15,0]
3699 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,14,2,3,4,5,6,15]
3702 ; AVX512BW-FCP-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3706 ; AVX512BW-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]
3712 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [0,1,13,0,4,5,6,7]
3714 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,13,4,5,6,7]
3717 ; AVX512BW-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3721 ; AVX512BW-FCP-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3]
3725 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,1,9,0,4,5,6,7]
3727 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [0,1,2,9,4,5,6,7]
3732 ; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm19[0,1,2,3],zmm28[4,5,6,7]
3745 ; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [0,8,1,9,0,8,1,9]
3746 ; AVX512BW-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]
3750 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm15 = xmm15[0],mem[0]
3752 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm12 {%k1}
3754 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,8,6,7]
3757 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm11 = xmm21[0],mem[0]
3759 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm8 {%k1}
3762 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm11 = [0,1,2,3,4,12,6,7]
3765 ; AVX512BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3770 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm20, %zmm5, %zmm5
3771 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [10,0,2,3,4,5,11,0]
3773 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,10,2,3,4,5,6,11]
3778 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm2, %zmm2
3797 ; AVX512DQ-BW: # %bb.0:
3798 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
3810 ; AVX512DQ-BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
3813 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm9 = [0,0,4,12]
3829 ; AVX512DQ-BW-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3834 ; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm13[0,1,2,3],zmm9[4,5,6,7]
3835 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm20 = [14,0,2,3,4,5,15,0]
3837 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,14,2,3,4,5,6,15]
3840 ; AVX512DQ-BW-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3844 ; AVX512DQ-BW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]
3850 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm23 = [0,1,13,0,4,5,6,7]
3852 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,13,4,5,6,7]
3855 ; AVX512DQ-BW-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3859 ; AVX512DQ-BW-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3]
3863 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,1,9,0,4,5,6,7]
3865 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm27 = [0,1,2,9,4,5,6,7]
3870 ; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm19[0,1,2,3],zmm28[4,5,6,7]
3883 ; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [0,8,1,9,0,8,1,9]
3884 ; AVX512DQ-BW-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]
3888 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm15 = xmm15[0],mem[0]
3890 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm12 {%k1}
3892 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,8,6,7]
3895 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm11 = xmm21[0],mem[0]
3897 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm8 {%k1}
3900 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm11 = [0,1,2,3,4,12,6,7]
3903 ; AVX512DQ-BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3908 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm20, %zmm5, %zmm5
3909 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm20 = [10,0,2,3,4,5,11,0]
3911 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,10,2,3,4,5,6,11]
3916 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm2, %zmm2
3935 ; AVX512DQ-BW-FCP: # %bb.0:
3936 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
3948 ; AVX512DQ-BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
3951 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm9 = [0,0,4,12]
3967 ; AVX512DQ-BW-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3972 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm13[0,1,2,3],zmm9[4,5,6,7]
3973 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [14,0,2,3,4,5,15,0]
3975 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,14,2,3,4,5,6,15]
3978 ; AVX512DQ-BW-FCP-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3982 ; AVX512DQ-BW-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]
3988 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [0,1,13,0,4,5,6,7]
3990 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,13,4,5,6,7]
3993 ; AVX512DQ-BW-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3997 ; AVX512DQ-BW-FCP-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3]
4001 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,1,9,0,4,5,6,7]
4003 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [0,1,2,9,4,5,6,7]
4008 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm19[0,1,2,3],zmm28[4,5,6,7]
4021 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm20 = [0,8,1,9,0,8,1,9]
4022 ; AVX512DQ-BW-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3]
4026 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm15 = xmm15[0],mem[0]
4028 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm15, %zmm0, %zmm12 {%k1}
4030 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,8,6,7]
4033 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm11 = xmm21[0],mem[0]
4035 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm0, %zmm8 {%k1}
4038 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm11 = [0,1,2,3,4,12,6,7]
4041 ; AVX512DQ-BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4046 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm20, %zmm5, %zmm5
4047 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [10,0,2,3,4,5,11,0]
4049 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,10,2,3,4,5,6,11]
4054 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm2, %zmm2
4077 %1 = shufflevector <16 x i64> %in.vec0, <16 x i64> %in.vec1, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
4078 %2 = shufflevector <16 x i64> %in.vec2, <16 x i64> %in.vec3, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
4079 %3 = shufflevector <16 x i64> %in.vec4, <16 x i64> %in.vec5, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
4080 %4 = shufflevector <32 x i64> %1, <32 x i64> %2, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
4081 %5 = shufflevector <32 x i64> %3, <32 x i64> poison, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
4082 %6 = shufflevector <64 x i64> %4, <64 x i64> %5, <96 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95>
4083 %interleaved.vec = shufflevector <96 x i64> %6, <96 x i64> poison, <96 x i32> <i32 0, i32 16, i32 32, i32 48, i32 64, i32 80, i32 1, i32 17, i32 33, i32 49, i32 65, i32 81, i32 2, i32 18, i32 34, i32 50, i32 66, i32 82, i32 3, i32 19, i32 35, i32 51, i32 67, i32 83, i32 4, i32 20, i32 36, i32 52, i32 68, i32 84, i32 5, i32 21, i32 37, i32 53, i32 69, i32 85, i32 6, i32 22, i32 38, i32 54, i32 70, i32 86, i32 7, i32 23, i32 39, i32 55, i32 71, i32 87, i32 8, i32 24, i32 40, i32 56, i32 72, i32 88, i32 9, i32 25, i32 41, i32 57, i32 73, i32 89, i32 10, i32 26, i32 42, i32 58, i32 74, i32 90, i32 11, i32 27, i32 43, i32 59, i32 75, i32 91, i32 12, i32 28, i32 44, i32 60, i32 76, i32 92, i32 13, i32 29, i32 45, i32 61, i32 77, i32 93, i32 14, i32 30, i32 46, i32 62, i32 78, i32 94, i32 15, i32 31, i32 47, i32 63, i32 79, i32 95>
4090 ; SSE: # %bb.0:
4091 ; SSE-NEXT: subq $1176, %rsp # imm = 0x498
4108 ; SSE-NEXT: movlhps {{.*#+}} xmm15 = xmm15[0],xmm2[0]
4109 ; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4111 ; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4113 ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm4[0]
4114 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4116 ; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4118 ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm6[0]
4119 ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4121 ; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4123 ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]
4124 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4126 ; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4128 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]
4129 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4131 ; SSE-NEXT: movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4133 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm5[0]
4134 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4136 ; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4138 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4139 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4141 ; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4144 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4145 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4147 ; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4151 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4152 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4154 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4158 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4159 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4161 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4165 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4166 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4168 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4172 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4173 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4175 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4179 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4180 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4182 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4186 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4187 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4189 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4193 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4194 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4196 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4200 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4201 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4203 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4207 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4208 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4210 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4214 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4215 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4217 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4221 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4222 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4224 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4228 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4229 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4231 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4235 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4236 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4238 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4242 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4243 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4245 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4249 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4250 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4252 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4256 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4257 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4259 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4263 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4264 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4266 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4270 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4271 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4273 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4277 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4278 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4280 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4284 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4285 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4287 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4291 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4292 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4294 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4298 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4299 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4301 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4305 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4306 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4308 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4312 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4313 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4315 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4319 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4320 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4322 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4326 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4327 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4329 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4333 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4334 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4336 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4340 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4341 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4343 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4347 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4350 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4354 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4355 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4357 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4361 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4362 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4364 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4368 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4369 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4374 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4375 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4380 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
4381 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4386 ; SSE-NEXT: movlhps {{.*#+}} xmm14 = xmm14[0],xmm0[0]
4391 ; SSE-NEXT: movlhps {{.*#+}} xmm10 = xmm10[0],xmm0[0]
4396 ; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]
4401 ; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm1[0]
4406 ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]
4411 ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm3[0]
4413 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
4429 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4431 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4433 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4435 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4437 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4439 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4441 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4443 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4447 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4449 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4451 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4453 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4455 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4457 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4459 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4461 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4463 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4465 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4467 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4469 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4471 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4473 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4475 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4477 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4479 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4481 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4483 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4485 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4487 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4489 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4491 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4493 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4495 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4497 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4499 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4501 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4503 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4505 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4507 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4509 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4511 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4513 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4515 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4517 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4519 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4521 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4523 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4525 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4527 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4529 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4531 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4533 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4535 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4537 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4539 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4541 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4543 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4545 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4547 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4549 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4551 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4553 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4555 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4557 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4559 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4561 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4563 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4565 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4567 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4569 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4571 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4573 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4575 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4577 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4579 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4581 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4583 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4585 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4587 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4589 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4591 ; SSE-NEXT: addq $1176, %rsp # imm = 0x498
4595 ; AVX: # %bb.0:
4596 ; AVX-NEXT: subq $1608, %rsp # imm = 0x648
4599 ; AVX-NEXT: vmovddup {{.*#+}} xmm2 = mem[0,0]
4601 ; AVX-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4605 ; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4608 ; AVX-NEXT: vblendpd {{.*#+}} ymm3 = ymm0[0,1],ymm3[2,3]
4609 ; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3]
4610 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4612 ; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4614 ; AVX-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4616 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4620 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
4621 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5],ymm3[6,7]
4622 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4623 ; AVX-NEXT: vmovddup {{.*#+}} xmm2 = mem[0,0]
4625 ; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4628 ; AVX-NEXT: vblendpd {{.*#+}} ymm3 = ymm1[0,1],ymm3[2,3]
4629 ; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3]
4630 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4632 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4635 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm2[0,1,2,3],ymm3[4,5,6,7]
4637 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1],ymm2[2,3],ymm3[4,5],ymm2[6,7]
4638 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4643 ; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm3[0,1],ymm2[2,3]
4644 ; AVX-NEXT: vmovddup {{.*#+}} xmm4 = mem[0,0]
4645 ; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2,3]
4646 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4648 ; AVX-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4650 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4653 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm2[0,1,2,3],ymm4[4,5,6,7]
4655 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1],ymm2[2,3],ymm4[4,5],ymm2[6,7]
4656 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4658 ; AVX-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4660 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4664 ; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm5[0,1],ymm2[2,3]
4665 ; AVX-NEXT: vmovddup {{.*#+}} xmm4 = mem[0,0]
4666 ; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2,3]
4667 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4669 ; AVX-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4671 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4674 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm2[0,1,2,3],ymm4[4,5,6,7]
4676 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1],ymm2[2,3],ymm4[4,5],ymm2[6,7]
4677 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4680 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4684 ; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm8[0,1],ymm2[2,3]
4685 ; AVX-NEXT: vmovddup {{.*#+}} xmm4 = mem[0,0]
4686 ; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2,3]
4687 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4689 ; AVX-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4691 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4694 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm2[0,1,2,3],ymm4[4,5,6,7]
4696 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1],ymm2[2,3],ymm4[4,5],ymm2[6,7]
4697 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4699 ; AVX-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4701 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4705 ; AVX-NEXT: vblendpd {{.*#+}} ymm4 = ymm2[0,1],ymm4[2,3]
4706 ; AVX-NEXT: vmovddup {{.*#+}} xmm6 = mem[0,0]
4707 ; AVX-NEXT: vblendpd {{.*#+}} ymm4 = ymm4[0],ymm6[1],ymm4[2,3]
4708 ; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4712 ; AVX-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4715 ; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm4[0,1,2,3],ymm6[4,5,6,7]
4717 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm6[0,1],ymm4[2,3],ymm6[4,5],ymm4[6,7]
4718 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4720 ; AVX-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4722 ; AVX-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4726 ; AVX-NEXT: vblendpd {{.*#+}} ymm6 = ymm4[0,1],ymm6[2,3]
4727 ; AVX-NEXT: vmovddup {{.*#+}} xmm7 = mem[0,0]
4728 ; AVX-NEXT: vblendpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2,3]
4729 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4731 ; AVX-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4733 ; AVX-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4736 ; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm6[0,1,2,3],ymm7[4,5,6,7]
4738 ; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1],ymm6[2,3],ymm7[4,5],ymm6[6,7]
4739 ; AVX-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4741 ; AVX-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4743 ; AVX-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4747 ; AVX-NEXT: vblendpd {{.*#+}} ymm6 = ymm9[0,1],ymm6[2,3]
4748 ; AVX-NEXT: vmovddup {{.*#+}} xmm7 = mem[0,0]
4749 ; AVX-NEXT: vblendpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2,3]
4750 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4752 ; AVX-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4754 ; AVX-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4757 ; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm6[0,1,2,3],ymm7[4,5,6,7]
4759 ; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1],ymm6[2,3],ymm7[4,5],ymm6[6,7]
4760 ; AVX-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4767 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[2],ymm7[3]
4768 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4775 ; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm6[0],ymm1[2],ymm6[3]
4776 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4783 ; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[2],ymm3[3]
4784 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4791 ; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[2],ymm3[3]
4792 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4799 ; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm7[0],ymm1[2],ymm7[3]
4800 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4807 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm7[0],ymm2[2],ymm7[3]
4808 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4815 ; AVX-NEXT: vshufpd {{.*#+}} ymm4 = ymm4[0],ymm7[0],ymm4[2],ymm7[3]
4816 ; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4823 ; AVX-NEXT: vshufpd {{.*#+}} ymm7 = ymm8[0],ymm7[0],ymm8[2],ymm7[3]
4824 ; AVX-NEXT: vmovupd %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4826 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm7 = xmm7[0],mem[0]
4827 ; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],mem[4,5,6,7]
4829 ; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],ymm8[6,7]
4830 ; AVX-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4834 ; AVX-NEXT: vblendpd {{.*#+}} ymm7 = ymm7[0,1],ymm8[2,3]
4835 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm7[0,1,2],ymm0[3]
4836 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4838 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
4839 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
4841 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm7[6,7]
4842 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4846 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm7[2,3]
4847 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm12[3]
4848 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4850 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
4851 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
4853 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm7[6,7]
4854 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4858 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm7[2,3]
4859 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm6[3]
4860 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4862 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
4863 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
4865 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm6[6,7]
4866 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4870 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm6[2,3]
4871 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm5[3]
4872 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4874 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
4875 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
4877 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm5[6,7]
4878 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4882 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm5[2,3]
4883 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm3[3]
4884 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4886 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
4887 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
4889 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7]
4890 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4894 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm3[2,3]
4895 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3]
4896 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4898 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
4899 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
4901 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
4902 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4906 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm15[2,3]
4907 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3]
4908 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4910 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
4911 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
4913 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm15[6,7]
4914 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4918 ; AVX-NEXT: vblendpd {{.*#+}} ymm12 = ymm15[0,1],ymm12[2,3]
4919 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm12[0,1,2],ymm4[3]
4920 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4921 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4922 ; AVX-NEXT: vmovlhps {{.*#+}} xmm14 = xmm0[0],xmm14[0]
4923 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4924 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm15 # 16-byte Folded Reload
4925 ; AVX-NEXT: # xmm15 = xmm0[0],mem[0]
4926 ; AVX-NEXT: vmovlhps {{.*#+}} xmm13 = xmm13[0],xmm11[0]
4927 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4928 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm12 # 16-byte Folded Reload
4929 ; AVX-NEXT: # xmm12 = xmm0[0],mem[0]
4930 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4931 ; AVX-NEXT: vmovlhps {{.*#+}} xmm11 = xmm0[0],xmm10[0]
4932 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4933 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm10 # 16-byte Folded Reload
4934 ; AVX-NEXT: # xmm10 = xmm0[0],mem[0]
4935 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4936 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm9 # 16-byte Folded Reload
4937 ; AVX-NEXT: # xmm9 = xmm0[0],mem[0]
4938 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4939 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm8 # 16-byte Folded Reload
4940 ; AVX-NEXT: # xmm8 = xmm0[0],mem[0]
4941 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4942 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm7 # 16-byte Folded Reload
4943 ; AVX-NEXT: # xmm7 = xmm0[0],mem[0]
4944 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4946 ; AVX-NEXT: # xmm6 = xmm0[0],mem[0]
4947 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4948 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload
4949 ; AVX-NEXT: # xmm5 = xmm0[0],mem[0]
4950 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4951 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm4 # 16-byte Folded Reload
4952 ; AVX-NEXT: # xmm4 = xmm0[0],mem[0]
4953 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4954 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload
4955 ; AVX-NEXT: # xmm3 = xmm0[0],mem[0]
4956 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4957 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload
4958 ; AVX-NEXT: # xmm2 = xmm0[0],mem[0]
4959 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4960 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload
4961 ; AVX-NEXT: # xmm1 = xmm0[0],mem[0]
4962 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4963 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
4964 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
4965 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
4982 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
4984 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
4986 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
4988 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
4990 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
4992 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
4994 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
4996 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
4998 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5000 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5002 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5004 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5006 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5008 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5010 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5012 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5014 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5016 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5018 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5020 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5022 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5024 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5026 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5028 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5030 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5032 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5034 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5036 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5038 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5040 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5042 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5044 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5046 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5048 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5050 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5052 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5054 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5056 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5058 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5060 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5062 ; AVX-NEXT: addq $1608, %rsp # imm = 0x648
5067 ; AVX2: # %bb.0:
5068 ; AVX2-NEXT: subq $1208, %rsp # imm = 0x4B8
5070 ; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5072 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5075 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
5077 ; AVX2-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5079 ; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5081 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5083 ; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5085 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm5[0,1],ymm3[0,1]
5086 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1],ymm2[2,3],ymm3[4,5,6,7]
5087 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5090 ; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5092 ; AVX2-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5094 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5096 ; AVX2-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5099 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
5100 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
5101 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5102 ; AVX2-NEXT: vmovddup {{.*#+}} xmm1 = xmm0[0,0]
5104 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm4[0,1],ymm2[0,1]
5105 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]
5106 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5109 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
5111 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5113 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
5114 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5116 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5118 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5120 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm2[0,1],ymm0[0,1]
5122 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
5123 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
5124 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5126 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5128 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5131 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
5133 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
5134 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5136 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5138 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5140 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5142 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
5144 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
5145 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
5146 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5148 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5150 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5153 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
5155 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
5156 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5158 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5160 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5162 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5164 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
5166 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
5167 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
5168 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5171 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5174 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
5176 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
5177 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5179 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5183 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
5185 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
5186 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
5187 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5192 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
5194 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
5195 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5197 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5201 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
5203 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
5204 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
5205 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5210 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
5212 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
5213 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5215 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5219 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[0,1],ymm0[0,1]
5221 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm0[0,0]
5222 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7]
5223 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5228 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm15[4,5,6,7]
5230 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
5231 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5232 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5233 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
5234 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
5235 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
5236 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5237 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5238 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5239 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
5240 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
5241 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
5242 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5243 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5244 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5245 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
5246 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
5247 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
5248 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5249 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5250 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5251 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
5252 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
5253 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
5254 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5255 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5256 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5258 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
5259 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
5260 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5261 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5264 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5265 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5268 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5269 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5272 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5273 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5277 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
5280 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
5281 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5283 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
5286 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
5287 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5291 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5295 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
5298 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
5299 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5301 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
5304 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
5305 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5309 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5313 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
5316 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
5317 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5319 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
5322 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
5323 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5330 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
5333 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
5334 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5336 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
5339 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
5340 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5347 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
5350 ; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm3[0,1,2,3,4,5],ymm4[6,7]
5352 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
5355 ; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
5362 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
5365 ; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm4[0,1,2,3,4,5],ymm5[6,7]
5367 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
5370 ; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
5377 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5380 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm12[6,7]
5382 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
5385 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
5392 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm2[0],ymm12[0],ymm2[2],ymm12[2]
5395 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
5397 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload
5400 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm12[2,3],ymm2[4,5,6,7]
5404 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
5405 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
5406 ; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],mem[6,7]
5407 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
5408 ; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3,4,5],mem[6,7]
5409 ; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],mem[6,7]
5410 ; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],mem[6,7]
5411 ; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],mem[6,7]
5412 ; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],mem[6,7]
5413 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]
5414 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
5418 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5423 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5428 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5433 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5436 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5438 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5440 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5443 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5445 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5447 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5450 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5452 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5454 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5457 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5459 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5461 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5463 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5465 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5467 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5469 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5471 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5473 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5475 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5477 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5479 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5481 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5483 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5485 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5487 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5489 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5491 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5493 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5495 ; AVX2-NEXT: addq $1208, %rsp # imm = 0x4B8
5500 ; AVX2-FP: # %bb.0:
5501 ; AVX2-FP-NEXT: subq $1208, %rsp # imm = 0x4B8
5503 ; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5505 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5508 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
5510 ; AVX2-FP-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5512 ; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5514 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5516 ; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5518 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm5[0,1],ymm3[0,1]
5519 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1],ymm2[2,3],ymm3[4,5,6,7]
5520 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5523 ; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5525 ; AVX2-FP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5527 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5529 ; AVX2-FP-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5532 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
5533 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
5534 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5535 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm1 = xmm0[0,0]
5537 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm4[0,1],ymm2[0,1]
5538 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]
5539 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5542 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
5544 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5546 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
5547 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5549 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5551 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5553 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm2[0,1],ymm0[0,1]
5555 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
5556 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
