Lines Matching +full:0 +full:x2e8
20 ; SSE: # %bb.0:
21 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
22 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10
26 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
27 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,3,2,3]
28 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm0[0,3,2,3,4,5,6,7]
34 ; SSE-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
37 ; SSE-NEXT: punpcklwd {{.*#+}} xmm6 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3]
39 ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm5[0,3,2,3,4,5,6,7]
41 ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,6,5,6,7]
48 ; SSE-NEXT: punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3]
52 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[0,3,2,3,4,5,6,7]
71 ; AVX: # %bb.0:
72 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
73 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10
75 ; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
82 ; AVX-NEXT: vpextrw $0, %xmm1, (%rsi)
83 ; AVX-NEXT: vpextrw $0, %xmm2, (%rdx)
84 ; AVX-NEXT: vpextrw $0, %xmm3, (%rcx)
85 ; AVX-NEXT: vpextrw $0, %xmm4, (%r8)
86 ; AVX-NEXT: vpextrw $0, %xmm5, (%r9)
87 ; AVX-NEXT: vpextrw $0, %xmm6, (%r10)
88 ; AVX-NEXT: vpextrw $0, %xmm0, (%rax)
92 ; AVX2: # %bb.0:
93 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
94 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10
96 ; AVX2-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
103 ; AVX2-NEXT: vpextrw $0, %xmm1, (%rsi)
104 ; AVX2-NEXT: vpextrw $0, %xmm2, (%rdx)
105 ; AVX2-NEXT: vpextrw $0, %xmm3, (%rcx)
106 ; AVX2-NEXT: vpextrw $0, %xmm4, (%r8)
107 ; AVX2-NEXT: vpextrw $0, %xmm5, (%r9)
108 ; AVX2-NEXT: vpextrw $0, %xmm6, (%r10)
109 ; AVX2-NEXT: vpextrw $0, %xmm0, (%rax)
113 ; AVX2-FP: # %bb.0:
114 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
115 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10
117 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
124 ; AVX2-FP-NEXT: vpextrw $0, %xmm1, (%rsi)
125 ; AVX2-FP-NEXT: vpextrw $0, %xmm2, (%rdx)
126 ; AVX2-FP-NEXT: vpextrw $0, %xmm3, (%rcx)
127 ; AVX2-FP-NEXT: vpextrw $0, %xmm4, (%r8)
128 ; AVX2-FP-NEXT: vpextrw $0, %xmm5, (%r9)
129 ; AVX2-FP-NEXT: vpextrw $0, %xmm6, (%r10)
130 ; AVX2-FP-NEXT: vpextrw $0, %xmm0, (%rax)
134 ; AVX2-FCP: # %bb.0:
135 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
136 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
138 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
145 ; AVX2-FCP-NEXT: vpextrw $0, %xmm1, (%rsi)
146 ; AVX2-FCP-NEXT: vpextrw $0, %xmm2, (%rdx)
147 ; AVX2-FCP-NEXT: vpextrw $0, %xmm3, (%rcx)
148 ; AVX2-FCP-NEXT: vpextrw $0, %xmm4, (%r8)
149 ; AVX2-FCP-NEXT: vpextrw $0, %xmm5, (%r9)
150 ; AVX2-FCP-NEXT: vpextrw $0, %xmm6, (%r10)
151 ; AVX2-FCP-NEXT: vpextrw $0, %xmm0, (%rax)
155 ; AVX512: # %bb.0:
156 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
157 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10
159 ; AVX512-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
166 ; AVX512-NEXT: vpextrw $0, %xmm1, (%rsi)
167 ; AVX512-NEXT: vpextrw $0, %xmm2, (%rdx)
168 ; AVX512-NEXT: vpextrw $0, %xmm3, (%rcx)
169 ; AVX512-NEXT: vpextrw $0, %xmm4, (%r8)
170 ; AVX512-NEXT: vpextrw $0, %xmm5, (%r9)
171 ; AVX512-NEXT: vpextrw $0, %xmm6, (%r10)
172 ; AVX512-NEXT: vpextrw $0, %xmm0, (%rax)
176 ; AVX512-FCP: # %bb.0:
177 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
178 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
180 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
187 ; AVX512-FCP-NEXT: vpextrw $0, %xmm1, (%rsi)
188 ; AVX512-FCP-NEXT: vpextrw $0, %xmm2, (%rdx)
189 ; AVX512-FCP-NEXT: vpextrw $0, %xmm3, (%rcx)
190 ; AVX512-FCP-NEXT: vpextrw $0, %xmm4, (%r8)
191 ; AVX512-FCP-NEXT: vpextrw $0, %xmm5, (%r9)
192 ; AVX512-FCP-NEXT: vpextrw $0, %xmm6, (%r10)
193 ; AVX512-FCP-NEXT: vpextrw $0, %xmm0, (%rax)
197 ; AVX512DQ: # %bb.0:
198 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
199 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10
201 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
208 ; AVX512DQ-NEXT: vpextrw $0, %xmm1, (%rsi)
209 ; AVX512DQ-NEXT: vpextrw $0, %xmm2, (%rdx)
210 ; AVX512DQ-NEXT: vpextrw $0, %xmm3, (%rcx)
211 ; AVX512DQ-NEXT: vpextrw $0, %xmm4, (%r8)
212 ; AVX512DQ-NEXT: vpextrw $0, %xmm5, (%r9)
213 ; AVX512DQ-NEXT: vpextrw $0, %xmm6, (%r10)
214 ; AVX512DQ-NEXT: vpextrw $0, %xmm0, (%rax)
218 ; AVX512DQ-FCP: # %bb.0:
219 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
220 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
222 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
229 ; AVX512DQ-FCP-NEXT: vpextrw $0, %xmm1, (%rsi)
230 ; AVX512DQ-FCP-NEXT: vpextrw $0, %xmm2, (%rdx)
231 ; AVX512DQ-FCP-NEXT: vpextrw $0, %xmm3, (%rcx)
232 ; AVX512DQ-FCP-NEXT: vpextrw $0, %xmm4, (%r8)
233 ; AVX512DQ-FCP-NEXT: vpextrw $0, %xmm5, (%r9)
234 ; AVX512DQ-FCP-NEXT: vpextrw $0, %xmm6, (%r10)
235 ; AVX512DQ-FCP-NEXT: vpextrw $0, %xmm0, (%rax)
239 ; AVX512BW: # %bb.0:
240 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
241 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
243 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
250 ; AVX512BW-NEXT: vpextrw $0, %xmm1, (%rsi)
251 ; AVX512BW-NEXT: vpextrw $0, %xmm2, (%rdx)
252 ; AVX512BW-NEXT: vpextrw $0, %xmm3, (%rcx)
253 ; AVX512BW-NEXT: vpextrw $0, %xmm4, (%r8)
254 ; AVX512BW-NEXT: vpextrw $0, %xmm5, (%r9)
255 ; AVX512BW-NEXT: vpextrw $0, %xmm6, (%r10)
256 ; AVX512BW-NEXT: vpextrw $0, %xmm0, (%rax)
260 ; AVX512BW-FCP: # %bb.0:
261 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
262 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
264 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
271 ; AVX512BW-FCP-NEXT: vpextrw $0, %xmm1, (%rsi)
272 ; AVX512BW-FCP-NEXT: vpextrw $0, %xmm2, (%rdx)
273 ; AVX512BW-FCP-NEXT: vpextrw $0, %xmm3, (%rcx)
274 ; AVX512BW-FCP-NEXT: vpextrw $0, %xmm4, (%r8)
275 ; AVX512BW-FCP-NEXT: vpextrw $0, %xmm5, (%r9)
276 ; AVX512BW-FCP-NEXT: vpextrw $0, %xmm6, (%r10)
277 ; AVX512BW-FCP-NEXT: vpextrw $0, %xmm0, (%rax)
281 ; AVX512DQ-BW: # %bb.0:
282 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
283 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
285 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
292 ; AVX512DQ-BW-NEXT: vpextrw $0, %xmm1, (%rsi)
293 ; AVX512DQ-BW-NEXT: vpextrw $0, %xmm2, (%rdx)
294 ; AVX512DQ-BW-NEXT: vpextrw $0, %xmm3, (%rcx)
295 ; AVX512DQ-BW-NEXT: vpextrw $0, %xmm4, (%r8)
296 ; AVX512DQ-BW-NEXT: vpextrw $0, %xmm5, (%r9)
297 ; AVX512DQ-BW-NEXT: vpextrw $0, %xmm6, (%r10)
298 ; AVX512DQ-BW-NEXT: vpextrw $0, %xmm0, (%rax)
302 ; AVX512DQ-BW-FCP: # %bb.0:
303 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
304 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
306 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
313 ; AVX512DQ-BW-FCP-NEXT: vpextrw $0, %xmm1, (%rsi)
314 ; AVX512DQ-BW-FCP-NEXT: vpextrw $0, %xmm2, (%rdx)
315 ; AVX512DQ-BW-FCP-NEXT: vpextrw $0, %xmm3, (%rcx)
316 ; AVX512DQ-BW-FCP-NEXT: vpextrw $0, %xmm4, (%r8)
317 ; AVX512DQ-BW-FCP-NEXT: vpextrw $0, %xmm5, (%r9)
318 ; AVX512DQ-BW-FCP-NEXT: vpextrw $0, %xmm6, (%r10)
319 ; AVX512DQ-BW-FCP-NEXT: vpextrw $0, %xmm0, (%rax)
322 %strided.vec0 = shufflevector <14 x i8> %wide.vec, <14 x i8> poison, <2 x i32> <i32 0, i32 7>
341 ; SSE: # %bb.0:
344 ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [65535,65535,0,65535,65535,65535,65535,65535]
350 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
351 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,65535,65535,65535,65535,65535,0,65535]
354 ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,0],xmm4[0,0]
355 ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[2,0],xmm4[2,3]
356 ; SSE-NEXT: movdqa {{.*#+}} xmm7 = [65535,0,65535,65535,0,65535,65535,65535]
360 ; SSE-NEXT: movdqa {{.*#+}} xmm6 = [65535,0,65535,65535,65535,0,65535,65535]
364 ; SSE-NEXT: movdqa {{.*#+}} xmm13 = [65535,65535,0,65535,65535,0,65535,65535]
369 ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm0[0,2,2,3]
370 ; SSE-NEXT: movdqa {{.*#+}} xmm11 = [65535,65535,65535,0,65535,65535,0,65535]
376 ; SSE-NEXT: movdqa {{.*#+}} xmm15 = [0,65535,65535,65535,65535,65535,65535,0]
377 ; SSE-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3],xmm5[4],xmm1[4],xmm5[5],xmm1[5],xmm5[6],xmm1[6],xmm5[7],xmm1[7]
382 ; SSE-NEXT: punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm1[0],xmm7[1],xmm1[1],xmm7[2],xmm1[2],xmm7[3],xmm1[3],xmm7[4],xmm1[4],xmm7[5],xmm1[5],xmm7[6],xmm1[6],xmm7[7],xmm1[7]
383 ; SSE-NEXT: pshuflw {{.*#+}} xmm3 = xmm7[0,2,2,3,4,5,6,7]
384 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,3,2,3]
385 ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm3[1,1,0,3,4,5,6,7]
386 ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [65535,0,65535,65535,65535,65535,65535,65535]
393 ; SSE-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3],xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
394 ; SSE-NEXT: punpcklwd {{.*#+}} xmm6 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
399 ; SSE-NEXT: punpcklbw {{.*#+}} xmm13 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3],xmm13[4],xmm1[4],xmm13[5],xmm1[5],xmm13[6],xmm1[6],xmm13[7],xmm1[7]
401 ; SSE-NEXT: pshuflw {{.*#+}} xmm4 = xmm4[0,2,2,3,4,5,6,7]
402 ; SSE-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
403 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
404 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rdi
405 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,1,3]
406 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[0,3,2,3,4,5,6,7]
407 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,3,2,3]
408 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[0,3,2,1,4,5,6,7]
410 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm15[0,3,2,3]
411 ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm5[1,0,3,2,4,5,6,7]
415 ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,6,5,6,7]
416 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[2,0,2,3]
419 ; SSE-NEXT: punpckldq {{.*#+}} xmm10 = xmm10[0],xmm12[0],xmm10[1],xmm12[1]
423 ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm7[0,2,2,3,4,5,6,7]
424 ; SSE-NEXT: punpcklbw {{.*#+}} xmm10 = xmm10[0],xmm1[0],xmm10[1],xmm1[1],xmm10[2],xmm1[2],xmm10[3],xmm1[3],xmm10[4],xmm1[4],xmm10[5],xmm1[5],xmm10[6],xmm1[6],xmm10[7],xmm1[7]
427 ; SSE-NEXT: punpcklwd {{.*#+}} xmm8 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
431 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
432 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,4,6,7]
434 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm11[0,1,2,1]
435 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,7,6,7]
448 ; AVX: # %bb.0:
449 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
450 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10
454 ; AVX-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[0,7,14],zero,xmm0[u,u,u,u,u,u,u,u,u,u,u,u]
459 ; AVX-NEXT: vmovd {{.*#+}} xmm4 = [2,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
461 ; AVX-NEXT: vpshufb {{.*#+}} xmm6 = xmm1[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
462 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
463 ; AVX-NEXT: vmovd {{.*#+}} xmm6 = [3,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
466 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
468 ; AVX-NEXT: vmovd {{.*#+}} xmm8 = [4,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
470 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
473 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
476 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
487 ; AVX2: # %bb.0:
488 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
489 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10
493 ; AVX2-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[0,7,14],zero,xmm0[u,u,u,u,u,u,u,u,u,u,u,u]
500 ; AVX2-NEXT: vpshufb {{.*#+}} xmm6 = xmm1[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
501 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
505 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
509 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
512 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
515 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
526 ; AVX2-FP: # %bb.0:
527 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
528 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10
532 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[0,7,14],zero,xmm0[u,u,u,u,u,u,u,u,u,u,u,u]
539 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm6 = xmm1[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
540 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
544 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
548 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
551 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
554 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
565 ; AVX2-FCP: # %bb.0:
566 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
567 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
571 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[0,7,14],zero,xmm0[u,u,u,u,u,u,u,u,u,u,u,u]
578 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm1[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
579 ; AVX2-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
583 ; AVX2-FCP-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
587 ; AVX2-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
590 ; AVX2-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
593 ; AVX2-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
604 ; AVX512: # %bb.0:
605 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
606 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10
610 ; AVX512-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[0,7,14],zero,xmm0[u,u,u,u,u,u,u,u,u,u,u,u]
615 ; AVX512-NEXT: vmovd {{.*#+}} xmm4 = [2,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
617 ; AVX512-NEXT: vpshufb {{.*#+}} xmm6 = xmm1[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
618 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
619 ; AVX512-NEXT: vmovd {{.*#+}} xmm6 = [3,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
622 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
624 ; AVX512-NEXT: vmovd {{.*#+}} xmm8 = [4,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
626 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
629 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
632 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
643 ; AVX512-FCP: # %bb.0:
644 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
645 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
649 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[0,7,14],zero,xmm0[u,u,u,u,u,u,u,u,u,u,u,u]
654 ; AVX512-FCP-NEXT: vmovd {{.*#+}} xmm4 = [2,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
656 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm1[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
657 ; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
658 ; AVX512-FCP-NEXT: vmovd {{.*#+}} xmm6 = [3,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
661 ; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
663 ; AVX512-FCP-NEXT: vmovd {{.*#+}} xmm8 = [4,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
665 ; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
668 ; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
671 ; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
682 ; AVX512DQ: # %bb.0:
683 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
684 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10
688 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[0,7,14],zero,xmm0[u,u,u,u,u,u,u,u,u,u,u,u]
693 ; AVX512DQ-NEXT: vmovd {{.*#+}} xmm4 = [2,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
695 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm6 = xmm1[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
696 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
697 ; AVX512DQ-NEXT: vmovd {{.*#+}} xmm6 = [3,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
700 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
702 ; AVX512DQ-NEXT: vmovd {{.*#+}} xmm8 = [4,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
704 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
707 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
710 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
721 ; AVX512DQ-FCP: # %bb.0:
722 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
723 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
727 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[0,7,14],zero,xmm0[u,u,u,u,u,u,u,u,u,u,u,u]
732 ; AVX512DQ-FCP-NEXT: vmovd {{.*#+}} xmm4 = [2,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
734 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm1[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
735 ; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
736 ; AVX512DQ-FCP-NEXT: vmovd {{.*#+}} xmm6 = [3,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
739 ; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
741 ; AVX512DQ-FCP-NEXT: vmovd {{.*#+}} xmm8 = [4,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
743 ; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
746 ; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
749 ; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
760 ; AVX512BW: # %bb.0:
761 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
762 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
766 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[0,7,14],zero,xmm0[u,u,u,u,u,u,u,u,u,u,u,u]
773 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm6 = xmm1[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
774 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
778 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
782 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
785 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
788 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
799 ; AVX512BW-FCP: # %bb.0:
800 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
801 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
805 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[0,7,14],zero,xmm0[u,u,u,u,u,u,u,u,u,u,u,u]
812 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm1[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
813 ; AVX512BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
817 ; AVX512BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
821 ; AVX512BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
824 ; AVX512BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
827 ; AVX512BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
838 ; AVX512DQ-BW: # %bb.0:
839 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
840 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
844 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[0,7,14],zero,xmm0[u,u,u,u,u,u,u,u,u,u,u,u]
851 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm6 = xmm1[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
852 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
856 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
860 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
863 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
866 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
877 ; AVX512DQ-BW-FCP: # %bb.0:
878 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
879 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
883 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[0,7,14],zero,xmm0[u,u,u,u,u,u,u,u,u,u,u,u]
890 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm1[0,7,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
891 ; AVX512DQ-BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
895 ; AVX512DQ-BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
899 ; AVX512DQ-BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
902 ; AVX512DQ-BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
905 ; AVX512DQ-BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
915 %strided.vec0 = shufflevector <28 x i8> %wide.vec, <28 x i8> poison, <4 x i32> <i32 0, i32 7, i32 14, i32 21>
934 ; SSE: # %bb.0:
939 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,65535,0,65535,65535,65535,0,65535]
947 ; SSE-NEXT: movdqa {{.*#+}} xmm7 = [65535,65535,65535,65535,0,65535,0,65535]
948 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
953 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm7[0,2,1,3]
954 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[0,3,2,3,4,5,6,7]
955 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,3,1,1]
956 ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm2[0,3,2,1,4,5,6,7]
958 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [255,255,255,255,255,0,0,0,255,255,255,255,255,255,255,255]
960 ; SSE-NEXT: movdqa {{.*#+}} xmm9 = [0,65535,65535,65535,65535,65535,65535,65535]
965 ; SSE-NEXT: punpcklbw {{.*#+}} xmm9 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3],xmm9[4],xmm4[4],xmm9[5],xmm4[5],xmm9[6],xmm4[6],xmm9[7],xmm4[7]
967 ; SSE-NEXT: movss {{.*#+}} xmm8 = xmm0[0],xmm8[1,2,3]
968 ; SSE-NEXT: movdqa {{.*#+}} xmm5 = [65535,0,65535,65535,0,65535,65535,65535]
969 ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
973 ; SSE-NEXT: movdqa {{.*#+}} xmm12 = [65535,65535,0,65535,65535,65535,65535,65535]
976 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
979 ; SSE-NEXT: shufps {{.*#+}} xmm14 = xmm14[1,0],xmm6[0,0]
980 ; SSE-NEXT: shufps {{.*#+}} xmm14 = xmm14[2,0],xmm6[2,3]
982 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
987 ; SSE-NEXT: punpcklwd {{.*#+}} xmm9 = xmm9[0],xmm15[0],xmm9[1],xmm15[1],xmm9[2],xmm15[2],xmm9[3],xmm15[3]
988 ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm9[0,1,2,3,6,5,6,7]
989 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,1,2,1]
990 ; SSE-NEXT: pshufhw {{.*#+}} xmm9 = xmm6[0,1,2,3,4,4,5,6]
996 ; SSE-NEXT: movdqa {{.*#+}} xmm7 = [65535,65535,65535,0,65535,65535,0,65535]
1004 ; SSE-NEXT: punpcklbw {{.*#+}} xmm9 = xmm9[0],xmm0[0],xmm9[1],xmm0[1],xmm9[2],xmm0[2],xmm9[3],xmm0[3],xmm9[4],xmm0[4],xmm9[5],xmm0[5],xmm9[6],xmm0[6],xmm9[7],xmm0[7]
1005 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,0,65535,65535,65535,65535,0,65535]
1012 ; SSE-NEXT: punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3],xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
1013 ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [65535,65535,65535,0,65535,65535,65535,65535]
1018 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
1019 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,7,6]
1023 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,3,2,3]
1024 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[1,0,3,2,4,5,6,7]
1025 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5]
1029 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
1034 ; SSE-NEXT: punpcklbw {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1],xmm10[2],xmm6[2],xmm10[3],xmm6[3],xmm10[4],xmm6[4],xmm10[5],xmm6[5],xmm10[6],xmm6[6],xmm10[7],xmm6[7]
1038 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
1039 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,7,6,7]
1040 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,3,3]
1041 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[2,1,0,3,4,5,6,7]
1044 ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [65535,65535,65535,65535,0,65535,65535,65535]
1048 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
1049 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,7]
1054 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,0,65535,65535,65535,0,65535,65535]
1061 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
1063 ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [65535,0,65535,0,65535,65535,65535,65535]
1066 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
1067 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rdi
1069 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm3[3,2,1,0,4,5,6,7]
1070 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,7,7,7]
1073 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
1074 ; SSE-NEXT: punpcklbw {{.*#+}} xmm12 = xmm12[0],xmm6[0],xmm12[1],xmm6[1],xmm12[2],xmm6[2],xmm12[3],xmm6[3],xmm12[4],xmm6[4],xmm12[5],xmm6[5],xmm12[6],xmm6[6],xmm12[7],xmm6[7]
1075 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm12[0,1,2,3,7,5,6,7]
1077 ; SSE-NEXT: pshuflw {{.*#+}} xmm3 = xmm12[0,3,2,3,4,5,6,7]
1078 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,2,0]
1079 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,4,7,6]
1084 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,65535,0,65535,65535,0,65535,65535]
1095 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3],xmm0[4],xmm4[4],xmm0[5],xmm4[5],xmm0[6],xmm4[6],xmm0[7],xmm4[7]
1097 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
1098 ; SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
1099 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
1100 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,4,6,5]
1101 ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [65535,65535,65535,65535,65535,65535,0,65535]
1106 ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm0[0,3,2,3]
1109 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,2,2,3]
1110 ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
1114 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
1115 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
1118 ; SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
1119 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [0,65535,65535,65535,65535,65535,65535,0]
1120 ; SSE-NEXT: punpcklbw {{.*#+}} xmm14 = xmm14[0],xmm4[0],xmm14[1],xmm4[1],xmm14[2],xmm4[2],xmm14[3],xmm4[3],xmm14[4],xmm4[4],xmm14[5],xmm4[5],xmm14[6],xmm4[6],xmm14[7],xmm4[7]
1124 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,3]
1125 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,4,7,6]
1127 ; SSE-NEXT: pshufd {{.*#+}} xmm11 = xmm1[0,3,2,3]
1133 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3],xmm0[4],xmm4[4],xmm0[5],xmm4[5],xmm0[6],xmm4[6],xmm0[7],xmm4[7]
1134 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,4,6,7]
1136 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm7[0,1,2,1]
1137 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,7,6,7]
1139 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
1140 ; SSE-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
1141 ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [65535,0,65535,65535,65535,65535,65535,65535]
1145 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,1,3]
1146 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,7]
1147 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
1148 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
1150 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,3,2,3]
1161 ; AVX: # %bb.0:
1162 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
1163 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10
1169 ; AVX-NEXT: vpshufb {{.*#+}} xmm5 = xmm0[0,7,14],zero,zero,xmm0[u,u,u,u,u,u,u,u,u,u,u]
1171 ; AVX-NEXT: vpalignr {{.*#+}} xmm5 = xmm2[3,4,5,6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2]
1172 ; AVX-NEXT: vbroadcastss {{.*#+}} xmm6 = [0,0,7,14,0,0,7,14,0,0,7,14,0,0,7,14]
1174 ; AVX-NEXT: vmovddup {{.*#+}} xmm7 = [255,255,255,255,255,0,0,0,255,255,255,255,255,0,0,0]
1175 ; AVX-NEXT: # xmm7 = mem[0,0]
1180 ; AVX-NEXT: vpalignr {{.*#+}} xmm8 = xmm2[4,5,6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2,3]
1184 ; AVX-NEXT: vpshufb {{.*#+}} xmm9 = zero,zero,xmm1[0,7,14,u,u,u,u,u,u,u,u,u,u,u]
1186 ; AVX-NEXT: vpalignr {{.*#+}} xmm9 = xmm2[5,6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4]
1192 ; AVX-NEXT: vpalignr {{.*#+}} xmm10 = xmm2[6,7,8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5]
1196 ; AVX-NEXT: vpshufb {{.*#+}} xmm9 = xmm2[u,u,u,u,0,7,14],zero,xmm2[u,u,u,u,u,u,u,u]
1198 ; AVX-NEXT: vmovd {{.*#+}} xmm9 = [4,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
1201 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
1202 ; AVX-NEXT: vpblendw {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,3],xmm10[4,5,6,7]
1208 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
1209 ; AVX-NEXT: vpblendw {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3],xmm11[4,5,6,7]
1210 ; AVX-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u,0,7,u,u,u,u,u,u,u,u,u,u,u,u]
1212 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
1215 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
1216 ; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5,6,7]
1227 ; AVX2: # %bb.0:
1228 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
1229 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10
1232 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
1236 ; AVX2-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,7,14],zero,zero,xmm2[3,10],zero,xmm2[u,u,u,u,u,u,u,u]
1238 ; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = ymm0[0],ymm1[1,2],ymm0[3,4],ymm1[5],ymm0[6,7,8],ymm1[9,10],ymm0[11,12],ymm1[13],ymm0[14,15]
1243 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm4 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
1246 ; AVX2-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,xmm5[0,7,14],zero,zero,xmm5[3,u,u,u,u,u,u,u,u]
1249 ; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6,7]
1254 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm6 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,0,0]
1258 ; AVX2-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,11],zero,zero,xmm6[0,7,14],zero,xmm6[u,u,u,u,u,u,u,u]
1260 ; AVX2-NEXT: vpblendw {{.*#+}} ymm7 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7,8],ymm0[9,10],ymm1[11,12],ymm0[13,14],ymm1[15]
1265 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm8 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,0]
1268 ; AVX2-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,11,0,7,u,u,u,u,u,u,u,u,u,u,u,u]
1270 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
1282 ; AVX2-FP: # %bb.0:
1283 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1284 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10
1287 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
1291 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,7,14],zero,zero,xmm2[3,10],zero,xmm2[u,u,u,u,u,u,u,u]
1293 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = ymm0[0],ymm1[1,2],ymm0[3,4],ymm1[5],ymm0[6,7,8],ymm1[9,10],ymm0[11,12],ymm1[13],ymm0[14,15]
1298 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm4 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
1301 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,xmm5[0,7,14],zero,zero,xmm5[3,u,u,u,u,u,u,u,u]
1304 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6,7]
1309 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm6 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,0,0]
1313 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,11],zero,zero,xmm6[0,7,14],zero,xmm6[u,u,u,u,u,u,u,u]
1315 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm7 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7,8],ymm0[9,10],ymm1[11,12],ymm0[13,14],ymm1[15]
1320 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm8 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,0]
1323 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,11,0,7,u,u,u,u,u,u,u,u,u,u,u,u]
1325 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
1337 ; AVX2-FCP: # %bb.0:
1338 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1339 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
1342 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
1346 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,7,14],zero,zero,xmm2[3,10],zero,xmm2[u,u,u,u,u,u,u,u]
1348 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm0[0],ymm1[1,2],ymm0[3,4],ymm1[5],ymm0[6,7,8],ymm1[9,10],ymm0[11,12],ymm1[13],ymm0[14,15]
1353 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm4 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
1356 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,xmm5[0,7,14],zero,zero,xmm5[3,u,u,u,u,u,u,u,u]
1359 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6,7]
1364 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm6 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,0,0]
1368 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,11],zero,zero,xmm6[0,7,14],zero,xmm6[u,u,u,u,u,u,u,u]
1370 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7,8],ymm0[9,10],ymm1[11,12],ymm0[13,14],ymm1[15]
1375 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm8 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,0]
1378 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,11,0,7,u,u,u,u,u,u,u,u,u,u,u,u]
1380 ; AVX2-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
1392 ; AVX512: # %bb.0:
1393 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
1394 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10
1397 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm2 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,65535,65535,65535,65535]
1401 ; AVX512-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,7,14],zero,zero,xmm2[3,10],zero,xmm2[u,u,u,u,u,u,u,u]
1403 ; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm0[0],ymm1[1,2],ymm0[3,4],ymm1[5],ymm0[6,7,8],ymm1[9,10],ymm0[11,12],ymm1[13],ymm0[14,15]
1408 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm4 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,65535,65535,65535]
1411 ; AVX512-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,xmm5[0,7,14],zero,zero,xmm5[3,u,u,u,u,u,u,u,u]
1414 ; AVX512-NEXT: vpblendd {{.*#+}} ymm5 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6,7]
1419 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm6 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
1423 ; AVX512-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,11],zero,zero,xmm6[0,7,14],zero,xmm6[u,u,u,u,u,u,u,u]
1425 ; AVX512-NEXT: vpblendw {{.*#+}} ymm7 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7,8],ymm0[9,10],ymm1[11,12],ymm0[13,14],ymm1[15]
1432 ; AVX512-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,11,0,7,u,u,u,u,u,u,u,u,u,u,u,u]
1434 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
1446 ; AVX512-FCP: # %bb.0:
1447 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1448 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
1451 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,65535,65535,65535,65535]
1455 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,7,14],zero,zero,xmm2[3,10],zero,xmm2[u,u,u,u,u,u,u,u]
1457 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm0[0],ymm1[1,2],ymm0[3,4],ymm1[5],ymm0[6,7,8],ymm1[9,10],ymm0[11,12],ymm1[13],ymm0[14,15]
1462 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,65535,65535,65535]
1465 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,xmm5[0,7,14],zero,zero,xmm5[3,u,u,u,u,u,u,u,u]
1468 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6,7]
1473 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
1477 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,11],zero,zero,xmm6[0,7,14],zero,xmm6[u,u,u,u,u,u,u,u]
1479 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7,8],ymm0[9,10],ymm1[11,12],ymm0[13,14],ymm1[15]
1486 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,11,0,7,u,u,u,u,u,u,u,u,u,u,u,u]
1488 ; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
1500 ; AVX512DQ: # %bb.0:
1501 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
1502 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10
1505 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm2 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,65535,65535,65535,65535]
1509 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,7,14],zero,zero,xmm2[3,10],zero,xmm2[u,u,u,u,u,u,u,u]
1511 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm0[0],ymm1[1,2],ymm0[3,4],ymm1[5],ymm0[6,7,8],ymm1[9,10],ymm0[11,12],ymm1[13],ymm0[14,15]
1516 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm4 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,65535,65535,65535]
1519 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,xmm5[0,7,14],zero,zero,xmm5[3,u,u,u,u,u,u,u,u]
1522 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm5 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6,7]
1527 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm6 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
1531 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,11],zero,zero,xmm6[0,7,14],zero,xmm6[u,u,u,u,u,u,u,u]
1533 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm7 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7,8],ymm0[9,10],ymm1[11,12],ymm0[13,14],ymm1[15]
1540 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,11,0,7,u,u,u,u,u,u,u,u,u,u,u,u]
1542 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
1554 ; AVX512DQ-FCP: # %bb.0:
1555 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1556 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
1559 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,65535,65535,65535,65535]
1563 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,7,14],zero,zero,xmm2[3,10],zero,xmm2[u,u,u,u,u,u,u,u]
1565 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm0[0],ymm1[1,2],ymm0[3,4],ymm1[5],ymm0[6,7,8],ymm1[9,10],ymm0[11,12],ymm1[13],ymm0[14,15]
1570 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,65535,65535,65535]
1573 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,xmm5[0,7,14],zero,zero,xmm5[3,u,u,u,u,u,u,u,u]
1576 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6,7]
1581 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
1585 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,11],zero,zero,xmm6[0,7,14],zero,xmm6[u,u,u,u,u,u,u,u]
1587 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7,8],ymm0[9,10],ymm1[11,12],ymm0[13,14],ymm1[15]
1594 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,11,0,7,u,u,u,u,u,u,u,u,u,u,u,u]
1596 ; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
1608 ; AVX512BW: # %bb.0:
1609 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
1610 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
1613 ; AVX512BW-NEXT: movw $290, %di # imm = 0x122
1618 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,7,14],zero,zero,xmm2[3,10],zero,xmm2[u,u,u,u,u,u,u,u]
1620 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm3 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5],ymm1[6,7,8],ymm0[9,10],ymm1[11,12],ymm0[13],ymm1[14,15]
1625 ; AVX512BW-NEXT: movw $580, %di # imm = 0x244
1629 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,xmm5[0,7,14],zero,zero,xmm5[3,u,u,u,u,u,u,u,u]
1632 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6,7]
1637 ; AVX512BW-NEXT: movw $4644, %di # imm = 0x1224
1642 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,11],zero,zero,xmm6[0,7,14],zero,xmm6[u,u,u,u,u,u,u,u]
1644 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm7 = ymm0[0],ymm1[1,2],ymm0[3,4],ymm1[5,6],ymm0[7,8],ymm1[9,10],ymm0[11,12],ymm1[13,14],ymm0[15]
1649 ; AVX512BW-NEXT: movw $9288, %di # imm = 0x2448
1653 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,11,0,7,u,u,u,u,u,u,u,u,u,u,u,u]
1655 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
1667 ; AVX512BW-FCP: # %bb.0:
1668 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1669 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