5557 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5559 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5561 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5564 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
5566 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
5567 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5569 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5571 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5573 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5575 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
5577 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
5578 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
5579 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5581 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5583 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5586 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
5588 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
5589 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5591 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5593 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5595 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5597 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
5599 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
5600 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
5601 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5604 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5607 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
5609 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
5610 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5612 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5616 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
5618 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
5619 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
5620 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5625 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
5627 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
5628 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5630 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5634 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
5636 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
5637 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
5638 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5643 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
5645 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
5646 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5648 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5652 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[0,1],ymm0[0,1]
5654 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm0[0,0]
5655 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7]
5656 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5661 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm15[4,5,6,7]
5663 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
5664 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5665 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5666 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
5667 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
5668 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
5669 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5670 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5671 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5672 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
5673 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
5674 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
5675 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5676 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5677 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5678 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
5679 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
5680 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
5681 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5682 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5683 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5684 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
5685 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
5686 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
5687 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5688 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5689 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5691 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
5692 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
5693 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5694 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5697 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5698 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5701 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5702 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5705 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5706 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5710 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
5713 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
5714 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5716 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
5719 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
5720 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5724 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5728 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
5731 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
5732 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5734 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
5737 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
5738 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5742 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5746 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
5749 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
5750 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5752 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
5755 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
5756 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5763 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
5766 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
5767 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5769 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
5772 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
5773 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5780 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
5783 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm3[0,1,2,3,4,5],ymm4[6,7]
5785 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
5788 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
5795 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
5798 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm4[0,1,2,3,4,5],ymm5[6,7]
5800 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
5803 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
5810 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
5813 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm12[6,7]
5815 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
5818 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
5825 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm2[0],ymm12[0],ymm2[2],ymm12[2]
5828 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
5830 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload
5833 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm12[2,3],ymm2[4,5,6,7]
5837 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
5838 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
5839 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],mem[6,7]
5840 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
5841 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3,4,5],mem[6,7]
5842 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],mem[6,7]
5843 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],mem[6,7]
5844 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],mem[6,7]
5845 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],mem[6,7]
5846 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]
5847 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
5851 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5856 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5861 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5866 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5869 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5871 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5873 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5876 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5878 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5880 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5883 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5885 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5887 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5890 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5892 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5894 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5896 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5898 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5900 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5902 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5904 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5906 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5908 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5910 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5912 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5914 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5916 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5918 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5920 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5922 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5924 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5926 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
5928 ; AVX2-FP-NEXT: addq $1208, %rsp # imm = 0x4B8
5933 ; AVX2-FCP: # %bb.0:
5934 ; AVX2-FCP-NEXT: subq $1208, %rsp # imm = 0x4B8
5936 ; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5938 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5941 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
5943 ; AVX2-FCP-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5945 ; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5947 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5949 ; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5951 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm5[0,1],ymm3[0,1]
5952 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1],ymm2[2,3],ymm3[4,5,6,7]
5953 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5956 ; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5958 ; AVX2-FCP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5960 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5962 ; AVX2-FCP-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5965 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
5966 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
5967 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5968 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm1 = xmm0[0,0]
5970 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm4[0,1],ymm2[0,1]
5971 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]
5972 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5975 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
5977 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5979 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
5980 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5982 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5984 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5986 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm2[0,1],ymm0[0,1]
5988 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
5989 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
5990 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5992 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5994 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5997 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
5999 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6000 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6002 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6004 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6006 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6008 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
6010 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
6011 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
6012 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6014 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6016 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6019 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
6021 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6022 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6024 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6026 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6028 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6030 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
6032 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
6033 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
6034 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6037 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6040 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
6042 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6043 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6045 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6049 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
6051 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
6052 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
6053 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6058 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
6060 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6061 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6063 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6067 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
6069 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
6070 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
6071 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6076 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
6078 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6079 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6081 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6085 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[0,1],ymm0[0,1]
6087 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm0[0,0]
6088 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7]
6089 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6094 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm15[4,5,6,7]
6096 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
6097 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6098 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6099 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
6100 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
6101 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
6102 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
6103 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6104 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6105 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
6106 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
6107 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
6108 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
6109 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6110 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6111 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
6112 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
6113 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
6114 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
6115 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6116 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6117 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
6118 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
6119 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
6120 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
6121 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6122 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6124 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
6125 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
6126 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
6127 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6130 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
6131 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6134 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
6135 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6138 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
6139 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6143 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
6146 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
6147 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6149 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
6152 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
6153 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6157 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6161 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
6164 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
6165 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6167 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
6170 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
6171 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6175 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6179 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
6182 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
6183 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6185 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
6188 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
6189 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6196 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
6199 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
6200 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6202 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
6205 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
6206 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6213 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
6216 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm3[0,1,2,3,4,5],ymm4[6,7]
6218 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
6221 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
6228 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
6231 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm4[0,1,2,3,4,5],ymm5[6,7]
6233 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
6236 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
6243 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
6246 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm12[6,7]
6248 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
6251 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
6258 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm2[0],ymm12[0],ymm2[2],ymm12[2]
6261 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
6263 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload
6266 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm12[2,3],ymm2[4,5,6,7]
6270 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
6271 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
6272 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],mem[6,7]
6273 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
6274 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3,4,5],mem[6,7]
6275 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],mem[6,7]
6276 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],mem[6,7]
6277 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],mem[6,7]
6278 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],mem[6,7]
6279 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]
6280 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
6284 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6289 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6294 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6299 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6302 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6304 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6306 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6309 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6311 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6313 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6316 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6318 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6320 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6323 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6325 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6327 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6329 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6331 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6333 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6335 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6337 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6339 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6341 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6343 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6345 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6347 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6349 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6351 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6353 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6355 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6357 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6359 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6361 ; AVX2-FCP-NEXT: addq $1208, %rsp # imm = 0x4B8
6366 ; AVX512: # %bb.0:
6367 ; AVX512-NEXT: subq $648, %rsp # imm = 0x288
6384 ; AVX512-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
6387 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6390 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6393 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6396 ; AVX512-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
6399 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6401 ; AVX512-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
6404 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6406 ; AVX512-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
6409 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6412 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6415 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6418 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6426 ; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm1 = [0,8,1,9,0,8,1,9]
6427 ; AVX512-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3]
6432 ; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6436 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6446 ; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
6455 ; AVX512-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]
6459 ; AVX512-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]
6464 ; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
6483 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
6485 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6486 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
6488 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6489 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
6494 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
6496 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
6498 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
6499 ; AVX512-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
6500 ; AVX512-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
6501 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
6503 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
6506 ; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm22[0,1,2,3],zmm26[4,5,6,7]
6512 ; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm28[0,1,2,3],zmm17[4,5,6,7]
6515 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,0,4,5,6,7]
6520 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,9,4,5,6,7]
6529 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,13,0,4,5,6,7]
6534 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,13,4,5,6,7]
6539 ; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm16[0,1,2,3],zmm10[4,5,6,7]
6540 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm5 = [14,0,2,3,4,5,15,0]
6545 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
6551 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
6553 ; AVX512-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm11 {%k1}
6555 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
6557 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
6558 ; AVX512-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm10 {%k1}
6560 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
6562 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
6563 ; AVX512-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm16 {%k1}
6565 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
6567 ; AVX512-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm19 {%k1}
6569 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,3,4,8,6,7]
6580 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm19, %zmm6
6584 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm19 = [10,0,2,3,4,5,11,0]
6588 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm24, %zmm0
6593 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm7, %zmm1
6594 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
6599 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm21, %zmm2
6600 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
6603 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm19 = [0,10,2,3,4,5,6,11]
6605 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,1,2,3,4,12,6,7]
6613 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
6638 ; AVX512-NEXT: addq $648, %rsp # imm = 0x288
6643 ; AVX512-FCP: # %bb.0:
6644 ; AVX512-FCP-NEXT: subq $648, %rsp # imm = 0x288
6661 ; AVX512-FCP-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
6664 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6667 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6670 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6673 ; AVX512-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
6676 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6678 ; AVX512-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
6681 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6683 ; AVX512-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
6686 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6689 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6692 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6695 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6703 ; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm1 = [0,8,1,9,0,8,1,9]
6704 ; AVX512-FCP-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3]
6709 ; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6713 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6723 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
6732 ; AVX512-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]
6736 ; AVX512-FCP-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]
6741 ; AVX512-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
6760 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
6762 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6763 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
6765 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6766 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
6771 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
6773 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
6775 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
6776 ; AVX512-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
6777 ; AVX512-FCP-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
6778 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
6780 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
6783 ; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm22[0,1,2,3],zmm26[4,5,6,7]
6789 ; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm28[0,1,2,3],zmm17[4,5,6,7]
6792 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,0,4,5,6,7]
6797 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,9,4,5,6,7]
6806 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,13,0,4,5,6,7]
6811 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,13,4,5,6,7]
6816 ; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm16[0,1,2,3],zmm10[4,5,6,7]
6817 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [14,0,2,3,4,5,15,0]
6822 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
6828 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
6830 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm11 {%k1}
6832 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
6834 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
6835 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm10 {%k1}
6837 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
6839 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
6840 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm16 {%k1}
6842 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
6844 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm19 {%k1}
6846 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,3,4,8,6,7]
6857 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm19, %zmm6
6861 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [10,0,2,3,4,5,11,0]
6865 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm24, %zmm0
6870 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm7, %zmm1
6871 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
6876 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm21, %zmm2
6877 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
6880 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [0,10,2,3,4,5,6,11]
6882 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,1,2,3,4,12,6,7]
6890 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
6915 ; AVX512-FCP-NEXT: addq $648, %rsp # imm = 0x288
6920 ; AVX512DQ: # %bb.0:
6921 ; AVX512DQ-NEXT: subq $648, %rsp # imm = 0x288
6938 ; AVX512DQ-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
6941 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6944 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6947 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6950 ; AVX512DQ-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
6953 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6955 ; AVX512DQ-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
6958 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6960 ; AVX512DQ-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
6963 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6966 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6969 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6972 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6980 ; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm1 = [0,8,1,9,0,8,1,9]
6981 ; AVX512DQ-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3]
6986 ; AVX512DQ-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
6990 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7000 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
7009 ; AVX512DQ-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]
7013 ; AVX512DQ-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]
7018 ; AVX512DQ-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
7037 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7039 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7040 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7042 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7043 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
7048 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7050 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7052 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7053 ; AVX512DQ-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
7054 ; AVX512DQ-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
7055 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7057 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7060 ; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm22[0,1,2,3],zmm26[4,5,6,7]
7066 ; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm28[0,1,2,3],zmm17[4,5,6,7]
7069 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,0,4,5,6,7]
7074 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,9,4,5,6,7]
7083 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,13,0,4,5,6,7]
7088 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,13,4,5,6,7]
7093 ; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm16[0,1,2,3],zmm10[4,5,6,7]
7094 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm5 = [14,0,2,3,4,5,15,0]
7099 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
7105 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7107 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm11 {%k1}
7109 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7111 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
7112 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm10 {%k1}
7114 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7116 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
7117 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm16 {%k1}
7119 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7121 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm19 {%k1}
7123 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,3,4,8,6,7]
7134 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm19, %zmm6
7138 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm19 = [10,0,2,3,4,5,11,0]
7142 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm24, %zmm0
7147 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm7, %zmm1
7148 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
7153 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm21, %zmm2
7154 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
7157 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm19 = [0,10,2,3,4,5,6,11]
7159 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,1,2,3,4,12,6,7]
7167 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
7192 ; AVX512DQ-NEXT: addq $648, %rsp # imm = 0x288
7197 ; AVX512DQ-FCP: # %bb.0:
7198 ; AVX512DQ-FCP-NEXT: subq $648, %rsp # imm = 0x288
7215 ; AVX512DQ-FCP-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
7218 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7221 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7224 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7227 ; AVX512DQ-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
7230 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7232 ; AVX512DQ-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
7235 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7237 ; AVX512DQ-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
7240 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7243 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7246 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7249 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7257 ; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm1 = [0,8,1,9,0,8,1,9]
7258 ; AVX512DQ-FCP-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3]
7263 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7267 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7277 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
7286 ; AVX512DQ-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]
7290 ; AVX512DQ-FCP-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]
7295 ; AVX512DQ-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
7314 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7316 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7317 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7319 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7320 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
7325 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7327 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7329 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7330 ; AVX512DQ-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
7331 ; AVX512DQ-FCP-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
7332 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7334 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7337 ; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm22[0,1,2,3],zmm26[4,5,6,7]
7343 ; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm28[0,1,2,3],zmm17[4,5,6,7]
7346 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,0,4,5,6,7]
7351 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,9,4,5,6,7]
7360 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,13,0,4,5,6,7]
7365 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,13,4,5,6,7]
7370 ; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm16[0,1,2,3],zmm10[4,5,6,7]
7371 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [14,0,2,3,4,5,15,0]
7376 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
7382 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7384 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm11 {%k1}
7386 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7388 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
7389 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm10 {%k1}
7391 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7393 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
7394 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm16 {%k1}
7396 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7398 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm19 {%k1}
7400 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,3,4,8,6,7]
7411 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm19, %zmm6
7415 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [10,0,2,3,4,5,11,0]
7419 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm24, %zmm0
7424 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm7, %zmm1
7425 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
7430 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm21, %zmm2
7431 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
7434 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [0,10,2,3,4,5,6,11]
7436 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,1,2,3,4,12,6,7]
7444 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
7469 ; AVX512DQ-FCP-NEXT: addq $648, %rsp # imm = 0x288
7474 ; AVX512BW: # %bb.0:
7475 ; AVX512BW-NEXT: subq $648, %rsp # imm = 0x288
7492 ; AVX512BW-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
7495 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7498 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7501 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7504 ; AVX512BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
7507 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7509 ; AVX512BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
7512 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7514 ; AVX512BW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
7517 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7520 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7523 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7526 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7534 ; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm1 = [0,8,1,9,0,8,1,9]
7535 ; AVX512BW-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3]
7540 ; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7544 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7554 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
7563 ; AVX512BW-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]
7567 ; AVX512BW-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]
7572 ; AVX512BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
7591 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7593 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7594 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7596 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7597 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
7602 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7604 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7606 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7607 ; AVX512BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
7608 ; AVX512BW-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
7609 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7611 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7614 ; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm22[0,1,2,3],zmm26[4,5,6,7]
7620 ; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm28[0,1,2,3],zmm17[4,5,6,7]
7623 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,0,4,5,6,7]
7628 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,9,4,5,6,7]
7637 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,13,0,4,5,6,7]
7642 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,13,4,5,6,7]
7647 ; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm16[0,1,2,3],zmm10[4,5,6,7]
7648 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [14,0,2,3,4,5,15,0]
7653 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
7659 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7661 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm11 {%k1}
7663 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7665 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
7666 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm10 {%k1}
7668 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7670 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
7671 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm16 {%k1}
7673 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7675 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm19 {%k1}
7677 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,3,4,8,6,7]
7688 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm19, %zmm6
7692 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm19 = [10,0,2,3,4,5,11,0]
7696 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm24, %zmm0
7701 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm7, %zmm1
7702 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
7707 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm21, %zmm2
7708 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
7711 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm19 = [0,10,2,3,4,5,6,11]
7713 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,1,2,3,4,12,6,7]
7721 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
7746 ; AVX512BW-NEXT: addq $648, %rsp # imm = 0x288
7751 ; AVX512BW-FCP: # %bb.0:
7752 ; AVX512BW-FCP-NEXT: subq $648, %rsp # imm = 0x288
7769 ; AVX512BW-FCP-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
7772 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7775 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7778 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7781 ; AVX512BW-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
7784 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7786 ; AVX512BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
7789 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7791 ; AVX512BW-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
7794 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7797 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7800 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7803 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7811 ; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm1 = [0,8,1,9,0,8,1,9]
7812 ; AVX512BW-FCP-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3]
7817 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7821 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7831 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
7840 ; AVX512BW-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]
7844 ; AVX512BW-FCP-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]
7849 ; AVX512BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
7868 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7870 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7871 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7873 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
7874 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
7879 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7881 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7883 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7884 ; AVX512BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
7885 ; AVX512BW-FCP-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
7886 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7888 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
7891 ; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm22[0,1,2,3],zmm26[4,5,6,7]
7897 ; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm28[0,1,2,3],zmm17[4,5,6,7]
7900 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,0,4,5,6,7]
7905 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,9,4,5,6,7]
7914 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,13,0,4,5,6,7]
7919 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,13,4,5,6,7]
7924 ; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm16[0,1,2,3],zmm10[4,5,6,7]
7925 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [14,0,2,3,4,5,15,0]
7930 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
7936 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7938 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm11 {%k1}
7940 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7942 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
7943 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm10 {%k1}
7945 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7947 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
7948 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm16 {%k1}
7950 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
7952 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm19 {%k1}
7954 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,3,4,8,6,7]
7965 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm19, %zmm6
7969 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [10,0,2,3,4,5,11,0]
7973 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm24, %zmm0
7978 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm7, %zmm1
7979 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
7984 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm21, %zmm2
7985 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
7988 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [0,10,2,3,4,5,6,11]
7990 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,1,2,3,4,12,6,7]
7998 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
8023 ; AVX512BW-FCP-NEXT: addq $648, %rsp # imm = 0x288
8028 ; AVX512DQ-BW: # %bb.0:
8029 ; AVX512DQ-BW-NEXT: subq $648, %rsp # imm = 0x288
8046 ; AVX512DQ-BW-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
8049 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8052 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8055 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8058 ; AVX512DQ-BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8061 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8063 ; AVX512DQ-BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8066 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8068 ; AVX512DQ-BW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8071 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8074 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8077 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8080 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8088 ; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm1 = [0,8,1,9,0,8,1,9]
8089 ; AVX512DQ-BW-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3]
8094 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8098 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8108 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
8117 ; AVX512DQ-BW-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]
8121 ; AVX512DQ-BW-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]
8126 ; AVX512DQ-BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8145 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8147 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8148 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8150 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8151 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
8156 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8158 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8160 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8161 ; AVX512DQ-BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
8162 ; AVX512DQ-BW-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
8163 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8165 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8168 ; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm22[0,1,2,3],zmm26[4,5,6,7]
8174 ; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm28[0,1,2,3],zmm17[4,5,6,7]
8177 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,0,4,5,6,7]
8182 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,9,4,5,6,7]
8191 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,13,0,4,5,6,7]
8196 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,13,4,5,6,7]
8201 ; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm16[0,1,2,3],zmm10[4,5,6,7]
8202 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [14,0,2,3,4,5,15,0]
8207 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
8213 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
8215 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm11 {%k1}
8217 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
8219 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
8220 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm10 {%k1}
8222 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
8224 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
8225 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm16 {%k1}
8227 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
8229 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm19 {%k1}
8231 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,3,4,8,6,7]
8242 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm19, %zmm6
8246 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm19 = [10,0,2,3,4,5,11,0]
8250 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm24, %zmm0
8255 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm7, %zmm1
8256 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
8261 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm21, %zmm2
8262 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
8265 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm19 = [0,10,2,3,4,5,6,11]
8267 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,1,2,3,4,12,6,7]
8275 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
8300 ; AVX512DQ-BW-NEXT: addq $648, %rsp # imm = 0x288
8305 ; AVX512DQ-BW-FCP: # %bb.0:
8306 ; AVX512DQ-BW-FCP-NEXT: subq $648, %rsp # imm = 0x288
8323 ; AVX512DQ-BW-FCP-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
8326 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8329 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8332 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8335 ; AVX512DQ-BW-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8338 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8340 ; AVX512DQ-BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8343 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8345 ; AVX512DQ-BW-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8348 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8351 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8354 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8357 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8365 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm1 = [0,8,1,9,0,8,1,9]
8366 ; AVX512DQ-BW-FCP-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3]
8371 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8375 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8385 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
8394 ; AVX512DQ-BW-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]
8398 ; AVX512DQ-BW-FCP-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]
8403 ; AVX512DQ-BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8422 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8424 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8425 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8427 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8428 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
8433 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8435 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8437 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8438 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
8439 ; AVX512DQ-BW-FCP-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
8440 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8442 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8445 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm22[0,1,2,3],zmm26[4,5,6,7]
8451 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm28[0,1,2,3],zmm17[4,5,6,7]
8454 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,9,0,4,5,6,7]
8459 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,9,4,5,6,7]
8468 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,13,0,4,5,6,7]
8473 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,13,4,5,6,7]
8478 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm16[0,1,2,3],zmm10[4,5,6,7]
8479 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [14,0,2,3,4,5,15,0]
8484 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,14,2,3,4,5,6,15]
8490 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
8492 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm11 {%k1}
8494 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
8496 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
8497 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm10 {%k1}
8499 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
8501 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
8502 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm16 {%k1}
8504 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm5 = xmm5[0],mem[0]
8506 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm0, %zmm19 {%k1}
8508 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,3,4,8,6,7]
8519 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm19, %zmm6
8523 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [10,0,2,3,4,5,11,0]
8527 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm24, %zmm0
8532 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm7, %zmm1
8533 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
8538 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm21, %zmm2
8539 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
8542 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [0,10,2,3,4,5,6,11]
8544 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,1,2,3,4,12,6,7]
8552 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
8577 ; AVX512DQ-BW-FCP-NEXT: addq $648, %rsp # imm = 0x288
8586 %1 = shufflevector <32 x i64> %in.vec0, <32 x i64> %in.vec1, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
8587 %2 = shufflevector <32 x i64> %in.vec2, <32 x i64> %in.vec3, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
8588 %3 = shufflevector <32 x i64> %in.vec4, <32 x i64> %in.vec5, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
8589 %4 = shufflevector <64 x i64> %1, <64 x i64> %2, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
8590 %5 = shufflevector <64 x i64> %3, <64 x i64> poison, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
8591 %6 = shufflevector <128 x i64> %4, <128 x i64> %5, <192 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191>
8592 %interleaved.vec = shufflevector <192 x i64> %6, <192 x i64> poison, <192 x i32> <i32 0, i32 32, i32 64, i32 96, i32 128, i32 160, i32 1, i32 33, i32 65, i32 97, i32 129, i32 161, i32 2, i32 34, i32 66, i32 98, i32 130, i32 162, i32 3, i32 35, i32 67, i32 99, i32 131, i32 163, i32 4, i32 36, i32 68, i32 100, i32 132, i32 164, i32 5, i32 37, i32 69, i32 101, i32 133, i32 165, i32 6, i32 38, i32 70, i32 102, i32 134, i32 166, i32 7, i32 39, i32 71, i32 103, i32 135, i32 167, i32 8, i32 40, i32 72, i32 104, i32 136, i32 168, i32 9, i32 41, i32 73, i32 105, i32 137, i32 169, i32 10, i32 42, i32 74, i32 106, i32 138, i32 170, i32 11, i32 43, i32 75, i32 107, i32 139, i32 171, i32 12, i32 44, i32 76, i32 108, i32 140, i32 172, i32 13, i32 45, i32 77, i32 109, i32 141, i32 173, i32 14, i32 46, i32 78, i32 110, i32 142, i32 174, i32 15, i32 47, i32 79, i32 111, i32 143, i32 175, i32 16, i32 48, i32 80, i32 112, i32 144, i32 176, i32 17, i32 49, i32 81, i32 113, i32 145, i32 177, i32 18, i32 50, i32 82, i32 114, i32 146, i32 178, i32 19, i32 51, i32 83, i32 115, i32 147, i32 179, i32 20, i32 52, i32 84, i32 116, i32 148, i32 180, i32 21, i32 53, i32 85, i32 117, i32 149, i32 181, i32 22, i32 54, i32 86, i32 118, i32 150, i32 182, i32 23, i32 55, i32 87, i32 119, i32 151, i32 183, i32 24, i32 56, i32 88, i32 120, i32 152, i32 184, i32 25, i32 57, i32 89, i32 121, i32 153, i32 185, i32 26, i32 58, i32 90, i32 122, i32 154, i32 186, i32 27, i32 59, i32 91, i32 123, i32 155, i32 187, i32 28, i32 60, i32 92, i32 124, i32 156, i32 188, i32 29, i32 61, i32 93, i32 125, i32 157, i32 189, i32 30, i32 62, i32 94, i32 126, i32 158, i32 190, i32 31, i32 63, i32 95, i32 127, i32 159, i32 191>
8599 ; SSE: # %bb.0:
8600 ; SSE-NEXT: subq $2712, %rsp # imm = 0xA98
8617 ; SSE-NEXT: movlhps {{.*#+}} xmm15 = xmm15[0],xmm2[0]
8618 ; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8620 ; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8622 ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm4[0]
8623 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8625 ; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8627 ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm6[0]
8628 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8630 ; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8632 ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]
8633 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8635 ; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8637 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]
8638 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8640 ; SSE-NEXT: movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8642 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm5[0]
8643 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8645 ; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8647 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8648 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8650 ; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8653 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8654 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8656 ; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8660 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8661 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8663 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8667 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8668 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8670 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8674 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8675 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8677 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8681 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8682 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8684 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8688 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8689 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8691 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8695 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8696 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8698 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8702 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8703 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8705 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8709 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8710 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8712 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8716 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8717 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8719 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8723 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8724 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8726 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8730 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8731 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8733 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8737 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8738 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8740 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8744 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8745 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8747 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8751 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8752 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8754 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8758 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8759 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8761 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8765 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8766 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8768 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8772 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8773 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8775 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8779 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8780 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8782 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8786 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8787 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8789 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8793 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8794 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8796 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8800 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8801 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8803 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8807 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8808 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8810 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8814 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8815 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8817 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8821 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8822 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8824 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8828 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8829 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8831 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8835 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8836 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8838 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8842 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8843 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8845 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8849 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8850 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8852 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8856 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8857 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8859 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8863 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8864 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8866 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8870 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8871 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8873 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8877 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8878 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8880 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8884 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8885 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8887 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8891 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8892 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8894 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8898 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8899 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8901 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8905 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8906 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8908 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8912 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8913 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8915 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8919 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8920 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8922 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8926 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8927 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8929 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8933 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8934 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8936 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8940 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8941 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8943 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8947 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8948 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8950 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8954 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8955 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8957 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8961 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8962 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8964 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8968 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8969 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8971 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8975 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8976 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8978 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8982 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8983 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8985 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8989 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8990 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8992 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8996 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
8997 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8999 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9003 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9004 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9006 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9010 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9011 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9013 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9017 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9018 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9020 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9024 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9025 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9027 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9031 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9032 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9034 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9038 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9039 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9041 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9045 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9046 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9048 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9052 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9053 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9055 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9059 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9060 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9062 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9066 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9067 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9069 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9073 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9074 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9076 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9080 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9081 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9083 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9087 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9088 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9090 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9094 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9095 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9097 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9101 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9102 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9104 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9108 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9109 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9111 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9115 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9116 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9118 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9122 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9123 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9125 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9129 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9130 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9132 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9136 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9137 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9139 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9143 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9144 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9146 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9150 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9151 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9153 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9157 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9158 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9160 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9164 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9165 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9167 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9171 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9172 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9174 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9178 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9179 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9181 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9185 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9186 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9188 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9192 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9195 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9199 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9200 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9202 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9206 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9207 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9209 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9213 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9214 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9219 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9220 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9225 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
9226 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9231 ; SSE-NEXT: movlhps {{.*#+}} xmm13 = xmm13[0],xmm0[0]
9236 ; SSE-NEXT: movlhps {{.*#+}} xmm10 = xmm10[0],xmm0[0]
9241 ; SSE-NEXT: movlhps {{.*#+}} xmm9 = xmm9[0],xmm0[0]
9246 ; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm1[0]
9251 ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]
9256 ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm3[0]
9258 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
9274 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9276 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9278 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9280 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9282 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9284 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9286 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9288 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9292 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9294 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9296 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9298 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9300 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9302 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9304 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9306 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9308 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9310 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9312 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9314 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9316 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9318 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9320 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9322 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9324 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9326 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9328 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9330 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9332 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9334 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9336 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9338 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9340 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9342 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9344 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9346 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9348 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9350 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9352 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9354 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9356 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9358 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9360 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9362 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9364 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9366 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9368 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9370 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9372 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9374 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9376 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9378 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9380 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9382 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9384 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9386 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9388 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9390 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9392 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9394 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9396 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9398 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9400 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9402 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9404 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9406 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9408 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9410 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9412 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9414 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9416 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9418 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9420 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9422 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9424 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9426 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9428 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9430 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9432 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9434 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9436 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9438 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9440 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9442 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9444 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9446 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9448 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9450 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9452 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9454 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9456 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9458 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9460 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9462 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9464 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9466 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9468 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9470 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9472 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9474 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9476 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9478 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9480 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9482 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9484 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9486 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9488 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9490 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9492 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9494 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9496 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9498 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9500 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9502 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9504 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9506 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9508 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9510 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9512 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9514 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9516 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9518 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9520 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9522 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9524 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9526 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9528 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9530 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9532 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9534 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9536 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9538 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9540 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9542 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9544 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9546 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9548 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9550 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9552 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9554 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9556 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9558 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9560 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9562 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9564 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9566 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9568 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9570 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9572 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9574 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9576 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9578 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9580 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9582 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9584 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9586 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9588 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9590 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9592 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9594 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9596 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9598 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9600 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9602 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9604 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9606 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9608 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9610 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9612 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9614 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9616 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9618 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9620 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9622 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9624 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9626 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9628 ; SSE-NEXT: addq $2712, %rsp # imm = 0xA98
9632 ; AVX: # %bb.0:
9633 ; AVX-NEXT: subq $3464, %rsp # imm = 0xD88
9635 ; AVX-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9637 ; AVX-NEXT: vmovddup {{.*#+}} xmm1 = mem[0,0]
9639 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9641 ; AVX-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9643 ; AVX-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9645 ; AVX-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9648 ; AVX-NEXT: vblendpd {{.*#+}} ymm3 = ymm0[0,1],ymm3[2,3]
9649 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3]
9650 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9652 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9654 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9658 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm4[4,5,6,7]
9659 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm3[2,3],ymm1[4,5],ymm3[6,7]
9660 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9661 ; AVX-NEXT: vmovddup {{.*#+}} xmm1 = mem[0,0]
9663 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9666 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm7[0,1,2,3],ymm3[4,5,6,7]
9667 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5,6,7]
9668 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9670 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9672 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9675 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm3[4,5,6,7]
9677 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5],ymm1[6,7]
9678 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9680 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9684 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm15[0,1],ymm1[2,3]
9685 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = mem[0,0]
9686 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2,3]
9687 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9691 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9694 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm3[4,5,6,7]
9696 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5],ymm1[6,7]
9697 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9699 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9701 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9705 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm9[0,1],ymm1[2,3]
9706 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = mem[0,0]
9707 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2,3]
9708 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9710 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9712 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9715 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm3[4,5,6,7]
9717 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5],ymm1[6,7]
9718 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9720 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9722 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9726 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9727 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
9728 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = mem[0,0]
9729 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm3[2,3],ymm1[4,5,6,7]
9730 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9732 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9734 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9737 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm3[4,5,6,7]
9739 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5],ymm1[6,7]
9740 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9742 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9744 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9748 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm8[0,1],ymm1[2,3]
9749 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = mem[0,0]
9750 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2,3]
9751 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9753 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9755 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9758 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm3[4,5,6,7]
9760 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5],ymm1[6,7]
9761 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9763 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9765 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9769 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm7[0,1],ymm1[2,3]
9770 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = mem[0,0]
9771 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2,3]
9772 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9774 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9776 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9779 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm3[4,5,6,7]
9781 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5],ymm1[6,7]
9782 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9784 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9786 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9790 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm11[0,1],ymm1[2,3]
9791 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = mem[0,0]
9792 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2,3]
9793 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9795 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9797 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9800 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm3[4,5,6,7]
9802 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5],ymm1[6,7]
9803 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9805 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9807 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9811 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9812 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
9813 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = mem[0,0]
9814 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm3[2,3],ymm1[4,5,6,7]
9815 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9817 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9819 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9822 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm3[4,5,6,7]
9824 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5],ymm1[6,7]
9825 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9827 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9829 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9833 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9834 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
9835 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = mem[0,0]
9836 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm3[2,3],ymm1[4,5,6,7]
9837 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9839 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9841 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9844 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm3[4,5,6,7]
9846 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5],ymm1[6,7]
9847 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9849 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9851 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9855 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9856 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
9857 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = mem[0,0]
9858 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm3[2,3],ymm1[4,5,6,7]
9859 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9861 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9863 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9866 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm3[4,5,6,7]
9868 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5],ymm1[6,7]
9869 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9871 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9873 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9877 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9878 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
9879 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = mem[0,0]
9880 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm3[2,3],ymm1[4,5,6,7]
9881 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9883 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9885 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9888 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm3[4,5,6,7]
9890 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5],ymm1[6,7]
9891 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9893 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9895 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9899 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm12[0,1],ymm1[2,3]
9900 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = mem[0,0]
9901 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2,3]
9902 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9904 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9906 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9909 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm3[4,5,6,7]
9911 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5],ymm1[6,7]
9912 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9914 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9916 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9920 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm13[0,1],ymm1[2,3]
9921 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = mem[0,0]
9922 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2,3]
9923 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9925 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9927 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9930 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm3[4,5,6,7]
9932 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5],ymm1[6,7]
9933 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9935 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9937 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9941 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm14[0,1],ymm1[2,3]
9942 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = mem[0,0]
9943 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2,3]
9944 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9946 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9948 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9951 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3],ymm3[4,5,6,7]
9953 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5],ymm1[6,7]
9954 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9956 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9958 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9962 ; AVX-NEXT: vblendpd {{.*#+}} ymm3 = ymm5[0,1],ymm3[2,3]
9963 ; AVX-NEXT: vmovddup {{.*#+}} xmm2 = mem[0,0]
9964 ; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm3[0],ymm2[1],ymm3[2,3]
9965 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9967 ; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9969 ; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9972 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm2[0,1,2,3],ymm3[4,5,6,7]
9974 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm2[2,3],ymm3[4,5],ymm2[6,7]
9975 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9982 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm2[0],ymm3[0],ymm2[2],ymm3[3]
9983 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9987 ; AVX-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm0 # 32-byte Folded Reload
9991 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[3]
9992 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9999 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[3]
10000 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10007 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm9[0],ymm0[0],ymm9[2],ymm0[3]
10008 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10012 ; AVX-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm1 # 32-byte Folded Reload
10016 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[3]
10017 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10024 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[3]
10025 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10032 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[3]
10033 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10040 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[3]
10041 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10045 ; AVX-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
10049 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[3]
10050 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10054 ; AVX-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
10057 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10059 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[3]
10060 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10064 ; AVX-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