1672 ; AVX512BW-FCP-NEXT: movw $290, %di # imm = 0x122
1677 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,7,14],zero,zero,xmm2[3,10],zero,xmm2[u,u,u,u,u,u,u,u]
1679 ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5],ymm1[6,7,8],ymm0[9,10],ymm1[11,12],ymm0[13],ymm1[14,15]
1684 ; AVX512BW-FCP-NEXT: movw $580, %di # imm = 0x244
1688 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,xmm5[0,7,14],zero,zero,xmm5[3,u,u,u,u,u,u,u,u]
1691 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6,7]
1696 ; AVX512BW-FCP-NEXT: movw $4644, %di # imm = 0x1224
1701 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,11],zero,zero,xmm6[0,7,14],zero,xmm6[u,u,u,u,u,u,u,u]
1703 ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm0[0],ymm1[1,2],ymm0[3,4],ymm1[5,6],ymm0[7,8],ymm1[9,10],ymm0[11,12],ymm1[13,14],ymm0[15]
1708 ; AVX512BW-FCP-NEXT: movw $9288, %di # imm = 0x2448
1712 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,11,0,7,u,u,u,u,u,u,u,u,u,u,u,u]
1714 ; AVX512BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
1726 ; AVX512DQ-BW: # %bb.0:
1727 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
1728 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
1731 ; AVX512DQ-BW-NEXT: movw $290, %di # imm = 0x122
1736 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,7,14],zero,zero,xmm2[3,10],zero,xmm2[u,u,u,u,u,u,u,u]
1738 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm3 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5],ymm1[6,7,8],ymm0[9,10],ymm1[11,12],ymm0[13],ymm1[14,15]
1743 ; AVX512DQ-BW-NEXT: movw $580, %di # imm = 0x244
1747 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,xmm5[0,7,14],zero,zero,xmm5[3,u,u,u,u,u,u,u,u]
1750 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6,7]
1755 ; AVX512DQ-BW-NEXT: movw $4644, %di # imm = 0x1224
1760 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,11],zero,zero,xmm6[0,7,14],zero,xmm6[u,u,u,u,u,u,u,u]
1762 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm7 = ymm0[0],ymm1[1,2],ymm0[3,4],ymm1[5,6],ymm0[7,8],ymm1[9,10],ymm0[11,12],ymm1[13,14],ymm0[15]
1767 ; AVX512DQ-BW-NEXT: movw $9288, %di # imm = 0x2448
1771 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,11,0,7,u,u,u,u,u,u,u,u,u,u,u,u]
1773 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
1785 ; AVX512DQ-BW-FCP: # %bb.0:
1786 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1787 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
1790 ; AVX512DQ-BW-FCP-NEXT: movw $290, %di # imm = 0x122
1795 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,7,14],zero,zero,xmm2[3,10],zero,xmm2[u,u,u,u,u,u,u,u]
1797 ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5],ymm1[6,7,8],ymm0[9,10],ymm1[11,12],ymm0[13],ymm1[14,15]
1802 ; AVX512DQ-BW-FCP-NEXT: movw $580, %di # imm = 0x244
1806 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,xmm5[0,7,14],zero,zero,xmm5[3,u,u,u,u,u,u,u,u]
1809 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6,7]
1814 ; AVX512DQ-BW-FCP-NEXT: movw $4644, %di # imm = 0x1224
1819 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,11],zero,zero,xmm6[0,7,14],zero,xmm6[u,u,u,u,u,u,u,u]
1821 ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm0[0],ymm1[1,2],ymm0[3,4],ymm1[5,6],ymm0[7,8],ymm1[9,10],ymm0[11,12],ymm1[13,14],ymm0[15]
1826 ; AVX512DQ-BW-FCP-NEXT: movw $9288, %di # imm = 0x2448
1830 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,11,0,7,u,u,u,u,u,u,u,u,u,u,u,u]
1832 ; AVX512DQ-BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
1843 %strided.vec0 = shufflevector <56 x i8> %wide.vec, <56 x i8> poison, <8 x i32> <i32 0, i32 7, i32 14, i32 21, i32 28, i32 35, i32 42, i32 49>
1862 ; SSE: # %bb.0:
1871 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,0,65535,65535,65535,0,65535,65535]
1880 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm13[0],xmm1[1],xmm13[1],xmm1[2],xmm13[2],xmm1[3],xmm13[3],xmm1[4],xmm13[4],xmm1[5],xmm13[5],xmm1[6],xmm13[6],xmm1[7],xmm13[7]
1881 ; SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
1882 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,5,6,7]
1883 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
1884 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,5,6]
1885 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,3,2,3]
1886 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,3,2,3,4,5,6,7]
1888 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [255,255,255,255,255,0,0,0,0,0,255,255,255,255,255,255]
1889 ; SSE-NEXT: movdqa {{.*#+}} xmm10 = [65535,65535,0,65535,65535,65535,0,65535]
1894 ; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1899 ; SSE-NEXT: movdqa {{.*#+}} xmm14 = [65535,65535,65535,65535,0,65535,0,65535]
1902 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm13[0],xmm3[1],xmm13[1],xmm3[2],xmm13[2],xmm3[3],xmm13[3],xmm3[4],xmm13[4],xmm3[5],xmm13[5],xmm3[6],xmm13[6],xmm3[7],xmm13[7]
1905 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[0,2,1,3]
1906 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,3,2,3,4,5,6,7]
1907 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,3,1,1]
1908 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,3,2,1,4,5,6,7]
1913 ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [65535,65535,65535,0,65535,65535,0,65535]
1916 ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1924 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
1925 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,4,7]
1927 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
1928 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,7,7]
1933 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1934 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
1935 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1937 ; SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
1938 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
1939 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,7]
1941 ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [65535,65535,65535,65535,65535,65535,65535,0]
1947 ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [65535,65535,65535,65535,65535,0,0,0]
1951 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1952 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,65535,0,65535,65535,0,65535,65535]
1960 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1],xmm1[2],xmm15[2],xmm1[3],xmm15[3],xmm1[4],xmm15[4],xmm1[5],xmm15[5],xmm1[6],xmm15[6],xmm1[7],xmm15[7]
1961 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,65535,65535,0,65535,65535,65535,65535]
1965 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,1,2,1]
1966 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,7,6]
1969 ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [255,255,255,255,255,0,0,0,0,255,255,255,255,255,255,255]
1974 ; SSE-NEXT: movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1980 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1],xmm1[2],xmm15[2],xmm1[3],xmm15[3],xmm1[4],xmm15[4],xmm1[5],xmm15[5],xmm1[6],xmm15[6],xmm1[7],xmm15[7]
1981 ; SSE-NEXT: movdqa {{.*#+}} xmm6 = [65535,0,65535,65535,65535,65535,0,65535]
1987 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,3,2,3]
1988 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[1,0,3,2,4,5,6,7]
1989 ; SSE-NEXT: pshufhw {{.*#+}} xmm13 = xmm1[0,1,2,3,5,5,5,5]
2002 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
2006 ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2007 ; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2012 ; SSE-NEXT: movdqa {{.*#+}} xmm14 = [65535,0,65535,65535,0,65535,65535,65535]
2015 ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2020 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2021 ; SSE-NEXT: pshufd {{.*#+}} xmm11 = xmm8[0,2,2,3]
2025 ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2028 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2031 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
2034 ; SSE-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
2037 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2038 ; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2040 ; SSE-NEXT: punpcklbw {{.*#+}} xmm10 = xmm10[0],xmm1[0],xmm10[1],xmm1[1],xmm10[2],xmm1[2],xmm10[3],xmm1[3],xmm10[4],xmm1[4],xmm10[5],xmm1[5],xmm10[6],xmm1[6],xmm10[7],xmm1[7]
2041 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm10[0,1,2,3,7,5,6,7]
2042 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,2,0]
2043 ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm2[0,1,2,3,6,4,6,5]
2044 ; SSE-NEXT: movdqa {{.*#+}} xmm12 = [65535,65535,65535,65535,65535,65535,0,65535]
2048 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2052 ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2053 ; SSE-NEXT: punpcklbw {{.*#+}} xmm14 = xmm14[0],xmm1[0],xmm14[1],xmm1[1],xmm14[2],xmm1[2],xmm14[3],xmm1[3],xmm14[4],xmm1[4],xmm14[5],xmm1[5],xmm14[6],xmm1[6],xmm14[7],xmm1[7]
2054 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,0,65535,65535,65535,65535,65535,65535]
2057 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2058 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2060 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2061 ; SSE-NEXT: punpcklwd {{.*#+}} xmm8 = xmm8[0],xmm5[0],xmm8[1],xmm5[1],xmm8[2],xmm5[2],xmm8[3],xmm5[3]
2063 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,65535,65535,65535,65535,65535,65535,0]
2066 ; SSE-NEXT: pshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
2067 ; SSE-NEXT: # xmm5 = mem[0,3,2,3]
2068 ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm5[0,0,3,2,4,5,6,7]
2069 ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,5,4,6,7]
2073 ; SSE-NEXT: movdqa {{.*#+}} xmm5 = [255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0]
2078 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2079 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,65535,65535,0,65535,65535,0,65535]
2082 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2087 ; SSE-NEXT: punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3],xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
2088 ; SSE-NEXT: movdqa {{.*#+}} xmm13 = [65535,0,65535,65535,65535,65535,0,65535]
2096 ; SSE-NEXT: movdqa {{.*#+}} xmm6 = [65535,65535,65535,65535,65535,65535,65535,0]
2099 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,3,2,3]
2100 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,1,0,3,4,5,6,7]
2101 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,5,6,7]
2107 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
2114 ; SSE-NEXT: punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3],xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
2117 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
2119 ; SSE-NEXT: punpckldq {{.*#+}} xmm11 = xmm11[0],xmm1[0],xmm11[1],xmm1[1]
2122 ; SSE-NEXT: punpcklbw {{.*#+}} xmm11 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3],xmm11[4],xmm8[4],xmm11[5],xmm8[5],xmm11[6],xmm8[6],xmm11[7],xmm8[7]
2123 ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [65535,65535,65535,65535,0,65535,65535,65535]
2127 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,2,1]
2128 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,7]
2131 ; SSE-NEXT: movdqa {{.*#+}} xmm13 = [255,255,255,255,255,0,0,0,0,255,255,255,255,255,255,255]
2134 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm7[0,2,1,3]
2135 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,7]
2136 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,3,3]
2137 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[2,1,0,3,4,5,6,7]
2150 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3],xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
2153 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm4[0,3,2,3,4,5,6,7]
2154 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
2155 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,4,7,6]
2158 ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [65535,0,65535,65535,65535,0,65535,65535]
2162 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
2167 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
2168 ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [65535,0,65535,0,65535,65535,65535,65535]
2174 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm4[3,2,1,0,4,5,6,7]
2175 ; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm0[0,1,2,3,7,7,7,7]
2180 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
2181 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2186 ; SSE-NEXT: punpcklbw {{.*#+}} xmm15 = xmm15[0],xmm2[0],xmm15[1],xmm2[1],xmm15[2],xmm2[2],xmm15[3],xmm2[3],xmm15[4],xmm2[4],xmm15[5],xmm2[5],xmm15[6],xmm2[6],xmm15[7],xmm2[7]
2187 ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [65535,0,65535,65535,65535,65535,0,65535]
2192 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
2194 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
2196 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,65535,65535,65535,65535,65535,65535,0]
2199 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm11[0,2,1,0,4,5,6,7]
2200 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,6,7]
2208 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,65535,0,65535,65535,0,65535,65535]
2218 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3],xmm4[4],xmm0[4],xmm4[5],xmm0[5],xmm4[6],xmm0[6],xmm4[7],xmm0[7]
2220 ; SSE-NEXT: pshuflw {{.*#+}} xmm4 = xmm4[0,2,2,3,4,5,6,7]
2221 ; SSE-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
2222 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
2229 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm10[0,1,3,3]
2230 ; SSE-NEXT: movss {{.*#+}} xmm2 = xmm4[0],xmm2[1,2,3]
2231 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
2233 ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [65535,0,65535,65535,65535,0,65535,65535]
2239 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3],xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
2245 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,0,3]
2246 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7]
2247 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,7]
2249 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,65535,65535,65535,65535,65535,65535,0]
2251 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm13[0,1,2,1]
2252 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,7]
2260 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
2261 ; SSE-NEXT: movdqa {{.*#+}} xmm13 = [65535,65535,65535,0,65535,65535,0,65535]
2263 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
2266 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm10[0],xmm2[1],xmm10[1],xmm2[2],xmm10[2],xmm2[3],xmm10[3],xmm2[4],xmm10[4],xmm2[5],xmm10[5],xmm2[6],xmm10[6],xmm2[7],xmm10[7]
2267 ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [0,65535,65535,65535,65535,65535,65535,0]
2271 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm4[0,1,0,3]
2272 ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm2[0,1,2,3,5,4,7,6]
2275 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm9[0,2,2,3]
2277 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
2280 ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
2284 ; SSE-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[0,2,2,3,4,5,6,7]
2285 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm10[0],xmm2[1],xmm10[1],xmm2[2],xmm10[2],xmm2[3],xmm10[3],xmm2[4],xmm10[4],xmm2[5],xmm10[5],xmm2[6],xmm10[6],xmm2[7],xmm10[7]
2288 ; SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
2290 ; SSE-NEXT: movss {{.*#+}} xmm6 = xmm2[0],xmm6[1,2,3]
2293 ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [65535,65535,0,65535,65535,0,65535,65535]
2298 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm10[0],xmm2[1],xmm10[1],xmm2[2],xmm10[2],xmm2[3],xmm10[3],xmm2[4],xmm10[4],xmm2[5],xmm10[5],xmm2[6],xmm10[6],xmm2[7],xmm10[7]
2300 ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [65535,65535,0,65535,0,65535,65535,65535]
2304 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
2305 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm12[0,1,0,3]
2306 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,7,6]
2308 ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [255,255,255,255,255,255,255,255,255,255,255,255,255,0,0,0]
2311 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm4[2,1,0,3]
2312 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[0,0,3,2,4,5,6,7]
2313 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,5,5,5]
2324 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm10[0],xmm4[1],xmm10[1],xmm4[2],xmm10[2],xmm4[3],xmm10[3],xmm4[4],xmm10[4],xmm4[5],xmm10[5],xmm4[6],xmm10[6],xmm4[7],xmm10[7]
2325 ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm4[0,1,2,3,6,4,6,7]
2327 ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm7[0,1,2,1]
2328 ; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,5,7,6,7]
2330 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
2332 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
2336 ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm14[0,2,1,3]
2337 ; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,7,6,7]
2338 ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,1,2,0]
2339 ; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,7,6,5]
2343 ; SSE-NEXT: movss {{.*#+}} xmm4 = xmm6[0],xmm4[1,2,3]
2344 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
2345 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
2347 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm0[0,2,2,3]
2348 ; SSE-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]
2350 ; SSE-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm10[0],xmm6[1],xmm10[1],xmm6[2],xmm10[2],xmm6[3],xmm10[3],xmm6[4],xmm10[4],xmm6[5],xmm10[5],xmm6[6],xmm10[6],xmm6[7],xmm10[7]
2352 ; SSE-NEXT: movdqa {{.*#+}} xmm9 = [65535,65535,65535,0,65535,0,65535,65535]
2357 ; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm6[0,1,0,3,4,5,6,7]
2360 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm12[0,1,0,3]
2361 ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,5,4,7]
2368 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
2370 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2375 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
2377 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
2383 ; AVX: # %bb.0:
2389 ; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[0,7,14],zero,zero,xmm2[u,u,u,u,u,u,u,u,u,u,u]
2394 ; AVX-NEXT: vmovq {{.*#+}} xmm5 = [255,255,255,255,255,0,0,0,0,0,0,0,0,0,0,0]
2396 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2406 ; AVX-NEXT: vmovq {{.*#+}} xmm11 = [255,255,255,255,255,0,0,0,0,0,0,0,0,0,0,0]
2409 ; AVX-NEXT: vpshufb {{.*#+}} xmm10 = xmm5[u,u,u,u,u,0,7,14,u,u,u,u,u,u,u,u]
2410 ; AVX-NEXT: vpunpckldq {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]
2412 ; AVX-NEXT: vpblendw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,5,6],xmm12[7]
2417 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2419 ; AVX-NEXT: vpshufb {{.*#+}} xmm13 = zero,zero,xmm7[0,7,14,u,u,u,u,u,u,u,u,u,u,u]
2428 ; AVX-NEXT: vpblendw {{.*#+}} xmm13 = xmm13[0,1,2,3,4,5,6],xmm12[7]
2440 ; AVX-NEXT: vpshufb {{.*#+}} xmm14 = xmm1[u,u,u,u,u,u,u,u,u],zero,zero,xmm1[0,7,14,u,u]
2442 ; AVX-NEXT: vpblendw {{.*#+}} xmm13 = xmm13[0,1,2,3,4,5,6],xmm12[7]
2446 ; AVX-NEXT: vmovd {{.*#+}} xmm13 = [4,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
2449 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
2451 ; AVX-NEXT: vpshufb {{.*#+}} xmm0 = xmm3[u,u,u,u,0,7,14],zero,zero,xmm3[u,u,u,u,u,u,u]
2453 ; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm14[0,1],xmm0[2,3,4,5,6,7]
2457 ; AVX-NEXT: vpblendw {{.*#+}} xmm12 = xmm14[0,1,2,3,4,5,6],xmm12[7]