10067 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10069 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[3]
10070 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10074 ; AVX-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
10077 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10079 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[3]
10080 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10087 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[3]
10088 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10095 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[3]
10096 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10103 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[3]
10104 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10111 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[3]
10112 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10114 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10115 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10117 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10118 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10122 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]
10123 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm4[3]
10124 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10126 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10127 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10129 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10130 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10134 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]
10135 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm6[3]
10136 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10138 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10139 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10141 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10142 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10146 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]
10147 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm15[3]
10148 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10150 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10151 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10153 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10154 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10158 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]
10159 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm3[3]
10160 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10162 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10163 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10165 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10166 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10170 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]
10171 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm9[3]
10172 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10174 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10175 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10177 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10178 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10182 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]
10183 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm8[3]
10184 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10186 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10187 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10189 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10190 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10194 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]
10195 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm7[3]
10196 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10198 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10199 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10201 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10202 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10206 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]
10207 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm10[3]
10208 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10210 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10211 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10213 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10214 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10218 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]
10219 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm11[3]
10220 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10222 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10223 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10225 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10226 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10230 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
10231 ; AVX-NEXT: vblendps $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
10232 ; AVX-NEXT: # ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
10233 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10235 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10236 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10238 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10239 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10243 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
10244 ; AVX-NEXT: vblendps $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
10245 ; AVX-NEXT: # ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
10246 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10248 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10249 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10251 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10252 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10256 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
10257 ; AVX-NEXT: vblendps $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
10258 ; AVX-NEXT: # ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
10259 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10261 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10262 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10264 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10265 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10269 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]
10270 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm12[3]
10271 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10273 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10274 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10276 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10277 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10281 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]
10282 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm13[3]
10283 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10285 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10286 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10288 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10289 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10293 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]
10294 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm14[3]
10295 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10297 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],mem[0]
10298 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10300 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10301 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10305 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]
10306 ; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1,2],ymm5[3]
10307 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10308 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10309 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10310 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10311 ; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10312 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10313 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10314 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10315 ; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10316 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10317 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10318 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10319 ; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10320 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10321 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10322 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10323 ; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10324 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10325 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10326 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10327 ; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10328 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10330 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10331 ; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10332 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10333 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10334 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10335 ; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10336 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10337 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10338 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10339 ; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10340 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10341 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10342 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10343 ; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10344 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10345 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10346 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10347 ; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10348 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10349 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10350 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10351 ; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10352 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10353 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10354 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10356 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10357 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10358 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10359 ; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10360 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10361 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10362 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10363 ; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10364 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10365 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10366 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10367 ; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10368 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10369 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10370 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10371 ; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10372 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10373 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm15 # 16-byte Folded Reload
10374 ; AVX-NEXT: # xmm15 = xmm0[0],mem[0]
10375 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10376 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm14 # 16-byte Folded Reload
10377 ; AVX-NEXT: # xmm14 = xmm0[0],mem[0]
10378 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10379 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm13 # 16-byte Folded Reload
10380 ; AVX-NEXT: # xmm13 = xmm0[0],mem[0]
10381 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10382 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm12 # 16-byte Folded Reload
10383 ; AVX-NEXT: # xmm12 = xmm0[0],mem[0]
10384 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10385 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm11 # 16-byte Folded Reload
10386 ; AVX-NEXT: # xmm11 = xmm0[0],mem[0]
10387 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10388 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm10 # 16-byte Folded Reload
10389 ; AVX-NEXT: # xmm10 = xmm0[0],mem[0]
10390 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10391 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm9 # 16-byte Folded Reload
10392 ; AVX-NEXT: # xmm9 = xmm0[0],mem[0]
10393 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10394 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm8 # 16-byte Folded Reload
10395 ; AVX-NEXT: # xmm8 = xmm0[0],mem[0]
10396 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10397 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm7 # 16-byte Folded Reload
10398 ; AVX-NEXT: # xmm7 = xmm0[0],mem[0]
10399 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10400 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm6 # 16-byte Folded Reload
10401 ; AVX-NEXT: # xmm6 = xmm0[0],mem[0]
10402 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10403 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload
10404 ; AVX-NEXT: # xmm5 = xmm0[0],mem[0]
10405 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10406 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm4 # 16-byte Folded Reload
10407 ; AVX-NEXT: # xmm4 = xmm0[0],mem[0]
10408 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10409 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload
10410 ; AVX-NEXT: # xmm3 = xmm0[0],mem[0]
10411 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10412 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload
10413 ; AVX-NEXT: # xmm2 = xmm0[0],mem[0]
10414 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10415 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload
10416 ; AVX-NEXT: # xmm1 = xmm0[0],mem[0]
10417 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10418 ; AVX-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
10419 ; AVX-NEXT: # xmm0 = xmm0[0],mem[0]
10420 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
10437 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10439 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10441 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10443 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10447 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10449 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10451 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10453 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10455 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10457 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10459 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10461 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10463 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10465 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10467 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10469 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10471 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10473 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10475 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10477 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10479 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10481 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10483 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10485 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10487 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10489 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10491 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10493 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10495 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10497 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10499 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10501 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10503 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10505 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10507 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10509 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10511 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10513 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10515 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10517 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10519 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10521 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10523 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10525 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10527 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10529 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10531 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10533 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10535 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10537 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10539 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10541 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10543 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10545 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10547 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10549 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10551 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10553 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10555 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10557 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10559 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10561 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10563 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10565 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10567 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10569 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10571 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10573 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10575 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10577 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10579 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10581 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10583 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10585 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10587 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10589 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10591 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10593 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10595 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10597 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10599 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10601 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10603 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10605 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10607 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10609 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10611 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10613 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10615 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10617 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10619 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10621 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10623 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10625 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10627 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10629 ; AVX-NEXT: addq $3464, %rsp # imm = 0xD88
10634 ; AVX2: # %bb.0:
10635 ; AVX2-NEXT: subq $2968, %rsp # imm = 0xB98
10637 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10639 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10641 ; AVX2-NEXT: vmovddup {{.*#+}} xmm1 = xmm0[0,0]
10643 ; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10645 ; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10647 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10649 ; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10651 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm4[0,1],ymm2[0,1]
10652 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]
10653 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10656 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10658 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10660 ; AVX2-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10663 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
10664 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
10665 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10667 ; AVX2-NEXT: vmovddup {{.*#+}} xmm1 = xmm0[0,0]
10669 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm3[0,1],ymm2[0,1]
10670 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]
10671 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10673 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10676 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
10678 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10680 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
10681 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10683 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10685 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10687 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm2[0,1],ymm0[0,1]
10689 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
10690 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
10691 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10693 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10695 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10698 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
10700 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10701 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10703 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10705 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10707 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10709 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
10711 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
10712 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
10713 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10715 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10717 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10720 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
10722 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10723 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10725 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10727 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10729 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10731 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
10733 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
10734 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
10735 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10737 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10739 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10742 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
10744 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10745 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10747 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10749 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10751 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10753 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
10755 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
10756 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
10757 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10759 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10761 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10764 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
10766 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10767 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10769 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10771 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10773 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10775 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
10777 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
10778 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
10779 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10781 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10783 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10786 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
10788 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10789 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10791 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10793 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10795 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10797 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
10799 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
10800 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
10801 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10803 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10805 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10808 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
10810 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10811 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10813 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10815 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10817 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10819 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
10821 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
10822 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
10823 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10825 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10827 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10830 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
10832 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10833 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10835 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10837 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10839 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10841 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
10843 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
10844 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
10845 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10847 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10849 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10852 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
10854 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10855 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10857 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10859 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10861 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10863 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
10865 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
10866 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
10867 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10869 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10871 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10874 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
10876 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10877 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10879 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10881 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10883 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10885 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
10887 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
10888 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
10889 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10891 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10893 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10896 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
10898 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10899 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10901 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10903 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10905 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10907 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
10909 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
10910 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
10911 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10913 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10917 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
10919 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10920 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10922 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10926 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
10928 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
10929 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
10930 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10935 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
10937 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10938 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10940 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10944 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
10946 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
10947 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
10948 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10953 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
10955 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
10956 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10958 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10962 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[0,1],ymm0[0,1]
10964 ; AVX2-NEXT: vmovddup {{.*#+}} xmm2 = xmm0[0,0]
10965 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7]
10966 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10971 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm14[4,5,6,7]
10973 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
10974 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10975 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10976 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
10977 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10978 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
10979 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
10980 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10981 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10982 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
10983 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10984 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
10985 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
10986 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10987 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10988 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
10989 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10990 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
10991 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
10992 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10993 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10994 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
10995 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10996 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
10997 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
10998 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10999 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11000 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11001 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11002 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11003 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11004 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11005 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11006 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11007 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11008 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11009 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11010 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11011 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11012 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11013 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11014 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11015 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11016 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11017 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11018 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11019 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11020 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11021 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11022 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11023 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11024 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11025 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11026 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11027 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11028 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11029 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11030 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11031 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11032 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11033 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11034 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11035 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11036 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11037 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11038 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11039 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11040 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11041 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11042 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11043 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11044 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11045 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11046 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11047 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11048 ; AVX2-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11049 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11051 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11052 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11055 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11056 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11059 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11060 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11063 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11064 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11068 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11071 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
11072 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11074 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11077 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11078 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11085 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11088 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
11089 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11091 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11094 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11095 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11102 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11105 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
11106 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11108 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11111 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11112 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11119 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11122 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
11123 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11125 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11128 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11129 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11133 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11137 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11140 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
11141 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11143 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11146 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11147 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11151 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11155 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11158 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
11159 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11161 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11164 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11165 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11169 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11173 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11176 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
11177 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11179 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11182 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11183 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11187 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11191 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11194 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
11195 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11197 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11200 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11201 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11205 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11209 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11212 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
11213 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11215 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11218 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11219 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11223 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11227 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11230 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
11231 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11233 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11236 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11237 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11241 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11245 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11248 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
11249 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11251 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11254 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11255 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11259 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11263 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11266 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
11267 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11269 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11272 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11273 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11277 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11281 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11284 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
11285 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11287 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11290 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11291 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11298 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11301 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
11302 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11304 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11307 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11308 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11315 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11318 ; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm4[0,1,2,3,4,5],ymm5[6,7]
11320 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11323 ; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11330 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
11333 ; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3,4,5],ymm8[6,7]
11335 ; AVX2-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
11338 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7]
11342 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
11343 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm12[0,1,2,3,4,5],mem[6,7]
11344 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11345 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm13[0,1,2,3,4,5],mem[6,7]
11346 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11347 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm14[0,1,2,3,4,5],mem[6,7]
11348 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11349 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11350 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]
11351 ; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11352 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11353 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm1[0,1,2,3,4,5],mem[6,7]
11354 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11355 ; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm1[0,1,2,3,4,5],mem[6,7]
11356 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11357 ; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm1[0,1,2,3,4,5],mem[6,7]
11358 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11359 ; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm1[0,1,2,3,4,5],mem[6,7]
11360 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11361 ; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm1[0,1,2,3,4,5],mem[6,7]
11362 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11363 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm1[0,1,2,3,4,5],mem[6,7]
11364 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11365 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],mem[6,7]
11366 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11367 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]
11368 ; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3,4,5],mem[6,7]
11369 ; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],mem[6,7]
11370 ; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],mem[6,7]
11371 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
11375 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11380 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11383 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11385 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11387 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11390 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11392 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11394 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11397 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11399 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11401 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11404 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11406 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11408 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11411 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11413 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11415 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11418 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11420 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11422 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11425 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11427 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11429 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11432 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11434 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11436 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11439 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11441 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11443 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11446 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11448 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11450 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11452 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11454 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11456 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11458 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11460 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11462 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11464 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11466 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11468 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11470 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11472 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11474 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11476 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11478 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11480 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11482 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11484 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11486 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11488 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11490 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11492 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11494 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11496 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11498 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11500 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11502 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11504 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11506 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11508 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11510 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11512 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11514 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11516 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11518 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11520 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11522 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11524 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11526 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11528 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11530 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11532 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11534 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11536 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11538 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11540 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11542 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11544 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11546 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11548 ; AVX2-NEXT: addq $2968, %rsp # imm = 0xB98
11553 ; AVX2-FP: # %bb.0:
11554 ; AVX2-FP-NEXT: subq $2968, %rsp # imm = 0xB98
11556 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11558 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11560 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm1 = xmm0[0,0]
11562 ; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11564 ; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11566 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11568 ; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11570 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm4[0,1],ymm2[0,1]
11571 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]
11572 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11575 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11577 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11579 ; AVX2-FP-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11582 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
11583 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
11584 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11586 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm1 = xmm0[0,0]
11588 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm3[0,1],ymm2[0,1]
11589 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]
11590 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11592 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11595 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
11597 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11599 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
11600 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11602 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11604 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11606 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm2[0,1],ymm0[0,1]
11608 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
11609 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
11610 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11612 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11614 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11617 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
11619 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
11620 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11622 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11624 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11626 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11628 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
11630 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
11631 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
11632 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11634 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11636 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11639 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
11641 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
11642 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11644 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11646 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11648 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11650 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
11652 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
11653 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
11654 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11656 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11658 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11661 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
11663 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
11664 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11666 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11668 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11670 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11672 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
11674 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
11675 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
11676 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11678 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11680 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11683 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
11685 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
11686 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11688 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11690 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11692 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11694 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
11696 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
11697 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
11698 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11700 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11702 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11705 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
11707 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
11708 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11710 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11712 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11714 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11716 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
11718 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
11719 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
11720 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11722 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11724 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11727 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
11729 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
11730 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11732 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11734 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11736 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11738 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
11740 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
11741 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
11742 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11744 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11746 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11749 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
11751 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
11752 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11754 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11756 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11758 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11760 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
11762 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
11763 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
11764 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11766 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11768 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11771 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
11773 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
11774 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11776 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11778 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11780 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11782 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
11784 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
11785 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
11786 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11788 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11790 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11793 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
11795 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
11796 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11798 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11800 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11802 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11804 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
11806 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
11807 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
11808 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11810 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11812 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11815 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
11817 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
11818 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11820 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11822 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11824 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11826 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
11828 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
11829 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
11830 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11832 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11836 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
11838 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
11839 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11841 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11845 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
11847 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
11848 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
11849 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11854 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
11856 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
11857 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11859 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11863 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
11865 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
11866 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
11867 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11872 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
11874 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
11875 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11877 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11881 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[0,1],ymm0[0,1]
11883 ; AVX2-FP-NEXT: vmovddup {{.*#+}} xmm2 = xmm0[0,0]
11884 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7]
11885 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11890 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm14[4,5,6,7]
11892 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
11893 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11894 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11895 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11896 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11897 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11898 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11899 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11900 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11901 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11902 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11903 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11904 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11905 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11906 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11907 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11908 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11909 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11910 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11911 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11912 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11913 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11914 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11915 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11916 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11917 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11918 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11919 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11920 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11921 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11922 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11923 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11924 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11925 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11926 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11927 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11928 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11929 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11930 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11931 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11932 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11933 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11934 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11935 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11936 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11937 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11938 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11939 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11940 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11941 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11942 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11943 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11944 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11945 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11946 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11947 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11948 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11949 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11950 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11951 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11952 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11953 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11954 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11955 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11956 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11957 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11958 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11959 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11960 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11961 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11962 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11963 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
11964 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11965 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11966 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11967 ; AVX2-FP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
11968 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11970 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11971 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11974 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11975 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11978 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11979 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11982 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
11983 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11987 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
11990 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
11991 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11993 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11996 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
11997 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12004 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12007 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12008 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12010 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12013 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12014 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12021 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12024 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12025 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12027 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12030 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12031 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12038 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12041 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12042 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12044 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12047 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12048 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12052 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12056 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12059 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12060 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12062 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12065 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12066 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12070 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12074 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12077 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12078 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12080 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12083 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12084 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12088 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12092 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12095 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12096 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12098 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12101 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12102 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12106 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12110 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12113 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12114 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12116 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12119 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12120 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12124 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12128 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12131 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12132 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12134 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12137 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12138 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12142 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12146 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12149 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12150 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12152 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12155 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12156 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12160 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12164 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12167 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12168 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12170 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12173 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12174 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12178 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12182 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12185 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12186 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12188 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12191 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12192 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12196 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12200 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12203 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12204 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12206 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12209 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12210 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12217 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12220 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12221 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12223 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12226 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12227 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12234 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12237 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm4[0,1,2,3,4,5],ymm5[6,7]
12239 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12242 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12249 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
12252 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3,4,5],ymm8[6,7]
12254 ; AVX2-FP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
12257 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7]
12261 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
12262 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm12[0,1,2,3,4,5],mem[6,7]
12263 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12264 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm13[0,1,2,3,4,5],mem[6,7]
12265 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12266 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm14[0,1,2,3,4,5],mem[6,7]
12267 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12268 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12269 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]
12270 ; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12271 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12272 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm1[0,1,2,3,4,5],mem[6,7]
12273 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12274 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm1[0,1,2,3,4,5],mem[6,7]
12275 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12276 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm1[0,1,2,3,4,5],mem[6,7]
12277 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12278 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm1[0,1,2,3,4,5],mem[6,7]
12279 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12280 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm1[0,1,2,3,4,5],mem[6,7]
12281 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12282 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm1[0,1,2,3,4,5],mem[6,7]
12283 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12284 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],mem[6,7]
12285 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12286 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]
12287 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3,4,5],mem[6,7]
12288 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],mem[6,7]
12289 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],mem[6,7]
12290 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
12294 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12299 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12302 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12304 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12306 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12309 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12311 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12313 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12316 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12318 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12320 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12323 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12325 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12327 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12330 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12332 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12334 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12337 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12339 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12341 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12344 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12346 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12348 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12351 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12353 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12355 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12358 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12360 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12362 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12365 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12367 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12369 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12371 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12373 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12375 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12377 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12379 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12381 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12383 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12385 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12387 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12389 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12391 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12393 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12395 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12397 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12399 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12401 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12403 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12405 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12407 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12409 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12411 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12413 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12415 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12417 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12419 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12421 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12423 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12425 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12427 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12429 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12431 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12433 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12435 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12437 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12439 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12441 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12443 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12445 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12447 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12449 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12451 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12453 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12455 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12457 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12459 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12461 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12463 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12465 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12467 ; AVX2-FP-NEXT: addq $2968, %rsp # imm = 0xB98
12472 ; AVX2-FCP: # %bb.0:
12473 ; AVX2-FCP-NEXT: subq $2968, %rsp # imm = 0xB98
12475 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12477 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12479 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm1 = xmm0[0,0]
12481 ; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12483 ; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12485 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12487 ; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12489 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm4[0,1],ymm2[0,1]
12490 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]
12491 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12494 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12496 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12498 ; AVX2-FCP-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12501 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
12502 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
12503 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12505 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm1 = xmm0[0,0]
12507 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm3[0,1],ymm2[0,1]
12508 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7]
12509 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12511 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12514 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
12516 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12518 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
12519 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12521 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12523 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12525 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm2[0,1],ymm0[0,1]
12527 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
12528 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
12529 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12531 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12533 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12536 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
12538 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
12539 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12541 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12543 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12545 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12547 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
12549 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
12550 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
12551 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12553 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12555 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12558 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
12560 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
12561 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12563 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12565 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12567 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12569 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
12571 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
12572 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
12573 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12575 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12577 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12580 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
12582 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
12583 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12585 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12587 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12589 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12591 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
12593 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
12594 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
12595 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12597 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12599 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12602 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
12604 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
12605 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12607 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12609 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12611 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12613 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
12615 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
12616 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
12617 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12619 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12621 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12624 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
12626 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
12627 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12629 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12631 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12633 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12635 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
12637 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
12638 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
12639 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12641 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12643 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12646 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
12648 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
12649 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12651 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12653 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12655 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12657 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
12659 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
12660 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
12661 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12663 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12665 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12668 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
12670 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
12671 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12673 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12675 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12677 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12679 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
12681 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
12682 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
12683 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12685 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12687 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12690 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
12692 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
12693 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12695 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12697 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12699 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12701 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
12703 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
12704 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
12705 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12707 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12709 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12712 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
12714 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
12715 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12717 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12719 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12721 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12723 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
12725 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
12726 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
12727 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12729 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12731 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12734 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
12736 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
12737 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12739 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12741 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12743 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12745 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
12747 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
12748 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
12749 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12751 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12755 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
12757 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
12758 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12760 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12764 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
12766 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
12767 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
12768 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12773 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
12775 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
12776 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12778 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12782 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[0,1],ymm0[0,1]
12784 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm1[0,0]
12785 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
12786 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12791 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
12793 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
12794 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12796 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12800 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[0,1],ymm0[0,1]
12802 ; AVX2-FCP-NEXT: vmovddup {{.*#+}} xmm2 = xmm0[0,0]
12803 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7]
12804 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12809 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm14[4,5,6,7]
12811 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
12812 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12813 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12814 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
12815 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12816 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
12817 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12818 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12819 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12820 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
12821 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12822 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
12823 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12824 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12825 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12826 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
12827 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12828 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
12829 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12830 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12831 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12832 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
12833 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12834 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
12835 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12836 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12837 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12838 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
12839 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12840 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
12841 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12842 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12843 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12844 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
12845 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12846 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
12847 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12848 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12849 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12850 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
12851 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12852 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
12853 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12854 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12855 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12856 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
12857 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12858 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
12859 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12860 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12861 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12862 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
12863 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12864 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
12865 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12866 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12867 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12868 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
12869 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12870 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
12871 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12872 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12873 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12874 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
12875 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12876 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
12877 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12878 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12879 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12880 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
12881 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12882 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 16-byte Folded Reload
12883 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12884 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12885 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
12886 ; AVX2-FCP-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
12887 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
12889 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12890 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12893 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12894 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12897 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12898 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12901 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
12902 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12906 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12909 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12910 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12912 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12915 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12916 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12923 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12926 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12927 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12929 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12932 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12933 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12940 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12943 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12944 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12946 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12949 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12950 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12957 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12960 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12961 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12963 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12966 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12967 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12971 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12975 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12978 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12979 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12981 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12984 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
12985 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12989 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12993 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
12996 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
12997 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12999 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
13002 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
13003 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13007 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13011 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
13014 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
13015 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13017 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
13020 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
13021 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13025 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13029 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
13032 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
13033 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13035 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
13038 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
13039 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13043 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13047 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
13050 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
13051 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13053 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
13056 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
13057 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13061 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13065 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
13068 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
13069 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13071 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
13074 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
13075 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13079 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13083 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
13086 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
13087 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13089 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
13092 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
13093 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13097 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13101 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
13104 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
13105 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13107 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
13110 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
13111 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13115 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13119 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
13122 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
13123 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13125 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
13128 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
13129 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13136 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
13139 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
13140 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13142 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
13145 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
13146 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13153 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
13156 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm4[0,1,2,3,4,5],ymm5[6,7]