2463 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm14[0],xmm0[0],xmm14[1],xmm0[1],xmm14[2],xmm0[2],xmm14[3],xmm0[3]
2467 ; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm14[2,3,4,5,6,7]
2471 ; AVX-NEXT: vmovddup {{.*#+}} xmm15 = [0,9,10,11,12,128,128,128,0,9,10,11,12,128,128,128]
2472 ; AVX-NEXT: # xmm15 = mem[0,0]
2474 ; AVX-NEXT: vpshufb {{.*#+}} xmm8 = xmm6[u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,xmm6[0,7,14]
2479 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm7[0],xmm2[1],xmm7[1],xmm2[2],xmm7[2],xmm2[3],xmm7[3]
2481 ; AVX-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,u,u,u],zero,zero,xmm4[0,7,14,u,u,u,u,u,u,u]
2483 ; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm3[2,3,4,5,6,7]
2493 ; AVX-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]
2494 ; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[10,11,12,13,14,15],xmm6[0,1,2,3,4,5,6,7,8,9]
2495 ; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,u,u,u,u,u,u,u,u,0,1,2,3,8,15]
2496 ; AVX-NEXT: vpblendw $31, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
2497 ; AVX-NEXT: # xmm1 = mem[0,1,2,3,4],xmm1[5,6,7]
2499 ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
2504 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
2506 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
2511 ; AVX2: # %bb.0:
2512 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
2513 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10
2516 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
2520 ; AVX2-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,7,14],zero,zero,xmm2[3,10],zero,zero,zero,xmm2[u,u,u,u,u,u]
2524 ; AVX2-NEXT: vpblendd {{.*#+}} xmm2 = xmm9[0],xmm10[1],xmm9[2],xmm10[3]
2529 ; AVX2-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3,4],xmm4[5,6,7]
2530 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm4 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
2538 ; AVX2-NEXT: vpblendw {{.*#+}} xmm7 = xmm5[0],xmm4[1],xmm5[2,3,4],xmm4[5],xmm5[6,7]
2539 ; AVX2-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,u,u,u,u,u,u,u,u,0,7,14],zero,zero,xmm7[3,10]
2544 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm8 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
2548 ; AVX2-NEXT: vpshufb {{.*#+}} xmm8 = zero,zero,xmm8[0,7,14],zero,zero,xmm8[3,10,u,u,u,u,u,u,u]
2550 ; AVX2-NEXT: vpblendw {{.*#+}} xmm11 = xmm5[0,1],xmm4[2],xmm5[3,4],xmm4[5],xmm5[6,7]
2555 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535]
2561 ; AVX2-NEXT: vpblendd {{.*#+}} xmm10 = xmm10[0],xmm9[1],xmm10[2],xmm9[3]
2563 ; AVX2-NEXT: vpshufb {{.*#+}} xmm12 = xmm2[u,u,u,u,u,u,u,u,u],zero,zero,xmm2[0,7,14],zero,zero
2566 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm11 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
2570 ; AVX2-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[4,11],zero,zero,xmm11[0,7,14],zero,zero,xmm11[u,u,u,u,u,u,u]
2576 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm11 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
2582 ; AVX2-NEXT: vpblendw {{.*#+}} xmm12 = xmm4[0],xmm5[1,2],xmm4[3],xmm5[4,5,6],xmm4[7]
2583 ; AVX2-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[u,u,u,u,u,u,u,u,u,4,11],zero,zero,xmm12[0,7,14]
2587 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm12 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
2591 ; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[4,11],zero,zero,xmm0[0,7,14,u,u,u,u,u,u,u]
2593 ; AVX2-NEXT: vpblendw {{.*#+}} xmm1 = xmm4[0],xmm5[1,2,3],xmm4[4],xmm5[5,6],xmm4[7]
2609 ; AVX2-FP: # %bb.0:
2610 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
2611 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10
2614 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
2618 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,7,14],zero,zero,xmm2[3,10],zero,zero,zero,xmm2[u,u,u,u,u,u]
2622 ; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm2 = xmm9[0],xmm10[1],xmm9[2],xmm10[3]
2627 ; AVX2-FP-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3,4],xmm4[5,6,7]
2628 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm4 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
2636 ; AVX2-FP-NEXT: vpblendw {{.*#+}} xmm7 = xmm5[0],xmm4[1],xmm5[2,3,4],xmm4[5],xmm5[6,7]
2637 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,u,u,u,u,u,u,u,u,0,7,14],zero,zero,xmm7[3,10]
2642 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm8 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
2646 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm8 = zero,zero,xmm8[0,7,14],zero,zero,xmm8[3,10,u,u,u,u,u,u,u]
2648 ; AVX2-FP-NEXT: vpblendw {{.*#+}} xmm11 = xmm5[0,1],xmm4[2],xmm5[3,4],xmm4[5],xmm5[6,7]
2653 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535]
2659 ; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm10 = xmm10[0],xmm9[1],xmm10[2],xmm9[3]
2661 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm12 = xmm2[u,u,u,u,u,u,u,u,u],zero,zero,xmm2[0,7,14],zero,zero
2664 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
2668 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[4,11],zero,zero,xmm11[0,7,14],zero,zero,xmm11[u,u,u,u,u,u,u]
2674 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
2680 ; AVX2-FP-NEXT: vpblendw {{.*#+}} xmm12 = xmm4[0],xmm5[1,2],xmm4[3],xmm5[4,5,6],xmm4[7]
2681 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[u,u,u,u,u,u,u,u,u,4,11],zero,zero,xmm12[0,7,14]
2685 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm12 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
2689 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[4,11],zero,zero,xmm0[0,7,14,u,u,u,u,u,u,u]
2691 ; AVX2-FP-NEXT: vpblendw {{.*#+}} xmm1 = xmm4[0],xmm5[1,2,3],xmm4[4],xmm5[5,6],xmm4[7]
2707 ; AVX2-FCP: # %bb.0:
2708 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
2709 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
2712 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
2716 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,7,14],zero,zero,xmm2[3,10],zero,zero,zero,xmm2[u,u,u,u,u,u]
2720 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm9[0],xmm10[1],xmm9[2],xmm10[3]
2725 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3,4],xmm4[5,6,7]
2726 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm4 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
2734 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} xmm7 = xmm5[0],xmm4[1],xmm5[2,3,4],xmm4[5],xmm5[6,7]
2735 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,u,u,u,u,u,u,u,u,0,7,14],zero,zero,xmm7[3,10]
2740 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm8 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
2744 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm8 = zero,zero,xmm8[0,7,14],zero,zero,xmm8[3,10,u,u,u,u,u,u,u]
2746 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} xmm11 = xmm5[0,1],xmm4[2],xmm5[3,4],xmm4[5],xmm5[6,7]
2751 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535]
2757 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm10 = xmm10[0],xmm9[1],xmm10[2],xmm9[3]
2759 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm12 = xmm2[u,u,u,u,u,u,u,u,u],zero,zero,xmm2[0,7,14],zero,zero
2762 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
2766 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[4,11],zero,zero,xmm11[0,7,14],zero,zero,xmm11[u,u,u,u,u,u,u]
2772 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
2778 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} xmm12 = xmm4[0],xmm5[1,2],xmm4[3],xmm5[4,5,6],xmm4[7]
2779 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[u,u,u,u,u,u,u,u,u,4,11],zero,zero,xmm12[0,7,14]
2783 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm12 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
2787 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[4,11],zero,zero,xmm0[0,7,14,u,u,u,u,u,u,u]
2789 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} xmm1 = xmm4[0],xmm5[1,2,3],xmm4[4],xmm5[5,6],xmm4[7]
2805 ; AVX512: # %bb.0:
2806 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
2807 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10
2813 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm5 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0]
2817 ; AVX512-NEXT: vpblendw {{.*#+}} xmm6 = xmm3[0,1,2,3,4],xmm4[5,6,7]
2820 ; AVX512-NEXT: vpblendd {{.*#+}} xmm7 = xmm3[0],xmm4[1],xmm3[2],xmm4[3]
2821 ; AVX512-NEXT: vmovdqa {{.*#+}} xmm8 = [0,7,14,128,128,3,10,128,128,128,6,13,128,128,2,9]
2824 ; AVX512-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4],xmm7[5,6,7]
2826 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm8 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
2833 ; AVX512-NEXT: vpblendw {{.*#+}} xmm6 = xmm4[0],xmm3[1],xmm4[2,3,4],xmm3[5],xmm4[6,7]
2834 ; AVX512-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u,u,u,u,u,u,u,u,0,7,14],zero,zero,xmm6[3,10]
2839 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535]
2844 ; AVX512-NEXT: vpshufb {{.*#+}} xmm10 = zero,zero,xmm10[0,7,14],zero,zero,xmm10[3,10,u,u,u,u,u,u,u]
2846 ; AVX512-NEXT: vpblendw {{.*#+}} xmm11 = xmm4[0,1],xmm3[2],xmm4[3,4],xmm3[5],xmm4[6,7]
2851 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm10 = [65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535]
2858 ; AVX512-NEXT: vpblendd {{.*#+}} xmm13 = xmm4[0],xmm3[1],xmm4[2],xmm3[3]
2860 ; AVX512-NEXT: vpshufb {{.*#+}} xmm15 = xmm0[u,u,u,u,u,u,u,u,u],zero,zero,xmm0[0,7,14],zero,zero
2866 ; AVX512-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[4,11],zero,zero,xmm8[0,7,14],zero,zero,xmm8[u,u,u,u,u,u,u]
2877 ; AVX512-NEXT: vpblendw {{.*#+}} xmm9 = xmm3[0],xmm4[1,2],xmm3[3],xmm4[4,5,6],xmm3[7]
2878 ; AVX512-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[u,u,u,u,u,u,u,u,u,4,11],zero,zero,xmm9[0,7,14]
2885 ; AVX512-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14,u,u,u,u,u,u,u]
2887 ; AVX512-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0],xmm4[1,2,3],xmm3[4],xmm4[5,6],xmm3[7]
2903 ; AVX512-FCP: # %bb.0:
2904 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
2905 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
2911 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0]
2915 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm6 = xmm3[0,1,2,3,4],xmm4[5,6,7]
2918 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} xmm7 = xmm3[0],xmm4[1],xmm3[2],xmm4[3]
2919 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [0,7,14,128,128,3,10,128,128,128,6,13,128,128,2,9]
2922 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4],xmm7[5,6,7]
2924 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
2931 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm6 = xmm4[0],xmm3[1],xmm4[2,3,4],xmm3[5],xmm4[6,7]
2932 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u,u,u,u,u,u,u,u,0,7,14],zero,zero,xmm6[3,10]
2937 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535]
2942 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm10 = zero,zero,xmm10[0,7,14],zero,zero,xmm10[3,10,u,u,u,u,u,u,u]
2944 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm11 = xmm4[0,1],xmm3[2],xmm4[3,4],xmm3[5],xmm4[6,7]
2949 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535]
2956 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} xmm13 = xmm4[0],xmm3[1],xmm4[2],xmm3[3]
2958 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm0[u,u,u,u,u,u,u,u,u],zero,zero,xmm0[0,7,14],zero,zero
2964 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[4,11],zero,zero,xmm8[0,7,14],zero,zero,xmm8[u,u,u,u,u,u,u]
2975 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm9 = xmm3[0],xmm4[1,2],xmm3[3],xmm4[4,5,6],xmm3[7]
2976 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[u,u,u,u,u,u,u,u,u,4,11],zero,zero,xmm9[0,7,14]
2983 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14,u,u,u,u,u,u,u]
2985 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0],xmm4[1,2,3],xmm3[4],xmm4[5,6],xmm3[7]
3001 ; AVX512DQ: # %bb.0:
3002 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
3003 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10
3009 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm5 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0]
3013 ; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm6 = xmm3[0,1,2,3,4],xmm4[5,6,7]
3016 ; AVX512DQ-NEXT: vpblendd {{.*#+}} xmm7 = xmm3[0],xmm4[1],xmm3[2],xmm4[3]
3017 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm8 = [0,7,14,128,128,3,10,128,128,128,6,13,128,128,2,9]
3020 ; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4],xmm7[5,6,7]
3022 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm8 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
3029 ; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm6 = xmm4[0],xmm3[1],xmm4[2,3,4],xmm3[5],xmm4[6,7]
3030 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u,u,u,u,u,u,u,u,0,7,14],zero,zero,xmm6[3,10]
3035 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535]
3040 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm10 = zero,zero,xmm10[0,7,14],zero,zero,xmm10[3,10,u,u,u,u,u,u,u]
3042 ; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm11 = xmm4[0,1],xmm3[2],xmm4[3,4],xmm3[5],xmm4[6,7]
3047 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm10 = [65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535]
3054 ; AVX512DQ-NEXT: vpblendd {{.*#+}} xmm13 = xmm4[0],xmm3[1],xmm4[2],xmm3[3]
3056 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm15 = xmm0[u,u,u,u,u,u,u,u,u],zero,zero,xmm0[0,7,14],zero,zero
3062 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[4,11],zero,zero,xmm8[0,7,14],zero,zero,xmm8[u,u,u,u,u,u,u]
3073 ; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm9 = xmm3[0],xmm4[1,2],xmm3[3],xmm4[4,5,6],xmm3[7]
3074 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[u,u,u,u,u,u,u,u,u,4,11],zero,zero,xmm9[0,7,14]
3081 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14,u,u,u,u,u,u,u]
3083 ; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0],xmm4[1,2,3],xmm3[4],xmm4[5,6],xmm3[7]
3099 ; AVX512DQ-FCP: # %bb.0:
3100 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
3101 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
3107 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0]
3111 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm6 = xmm3[0,1,2,3,4],xmm4[5,6,7]
3114 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} xmm7 = xmm3[0],xmm4[1],xmm3[2],xmm4[3]
3115 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [0,7,14,128,128,3,10,128,128,128,6,13,128,128,2,9]
3118 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4],xmm7[5,6,7]
3120 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
3127 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm6 = xmm4[0],xmm3[1],xmm4[2,3,4],xmm3[5],xmm4[6,7]
3128 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u,u,u,u,u,u,u,u,0,7,14],zero,zero,xmm6[3,10]
3133 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535]
3138 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm10 = zero,zero,xmm10[0,7,14],zero,zero,xmm10[3,10,u,u,u,u,u,u,u]
3140 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm11 = xmm4[0,1],xmm3[2],xmm4[3,4],xmm3[5],xmm4[6,7]
3145 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535]
3152 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} xmm13 = xmm4[0],xmm3[1],xmm4[2],xmm3[3]
3154 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm0[u,u,u,u,u,u,u,u,u],zero,zero,xmm0[0,7,14],zero,zero
3160 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[4,11],zero,zero,xmm8[0,7,14],zero,zero,xmm8[u,u,u,u,u,u,u]
3171 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm9 = xmm3[0],xmm4[1,2],xmm3[3],xmm4[4,5,6],xmm3[7]
3172 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[u,u,u,u,u,u,u,u,u,4,11],zero,zero,xmm9[0,7,14]
3179 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14,u,u,u,u,u,u,u]
3181 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0],xmm4[1,2,3],xmm3[4],xmm4[5,6],xmm3[7]
3197 ; AVX512BW: # %bb.0:
3198 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
3199 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
3205 ; AVX512BW-NEXT: movw $-28382, %r11w # imm = 0x9122
3210 ; AVX512BW-NEXT: vpblendw {{.*#+}} xmm6 = xmm3[0,1,2,3,4],xmm4[5,6,7]
3213 ; AVX512BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm3[0],xmm4[1],xmm3[2],xmm4[3]
3214 ; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm8 = [0,7,14,128,128,3,10,128,128,128,6,13,128,128,2,9]
3217 ; AVX512BW-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4],xmm7[5,6,7]
3219 ; AVX512BW-NEXT: movw $4644, %di # imm = 0x1224
3226 ; AVX512BW-NEXT: vpblendw {{.*#+}} xmm7 = xmm4[0],xmm3[1],xmm4[2,3,4],xmm3[5],xmm4[6,7]
3227 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,u,u,u,u,u,u,u,u,0,7,14],zero,zero,xmm7[3,10]
3230 ; AVX512BW-NEXT: movw $-512, %di # imm = 0xFE00
3233 ; AVX512BW-NEXT: movw $8772, %di # imm = 0x2244
3238 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,xmm7[0,7,14],zero,zero,xmm7[3,10,u,u,u,u,u,u,u]
3240 ; AVX512BW-NEXT: vpblendw {{.*#+}} xmm8 = xmm4[0,1],xmm3[2],xmm4[3,4],xmm3[5],xmm4[6,7]
3245 ; AVX512BW-NEXT: movw $9288, %di # imm = 0x2448
3252 ; AVX512BW-NEXT: vpblendd {{.*#+}} xmm9 = xmm4[0],xmm3[1],xmm4[2],xmm3[3]
3254 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm11 = xmm0[u,u,u,u,u,u,u,u,u],zero,zero,xmm0[0,7,14],zero,zero
3260 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[4,11],zero,zero,xmm10[0,7,14],zero,zero,xmm10[u,u,u,u,u,u,u]
3271 ; AVX512BW-NEXT: vpblendw {{.*#+}} xmm11 = xmm3[0],xmm4[1,2],xmm3[3],xmm4[4,5,6],xmm3[7]
3272 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[u,u,u,u,u,u,u,u,u,4,11],zero,zero,xmm11[0,7,14]
3279 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,xmm1[4,11],zero,zero,xmm1[0,7,14,u,u,u,u,u,u,u]
3281 ; AVX512BW-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0],xmm4[1,2,3],xmm3[4],xmm4[5,6],xmm3[7]
3297 ; AVX512BW-FCP: # %bb.0:
3298 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
3299 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
3305 ; AVX512BW-FCP-NEXT: movw $-28382, %r11w # imm = 0x9122
3310 ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} xmm6 = xmm3[0,1,2,3,4],xmm4[5,6,7]
3313 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} xmm7 = xmm3[0],xmm4[1],xmm3[2],xmm4[3]
3314 ; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [0,7,14,128,128,3,10,128,128,128,6,13,128,128,2,9]
3317 ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4],xmm7[5,6,7]
3319 ; AVX512BW-FCP-NEXT: movw $4644, %di # imm = 0x1224
3326 ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} xmm7 = xmm4[0],xmm3[1],xmm4[2,3,4],xmm3[5],xmm4[6,7]
3327 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,u,u,u,u,u,u,u,u,0,7,14],zero,zero,xmm7[3,10]
3330 ; AVX512BW-FCP-NEXT: movw $-512, %di # imm = 0xFE00
3333 ; AVX512BW-FCP-NEXT: movw $8772, %di # imm = 0x2244
3338 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,xmm7[0,7,14],zero,zero,xmm7[3,10,u,u,u,u,u,u,u]
3340 ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} xmm8 = xmm4[0,1],xmm3[2],xmm4[3,4],xmm3[5],xmm4[6,7]
3345 ; AVX512BW-FCP-NEXT: movw $9288, %di # imm = 0x2448
3352 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} xmm9 = xmm4[0],xmm3[1],xmm4[2],xmm3[3]
3354 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm11 = xmm0[u,u,u,u,u,u,u,u,u],zero,zero,xmm0[0,7,14],zero,zero
3360 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[4,11],zero,zero,xmm10[0,7,14],zero,zero,xmm10[u,u,u,u,u,u,u]
3371 ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} xmm11 = xmm3[0],xmm4[1,2],xmm3[3],xmm4[4,5,6],xmm3[7]
3372 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[u,u,u,u,u,u,u,u,u,4,11],zero,zero,xmm11[0,7,14]
3379 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,xmm1[4,11],zero,zero,xmm1[0,7,14,u,u,u,u,u,u,u]