13158 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
13161 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
13168 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm1[0],ymm2[0],ymm1[2],ymm2[2]
13171 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3,4,5],ymm8[6,7]
13173 ; AVX2-FCP-NEXT: vperm2f128 $19, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
13176 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7]
13180 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
13181 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm12[0,1,2,3,4,5],mem[6,7]
13182 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13183 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm13[0,1,2,3,4,5],mem[6,7]
13184 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13185 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm14[0,1,2,3,4,5],mem[6,7]
13186 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13187 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
13188 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]
13189 ; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13190 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
13191 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm1[0,1,2,3,4,5],mem[6,7]
13192 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
13193 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm1[0,1,2,3,4,5],mem[6,7]
13194 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
13195 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm1[0,1,2,3,4,5],mem[6,7]
13196 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
13197 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm1[0,1,2,3,4,5],mem[6,7]
13198 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
13199 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm1[0,1,2,3,4,5],mem[6,7]
13200 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
13201 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm1[0,1,2,3,4,5],mem[6,7]
13202 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
13203 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],mem[6,7]
13204 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
13205 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]
13206 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3,4,5],mem[6,7]
13207 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],mem[6,7]
13208 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],mem[6,7]
13209 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
13213 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13218 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13221 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13223 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13225 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13228 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13230 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13232 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13235 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13237 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13239 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13242 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13244 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13246 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13249 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13251 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13253 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13256 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13258 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13260 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13263 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13265 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13267 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13270 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13272 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13274 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13277 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13279 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13281 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13284 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13286 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13288 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13290 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13292 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13294 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13296 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13298 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13300 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13302 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13304 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13306 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13308 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13310 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13312 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13314 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13316 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13318 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13320 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13322 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13324 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13326 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13328 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13330 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13332 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13334 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13336 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13338 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13340 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13342 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13344 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13346 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13348 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13350 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13352 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13354 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13356 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13358 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13360 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13362 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13364 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13366 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13368 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13370 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13372 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13374 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13376 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13378 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13380 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13382 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13384 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
13386 ; AVX2-FCP-NEXT: addq $2968, %rsp # imm = 0xB98
13391 ; AVX512: # %bb.0:
13392 ; AVX512-NEXT: subq $3720, %rsp # imm = 0xE88
13410 ; AVX512-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]
13415 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13418 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13421 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13424 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13427 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13430 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13432 ; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13434 ; AVX512-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
13437 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13439 ; AVX512-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
13442 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13444 ; AVX512-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
13447 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13450 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13451 ; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,8,1,9,0,8,1,9]
13452 ; AVX512-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
13454 ; AVX512-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13457 ; AVX512-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13460 ; AVX512-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13462 ; AVX512-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13465 ; AVX512-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13468 ; AVX512-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13471 ; AVX512-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13473 ; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13476 ; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13479 ; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13482 ; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13484 ; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13487 ; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13490 ; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13493 ; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13495 ; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13498 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13501 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13504 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13506 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13509 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13512 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13516 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13518 ; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13533 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13538 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13543 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13548 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13553 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13556 ; AVX512-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13558 ; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13563 ; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13565 ; AVX512-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13567 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13568 ; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
13575 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13586 ; AVX512-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
13589 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13591 ; AVX512-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]
13594 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13596 ; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
13598 ; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13601 ; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13605 ; AVX512-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13608 ; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13613 ; AVX512-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13619 ; AVX512-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13622 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13628 ; AVX512-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13637 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13646 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13650 ; AVX512-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13654 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
13656 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
13657 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13659 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
13661 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
13663 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
13665 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
13667 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
13671 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13672 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
13674 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13675 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
13677 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13678 ; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm0[0,1,2,3],zmm29[4,5,6,7]
13679 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13680 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
13682 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13684 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13686 ; AVX512-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
13688 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
13689 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13691 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13693 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13694 ; AVX512-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm15 # 64-byte Folded Reload
13695 ; AVX512-NEXT: # zmm15 = zmm0[0,1,2,3],mem[4,5,6,7]
13696 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13698 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13700 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13701 ; AVX512-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm13 # 64-byte Folded Reload
13702 ; AVX512-NEXT: # zmm13 = zmm0[0,1,2,3],mem[4,5,6,7]
13703 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
13704 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13706 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13708 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13709 ; AVX512-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
13710 ; AVX512-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
13711 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13713 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13715 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13716 ; AVX512-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
13717 ; AVX512-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
13718 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13719 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13721 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13723 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13724 ; AVX512-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
13725 ; AVX512-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
13726 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13727 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13730 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,9,0,4,5,6,7]
13732 ; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13735 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13738 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13741 ; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13744 ; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13747 ; AVX512-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13753 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
13755 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,13,0,4,5,6,7]
13757 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13759 ; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13761 ; AVX512-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13763 ; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13765 ; AVX512-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13767 ; AVX512-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13770 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
13771 ; AVX512-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm3 # 64-byte Folded Reload
13772 ; AVX512-NEXT: # zmm3 = zmm1[0,1,2,3],mem[4,5,6,7]
13773 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm1 = [14,0,2,3,4,5,15,0]
13775 ; AVX512-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13779 ; AVX512-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13782 ; AVX512-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13783 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
13785 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
13788 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13791 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
13792 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm3
13795 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
13796 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm5, %zmm5
13800 ; AVX512-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13801 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm1 = [10,0,2,3,4,5,11,0]
13803 ; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13806 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
13807 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm31
13809 ; AVX512-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13813 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13814 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm30
13816 ; AVX512-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13820 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13821 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm26
13823 ; AVX512-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13827 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13828 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm25
13830 ; AVX512-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13834 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13835 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm19
13837 ; AVX512-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13841 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13842 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm16
13844 ; AVX512-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13847 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13849 ; AVX512-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13851 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,9,4,5,6,7]
13852 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13854 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13856 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13858 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13860 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13862 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13864 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13866 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13868 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13870 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13872 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13874 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13877 ; AVX512-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13880 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,13,4,5,6,7]
13881 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13883 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13884 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13886 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13887 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13889 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13890 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13892 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13893 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13895 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13896 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13898 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13900 ; AVX512-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13902 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,14,2,3,4,5,6,15]
13903 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13905 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13908 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13910 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13912 ; AVX512-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13913 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13915 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13917 ; AVX512-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13919 ; AVX512-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13920 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13922 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13924 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
13926 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13927 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm0 {%k1}
13929 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
13931 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
13932 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm14 {%k1}
13934 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
13936 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
13937 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm15 {%k1}
13939 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
13941 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
13942 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm10 {%k1}
13944 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
13946 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
13947 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm12 {%k1}
13949 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
13951 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
13952 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm2 {%k1}
13954 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
13956 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
13957 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm9 {%k1}
13959 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
13961 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
13962 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm1 {%k1}
13964 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,1,2,3,4,8,6,7]
13980 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,10,2,3,4,5,6,11]
13981 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
13983 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,12,6,7]
13984 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
13986 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
13988 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
13991 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
13994 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
13997 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
14000 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
14003 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
14006 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14008 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
14009 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
14015 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14017 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14021 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14023 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14025 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14029 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14031 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14033 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14037 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14039 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14041 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14045 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14047 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14049 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14053 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14057 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14061 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14063 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14065 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14069 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14079 ; AVX512-NEXT: addq $3720, %rsp # imm = 0xE88
14084 ; AVX512-FCP: # %bb.0:
14085 ; AVX512-FCP-NEXT: subq $3720, %rsp # imm = 0xE88
14103 ; AVX512-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]
14108 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14111 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14114 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14117 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14120 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14123 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14125 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14127 ; AVX512-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
14130 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14132 ; AVX512-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
14135 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14137 ; AVX512-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
14140 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14143 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14144 ; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,8,1,9,0,8,1,9]
14145 ; AVX512-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
14147 ; AVX512-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14150 ; AVX512-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14153 ; AVX512-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14155 ; AVX512-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14158 ; AVX512-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14161 ; AVX512-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14164 ; AVX512-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14166 ; AVX512-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14169 ; AVX512-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14172 ; AVX512-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14175 ; AVX512-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14177 ; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14180 ; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14183 ; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14186 ; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14188 ; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14191 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14194 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14197 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14199 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14202 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14205 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14209 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14211 ; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14226 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14231 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14236 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14241 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14246 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14249 ; AVX512-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14251 ; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14256 ; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14258 ; AVX512-FCP-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14260 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14261 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
14268 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14279 ; AVX512-FCP-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
14282 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14284 ; AVX512-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]
14287 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14289 ; AVX512-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
14291 ; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14294 ; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14298 ; AVX512-FCP-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14301 ; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14306 ; AVX512-FCP-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14312 ; AVX512-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14315 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14321 ; AVX512-FCP-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14330 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14339 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14343 ; AVX512-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14347 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
14349 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
14350 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14352 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
14354 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
14356 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
14358 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
14360 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
14364 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14365 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
14367 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14368 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
14370 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14371 ; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm0[0,1,2,3],zmm29[4,5,6,7]
14372 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14373 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
14375 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14377 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14379 ; AVX512-FCP-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
14381 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
14382 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14384 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14386 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14387 ; AVX512-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm15 # 64-byte Folded Reload
14388 ; AVX512-FCP-NEXT: # zmm15 = zmm0[0,1,2,3],mem[4,5,6,7]
14389 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14391 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14393 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14394 ; AVX512-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm13 # 64-byte Folded Reload
14395 ; AVX512-FCP-NEXT: # zmm13 = zmm0[0,1,2,3],mem[4,5,6,7]
14396 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
14397 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14399 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14401 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14402 ; AVX512-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
14403 ; AVX512-FCP-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
14404 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14406 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14408 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14409 ; AVX512-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
14410 ; AVX512-FCP-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
14411 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14412 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14414 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14416 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14417 ; AVX512-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
14418 ; AVX512-FCP-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
14419 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14420 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14423 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,9,0,4,5,6,7]
14425 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14428 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14431 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14434 ; AVX512-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14437 ; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14440 ; AVX512-FCP-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14446 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
14448 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,13,0,4,5,6,7]
14450 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14452 ; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14454 ; AVX512-FCP-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14456 ; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14458 ; AVX512-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14460 ; AVX512-FCP-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14463 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
14464 ; AVX512-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm3 # 64-byte Folded Reload
14465 ; AVX512-FCP-NEXT: # zmm3 = zmm1[0,1,2,3],mem[4,5,6,7]
14466 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [14,0,2,3,4,5,15,0]
14468 ; AVX512-FCP-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14472 ; AVX512-FCP-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14475 ; AVX512-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14476 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
14478 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
14481 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14484 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
14485 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm3
14488 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
14489 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm5, %zmm5
14493 ; AVX512-FCP-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14494 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [10,0,2,3,4,5,11,0]
14496 ; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14499 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
14500 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm31
14502 ; AVX512-FCP-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14506 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14507 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm30
14509 ; AVX512-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14513 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14514 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm26
14516 ; AVX512-FCP-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14520 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14521 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm25
14523 ; AVX512-FCP-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14527 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14528 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm19
14530 ; AVX512-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14534 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14535 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm16
14537 ; AVX512-FCP-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14540 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14542 ; AVX512-FCP-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14544 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,9,4,5,6,7]
14545 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14547 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14549 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14551 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14553 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14555 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14557 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14559 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14561 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14563 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14565 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14567 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14570 ; AVX512-FCP-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14573 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,13,4,5,6,7]
14574 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14576 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14577 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14579 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14580 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14582 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14583 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14585 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14586 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14588 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14589 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14591 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14593 ; AVX512-FCP-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14595 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,14,2,3,4,5,6,15]
14596 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14598 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14601 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14603 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14605 ; AVX512-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14606 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14608 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14610 ; AVX512-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14612 ; AVX512-FCP-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14613 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14615 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14617 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
14619 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14620 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm0 {%k1}
14622 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
14624 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
14625 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm14 {%k1}
14627 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
14629 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
14630 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm15 {%k1}
14632 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
14634 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
14635 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm10 {%k1}
14637 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
14639 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
14640 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm12 {%k1}
14642 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
14644 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
14645 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm2 {%k1}
14647 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
14649 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
14650 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm9 {%k1}
14652 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
14654 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
14655 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm1 {%k1}
14657 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,1,2,3,4,8,6,7]
14673 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,10,2,3,4,5,6,11]
14674 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
14676 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,12,6,7]
14677 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
14679 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
14681 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
14684 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
14687 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
14690 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
14693 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
14696 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
14699 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14701 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
14702 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
14708 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14710 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14714 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14716 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14718 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14722 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14724 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14726 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14730 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14732 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14734 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14738 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14740 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14742 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14746 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14750 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14754 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14756 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14758 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14762 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
14772 ; AVX512-FCP-NEXT: addq $3720, %rsp # imm = 0xE88
14777 ; AVX512DQ: # %bb.0:
14778 ; AVX512DQ-NEXT: subq $3720, %rsp # imm = 0xE88
14796 ; AVX512DQ-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]
14801 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14804 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14807 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14810 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14813 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14816 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14818 ; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14820 ; AVX512DQ-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
14823 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14825 ; AVX512DQ-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
14828 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14830 ; AVX512DQ-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
14833 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14836 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14837 ; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,8,1,9,0,8,1,9]
14838 ; AVX512DQ-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
14840 ; AVX512DQ-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14843 ; AVX512DQ-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14846 ; AVX512DQ-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14848 ; AVX512DQ-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14851 ; AVX512DQ-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14854 ; AVX512DQ-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14857 ; AVX512DQ-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14859 ; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14862 ; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14865 ; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14868 ; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14870 ; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14873 ; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14876 ; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14879 ; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14881 ; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14884 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14887 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14890 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14892 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14895 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14898 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14902 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14904 ; AVX512DQ-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14919 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14924 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14929 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14934 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14939 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14942 ; AVX512DQ-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14944 ; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14949 ; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14951 ; AVX512DQ-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14953 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14954 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
14961 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14972 ; AVX512DQ-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
14975 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14977 ; AVX512DQ-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]
14980 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14982 ; AVX512DQ-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
14984 ; AVX512DQ-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14987 ; AVX512DQ-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14991 ; AVX512DQ-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14994 ; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
14999 ; AVX512DQ-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15005 ; AVX512DQ-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15008 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15014 ; AVX512DQ-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15023 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15032 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15036 ; AVX512DQ-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15040 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
15042 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
15043 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15045 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
15047 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
15049 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
15051 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
15053 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
15057 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15058 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
15060 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15061 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
15063 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15064 ; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm0[0,1,2,3],zmm29[4,5,6,7]
15065 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15066 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
15068 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15070 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15072 ; AVX512DQ-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
15074 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
15075 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15077 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15079 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15080 ; AVX512DQ-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm15 # 64-byte Folded Reload
15081 ; AVX512DQ-NEXT: # zmm15 = zmm0[0,1,2,3],mem[4,5,6,7]
15082 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15084 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15086 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15087 ; AVX512DQ-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm13 # 64-byte Folded Reload
15088 ; AVX512DQ-NEXT: # zmm13 = zmm0[0,1,2,3],mem[4,5,6,7]
15089 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
15090 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15092 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15094 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15095 ; AVX512DQ-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
15096 ; AVX512DQ-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
15097 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15099 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15101 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15102 ; AVX512DQ-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
15103 ; AVX512DQ-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
15104 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15105 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15107 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15109 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15110 ; AVX512DQ-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
15111 ; AVX512DQ-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
15112 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15113 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15116 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,9,0,4,5,6,7]
15118 ; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15121 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15124 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15127 ; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15130 ; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15133 ; AVX512DQ-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15139 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
15141 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,13,0,4,5,6,7]
15143 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15145 ; AVX512DQ-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15147 ; AVX512DQ-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15149 ; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15151 ; AVX512DQ-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15153 ; AVX512DQ-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15156 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
15157 ; AVX512DQ-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm3 # 64-byte Folded Reload
15158 ; AVX512DQ-NEXT: # zmm3 = zmm1[0,1,2,3],mem[4,5,6,7]
15159 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm1 = [14,0,2,3,4,5,15,0]
15161 ; AVX512DQ-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15165 ; AVX512DQ-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15168 ; AVX512DQ-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15169 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
15171 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
15174 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15177 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
15178 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm3
15181 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
15182 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm5, %zmm5
15186 ; AVX512DQ-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15187 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm1 = [10,0,2,3,4,5,11,0]
15189 ; AVX512DQ-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15192 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
15193 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm31
15195 ; AVX512DQ-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15199 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15200 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm30
15202 ; AVX512DQ-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15206 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15207 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm26
15209 ; AVX512DQ-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15213 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15214 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm25
15216 ; AVX512DQ-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15220 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15221 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm19
15223 ; AVX512DQ-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15227 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15228 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm16
15230 ; AVX512DQ-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15233 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15235 ; AVX512DQ-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15237 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,9,4,5,6,7]
15238 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15240 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15242 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15244 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15246 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15248 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15250 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15252 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15254 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15256 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15258 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15260 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15263 ; AVX512DQ-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15266 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,13,4,5,6,7]
15267 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15269 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15270 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15272 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15273 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15275 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15276 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15278 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15279 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15281 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15282 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15284 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15286 ; AVX512DQ-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15288 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,14,2,3,4,5,6,15]
15289 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15291 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15294 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15296 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15298 ; AVX512DQ-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15299 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15301 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15303 ; AVX512DQ-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15305 ; AVX512DQ-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15306 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15308 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15310 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
15312 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15313 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm0 {%k1}
15315 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
15317 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
15318 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm14 {%k1}
15320 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
15322 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
15323 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm15 {%k1}
15325 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
15327 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
15328 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm10 {%k1}
15330 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
15332 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
15333 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm12 {%k1}
15335 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
15337 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
15338 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm2 {%k1}
15340 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
15342 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
15343 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm9 {%k1}
15345 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
15347 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
15348 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm1 {%k1}
15350 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,1,2,3,4,8,6,7]
15366 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,10,2,3,4,5,6,11]
15367 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
15369 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,12,6,7]
15370 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
15372 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
15374 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
15377 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
15380 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
15383 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
15386 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
15389 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
15392 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15394 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
15395 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
15401 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15403 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15407 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15409 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15411 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15415 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15417 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15419 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15423 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15425 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15427 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15431 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15433 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15435 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15439 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15443 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15447 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15449 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15451 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15455 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15465 ; AVX512DQ-NEXT: addq $3720, %rsp # imm = 0xE88
15470 ; AVX512DQ-FCP: # %bb.0:
15471 ; AVX512DQ-FCP-NEXT: subq $3720, %rsp # imm = 0xE88
15489 ; AVX512DQ-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]
15494 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15497 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15500 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15503 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15506 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15509 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15511 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15513 ; AVX512DQ-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
15516 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15518 ; AVX512DQ-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
15521 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15523 ; AVX512DQ-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
15526 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15529 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15530 ; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,8,1,9,0,8,1,9]
15531 ; AVX512DQ-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
15533 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15536 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15539 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15541 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15544 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15547 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15550 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15552 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15555 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15558 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15561 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15563 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15566 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15569 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15572 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15574 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15577 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15580 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15583 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15585 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15588 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15591 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15595 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15597 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15612 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15617 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15622 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15627 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15632 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15635 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15637 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15642 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15644 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15646 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15647 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
15654 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15665 ; AVX512DQ-FCP-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
15668 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15670 ; AVX512DQ-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]
15673 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15675 ; AVX512DQ-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
15677 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15680 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15684 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15687 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15692 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15698 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15701 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15707 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15716 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15725 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15729 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15733 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
15735 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
15736 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15738 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
15740 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
15742 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
15744 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
15746 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
15750 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15751 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
15753 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15754 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
15756 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15757 ; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm0[0,1,2,3],zmm29[4,5,6,7]
15758 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15759 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
15761 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15763 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15765 ; AVX512DQ-FCP-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
15767 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
15768 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15770 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15772 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15773 ; AVX512DQ-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm15 # 64-byte Folded Reload
15774 ; AVX512DQ-FCP-NEXT: # zmm15 = zmm0[0,1,2,3],mem[4,5,6,7]
15775 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15777 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15779 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15780 ; AVX512DQ-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm13 # 64-byte Folded Reload
15781 ; AVX512DQ-FCP-NEXT: # zmm13 = zmm0[0,1,2,3],mem[4,5,6,7]
15782 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
15783 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15785 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15787 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15788 ; AVX512DQ-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
15789 ; AVX512DQ-FCP-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
15790 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15792 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15794 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15795 ; AVX512DQ-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
15796 ; AVX512DQ-FCP-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
15797 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15798 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15800 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15802 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15803 ; AVX512DQ-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
15804 ; AVX512DQ-FCP-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
15805 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15806 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15809 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,9,0,4,5,6,7]
15811 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15814 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15817 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15820 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15823 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15826 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15832 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
15834 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,13,0,4,5,6,7]
15836 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15838 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15840 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15842 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15844 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15846 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15849 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
15850 ; AVX512DQ-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm3 # 64-byte Folded Reload
15851 ; AVX512DQ-FCP-NEXT: # zmm3 = zmm1[0,1,2,3],mem[4,5,6,7]
15852 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [14,0,2,3,4,5,15,0]
15854 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15858 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15861 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15862 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
15864 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
15867 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15870 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
15871 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm3
15874 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
15875 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm5, %zmm5
15879 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15880 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [10,0,2,3,4,5,11,0]
15882 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15885 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
15886 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm31
15888 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15892 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15893 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm30
15895 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15899 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15900 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm26
15902 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15906 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15907 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm25
15909 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15913 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15914 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm19
15916 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15920 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15921 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm16
15923 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15926 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15928 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15930 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,9,4,5,6,7]
15931 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15933 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15935 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15937 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15939 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15941 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15943 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15945 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15947 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15949 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15951 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15953 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15956 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15959 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,13,4,5,6,7]
15960 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15962 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15963 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15965 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15966 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15968 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15969 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15971 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15972 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15974 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15975 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15977 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15979 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15981 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,14,2,3,4,5,6,15]
15982 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15984 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15987 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15989 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15991 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15992 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
15994 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15996 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15998 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
15999 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16001 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16003 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16005 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16006 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm0 {%k1}
16008 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16010 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
16011 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm14 {%k1}
16013 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16015 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
16016 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm15 {%k1}
16018 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16020 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
16021 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm10 {%k1}
16023 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16025 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
16026 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm12 {%k1}
16028 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16030 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
16031 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm2 {%k1}
16033 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16035 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
16036 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm9 {%k1}
16038 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16040 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
16041 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm1 {%k1}
16043 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,1,2,3,4,8,6,7]
16059 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,10,2,3,4,5,6,11]
16060 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
16062 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,12,6,7]
16063 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
16065 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
16067 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
16070 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
16073 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
16076 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
16079 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
16082 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
16085 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16087 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
16088 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
16094 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16096 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16100 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16102 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16104 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16108 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16110 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16112 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16116 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16118 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16120 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16124 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16126 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16128 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16132 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16136 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16140 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16142 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16144 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16148 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16158 ; AVX512DQ-FCP-NEXT: addq $3720, %rsp # imm = 0xE88
16163 ; AVX512BW: # %bb.0:
16164 ; AVX512BW-NEXT: subq $3720, %rsp # imm = 0xE88
16182 ; AVX512BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]
16187 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16190 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16193 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16196 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16199 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16202 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16204 ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16206 ; AVX512BW-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
16209 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16211 ; AVX512BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
16214 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16216 ; AVX512BW-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
16219 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16222 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16223 ; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,8,1,9,0,8,1,9]
16224 ; AVX512BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
16226 ; AVX512BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16229 ; AVX512BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16232 ; AVX512BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16234 ; AVX512BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16237 ; AVX512BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16240 ; AVX512BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16243 ; AVX512BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16245 ; AVX512BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16248 ; AVX512BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16251 ; AVX512BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16254 ; AVX512BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16256 ; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16259 ; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16262 ; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16265 ; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16267 ; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16270 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16273 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16276 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16278 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16281 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16284 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16288 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16290 ; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16305 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16310 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16315 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16320 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16325 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16328 ; AVX512BW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16330 ; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16335 ; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16337 ; AVX512BW-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16339 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16340 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
16347 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16358 ; AVX512BW-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
16361 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16363 ; AVX512BW-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]
16366 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16368 ; AVX512BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
16370 ; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16373 ; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16377 ; AVX512BW-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16380 ; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16385 ; AVX512BW-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16391 ; AVX512BW-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16394 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16400 ; AVX512BW-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16409 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16418 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16422 ; AVX512BW-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16426 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
16428 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
16429 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16431 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
16433 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
16435 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
16437 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
16439 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
16443 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16444 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
16446 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16447 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
16449 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16450 ; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm0[0,1,2,3],zmm29[4,5,6,7]
16451 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16452 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
16454 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16456 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16458 ; AVX512BW-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
16460 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
16461 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16463 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16465 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16466 ; AVX512BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm15 # 64-byte Folded Reload
16467 ; AVX512BW-NEXT: # zmm15 = zmm0[0,1,2,3],mem[4,5,6,7]
16468 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16470 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16472 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16473 ; AVX512BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm13 # 64-byte Folded Reload
16474 ; AVX512BW-NEXT: # zmm13 = zmm0[0,1,2,3],mem[4,5,6,7]
16475 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
16476 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16478 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16480 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16481 ; AVX512BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
16482 ; AVX512BW-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
16483 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16485 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16487 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16488 ; AVX512BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
16489 ; AVX512BW-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
16490 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16491 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16493 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16495 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16496 ; AVX512BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
16497 ; AVX512BW-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
16498 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16499 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16502 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,9,0,4,5,6,7]
16504 ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16507 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16510 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16513 ; AVX512BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16516 ; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16519 ; AVX512BW-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16525 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
16527 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,13,0,4,5,6,7]
16529 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16531 ; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16533 ; AVX512BW-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16535 ; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16537 ; AVX512BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16539 ; AVX512BW-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16542 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
16543 ; AVX512BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm3 # 64-byte Folded Reload
16544 ; AVX512BW-NEXT: # zmm3 = zmm1[0,1,2,3],mem[4,5,6,7]
16545 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [14,0,2,3,4,5,15,0]
16547 ; AVX512BW-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16551 ; AVX512BW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16554 ; AVX512BW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16555 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
16557 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
16560 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16563 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
16564 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm3
16567 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
16568 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm5, %zmm5
16572 ; AVX512BW-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16573 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [10,0,2,3,4,5,11,0]
16575 ; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16578 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
16579 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm31
16581 ; AVX512BW-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16585 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16586 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm30
16588 ; AVX512BW-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16592 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16593 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm26
16595 ; AVX512BW-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16599 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16600 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm25
16602 ; AVX512BW-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16606 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16607 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm19
16609 ; AVX512BW-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16613 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16614 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm16
16616 ; AVX512BW-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16619 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16621 ; AVX512BW-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16623 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,9,4,5,6,7]
16624 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16626 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16628 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16630 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16632 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16634 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16636 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16638 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16640 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16642 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16644 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16646 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16649 ; AVX512BW-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16652 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,13,4,5,6,7]
16653 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16655 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16656 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16658 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16659 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16661 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16662 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16664 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16665 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16667 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16668 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16670 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16672 ; AVX512BW-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16674 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,14,2,3,4,5,6,15]
16675 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16677 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16680 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16682 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16684 ; AVX512BW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16685 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16687 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16689 ; AVX512BW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16691 ; AVX512BW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16692 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16694 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16696 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16698 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16699 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm0 {%k1}
16701 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16703 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
16704 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm14 {%k1}
16706 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16708 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
16709 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm15 {%k1}
16711 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16713 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
16714 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm10 {%k1}
16716 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16718 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
16719 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm12 {%k1}
16721 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16723 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
16724 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm2 {%k1}
16726 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16728 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
16729 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm9 {%k1}
16731 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
16733 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
16734 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm1 {%k1}
16736 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,1,2,3,4,8,6,7]
16752 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,10,2,3,4,5,6,11]
16753 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
16755 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,12,6,7]
16756 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
16758 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
16760 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
16763 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
16766 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
16769 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
16772 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
16775 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
16778 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16780 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
16781 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
16787 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16789 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16793 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16795 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16797 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16801 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16803 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16805 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16809 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16811 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16813 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16817 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16819 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16821 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16825 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16829 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16833 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16835 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16837 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16841 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
16851 ; AVX512BW-NEXT: addq $3720, %rsp # imm = 0xE88
16856 ; AVX512BW-FCP: # %bb.0:
16857 ; AVX512BW-FCP-NEXT: subq $3720, %rsp # imm = 0xE88
16875 ; AVX512BW-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]
16880 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16883 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16886 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16889 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16892 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16895 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16897 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16899 ; AVX512BW-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
16902 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16904 ; AVX512BW-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
16907 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16909 ; AVX512BW-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
16912 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16915 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16916 ; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,8,1,9,0,8,1,9]
16917 ; AVX512BW-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
16919 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16922 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16925 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16927 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16930 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16933 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16936 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16938 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16941 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16944 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16947 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16949 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16952 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16955 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16958 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16960 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16963 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16966 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16969 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16971 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16974 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16977 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16981 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16983 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
16998 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17003 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17008 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17013 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17018 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17021 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17023 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17028 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17030 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17032 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17033 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
17040 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17051 ; AVX512BW-FCP-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
17054 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17056 ; AVX512BW-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]
17059 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17061 ; AVX512BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
17063 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17066 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17070 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17073 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17078 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17084 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17087 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17093 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17102 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17111 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17115 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17119 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
17121 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
17122 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17124 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
17126 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
17128 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
17130 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
17132 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
17136 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17137 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
17139 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17140 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
17142 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17143 ; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm0[0,1,2,3],zmm29[4,5,6,7]
17144 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17145 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
17147 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17149 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17151 ; AVX512BW-FCP-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
17153 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
17154 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17156 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17158 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17159 ; AVX512BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm15 # 64-byte Folded Reload
17160 ; AVX512BW-FCP-NEXT: # zmm15 = zmm0[0,1,2,3],mem[4,5,6,7]
17161 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17163 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17165 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17166 ; AVX512BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm13 # 64-byte Folded Reload
17167 ; AVX512BW-FCP-NEXT: # zmm13 = zmm0[0,1,2,3],mem[4,5,6,7]
17168 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
17169 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17171 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17173 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17174 ; AVX512BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
17175 ; AVX512BW-FCP-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