3381 ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0],xmm4[1,2,3],xmm3[4],xmm4[5,6],xmm3[7]
3397 ; AVX512DQ-BW: # %bb.0:
3398 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
3399 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
3405 ; AVX512DQ-BW-NEXT: movw $-28382, %r11w # imm = 0x9122
3410 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} xmm6 = xmm3[0,1,2,3,4],xmm4[5,6,7]
3413 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm3[0],xmm4[1],xmm3[2],xmm4[3]
3414 ; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm8 = [0,7,14,128,128,3,10,128,128,128,6,13,128,128,2,9]
3417 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4],xmm7[5,6,7]
3419 ; AVX512DQ-BW-NEXT: movw $4644, %di # imm = 0x1224
3426 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} xmm7 = xmm4[0],xmm3[1],xmm4[2,3,4],xmm3[5],xmm4[6,7]
3427 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,u,u,u,u,u,u,u,u,0,7,14],zero,zero,xmm7[3,10]
3430 ; AVX512DQ-BW-NEXT: movw $-512, %di # imm = 0xFE00
3433 ; AVX512DQ-BW-NEXT: movw $8772, %di # imm = 0x2244
3438 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,xmm7[0,7,14],zero,zero,xmm7[3,10,u,u,u,u,u,u,u]
3440 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} xmm8 = xmm4[0,1],xmm3[2],xmm4[3,4],xmm3[5],xmm4[6,7]
3445 ; AVX512DQ-BW-NEXT: movw $9288, %di # imm = 0x2448
3452 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm9 = xmm4[0],xmm3[1],xmm4[2],xmm3[3]
3454 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm11 = xmm0[u,u,u,u,u,u,u,u,u],zero,zero,xmm0[0,7,14],zero,zero
3460 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[4,11],zero,zero,xmm10[0,7,14],zero,zero,xmm10[u,u,u,u,u,u,u]
3471 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} xmm11 = xmm3[0],xmm4[1,2],xmm3[3],xmm4[4,5,6],xmm3[7]
3472 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[u,u,u,u,u,u,u,u,u,4,11],zero,zero,xmm11[0,7,14]
3479 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,xmm1[4,11],zero,zero,xmm1[0,7,14,u,u,u,u,u,u,u]
3481 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0],xmm4[1,2,3],xmm3[4],xmm4[5,6],xmm3[7]
3497 ; AVX512DQ-BW-FCP: # %bb.0:
3498 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
3499 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
3505 ; AVX512DQ-BW-FCP-NEXT: movw $-28382, %r11w # imm = 0x9122
3510 ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} xmm6 = xmm3[0,1,2,3,4],xmm4[5,6,7]
3513 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} xmm7 = xmm3[0],xmm4[1],xmm3[2],xmm4[3]
3514 ; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [0,7,14,128,128,3,10,128,128,128,6,13,128,128,2,9]
3517 ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3,4],xmm7[5,6,7]
3519 ; AVX512DQ-BW-FCP-NEXT: movw $4644, %di # imm = 0x1224
3526 ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} xmm7 = xmm4[0],xmm3[1],xmm4[2,3,4],xmm3[5],xmm4[6,7]
3527 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,u,u,u,u,u,u,u,u,0,7,14],zero,zero,xmm7[3,10]
3530 ; AVX512DQ-BW-FCP-NEXT: movw $-512, %di # imm = 0xFE00
3533 ; AVX512DQ-BW-FCP-NEXT: movw $8772, %di # imm = 0x2244
3538 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,xmm7[0,7,14],zero,zero,xmm7[3,10,u,u,u,u,u,u,u]
3540 ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} xmm8 = xmm4[0,1],xmm3[2],xmm4[3,4],xmm3[5],xmm4[6,7]
3545 ; AVX512DQ-BW-FCP-NEXT: movw $9288, %di # imm = 0x2448
3552 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} xmm9 = xmm4[0],xmm3[1],xmm4[2],xmm3[3]
3554 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm11 = xmm0[u,u,u,u,u,u,u,u,u],zero,zero,xmm0[0,7,14],zero,zero
3560 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[4,11],zero,zero,xmm10[0,7,14],zero,zero,xmm10[u,u,u,u,u,u,u]
3571 ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} xmm11 = xmm3[0],xmm4[1,2],xmm3[3],xmm4[4,5,6],xmm3[7]
3572 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[u,u,u,u,u,u,u,u,u,4,11],zero,zero,xmm11[0,7,14]
3579 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,xmm1[4,11],zero,zero,xmm1[0,7,14,u,u,u,u,u,u,u]
3581 ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0],xmm4[1,2,3],xmm3[4],xmm4[5,6],xmm3[7]
3596 %strided.vec0 = shufflevector <112 x i8> %wide.vec, <112 x i8> poison, <16 x i32> <i32 0, i32 7, i32 14, i32 21, i32 28, i32 35, i32 42, i32 49, i32 56, i32 63, i32 70, i32 77, i32 84, i32 91, i32 98, i32 105>
3615 ; SSE: # %bb.0:
3616 ; SSE-NEXT: subq $648, %rsp # imm = 0x288
3619 ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3624 ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3626 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3627 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,0,65535,65535,65535,0,65535,65535]
3637 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm10[0],xmm1[1],xmm10[1],xmm1[2],xmm10[2],xmm1[3],xmm10[3],xmm1[4],xmm10[4],xmm1[5],xmm10[5],xmm1[6],xmm10[6],xmm1[7],xmm10[7]
3638 ; SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
3639 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,5,6,7]
3640 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
3641 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm1[0,1,2,3,4,4,5,6]
3642 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,3,2,3]
3643 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,3,2,3,4,5,6,7]
3645 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [255,255,255,255,255,0,0,0,0,0,255,255,255,255,255,255]
3646 ; SSE-NEXT: movdqa {{.*#+}} xmm7 = [65535,65535,0,65535,65535,65535,0,65535]
3649 ; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3652 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3658 ; SSE-NEXT: movdqa {{.*#+}} xmm7 = [65535,65535,65535,65535,0,65535,0,65535]
3661 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm10[0],xmm3[1],xmm10[1],xmm3[2],xmm10[2],xmm3[3],xmm10[3],xmm3[4],xmm10[4],xmm3[5],xmm10[5],xmm3[6],xmm10[6],xmm3[7],xmm10[7]
3665 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,2,1,3]
3666 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,3,4,5,6,7]
3667 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,3,1,1]
3668 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7]
3674 ; SSE-NEXT: movdqa {{.*#+}} xmm7 = [65535,65535,65535,0,65535,65535,0,65535]
3677 ; SSE-NEXT: movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3682 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm10[0],xmm2[1],xmm10[1],xmm2[2],xmm10[2],xmm2[3],xmm10[3],xmm2[4],xmm10[4],xmm2[5],xmm10[5],xmm2[6],xmm10[6],xmm2[7],xmm10[7]
3683 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,4,7]
3685 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,2,1]
3686 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,7,7]
3691 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3692 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm10[0],xmm3[1],xmm10[1],xmm3[2],xmm10[2],xmm3[3],xmm10[3],xmm3[4],xmm10[4],xmm3[5],xmm10[5],xmm3[6],xmm10[6],xmm3[7],xmm10[7]
3693 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3694 ; SSE-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
3695 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,2,1]
3696 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,7]
3698 ; SSE-NEXT: movdqa {{.*#+}} xmm5 = [65535,65535,65535,65535,65535,65535,65535,0]
3705 ; SSE-NEXT: movdqa {{.*#+}} xmm5 = [65535,65535,65535,65535,65535,0,0,0]
3710 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3712 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3716 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3721 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm10[0],xmm2[1],xmm10[1],xmm2[2],xmm10[2],xmm2[3],xmm10[3],xmm2[4],xmm10[4],xmm2[5],xmm10[5],xmm2[6],xmm10[6],xmm2[7],xmm10[7]
3722 ; SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
3723 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,6,7]
3724 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
3725 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,5,6]
3726 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,3,2,3]
3727 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,3,4,5,6,7]
3732 ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3736 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3743 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm10[0],xmm3[1],xmm10[1],xmm3[2],xmm10[2],xmm3[3],xmm10[3],xmm3[4],xmm10[4],xmm3[5],xmm10[5],xmm3[6],xmm10[6],xmm3[7],xmm10[7]
3746 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,2,1,3]
3747 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,3,4,5,6,7]
3748 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,3,1,1]
3749 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7]
3757 ; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3763 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3],xmm0[4],xmm10[4],xmm0[5],xmm10[5],xmm0[6],xmm10[6],xmm0[7],xmm10[7]
3764 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,4,7]
3766 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
3767 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,7,7]
3772 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3773 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm10[0],xmm2[1],xmm10[1],xmm2[2],xmm10[2],xmm2[3],xmm10[3],xmm2[4],xmm10[4],xmm2[5],xmm10[5],xmm2[6],xmm10[6],xmm2[7],xmm10[7]
3774 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3775 ; SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
3776 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
3777 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,7]
3787 ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3788 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,65535,0,65535,65535,0,65535,65535]
3790 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
3791 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
3796 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,65535,65535,0,65535,65535,65535,65535]
3799 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm10[0],xmm1[1],xmm10[1],xmm1[2],xmm10[2],xmm1[3],xmm10[3],xmm1[4],xmm10[4],xmm1[5],xmm10[5],xmm1[6],xmm10[6],xmm1[7],xmm10[7]
3802 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
3803 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,7,6]
3806 ; SSE-NEXT: movdqa {{.*#+}} xmm5 = [255,255,255,255,255,0,0,0,0,255,255,255,255,255,255,255]
3816 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm10[0],xmm1[1],xmm10[1],xmm1[2],xmm10[2],xmm1[3],xmm10[3],xmm1[4],xmm10[4],xmm1[5],xmm10[5],xmm1[6],xmm10[6],xmm1[7],xmm10[7]
3817 ; SSE-NEXT: movdqa {{.*#+}} xmm7 = [65535,0,65535,65535,65535,65535,0,65535]
3823 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,3,2,3]
3824 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[1,0,3,2,4,5,6,7]
3825 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5]
3830 ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [65535,65535,0,65535,65535,65535,0,65535]
3832 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
3839 ; SSE-NEXT: movdqa {{.*#+}} xmm11 = [65535,65535,65535,65535,0,65535,0,65535]
3842 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm10[0],xmm3[1],xmm10[1],xmm3[2],xmm10[2],xmm3[3],xmm10[3],xmm3[4],xmm10[4],xmm3[5],xmm10[5],xmm3[6],xmm10[6],xmm3[7],xmm10[7]
3845 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
3847 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
3848 ; SSE-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
3852 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[0,3,2,3]
3853 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[0,0,3,2,4,5,6,7]
3854 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,4,6,7]
3858 ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0]
3863 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3864 ; SSE-NEXT: movdqa {{.*#+}} xmm5 = [65535,65535,0,65535,65535,0,65535,65535]
3866 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
3868 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
3875 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
3880 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,2,1]
3881 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,7,6]
3893 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3],xmm0[4],xmm5[4],xmm0[5],xmm5[5],xmm0[6],xmm5[6],xmm0[7],xmm5[7]
3894 ; SSE-NEXT: movdqa {{.*#+}} xmm7 = [65535,0,65535,65535,65535,65535,0,65535]
3900 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,3,2,3]
3901 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[1,0,3,2,4,5,6,7]
3902 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,5,5]
3906 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3909 ; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3916 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
3921 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3923 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
3925 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
3933 ; SSE-NEXT: movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3934 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,0,65535,65535,0,65535,65535,65535]
3940 ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3941 ; SSE-NEXT: pshufd {{.*#+}} xmm11 = xmm1[0,2,2,3]
3942 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3945 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3946 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
3948 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
3950 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
3956 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3957 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm10[0,2,2,3]
3958 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3959 ; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3961 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3962 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3963 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3964 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3967 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3968 ; SSE-NEXT: movdqa {{.*#+}} xmm7 = [65535,0,65535,65535,65535,65535,65535,65535]
3970 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
3972 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3975 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
3979 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
3980 ; SSE-NEXT: punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
3981 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
3984 ; SSE-NEXT: punpcklwd {{.*#+}} xmm15 = xmm15[0],xmm6[0],xmm15[1],xmm6[1],xmm15[2],xmm6[2],xmm15[3],xmm6[3]
3986 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3990 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3991 ; SSE-NEXT: movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3993 ; SSE-NEXT: punpcklbw {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3],xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7]
3994 ; SSE-NEXT: pshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,7,5,6,7]
3995 ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm9[0,1,2,0]
3996 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm9[0,1,2,3,6,4,6,5]
3997 ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [65535,65535,65535,65535,65535,65535,0,65535]
3999 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4000 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4001 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4002 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3],xmm0[4],xmm10[4],xmm0[5],xmm10[5],xmm0[6],xmm10[6],xmm0[7],xmm10[7]
4004 ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm0[0,1,2,3,7,5,6,7]
4005 ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm10[0,1,2,0]
4006 ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,6,4,6,5]
4009 ; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4011 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4013 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4016 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4017 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4018 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4021 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4022 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4023 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4024 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm9[0],xmm0[1],xmm9[1],xmm0[2],xmm9[2],xmm0[3],xmm9[3],xmm0[4],xmm9[4],xmm0[5],xmm9[5],xmm0[6],xmm9[6],xmm0[7],xmm9[7]
4026 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4028 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4029 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
4030 ; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4031 ; SSE-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm9[0],xmm6[1],xmm9[1],xmm6[2],xmm9[2],xmm6[3],xmm9[3],xmm6[4],xmm9[4],xmm6[5],xmm9[5],xmm6[6],xmm9[6],xmm6[7],xmm9[7]
4033 ; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4035 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4036 ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4037 ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4039 ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4040 ; SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3]
4042 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,65535,65535,65535,65535,65535,65535,0]
4045 ; SSE-NEXT: pshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
4046 ; SSE-NEXT: # xmm1 = mem[0,3,2,3]
4047 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,0,3,2,4,5,6,7]
4048 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,4,6,7]
4053 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0]
4056 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
4060 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4061 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,65535,65535,0,65535,65535,0,65535]
4063 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
4064 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
4068 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm9[0],xmm1[1],xmm9[1],xmm1[2],xmm9[2],xmm1[3],xmm9[3],xmm1[4],xmm9[4],xmm1[5],xmm9[5],xmm1[6],xmm9[6],xmm1[7],xmm9[7]
4069 ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [65535,0,65535,65535,65535,65535,0,65535]
4078 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[0,3,2,3]
4079 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,1,0,3,4,5,6,7]
4080 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,5,6,7]
4086 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
4087 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
4094 ; SSE-NEXT: punpcklbw {{.*#+}} xmm13 = xmm13[0],xmm6[0],xmm13[1],xmm6[1],xmm13[2],xmm6[2],xmm13[3],xmm6[3],xmm13[4],xmm6[4],xmm13[5],xmm6[5],xmm13[6],xmm6[6],xmm13[7],xmm6[7]
4098 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
4100 ; SSE-NEXT: punpckldq {{.*#+}} xmm11 = xmm11[0],xmm1[0],xmm11[1],xmm1[1]
4103 ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [65535,65535,65535,65535,0,65535,65535,65535]
4106 ; SSE-NEXT: punpcklbw {{.*#+}} xmm11 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3],xmm11[4],xmm6[4],xmm11[5],xmm6[5],xmm11[6],xmm6[6],xmm11[7],xmm6[7]
4109 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm11[0,1,2,1]
4110 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,7]
4113 ; SSE-NEXT: movdqa {{.*#+}} xmm6 = [255,255,255,255,255,0,0,0,0,255,255,255,255,255,255,255]
4116 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm13[0,2,1,3]
4117 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,7]
4118 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,3,3]
4119 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[2,1,0,3,4,5,6,7]
4126 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4127 ; SSE-NEXT: movdqa {{.*#+}} xmm10 = [65535,65535,65535,0,65535,65535,0,65535]
4129 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