17176 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17178 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17180 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17181 ; AVX512BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
17182 ; AVX512BW-FCP-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
17183 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17184 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17186 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17188 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17189 ; AVX512BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
17190 ; AVX512BW-FCP-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
17191 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17192 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17195 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,9,0,4,5,6,7]
17197 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17200 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17203 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17206 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17209 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17212 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17218 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
17220 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,13,0,4,5,6,7]
17222 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17224 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17226 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17228 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17230 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17232 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17235 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
17236 ; AVX512BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm3 # 64-byte Folded Reload
17237 ; AVX512BW-FCP-NEXT: # zmm3 = zmm1[0,1,2,3],mem[4,5,6,7]
17238 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [14,0,2,3,4,5,15,0]
17240 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17244 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17247 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17248 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
17250 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
17253 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17256 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
17257 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm3
17260 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
17261 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm5, %zmm5
17265 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17266 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [10,0,2,3,4,5,11,0]
17268 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17271 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
17272 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm31
17274 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17278 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17279 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm30
17281 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17285 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17286 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm26
17288 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17292 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17293 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm25
17295 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17299 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17300 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm19
17302 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17306 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17307 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm16
17309 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17312 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17314 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17316 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,9,4,5,6,7]
17317 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17319 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17321 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17323 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17325 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17327 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17329 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17331 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17333 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17335 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17337 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17339 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17342 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17345 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,13,4,5,6,7]
17346 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17348 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17349 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17351 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17352 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17354 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17355 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17357 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17358 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17360 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17361 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17363 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17365 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17367 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,14,2,3,4,5,6,15]
17368 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17370 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17373 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17375 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17377 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17378 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17380 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17382 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17384 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17385 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17387 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17389 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
17391 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17392 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm0 {%k1}
17394 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
17396 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
17397 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm14 {%k1}
17399 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
17401 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
17402 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm15 {%k1}
17404 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
17406 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
17407 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm10 {%k1}
17409 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
17411 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
17412 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm12 {%k1}
17414 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
17416 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
17417 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm2 {%k1}
17419 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
17421 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
17422 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm9 {%k1}
17424 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
17426 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
17427 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm1 {%k1}
17429 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,1,2,3,4,8,6,7]
17445 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,10,2,3,4,5,6,11]
17446 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
17448 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,12,6,7]
17449 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
17451 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
17453 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
17456 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
17459 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
17462 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
17465 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
17468 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
17471 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17473 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
17474 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
17480 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17482 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17486 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17488 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17490 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17494 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17496 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17498 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17502 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17504 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17506 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17510 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17512 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17514 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17518 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17522 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17526 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17528 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17530 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17534 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17544 ; AVX512BW-FCP-NEXT: addq $3720, %rsp # imm = 0xE88
17549 ; AVX512DQ-BW: # %bb.0:
17550 ; AVX512DQ-BW-NEXT: subq $3720, %rsp # imm = 0xE88
17568 ; AVX512DQ-BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]
17573 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17576 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17579 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17582 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17585 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17588 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17590 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17592 ; AVX512DQ-BW-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
17595 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17597 ; AVX512DQ-BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
17600 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17602 ; AVX512DQ-BW-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
17605 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17608 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17609 ; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,8,1,9,0,8,1,9]
17610 ; AVX512DQ-BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
17612 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17615 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17618 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17620 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17623 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17626 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17629 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17631 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17634 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17637 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17640 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17642 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17645 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17648 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17651 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17653 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17656 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17659 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17662 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17664 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17667 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17670 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17674 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17676 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17691 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17696 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17701 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17706 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17711 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17714 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17716 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17721 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17723 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17725 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17726 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
17733 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17744 ; AVX512DQ-BW-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
17747 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17749 ; AVX512DQ-BW-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]
17752 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17754 ; AVX512DQ-BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
17756 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17759 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17763 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17766 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17771 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17777 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17780 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17786 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17795 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17804 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17808 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17812 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
17814 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
17815 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17817 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
17819 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
17821 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
17823 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
17825 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
17829 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17830 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
17832 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17833 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
17835 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17836 ; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm0[0,1,2,3],zmm29[4,5,6,7]
17837 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17838 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
17840 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17842 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17844 ; AVX512DQ-BW-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
17846 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
17847 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17849 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17851 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17852 ; AVX512DQ-BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm15 # 64-byte Folded Reload
17853 ; AVX512DQ-BW-NEXT: # zmm15 = zmm0[0,1,2,3],mem[4,5,6,7]
17854 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17856 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17858 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17859 ; AVX512DQ-BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm13 # 64-byte Folded Reload
17860 ; AVX512DQ-BW-NEXT: # zmm13 = zmm0[0,1,2,3],mem[4,5,6,7]
17861 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
17862 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17864 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17866 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17867 ; AVX512DQ-BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
17868 ; AVX512DQ-BW-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
17869 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17871 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17873 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17874 ; AVX512DQ-BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
17875 ; AVX512DQ-BW-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
17876 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17877 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17879 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17881 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17882 ; AVX512DQ-BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
17883 ; AVX512DQ-BW-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
17884 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17885 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17888 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,9,0,4,5,6,7]
17890 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17893 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17896 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17899 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17902 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17905 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17911 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
17913 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,13,0,4,5,6,7]
17915 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17917 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17919 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17921 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17923 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17925 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17928 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
17929 ; AVX512DQ-BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm3 # 64-byte Folded Reload
17930 ; AVX512DQ-BW-NEXT: # zmm3 = zmm1[0,1,2,3],mem[4,5,6,7]
17931 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [14,0,2,3,4,5,15,0]
17933 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17937 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17940 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17941 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
17943 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
17946 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17949 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
17950 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm3
17953 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
17954 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm5, %zmm5
17958 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17959 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [10,0,2,3,4,5,11,0]
17961 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17964 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
17965 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm31
17967 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17971 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17972 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm30
17974 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17978 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17979 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm26
17981 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17985 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17986 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm25
17988 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17992 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
17993 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm19
17995 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
17999 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18000 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm16
18002 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18005 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18007 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18009 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,9,4,5,6,7]
18010 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18012 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18014 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18016 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18018 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18020 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18022 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18024 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18026 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18028 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18030 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18032 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18035 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18038 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,13,4,5,6,7]
18039 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18041 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18042 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18044 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18045 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18047 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18048 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18050 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18051 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18053 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18054 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18056 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18058 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18060 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,14,2,3,4,5,6,15]
18061 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18063 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18066 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18068 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18070 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18071 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18073 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18075 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18077 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18078 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18080 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18082 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18084 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18085 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm0 {%k1}
18087 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18089 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
18090 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm14 {%k1}
18092 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18094 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
18095 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm15 {%k1}
18097 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18099 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
18100 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm10 {%k1}
18102 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18104 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
18105 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm12 {%k1}
18107 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18109 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
18110 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm2 {%k1}
18112 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18114 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
18115 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm9 {%k1}
18117 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18119 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
18120 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm1 {%k1}
18122 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,1,2,3,4,8,6,7]
18138 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,10,2,3,4,5,6,11]
18139 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
18141 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,12,6,7]
18142 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
18144 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
18146 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
18149 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
18152 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
18155 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
18158 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
18161 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
18164 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18166 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
18167 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
18173 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18175 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18179 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18181 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18183 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18187 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18189 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18191 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18195 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18197 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18199 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18203 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18205 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18207 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18211 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18215 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18219 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18221 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18223 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18227 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18237 ; AVX512DQ-BW-NEXT: addq $3720, %rsp # imm = 0xE88
18242 ; AVX512DQ-BW-FCP: # %bb.0:
18243 ; AVX512DQ-BW-FCP-NEXT: subq $3720, %rsp # imm = 0xE88
18261 ; AVX512DQ-BW-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]
18266 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18269 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18272 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18275 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18278 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18281 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18283 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18285 ; AVX512DQ-BW-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
18288 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18290 ; AVX512DQ-BW-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
18293 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18295 ; AVX512DQ-BW-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
18298 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18301 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18302 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,8,1,9,0,8,1,9]
18303 ; AVX512DQ-BW-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
18305 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18308 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18311 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18313 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18316 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18319 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18322 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18324 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18327 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18330 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18333 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18335 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18338 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18341 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18344 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18346 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18349 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18352 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18355 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18357 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18360 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18363 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18367 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18369 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18384 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18389 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18394 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18399 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18404 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18407 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18409 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18414 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18416 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18418 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18419 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,0,4,12]
18426 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18437 ; AVX512DQ-BW-FCP-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3]
18440 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18442 ; AVX512DQ-BW-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]
18445 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18447 ; AVX512DQ-BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
18449 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18452 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18456 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18459 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18464 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18470 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18473 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18479 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18488 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18497 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18501 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18505 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
18507 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
18508 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18510 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
18512 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
18514 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
18516 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
18518 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
18522 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18523 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
18525 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18526 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
18528 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18529 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm0[0,1,2,3],zmm29[4,5,6,7]
18530 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18531 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
18533 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18535 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18537 ; AVX512DQ-BW-FCP-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
18539 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
18540 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18542 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18544 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18545 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm15 # 64-byte Folded Reload
18546 ; AVX512DQ-BW-FCP-NEXT: # zmm15 = zmm0[0,1,2,3],mem[4,5,6,7]
18547 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18549 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18551 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18552 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm13 # 64-byte Folded Reload
18553 ; AVX512DQ-BW-FCP-NEXT: # zmm13 = zmm0[0,1,2,3],mem[4,5,6,7]
18554 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
18555 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18557 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18559 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18560 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
18561 ; AVX512DQ-BW-FCP-NEXT: # zmm12 = zmm0[0,1,2,3],mem[4,5,6,7]
18562 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18564 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18566 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18567 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
18568 ; AVX512DQ-BW-FCP-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
18569 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18570 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18572 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18574 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18575 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
18576 ; AVX512DQ-BW-FCP-NEXT: # zmm0 = zmm0[0,1,2,3],mem[4,5,6,7]
18577 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18578 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18581 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,9,0,4,5,6,7]
18583 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18586 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18589 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18592 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18595 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18598 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18604 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
18606 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,13,0,4,5,6,7]
18608 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18610 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18612 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18614 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18616 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18618 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18621 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
18622 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm3 # 64-byte Folded Reload
18623 ; AVX512DQ-BW-FCP-NEXT: # zmm3 = zmm1[0,1,2,3],mem[4,5,6,7]
18624 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [14,0,2,3,4,5,15,0]
18626 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18630 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18633 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18634 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
18636 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
18639 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18642 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
18643 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm3
18646 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
18647 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm5, %zmm5
18651 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18652 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [10,0,2,3,4,5,11,0]
18654 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18657 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
18658 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm31
18660 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18664 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18665 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm30
18667 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18671 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18672 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm26
18674 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18678 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18679 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm25
18681 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18685 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18686 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm19
18688 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18692 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18693 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm0, %zmm16
18695 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18698 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18700 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18702 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,9,4,5,6,7]
18703 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18705 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18707 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18709 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18711 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18713 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18715 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18717 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18719 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18721 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18723 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18725 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18728 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18731 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,1,2,13,4,5,6,7]
18732 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18734 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18735 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18737 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18738 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18740 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18741 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18743 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18744 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18746 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18747 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18749 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18751 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18753 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,14,2,3,4,5,6,15]
18754 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18756 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18759 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18761 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18763 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18764 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18766 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18768 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18770 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18771 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18773 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18775 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18777 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18778 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm0 {%k1}
18780 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18782 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
18783 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm14 {%k1}
18785 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18787 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
18788 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm15 {%k1}
18790 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18792 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
18793 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm10 {%k1}
18795 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18797 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
18798 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm12 {%k1}
18800 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18802 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
18803 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm2 {%k1}
18805 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18807 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
18808 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm9 {%k1}
18810 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],mem[0]
18812 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
18813 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm0, %zmm1 {%k1}
18815 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,1,2,3,4,8,6,7]
18831 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,10,2,3,4,5,6,11]
18832 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
18834 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,12,6,7]
18835 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
18837 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
18839 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
18842 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
18845 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
18848 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
18851 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
18854 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
18857 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18859 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
18860 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
18866 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18868 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18872 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18874 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18876 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18880 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18882 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18884 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18888 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18890 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18892 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18896 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18898 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18900 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18904 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18908 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18912 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18914 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18916 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18920 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
18930 ; AVX512DQ-BW-FCP-NEXT: addq $3720, %rsp # imm = 0xE88
18939 %1 = shufflevector <64 x i64> %in.vec0, <64 x i64> %in.vec1, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
18940 %2 = shufflevector <64 x i64> %in.vec2, <64 x i64> %in.vec3, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
18941 %3 = shufflevector <64 x i64> %in.vec4, <64 x i64> %in.vec5, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
18942 %4 = shufflevector <128 x i64> %1, <128 x i64> %2, <256 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223, i32 224, i32 225, i32 226, i32 227, i32 228, i32 229, i32 230, i32 231, i32 232, i32 233, i32 234, i32 235, i32 236, i32 237, i32 238, i32 239, i32 240, i32 241, i32 242, i32 243, i32 244, i32 245, i32 246, i32 247, i32 248, i32 249, i32 250, i32 251, i32 252, i32 253, i32 254, i32 255>
18943 %5 = shufflevector <128 x i64> %3, <128 x i64> poison, <256 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
18944 %6 = shufflevector <256 x i64> %4, <256 x i64> %5, <384 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223, i32 224, i32 225, i32 226, i32 227, i32 228, i32 229, i32 230, i32 231, i32 232, i32 233, i32 234, i32 235, i32 236, i32 237, i32 238, i32 239, i32 240, i32 241, i32 242, i32 243, i32 244, i32 245, i32 246, i32 247, i32 248, i32 249, i32 250, i32 251, i32 252, i32 253, i32 254, i32 255, i32 256, i32 257, i32 258, i32 259, i32 260, i32 261, i32 262, i32 263, i32 264, i32 265, i32 266, i32 267, i32 268, i32 269, i32 270, i32 271, i32 272, i32 273, i32 274, i32 275, i32 276, i32 277, i32 278, i32 279, i32 280, i32 281, i32 282, i32 283, i32 284, i32 285, i32 286, i32 287, i32 288, i32 289, i32 290, i32 291, i32 292, i32 293, i32 294, i32 295, i32 296, i32 297, i32 298, i32 299, i32 300, i32 301, i32 302, i32 303, i32 304, i32 305, i32 306, i32 307, i32 308, i32 309, i32 310, i32 311, i32 312, i32 313, i32 314, i32 315, i32 316, i32 317, i32 318, i32 319, i32 320, i32 321, i32 322, i32 323, i32 324, i32 325, i32 326, i32 327, i32 328, i32 329, i32 330, i32 331, i32 332, i32 333, i32 334, i32 335, i32 336, i32 337, i32 338, i32 339, i32 340, i32 341, i32 342, i32 343, i32 344, i32 345, i32 346, i32 347, i32 348, i32 349, i32 350, i32 351, i32 352, i32 353, i32 354, i32 355, i32 356, i32 357, i32 358, i32 359, i32 360, i32 361, i32 362, i32 363, i32 364, i32 365, i32 366, i32 367, i32 368, i32 369, i32 370, i32 371, i32 372, i32 373, i32 374, i32 375, i32 376, i32 377, i32 378, i32 379, i32 380, i32 381, i32 382, i32 383>
18945 %interleaved.vec = shufflevector <384 x i64> %6, <384 x i64> poison, <384 x i32> <i32 0, i32 64, i32 128, i32 192, i32 256, i32 320, i32 1, i32 65, i32 129, i32 193, i32 257, i32 321, i32 2, i32 66, i32 130, i32 194, i32 258, i32 322, i32 3, i32 67, i32 131, i32 195, i32 259, i32 323, i32 4, i32 68, i32 132, i32 196, i32 260, i32 324, i32 5, i32 69, i32 133, i32 197, i32 261, i32 325, i32 6, i32 70, i32 134, i32 198, i32 262, i32 326, i32 7, i32 71, i32 135, i32 199, i32 263, i32 327, i32 8, i32 72, i32 136, i32 200, i32 264, i32 328, i32 9, i32 73, i32 137, i32 201, i32 265, i32 329, i32 10, i32 74, i32 138, i32 202, i32 266, i32 330, i32 11, i32 75, i32 139, i32 203, i32 267, i32 331, i32 12, i32 76, i32 140, i32 204, i32 268, i32 332, i32 13, i32 77, i32 141, i32 205, i32 269, i32 333, i32 14, i32 78, i32 142, i32 206, i32 270, i32 334, i32 15, i32 79, i32 143, i32 207, i32 271, i32 335, i32 16, i32 80, i32 144, i32 208, i32 272, i32 336, i32 17, i32 81, i32 145, i32 209, i32 273, i32 337, i32 18, i32 82, i32 146, i32 210, i32 274, i32 338, i32 19, i32 83, i32 147, i32 211, i32 275, i32 339, i32 20, i32 84, i32 148, i32 212, i32 276, i32 340, i32 21, i32 85, i32 149, i32 213, i32 277, i32 341, i32 22, i32 86, i32 150, i32 214, i32 278, i32 342, i32 23, i32 87, i32 151, i32 215, i32 279, i32 343, i32 24, i32 88, i32 152, i32 216, i32 280, i32 344, i32 25, i32 89, i32 153, i32 217, i32 281, i32 345, i32 26, i32 90, i32 154, i32 218, i32 282, i32 346, i32 27, i32 91, i32 155, i32 219, i32 283, i32 347, i32 28, i32 92, i32 156, i32 220, i32 284, i32 348, i32 29, i32 93, i32 157, i32 221, i32 285, i32 349, i32 30, i32 94, i32 158, i32 222, i32 286, i32 350, i32 31, i32 95, i32 159, i32 223, i32 287, i32 351, i32 32, i32 96, i32 160, i32 224, i32 288, i32 352, i32 33, i32 97, i32 161, i32 225, i32 289, i32 353, i32 34, i32 98, i32 162, i32 226, i32 290, i32 354, i32 35, i32 99, i32 163, i32 227, i32 291, i32 355, i32 36, i32 100, i32 164, i32 228, i32 292, i32 356, i32 37, i32 101, i32 165, i32 229, i32 293, i32 357, i32 38, i32 102, i32 166, i32 230, i32 294, i32 358, i32 39, i32 103, i32 167, i32 231, i32 295, i32 359, i32 40, i32 104, i32 168, i32 232, i32 296, i32 360, i32 41, i32 105, i32 169, i32 233, i32 297, i32 361, i32 42, i32 106, i32 170, i32 234, i32 298, i32 362, i32 43, i32 107, i32 171, i32 235, i32 299, i32 363, i32 44, i32 108, i32 172, i32 236, i32 300, i32 364, i32 45, i32 109, i32 173, i32 237, i32 301, i32 365, i32 46, i32 110, i32 174, i32 238, i32 302, i32 366, i32 47, i32 111, i32 175, i32 239, i32 303, i32 367, i32 48, i32 112, i32 176, i32 240, i32 304, i32 368, i32 49, i32 113, i32 177, i32 241, i32 305, i32 369, i32 50, i32 114, i32 178, i32 242, i32 306, i32 370, i32 51, i32 115, i32 179, i32 243, i32 307, i32 371, i32 52, i32 116, i32 180, i32 244, i32 308, i32 372, i32 53, i32 117, i32 181, i32 245, i32 309, i32 373, i32 54, i32 118, i32 182, i32 246, i32 310, i32 374, i32 55, i32 119, i32 183, i32 247, i32 311, i32 375, i32 56, i32 120, i32 184, i32 248, i32 312, i32 376, i32 57, i32 121, i32 185, i32 249, i32 313, i32 377, i32 58, i32 122, i32 186, i32 250, i32 314, i32 378, i32 59, i32 123, i32 187, i32 251, i32 315, i32 379, i32 60, i32 124, i32 188, i32 252, i32 316, i32 380, i32 61, i32 125, i32 189, i32 253, i32 317, i32 381, i32 62, i32 126, i32 190, i32 254, i32 318, i32 382, i32 63, i32 127, i32 191, i32 255, i32 319, i32 383>