4130 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
4135 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
4143 ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [65535,65535,65535,65535,65535,65535,65535,0]
4146 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,3,2,3]
4147 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,1,0,3,4,5,6,7]
4148 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,5,6,7]
4155 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
4156 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
4163 ; SSE-NEXT: punpcklbw {{.*#+}} xmm12 = xmm12[0],xmm6[0],xmm12[1],xmm6[1],xmm12[2],xmm6[2],xmm12[3],xmm6[3],xmm12[4],xmm6[4],xmm12[5],xmm6[5],xmm12[6],xmm6[6],xmm12[7],xmm6[7]
4166 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
4168 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
4169 ; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
4172 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
4176 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
4177 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,7]
4182 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm12[0,2,1,3]
4183 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,7]
4184 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,3,3]
4185 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[2,1,0,3,4,5,6,7]
4191 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4194 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
4199 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
4201 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm2[0,3,2,3,4,5,6,7]
4202 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
4203 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,7,6]
4208 ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [65535,0,65535,65535,65535,0,65535,65535]
4210 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
4212 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
4217 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
4218 ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [65535,0,65535,0,65535,65535,65535,65535]
4224 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm2[3,2,1,0,4,5,6,7]
4225 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,7,7,7,7]
4229 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
4230 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4237 ; SSE-NEXT: punpcklbw {{.*#+}} xmm14 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3],xmm14[4],xmm6[4],xmm14[5],xmm6[5],xmm14[6],xmm6[6],xmm14[7],xmm6[7]
4240 ; SSE-NEXT: pshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
4242 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
4244 ; SSE-NEXT: movdqa {{.*#+}} xmm10 = [65535,65535,65535,65535,65535,65535,65535,0]
4247 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm14[0,2,1,0,4,5,6,7]
4248 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,6,7]
4256 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4257 ; SSE-NEXT: movdqa {{.*#+}} xmm11 = [65535,65535,65535,0,65535,65535,0,65535]
4260 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
4265 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
4267 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,3,4,5,6,7]
4268 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
4269 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,7,6]
4272 ; SSE-NEXT: movdqa {{.*#+}} xmm8 = [65535,0,65535,65535,65535,0,65535,65535]
4276 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
4281 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
4287 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm2[3,2,1,0,4,5,6,7]
4288 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,7,7,7]
4293 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
4294 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
4300 ; SSE-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3],xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
4301 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,0,65535,65535,65535,65535,0,65535]
4306 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
4308 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,0,0,0]
4312 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm5[0,2,1,0,4,5,6,7]
4313 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,7,6,6,7]
4323 ; SSE-NEXT: movdqa {{.*#+}} xmm15 = [65535,65535,0,65535,65535,0,65535,65535]
4333 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
4335 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[0,2,2,3,4,5,6,7]
4336 ; SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
4337 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4339 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
4341 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
4345 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[0,1,3,3]
4346 ; SSE-NEXT: movss {{.*#+}} xmm0 = xmm2[0],xmm0[1,2,3]
4347 ; SSE-NEXT: movdqa {{.*#+}} xmm9 = [65535,0,65535,65535,65535,0,65535,65535]
4350 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
4355 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
4361 ; SSE-NEXT: pshufd $100, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
4362 ; SSE-NEXT: # xmm2 = mem[0,1,2,1]
4363 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,7]
4367 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm5[0,1,0,3]
4368 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[0,3,2,1,4,5,6,7]
4369 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,7]
4388 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
4390 ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm5[0,2,2,3,4,5,6,7]
4391 ; SSE-NEXT: punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
4392 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4394 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
4396 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
4400 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,1,3,3]
4401 ; SSE-NEXT: movss {{.*#+}} xmm0 = xmm5[0],xmm0[1,2,3]
4402 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
4411 ; SSE-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3],xmm5[4],xmm1[4],xmm5[5],xmm1[5],xmm5[6],xmm1[6],xmm5[7],xmm1[7]
4416 ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,1,0,3]
4417 ; SSE-NEXT: pshuflw {{.*#+}} xmm4 = xmm4[0,3,2,1,4,5,6,7]
4418 ; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,7,6,7]
4421 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm13[0,1,2,1]
4422 ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,4,7]
4430 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4431 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,65535,65535,0,65535,65535,0,65535]
4433 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
4436 ; SSE-NEXT: movdqa {{.*#+}} xmm10 = [0,65535,65535,65535,65535,65535,65535,0]
4439 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
4442 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,3]
4443 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm0[0,1,2,3,5,4,7,6]
4447 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4450 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
4451 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm13[0,2,2,3]
4452 ; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm0[0],xmm5[1],xmm0[1]
4456 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
4457 ; SSE-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3],xmm5[4],xmm1[4],xmm5[5],xmm1[5],xmm5[6],xmm1[6],xmm5[7],xmm1[7]
4460 ; SSE-NEXT: punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
4462 ; SSE-NEXT: movss {{.*#+}} xmm8 = xmm5[0],xmm8[1,2,3]
4463 ; SSE-NEXT: movdqa {{.*#+}} xmm6 = [65535,65535,0,65535,65535,0,65535,65535]
4470 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
4471 ; SSE-NEXT: movdqa {{.*#+}} xmm6 = [65535,65535,0,65535,0,65535,65535,65535]
4477 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4478 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
4479 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,3]
4480 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,7,6]
4482 ; SSE-NEXT: movdqa {{.*#+}} xmm7 = [255,255,255,255,255,255,255,255,255,255,255,255,255,0,0,0]
4485 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[2,1,0,3]
4486 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,3,2,4,5,6,7]
4487 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,5,5]
4491 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0]
4496 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4498 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
4501 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
4505 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[0,1,0,3]
4506 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm0[0,1,2,3,5,4,7,6]
4511 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
4512 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[0,2,2,3]
4514 ; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4517 ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1]
4521 ; SSE-NEXT: pshuflw {{.*#+}} xmm10 = xmm10[0,2,2,3,4,5,6,7]
4522 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
4525 ; SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3]
4527 ; SSE-NEXT: movss {{.*#+}} xmm8 = xmm0[0],xmm8[1,2,3]
4530 ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [65535,65535,0,65535,65535,0,65535,65535]
4536 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
4541 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4542 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
4543 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,3]
4544 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,7,6]
4548 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[2,1,0,3]
4549 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,3,2,4,5,6,7]
4550 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,5,5]
4554 ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0]
4563 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm9[0],xmm0[1],xmm9[1],xmm0[2],xmm9[2],xmm0[3],xmm9[3],xmm0[4],xmm9[4],xmm0[5],xmm9[5],xmm0[6],xmm9[6],xmm0[7],xmm9[7]
4564 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,4,6,7]
4566 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm1[0,1,2,1]
4567 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,5,7,6,7]
4569 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
4572 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
4575 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
4577 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,2,1,3]
4578 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,7,6,7]
4579 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,1,2,0]
4580 ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm8[0,1,2,3,4,7,6,5]
4584 ; SSE-NEXT: movss {{.*#+}} xmm10 = xmm0[0],xmm10[1,2,3]
4585 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4586 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
4587 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,3]
4588 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,4,7]
4592 ; SSE-NEXT: pshufd $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
4594 ; SSE-NEXT: pshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload
4595 ; SSE-NEXT: # xmm11 = mem[0,2,2,3]
4596 ; SSE-NEXT: punpckldq {{.*#+}} xmm11 = xmm11[0],xmm0[0],xmm11[1],xmm0[1]
4598 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
4599 ; SSE-NEXT: movdqa {{.*#+}} xmm9 = [65535,65535,65535,0,65535,0,65535,65535]
4606 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,1,0,3,4,5,6,7]
4615 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
4618 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm11[0],xmm0[1],xmm11[1],xmm0[2],xmm11[2],xmm0[3],xmm11[3],xmm0[4],xmm11[4],xmm0[5],xmm11[5],xmm0[6],xmm11[6],xmm0[7],xmm11[7]
4619 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,4,6,7]
4621 ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm2[0,1,2,1]
4622 ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,5,7,6,7]
4624 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
4627 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
4629 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
4631 ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm3[0,2,1,3]
4632 ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,4,7,6,7]
4633 ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm10[0,1,2,0]
4634 ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,4,7,6,5]
4638 ; SSE-NEXT: movss {{.*#+}} xmm10 = xmm0[0],xmm10[1,2,3]
4639 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
4640 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
4642 ; SSE-NEXT: pshufd {{.*#+}} xmm11 = xmm13[0,2,2,3]
4643 ; SSE-NEXT: punpckldq {{.*#+}} xmm11 = xmm11[0],xmm0[0],xmm11[1],xmm0[1]
4645 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm12[0],xmm0[1],xmm12[1],xmm0[2],xmm12[2],xmm0[3],xmm12[3],xmm0[4],xmm12[4],xmm0[5],xmm12[5],xmm0[6],xmm12[6],xmm0[7],xmm12[7]
4651 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,1,0,3,4,5,6,7]
4654 ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm2[0,1,0,3]
4655 ; SSE-NEXT: pshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,5,4,7]
4662 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4664 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4666 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4668 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4670 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4672 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4676 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4680 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
4683 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
4686 ; SSE-NEXT: addq $648, %rsp # imm = 0x288
4690 ; AVX: # %bb.0:
4702 ; AVX-NEXT: vmovdqa {{.*#+}} xmm0 = [u,u,u,0,0,0,0,255,255,255,255,255,u,u,u,u]
4704 ; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4710 ; AVX-NEXT: vpshufb {{.*#+}} xmm2 = xmm10[0,7,14],zero,zero,xmm10[u,u,u,u,u,u,u,u,u,u,u]
4715 ; AVX-NEXT: vmovq {{.*#+}} xmm3 = [255,255,255,255,255,0,0,0,0,0,0,0,0,0,0,0]
4717 ; AVX-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4719 ; AVX-NEXT: vpshufb {{.*#+}} xmm2 = xmm7[u,u,u,u,u,u,u],zero,zero,xmm7[0,7,14,u,u,u,u]
4725 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4732 ; AVX-NEXT: vmovq {{.*#+}} xmm0 = [255,255,255,255,255,0,0,0,0,0,0,0,0,0,0,0]
4736 ; AVX-NEXT: vpshufb {{.*#+}} xmm2 = xmm9[u,u,0,7,14],zero,zero,xmm9[u,u,u,u,u,u,u,u,u]
4741 ; AVX-NEXT: vmovq {{.*#+}} xmm3 = [0,0,255,255,255,255,255,0,0,0,0,0,0,0,0,0]
4743 ; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4745 ; AVX-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4746 ; AVX-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm11[0,7,14,u,u,u,u,u,u,u,u,u,u,u]
4748 ; AVX-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4753 ; AVX-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4761 ; AVX-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4768 ; AVX-NEXT: vmovq {{.*#+}} xmm0 = [0,0,255,255,255,255,255,0,0,0,0,0,0,0,0,0]
4771 ; AVX-NEXT: vpshufb {{.*#+}} xmm3 = xmm8[u,u],zero,zero,xmm8[0,7,14,u,u,u,u,u,u,u,u,u]
4787 ; AVX-NEXT: vpshufb {{.*#+}} xmm13 = xmm5[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
4791 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4796 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4798 ; AVX-NEXT: vpshufb {{.*#+}} xmm3 = xmm1[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm1[0,7,14]
4801 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4804 ; AVX-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
4806 ; AVX-NEXT: vpshufb {{.*#+}} xmm4 = xmm6[u,u,u,u,u,u,0,7,14],zero,zero,xmm6[u,u,u,u,u]
4808 ; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0],xmm2[1,2],xmm3[3,4,5,6,7]
4813 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
4814 ; AVX-NEXT: vbroadcastss {{.*#+}} xmm10 = [0,0,4,11,0,0,4,11,0,0,4,11,0,0,4,11]
4817 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
4818 ; AVX-NEXT: vpblendw $63, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm3 # 16-byte Folded Reload
4819 ; AVX-NEXT: # xmm3 = mem[0,1,2,3,4,5],xmm2[6,7]
4820 ; AVX-NEXT: vbroadcastss {{.*#+}} xmm7 = [0,0,6,13,0,0,6,13,0,0,6,13,0,0,6,13]
4825 ; AVX-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm8[0],xmm2[1],xmm8[1]
4827 ; AVX-NEXT: vpalignr {{.*#+}} xmm2 = xmm2[10,11,12,13,14,15],xmm8[0,1,2,3,4,5,6,7,8,9]
4828 ; AVX-NEXT: vpshufb {{.*#+}} xmm12 = xmm2[u,u,u,u,u,u,u,u,u,u,0,1,2,3,8,15]
4830 ; AVX-NEXT: vpshufb {{.*#+}} xmm13 = xmm2[0,7,14,u,u,u,u,u,u,u,u,u,u,u,u,u]
4832 ; AVX-NEXT: vmovaps {{.*#+}} ymm13 = [65535,65535,65535,65535,65535,0,0,0,0,0,65535,65535,65535,65535,65535,65535]
4833 ; AVX-NEXT: vandps {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm0 # 32-byte Folded Reload
4837 ; AVX-NEXT: vmovaps {{.*#+}} ymm12 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0,0,0,0,0]
4841 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4844 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
4845 ; AVX-NEXT: vpblendw $63, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm4 # 16-byte Folded Reload
4846 ; AVX-NEXT: # xmm4 = mem[0,1,2,3,4,5],xmm0[6,7]
4848 ; AVX-NEXT: vpshufb {{.*#+}} xmm13 = xmm6[u,u,u,u,u,0,7,14,u,u,u,u,u,u,u,u]
4849 ; AVX-NEXT: vpunpckldq {{.*#+}} xmm13 = xmm13[0],xmm3[0],xmm13[1],xmm3[1]
4851 ; AVX-NEXT: vpblendw {{.*#+}} xmm13 = xmm13[0,1,2,3,4,5,6],xmm3[7]
4856 ; AVX-NEXT: vmovaps {{.*#+}} ymm14 = [0,0,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
4864 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4867 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
4868 ; AVX-NEXT: vpblendw $63, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
4869 ; AVX-NEXT: # xmm0 = mem[0,1,2,3,4,5],xmm0[6,7]
4873 ; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,6],xmm3[7]
4876 ; AVX-NEXT: vmovd {{.*#+}} xmm12 = [2,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
4879 ; AVX-NEXT: vmovaps {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
4880 ; AVX-NEXT: vandps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload
4884 ; AVX-NEXT: vmovaps {{.*#+}} ymm13 = [65535,65535,65535,65535,65535,65535,65535,65535,65535,0,0,0,0,0,0,0]
4888 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4890 ; AVX-NEXT: vpshufb {{.*#+}} xmm7 = xmm9[u,u,u,u,u,u,u,u,u],zero,zero,xmm9[0,7,14,u,u]
4892 ; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,5,6],xmm3[7]
4895 ; AVX-NEXT: vmovd {{.*#+}} xmm4 = [3,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
4898 ; AVX-NEXT: vandps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm10 # 32-byte Folded Reload
4902 ; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm10 # 16-byte Folded Reload
4907 ; AVX-NEXT: vmovd {{.*#+}} xmm14 = [4,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
4908 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
4910 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm10[0],xmm7[0],xmm10[1],xmm7[1],xmm10[2],xmm7[2],xmm10[3],xmm7[3]
4911 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
4913 ; AVX-NEXT: vpshufb {{.*#+}} xmm12 = xmm15[u,u,u,u,0,7,14],zero,zero,xmm15[u,u,u,u,u,u,u]
4915 ; AVX-NEXT: vpblendw {{.*#+}} xmm7 = xmm7[0,1],xmm10[2,3,4,5,6,7]
4919 ; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm10[0,1,2,3,4,5,6],xmm3[7]
4928 ; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm7 # 16-byte Folded Reload
4931 ; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
4933 ; AVX-NEXT: vmovd {{.*#+}} xmm7 = [5,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
4936 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm10[0],xmm4[0],xmm10[1],xmm4[1],xmm10[2],xmm4[2],xmm10[3],xmm4[3]
4940 ; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0,1],xmm10[2,3,4,5,6,7]
4944 ; AVX-NEXT: vmovddup {{.*#+}} xmm12 = [0,9,10,11,12,128,128,128,0,9,10,11,12,128,128,128]
4945 ; AVX-NEXT: # xmm12 = mem[0,0]
4947 ; AVX-NEXT: vpshufb {{.*#+}} xmm5 = xmm8[u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,xmm8[0,7,14]
4955 ; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 16-byte Folded Reload
4959 ; AVX-NEXT: vmovd {{.*#+}} xmm7 = [6,13,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
4961 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm10[0],xmm5[0],xmm10[1],xmm5[1],xmm10[2],xmm5[2],xmm10[3],xmm5[3]
4963 ; AVX-NEXT: vpshufb {{.*#+}} xmm11 = xmm0[u,u,u,u],zero,zero,xmm0[0,7,14,u,u,u,u,u,u,u]
4965 ; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1],xmm10[2,3,4,5,6,7]
4978 ; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 16-byte Folded Reload
4981 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
4983 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
4985 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
4989 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
4991 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
4993 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
5000 ; AVX2: # %bb.0:
5008 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm15 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
5012 ; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,7,14],zero,zero,xmm0[3,10],zero,zero,zero,xmm0[u,u,u,u,u,u]
5014 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm14 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
5017 ; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm4[2],ymm1[3,4],ymm4[5],ymm1[6,7,8,9],ymm4[10],ymm1[11,12],ymm4[13],ymm1[14,15]
5019 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm4 = [65535,65535,65535,65535,65535,0,0,0,0,0,0,0,0,0,0,0]
5021 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm1 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
5027 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
5035 ; AVX2-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1],ymm8[2],ymm5[3,4,5],ymm8[6],ymm5[7,8,9],ymm8[10],ymm5[11,12,13],ymm8[14],ymm5[15]
5036 ; AVX2-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,u,0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
5037 ; AVX2-NEXT: vpmovsxdq {{.*#+}} ymm8 = [0,18446744073709551360,16777215,0]
5043 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm12[0],xmm9[1],xmm12[1],xmm9[2],xmm12[2],xmm9[3],xmm12[3]
5046 ; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6],ymm9[7]
5047 ; AVX2-NEXT: vpmovsxdq {{.*#+}} ymm9 = [18446744073709551615,18446744073709551615,16777215,0]
5049 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5053 ; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,u],zero,zero,xmm0[4,11],zero,zero,xmm0[0,7,14,u,u,u,u]
5057 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm12[0],xmm1[0],xmm12[1],xmm1[1],xmm12[2],xmm1[2],xmm12[3],xmm1[3]
5060 ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm1[7]
5069 ; AVX2-NEXT: vpshufb {{.*#+}} xmm8 = xmm4[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
5073 ; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [0,0,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,0,0,0,0,0]
5074 ; AVX2-NEXT: # ymm0 = mem[0,1,0,1]
5076 ; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5077 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm9 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
5081 ; AVX2-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u],zero,zero,xmm1[0,7,14],zero,zero,xmm1[3,10,u,u,u,u,u]
5089 ; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5090 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm8 = [0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535]
5097 ; AVX2-NEXT: vpshufb {{.*#+}} xmm14 = xmm5[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm5[0,7,14]
5102 ; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5104 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm12 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
5108 ; AVX2-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[u,u,4,11],zero,zero,xmm10[0,7,14],zero,zero,xmm10[u,u,u,u,u]
5116 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5117 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm0 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
5122 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
5124 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm0 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
5132 ; AVX2-NEXT: vpshufb {{.*#+}} xmm11 = zero,zero,xmm11[0,7,14],zero,zero,xmm11[3,10,u,u,u,u,u,u,u]
5135 ; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm11[3],ymm0[4,5],ymm11[6],ymm0[7,8,9,10],ymm11[11],ymm0[12,13],ymm11[14],ymm0[15]
5144 ; AVX2-NEXT: vpblendw {{.*#+}} ymm9 = ymm13[0],ymm9[1,2],ymm13[3],ymm9[4,5,6],ymm13[7,8],ymm9[9,10],ymm13[11],ymm9[12,13,14],ymm13[15]
5145 ; AVX2-NEXT: vpshufb {{.*#+}} ymm9 = ymm9[u,u,u,u,u,u,u,u,u,2,9,0,7,14,5,12,19,26,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
5149 ; AVX2-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,11],zero,zero,xmm1[0,7,14],zero,zero,xmm1[u,u,u,u,u,u,u]
5152 ; AVX2-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0],ymm8[1,2,3],ymm9[4],ymm8[5,6],ymm9[7,8],ymm8[9,10,11],ymm9[12],ymm8[13,14],ymm9[15]
5160 ; AVX2-NEXT: vpblendw {{.*#+}} ymm6 = ymm7[0],ymm6[1],ymm7[2,3],ymm6[4],ymm7[5,6,7,8],ymm6[9],ymm7[10,11],ymm6[12],ymm7[13,14,15]
5161 ; AVX2-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,u,u,u,u,u,u,4,11,2,9,0,7,14,21,28,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
5165 ; AVX2-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,xmm7[4,11],zero,zero,xmm7[0,7,14,u,u,u,u,u,u,u]
5168 ; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm7[1],ymm2[2,3,4],ymm7[5],ymm2[6,7,8],ymm7[9],ymm2[10,11,12],ymm7[13],ymm2[14,15]
5173 ; AVX2-NEXT: vpshufb {{.*#+}} xmm7 = xmm14[u,u,0,7,14],zero,zero,xmm14[3,10],zero,zero,zero,xmm14[u,u,u,u]
5177 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
5180 ; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5,6],ymm4[7]
5181 ; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = ymm0[0],ymm4[1,2,3,4,5,6,7],ymm0[8],ymm4[9,10,11,12,13,14,15]
5182 ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
5183 ; AVX2-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm4 # 32-byte Folded Reload
5184 ; AVX2-NEXT: # ymm4 = ymm3[0],mem[1,2,3,4,5,6,7],ymm3[8],mem[9,10,11,12,13,14,15]
5185 ; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]
5186 ; AVX2-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm4 # 32-byte Folded Reload
5187 ; AVX2-NEXT: # ymm4 = ymm9[0],mem[1,2,3,4,5,6,7],ymm9[8],mem[9,10,11,12,13,14,15]
5188 ; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3],ymm4[4,5,6,7]
5189 ; AVX2-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload
5190 ; AVX2-NEXT: # ymm5 = ymm1[0],mem[1,2,3,4,5,6,7],ymm1[8],mem[9,10,11,12,13,14,15]
5191 ; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm5[4,5,6,7]
5192 ; AVX2-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm5 # 32-byte Folded Reload
5193 ; AVX2-NEXT: # ymm5 = ymm2[0],mem[1,2,3,4,5,6,7],ymm2[8],mem[9,10,11,12,13,14,15]
5194 ; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm5[4,5,6,7]
5195 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
5202 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
5204 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
5211 ; AVX2-FP: # %bb.0:
5219 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm15 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
5223 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,7,14],zero,zero,xmm0[3,10],zero,zero,zero,xmm0[u,u,u,u,u,u]
5225 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm14 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
5228 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm4[2],ymm1[3,4],ymm4[5],ymm1[6,7,8,9],ymm4[10],ymm1[11,12],ymm4[13],ymm1[14,15]
5230 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm4 = [65535,65535,65535,65535,65535,0,0,0,0,0,0,0,0,0,0,0]
5232 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
5238 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
5246 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1],ymm8[2],ymm5[3,4,5],ymm8[6],ymm5[7,8,9],ymm8[10],ymm5[11,12,13],ymm8[14],ymm5[15]
5247 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,u,0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
5248 ; AVX2-FP-NEXT: vpmovsxdq {{.*#+}} ymm8 = [0,18446744073709551360,16777215,0]
5254 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm12[0],xmm9[1],xmm12[1],xmm9[2],xmm12[2],xmm9[3],xmm12[3]
5257 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6],ymm9[7]
5258 ; AVX2-FP-NEXT: vpmovsxdq {{.*#+}} ymm9 = [18446744073709551615,18446744073709551615,16777215,0]
5260 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5264 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,u],zero,zero,xmm0[4,11],zero,zero,xmm0[0,7,14,u,u,u,u]
5268 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm12[0],xmm1[0],xmm12[1],xmm1[1],xmm12[2],xmm1[2],xmm12[3],xmm1[3]
5271 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm1[7]
5280 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm8 = xmm4[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
5284 ; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [0,0,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,0,0,0,0,0]
5285 ; AVX2-FP-NEXT: # ymm0 = mem[0,1,0,1]
5287 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5288 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm9 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
5292 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u],zero,zero,xmm1[0,7,14],zero,zero,xmm1[3,10,u,u,u,u,u]
5300 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5301 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm8 = [0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535]
5308 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm14 = xmm5[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm5[0,7,14]
5313 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5315 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm12 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
5319 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[u,u,4,11],zero,zero,xmm10[0,7,14],zero,zero,xmm10[u,u,u,u,u]
5327 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5328 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm0 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
5333 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
5335 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm0 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
5343 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm11 = zero,zero,xmm11[0,7,14],zero,zero,xmm11[3,10,u,u,u,u,u,u,u]
5346 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm11[3],ymm0[4,5],ymm11[6],ymm0[7,8,9,10],ymm11[11],ymm0[12,13],ymm11[14],ymm0[15]
5355 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm9 = ymm13[0],ymm9[1,2],ymm13[3],ymm9[4,5,6],ymm13[7,8],ymm9[9,10],ymm13[11],ymm9[12,13,14],ymm13[15]
5356 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm9 = ymm9[u,u,u,u,u,u,u,u,u,2,9,0,7,14,5,12,19,26,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
5360 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,11],zero,zero,xmm1[0,7,14],zero,zero,xmm1[u,u,u,u,u,u,u]
5363 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0],ymm8[1,2,3],ymm9[4],ymm8[5,6],ymm9[7,8],ymm8[9,10,11],ymm9[12],ymm8[13,14],ymm9[15]
5371 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm6 = ymm7[0],ymm6[1],ymm7[2,3],ymm6[4],ymm7[5,6,7,8],ymm6[9],ymm7[10,11],ymm6[12],ymm7[13,14,15]
5372 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,u,u,u,u,u,u,4,11,2,9,0,7,14,21,28,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
5376 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,xmm7[4,11],zero,zero,xmm7[0,7,14,u,u,u,u,u,u,u]
5379 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm7[1],ymm2[2,3,4],ymm7[5],ymm2[6,7,8],ymm7[9],ymm2[10,11,12],ymm7[13],ymm2[14,15]
5384 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm7 = xmm14[u,u,0,7,14],zero,zero,xmm14[3,10],zero,zero,zero,xmm14[u,u,u,u]
5388 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
5391 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5,6],ymm4[7]
5392 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = ymm0[0],ymm4[1,2,3,4,5,6,7],ymm0[8],ymm4[9,10,11,12,13,14,15]
5393 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
5394 ; AVX2-FP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm4 # 32-byte Folded Reload
5395 ; AVX2-FP-NEXT: # ymm4 = ymm3[0],mem[1,2,3,4,5,6,7],ymm3[8],mem[9,10,11,12,13,14,15]
5396 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]
5397 ; AVX2-FP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm4 # 32-byte Folded Reload
5398 ; AVX2-FP-NEXT: # ymm4 = ymm9[0],mem[1,2,3,4,5,6,7],ymm9[8],mem[9,10,11,12,13,14,15]
5399 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3],ymm4[4,5,6,7]
5400 ; AVX2-FP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload
5401 ; AVX2-FP-NEXT: # ymm5 = ymm1[0],mem[1,2,3,4,5,6,7],ymm1[8],mem[9,10,11,12,13,14,15]
5402 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm5[4,5,6,7]
5403 ; AVX2-FP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm5 # 32-byte Folded Reload
5404 ; AVX2-FP-NEXT: # ymm5 = ymm2[0],mem[1,2,3,4,5,6,7],ymm2[8],mem[9,10,11,12,13,14,15]
5405 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm5[4,5,6,7]
5406 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
5413 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
5415 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
5422 ; AVX2-FCP: # %bb.0:
5431 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm13 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
5435 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,7,14],zero,zero,xmm0[3,10],zero,zero,zero,xmm0[u,u,u,u,u,u]
5437 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
5440 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
5442 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,65535,65535,65535,65535,0,0,0,0,0,0,0,0,0,0,0]
5444 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
5451 ; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,0,0,0,1,2,4,6]
5453 ; AVX2-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5455 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6],ymm2[7]
5456 ; AVX2-FCP-NEXT: vpmovsxdq {{.*#+}} ymm2 = [18446744073709551615,18446744073709551615,16777215,0]
5459 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm12 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
5467 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm7[2],ymm1[3,4,5],ymm7[6],ymm1[7,8,9],ymm7[10],ymm1[11,12,13],ymm7[14],ymm1[15]
5468 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
5469 ; AVX2-FCP-NEXT: vpmovsxdq {{.*#+}} ymm7 = [0,18446744073709551360,16777215,0]
5474 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,u],zero,zero,xmm1[4,11],zero,zero,xmm1[0,7,14,u,u,u,u]
5477 ; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,0,0,0,1,3,4,6]
5480 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6],ymm7[7]
5482 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5491 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm10 = xmm0[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
5495 ; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [0,0,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,0,0,0,0,0]
5496 ; AVX2-FCP-NEXT: # ymm1 = mem[0,1,0,1]
5498 ; AVX2-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5499 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm13 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
5503 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,u],zero,zero,xmm7[0,7,14],zero,zero,xmm7[3,10,u,u,u,u,u]
5511 ; AVX2-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
5512 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm10 = [0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535]
5519 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm12 = xmm2[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm2[0,7,14]
5524 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm7 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
5526 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm7 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
5530 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[u,u,4,11],zero,zero,xmm8[0,7,14],zero,zero,xmm8[u,u,u,u,u]
5542 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
5544 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm0 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
5552 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm2[0,7,14],zero,zero,xmm2[3,10,u,u,u,u,u,u,u]
5555 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm4[3],ymm0[4,5],ymm4[6],ymm0[7,8,9,10],ymm4[11],ymm0[12,13],ymm4[14],ymm0[15]
5564 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm13[1,2],ymm4[3],ymm13[4,5,6],ymm4[7,8],ymm13[9,10],ymm4[11],ymm13[12,13,14],ymm4[15]
5565 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,2,9,0,7,14,5,12,19,26,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
5569 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm9[4,11],zero,zero,xmm9[0,7,14],zero,zero,xmm9[u,u,u,u,u,u,u]
5572 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0],ymm10[1,2,3],ymm8[4],ymm10[5,6],ymm8[7,8],ymm10[9,10,11],ymm8[12],ymm10[13,14],ymm8[15]
5580 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0],ymm8[1],ymm6[2,3],ymm8[4],ymm6[5,6,7,8],ymm8[9],ymm6[10,11],ymm8[12],ymm6[13,14,15]
5581 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,u,u,u,u,u,u,4,11,2,9,0,7,14,21,28,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
5585 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,xmm1[4,11],zero,zero,xmm1[0,7,14,u,u,u,u,u,u,u]
5588 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm6[1],ymm3[2,3,4],ymm6[5],ymm3[6,7,8],ymm6[9],ymm3[10,11,12],ymm6[13],ymm3[14,15]
5593 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm12[u,u,0,7,14],zero,zero,xmm12[3,10],zero,zero,zero,xmm12[u,u,u,u]
5595 ; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,0,0,0,1,3,5,6]
5596 ; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload
5599 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5,6],ymm6[7]
5600 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm2[0],ymm3[1,2,3,4,5,6,7],ymm2[8],ymm3[9,10,11,12,13,14,15]
5601 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
5602 ; AVX2-FCP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm3 # 32-byte Folded Reload
5603 ; AVX2-FCP-NEXT: # ymm3 = ymm4[0],mem[1,2,3,4,5,6,7],ymm4[8],mem[9,10,11,12,13,14,15]
5604 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]
5605 ; AVX2-FCP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload
5606 ; AVX2-FCP-NEXT: # ymm4 = ymm0[0],mem[1,2,3,4,5,6,7],ymm0[8],mem[9,10,11,12,13,14,15]
5607 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
5608 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm15[1,2,3,4,5,6,7],ymm5[8],ymm15[9,10,11,12,13,14,15]
5609 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
5610 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm1[0],ymm14[1,2,3,4,5,6,7],ymm1[8],ymm14[9,10,11,12,13,14,15]
5611 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm5[4,5,6,7]
5614 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
5619 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
5621 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
5628 ; AVX512: # %bb.0:
5629 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
5630 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10
5631 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm0 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535]
5642 ; AVX512-NEXT: vpbroadcastd {{.*#+}} xmm5 = [0,0,4,11,0,0,4,11,0,0,4,11,0,0,4,11]
5647 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
5649 ; AVX512-NEXT: vpblendd {{.*#+}} ymm12 = ymm1[0,1,2,3,4,5,6],ymm6[7]
5650 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm14 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0]
5658 ; AVX512-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[0,7,14],zero,zero,xmm9[3,10],zero,zero,zero,xmm9[u,u,u,u,u,u]
5660 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm11 = [65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535]
5665 ; AVX512-NEXT: vpblendw {{.*#+}} ymm15 = ymm15[0,1],ymm10[2],ymm15[3,4],ymm10[5],ymm15[6,7,8,9],ymm10[10],ymm15[11,12],ymm10[13],ymm15[14,15]
5668 ; AVX512-NEXT: vpmovsxdq {{.*#+}} ymm16 = [18446744073709551615,18446744073709551615,16777215,0]
5675 ; AVX512-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[u,u,u],zero,zero,xmm12[4,11],zero,zero,xmm12[0,7,14,u,u,u,u]
5680 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm13[0],xmm8[1],xmm13[1],xmm8[2],xmm13[2],xmm8[3],xmm13[3]
5682 ; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3,4,5,6],ymm8[7]
5683 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm13 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
5692 ; AVX512-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm10[2],ymm12[3,4,5],ymm10[6],ymm12[7,8,9],ymm10[10],ymm12[11,12,13],ymm10[14],ymm12[15]
5693 ; AVX512-NEXT: vpshufb {{.*#+}} ymm12 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm12[0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
5701 ; AVX512-NEXT: vpshufb {{.*#+}} xmm8 = zero,zero,xmm8[0,7,14],zero,zero,xmm8[3,10,u,u,u,u,u,u,u]
5705 ; AVX512-NEXT: vpblendw {{.*#+}} ymm15 = ymm15[0,1,2],ymm10[3],ymm15[4,5],ymm10[6],ymm15[7,8,9,10],ymm10[11],ymm15[12,13],ymm10[14],ymm15[15]
5712 ; AVX512-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,0,7,14],zero,zero,xmm14[3,10],zero,zero,zero,xmm14[u,u,u,u]
5718 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm14[0],xmm12[1],xmm14[1],xmm12[2],xmm14[2],xmm12[3],xmm14[3]
5720 ; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5,6],ymm12[7]
5721 ; AVX512-NEXT: vpblendw {{.*#+}} ymm8 = ymm15[0],ymm8[1,2,3,4,5,6,7],ymm15[8],ymm8[9,10,11,12,13,14,15]
5722 ; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm15[0,1,2,3],ymm8[4,5,6,7]
5732 ; AVX512-NEXT: vpshufb {{.*#+}} xmm14 = xmm4[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
5745 ; AVX512-NEXT: vpblendw {{.*#+}} ymm14 = ymm10[0],ymm14[1,2],ymm10[3],ymm14[4,5,6],ymm10[7,8],ymm14[9,10],ymm10[11],ymm14[12,13,14],ymm10[15]
5746 ; AVX512-NEXT: vpshufb {{.*#+}} ymm14 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm14[2,9,0,7,14,5,12,19,26],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
5748 ; AVX512-NEXT: vpblendw {{.*#+}} ymm8 = ymm14[0],ymm12[1,2,3,4,5,6,7],ymm14[8],ymm12[9,10,11,12,13,14,15]
5749 ; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm8[4,5,6,7]
5755 ; AVX512-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[u,u],zero,zero,xmm8[0,7,14],zero,zero,xmm8[3,10,u,u,u,u,u]
5767 ; AVX512-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[4,11],zero,zero,xmm8[0,7,14],zero,zero,xmm8[u,u,u,u,u,u,u]
5771 ; AVX512-NEXT: vpblendw {{.*#+}} ymm14 = ymm10[0],ymm14[1,2,3],ymm10[4],ymm14[5,6],ymm10[7,8],ymm14[9,10,11],ymm10[12],ymm14[13,14],ymm10[15]
5774 ; AVX512-NEXT: vpblendw {{.*#+}} ymm8 = ymm14[0],ymm12[1,2,3,4,5,6,7],ymm14[8],ymm12[9,10,11,12,13,14,15]
5775 ; AVX512-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm8[4,5,6,7]
5784 ; AVX512-NEXT: vpshufb {{.*#+}} xmm15 = xmm5[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm5[0,7,14]
5796 ; AVX512-NEXT: vpblendw {{.*#+}} ymm15 = ymm15[0],ymm10[1],ymm15[2,3],ymm10[4],ymm15[5,6,7,8],ymm10[9],ymm15[10,11],ymm10[12],ymm15[13,14,15]
5797 ; AVX512-NEXT: vpshufb {{.*#+}} ymm15 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm15[4,11,2,9,0,7,14,21,28],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
5799 ; AVX512-NEXT: vpblendw {{.*#+}} ymm8 = ymm15[0],ymm12[1,2,3,4,5,6,7],ymm15[8],ymm12[9,10,11,12,13,14,15]
5800 ; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm15[0,1,2,3],ymm8[4,5,6,7]
5804 ; AVX512-NEXT: vpshufb {{.*#+}} xmm3 = xmm13[u,u,4,11],zero,zero,xmm13[0,7,14],zero,zero,xmm13[u,u,u,u,u]
5815 ; AVX512-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,xmm4[4,11],zero,zero,xmm4[0,7,14,u,u,u,u,u,u,u]
5818 ; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm10[1],ymm0[2,3,4],ymm10[5],ymm0[6,7,8],ymm10[9],ymm0[10,11,12],ymm10[13],ymm0[14,15]
5821 ; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm0[0],ymm3[1,2,3,4,5,6,7],ymm0[8],ymm3[9,10,11,12,13,14,15]
5822 ; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
5834 ; AVX512-FCP: # %bb.0:
5835 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
5836 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
5837 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535]
5847 ; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,1,2,4,6]
5851 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm1[0,1,2,3,4,5,6],ymm4[7]
5852 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0]
5860 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[0,7,14],zero,zero,xmm7[3,10],zero,zero,zero,xmm7[u,u,u,u,u,u]
5862 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535]
5867 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm11[0,1],ymm8[2],ymm11[3,4],ymm8[5],ymm11[6,7,8,9],ymm8[10],ymm11[11,12],ymm8[13],ymm11[14,15]
5870 ; AVX512-FCP-NEXT: vpmovsxdq {{.*#+}} ymm16 = [18446744073709551615,18446744073709551615,16777215,0]
5877 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u,u],zero,zero,xmm6[4,11],zero,zero,xmm6[0,7,14,u,u,u,u]
5880 ; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,0,0,0,1,3,4,6]
5883 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5,6],ymm10[7]
5884 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
5893 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm10 = ymm10[0,1],ymm8[2],ymm10[3,4,5],ymm8[6],ymm10[7,8,9],ymm8[10],ymm10[11,12,13],ymm8[14],ymm10[15]
5894 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm10[0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
5901 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm6 = zero,zero,xmm6[0,7,14],zero,zero,xmm6[3,10,u,u,u,u,u,u,u]
5905 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm14 = ymm14[0,1,2],ymm8[3],ymm14[4,5],ymm8[6],ymm14[7,8,9,10],ymm8[11],ymm14[12,13],ymm8[14],ymm14[15]
5912 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm13 = xmm13[u,u,0,7,14],zero,zero,xmm13[3,10],zero,zero,zero,xmm13[u,u,u,u]
5915 ; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm13 = [0,0,0,0,1,3,5,6]
5918 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5,6],ymm12[7]
5919 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm14[0],ymm6[1,2,3,4,5,6,7],ymm14[8],ymm6[9,10,11,12,13,14,15]
5920 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3],ymm6[4,5,6,7]
5932 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm12[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
5945 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm15 = ymm8[0],ymm15[1,2],ymm8[3],ymm15[4,5,6],ymm8[7,8],ymm15[9,10],ymm8[11],ymm15[12,13,14],ymm8[15]
5946 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm15 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm15[2,9,0,7,14,5,12,19,26],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
5948 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm15[0],ymm13[1,2,3,4,5,6,7],ymm15[8],ymm13[9,10,11,12,13,14,15]
5949 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm15[0,1,2,3],ymm6[4,5,6,7]
5955 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u],zero,zero,xmm6[0,7,14],zero,zero,xmm6[3,10,u,u,u,u,u]
5967 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,11],zero,zero,xmm6[0,7,14],zero,zero,xmm6[u,u,u,u,u,u,u]
5971 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm15 = ymm8[0],ymm15[1,2,3],ymm8[4],ymm15[5,6],ymm8[7,8],ymm15[9,10,11],ymm8[12],ymm15[13,14],ymm8[15]
5974 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm15[0],ymm13[1,2,3,4,5,6,7],ymm15[8],ymm13[9,10,11,12,13,14,15]
5975 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm15[0,1,2,3],ymm6[4,5,6,7]
5985 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm13 = xmm14[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm14[0,7,14]
5997 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm15 = ymm15[0],ymm8[1],ymm15[2,3],ymm8[4],ymm15[5,6,7,8],ymm8[9],ymm15[10,11],ymm8[12],ymm15[13,14,15]
5998 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm15 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm15[4,11,2,9,0,7,14,21,28],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
6000 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm15[0],ymm13[1,2,3,4,5,6,7],ymm15[8],ymm13[9,10,11,12,13,14,15]
6001 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm15[0,1,2,3],ymm6[4,5,6,7]
6005 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm11[u,u,4,11],zero,zero,xmm11[0,7,14],zero,zero,xmm11[u,u,u,u,u]
6016 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,xmm4[4,11],zero,zero,xmm4[0,7,14,u,u,u,u,u,u,u]
6019 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm8[1],ymm0[2,3,4],ymm8[5],ymm0[6,7,8],ymm8[9],ymm0[10,11,12],ymm8[13],ymm0[14,15]
6022 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm0[0],ymm3[1,2,3,4,5,6,7],ymm0[8],ymm3[9,10,11,12,13,14,15]
6023 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6035 ; AVX512DQ: # %bb.0:
6036 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
6037 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10
6038 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm0 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535]
6049 ; AVX512DQ-NEXT: vpbroadcastd {{.*#+}} xmm5 = [0,0,4,11,0,0,4,11,0,0,4,11,0,0,4,11]
6054 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
6056 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm12 = ymm1[0,1,2,3,4,5,6],ymm6[7]
6057 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm14 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0]
6065 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[0,7,14],zero,zero,xmm9[3,10],zero,zero,zero,xmm9[u,u,u,u,u,u]
6067 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm11 = [65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535]
6072 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm15 = ymm15[0,1],ymm10[2],ymm15[3,4],ymm10[5],ymm15[6,7,8,9],ymm10[10],ymm15[11,12],ymm10[13],ymm15[14,15]
6075 ; AVX512DQ-NEXT: vpmovsxdq {{.*#+}} ymm16 = [18446744073709551615,18446744073709551615,16777215,0]
6082 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[u,u,u],zero,zero,xmm12[4,11],zero,zero,xmm12[0,7,14,u,u,u,u]
6087 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm13[0],xmm8[1],xmm13[1],xmm8[2],xmm13[2],xmm8[3],xmm13[3]
6089 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3,4,5,6],ymm8[7]
6090 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm13 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
6099 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm10[2],ymm12[3,4,5],ymm10[6],ymm12[7,8,9],ymm10[10],ymm12[11,12,13],ymm10[14],ymm12[15]
6100 ; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm12 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm12[0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
6108 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm8 = zero,zero,xmm8[0,7,14],zero,zero,xmm8[3,10,u,u,u,u,u,u,u]
6112 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm15 = ymm15[0,1,2],ymm10[3],ymm15[4,5],ymm10[6],ymm15[7,8,9,10],ymm10[11],ymm15[12,13],ymm10[14],ymm15[15]
6119 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,0,7,14],zero,zero,xmm14[3,10],zero,zero,zero,xmm14[u,u,u,u]
6125 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm14[0],xmm12[1],xmm14[1],xmm12[2],xmm14[2],xmm12[3],xmm14[3]
6127 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5,6],ymm12[7]
6128 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm8 = ymm15[0],ymm8[1,2,3,4,5,6,7],ymm15[8],ymm8[9,10,11,12,13,14,15]
6129 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm15[0,1,2,3],ymm8[4,5,6,7]
6139 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm14 = xmm4[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
6152 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm14 = ymm10[0],ymm14[1,2],ymm10[3],ymm14[4,5,6],ymm10[7,8],ymm14[9,10],ymm10[11],ymm14[12,13,14],ymm10[15]
6153 ; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm14 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm14[2,9,0,7,14,5,12,19,26],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
6155 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm8 = ymm14[0],ymm12[1,2,3,4,5,6,7],ymm14[8],ymm12[9,10,11,12,13,14,15]
6156 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm8[4,5,6,7]
6162 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[u,u],zero,zero,xmm8[0,7,14],zero,zero,xmm8[3,10,u,u,u,u,u]
6174 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[4,11],zero,zero,xmm8[0,7,14],zero,zero,xmm8[u,u,u,u,u,u,u]
6178 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm14 = ymm10[0],ymm14[1,2,3],ymm10[4],ymm14[5,6],ymm10[7,8],ymm14[9,10,11],ymm10[12],ymm14[13,14],ymm10[15]
6181 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm8 = ymm14[0],ymm12[1,2,3,4,5,6,7],ymm14[8],ymm12[9,10,11,12,13,14,15]
6182 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm8[4,5,6,7]
6191 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm15 = xmm5[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm5[0,7,14]
6203 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm15 = ymm15[0],ymm10[1],ymm15[2,3],ymm10[4],ymm15[5,6,7,8],ymm10[9],ymm15[10,11],ymm10[12],ymm15[13,14,15]
6204 ; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm15 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm15[4,11,2,9,0,7,14,21,28],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
6206 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm8 = ymm15[0],ymm12[1,2,3,4,5,6,7],ymm15[8],ymm12[9,10,11,12,13,14,15]
6207 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm15[0,1,2,3],ymm8[4,5,6,7]
6211 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm3 = xmm13[u,u,4,11],zero,zero,xmm13[0,7,14],zero,zero,xmm13[u,u,u,u,u]
6222 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,xmm4[4,11],zero,zero,xmm4[0,7,14,u,u,u,u,u,u,u]
6225 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm10[1],ymm0[2,3,4],ymm10[5],ymm0[6,7,8],ymm10[9],ymm0[10,11,12],ymm10[13],ymm0[14,15]
6228 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm0[0],ymm3[1,2,3,4,5,6,7],ymm0[8],ymm3[9,10,11,12,13,14,15]
6229 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6241 ; AVX512DQ-FCP: # %bb.0:
6242 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
6243 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
6244 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535]
6254 ; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,1,2,4,6]
6258 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm1[0,1,2,3,4,5,6],ymm4[7]
6259 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0]
6267 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[0,7,14],zero,zero,xmm7[3,10],zero,zero,zero,xmm7[u,u,u,u,u,u]
6269 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535]
6274 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm11[0,1],ymm8[2],ymm11[3,4],ymm8[5],ymm11[6,7,8,9],ymm8[10],ymm11[11,12],ymm8[13],ymm11[14,15]
6277 ; AVX512DQ-FCP-NEXT: vpmovsxdq {{.*#+}} ymm16 = [18446744073709551615,18446744073709551615,16777215,0]
6284 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u,u],zero,zero,xmm6[4,11],zero,zero,xmm6[0,7,14,u,u,u,u]
6287 ; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,0,0,0,1,3,4,6]
6290 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5,6],ymm10[7]
6291 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
6300 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm10 = ymm10[0,1],ymm8[2],ymm10[3,4,5],ymm8[6],ymm10[7,8,9],ymm8[10],ymm10[11,12,13],ymm8[14],ymm10[15]
6301 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm10[0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
6308 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm6 = zero,zero,xmm6[0,7,14],zero,zero,xmm6[3,10,u,u,u,u,u,u,u]
6312 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm14 = ymm14[0,1,2],ymm8[3],ymm14[4,5],ymm8[6],ymm14[7,8,9,10],ymm8[11],ymm14[12,13],ymm8[14],ymm14[15]
6319 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm13 = xmm13[u,u,0,7,14],zero,zero,xmm13[3,10],zero,zero,zero,xmm13[u,u,u,u]
6322 ; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm13 = [0,0,0,0,1,3,5,6]
6325 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5,6],ymm12[7]
6326 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm14[0],ymm6[1,2,3,4,5,6,7],ymm14[8],ymm6[9,10,11,12,13,14,15]
6327 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3],ymm6[4,5,6,7]
6339 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm12[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
6352 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm15 = ymm8[0],ymm15[1,2],ymm8[3],ymm15[4,5,6],ymm8[7,8],ymm15[9,10],ymm8[11],ymm15[12,13,14],ymm8[15]
6353 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm15 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm15[2,9,0,7,14,5,12,19,26],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
6355 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm15[0],ymm13[1,2,3,4,5,6,7],ymm15[8],ymm13[9,10,11,12,13,14,15]
6356 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm15[0,1,2,3],ymm6[4,5,6,7]
6362 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u],zero,zero,xmm6[0,7,14],zero,zero,xmm6[3,10,u,u,u,u,u]
6374 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,11],zero,zero,xmm6[0,7,14],zero,zero,xmm6[u,u,u,u,u,u,u]
6378 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm15 = ymm8[0],ymm15[1,2,3],ymm8[4],ymm15[5,6],ymm8[7,8],ymm15[9,10,11],ymm8[12],ymm15[13,14],ymm8[15]
6381 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm15[0],ymm13[1,2,3,4,5,6,7],ymm15[8],ymm13[9,10,11,12,13,14,15]
6382 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm15[0,1,2,3],ymm6[4,5,6,7]
6392 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm13 = xmm14[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm14[0,7,14]
6404 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm15 = ymm15[0],ymm8[1],ymm15[2,3],ymm8[4],ymm15[5,6,7,8],ymm8[9],ymm15[10,11],ymm8[12],ymm15[13,14,15]
6405 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm15 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm15[4,11,2,9,0,7,14,21,28],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
6407 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm15[0],ymm13[1,2,3,4,5,6,7],ymm15[8],ymm13[9,10,11,12,13,14,15]
6408 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm15[0,1,2,3],ymm6[4,5,6,7]
6412 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm11[u,u,4,11],zero,zero,xmm11[0,7,14],zero,zero,xmm11[u,u,u,u,u]
6423 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,xmm4[4,11],zero,zero,xmm4[0,7,14,u,u,u,u,u,u,u]
6426 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm8[1],ymm0[2,3,4],ymm8[5],ymm0[6,7,8],ymm8[9],ymm0[10,11,12],ymm8[13],ymm0[14,15]
6429 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm0[0],ymm3[1,2,3,4,5,6,7],ymm0[8],ymm3[9,10,11,12,13,14,15]
6430 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6442 ; AVX512BW: # %bb.0:
6443 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
6444 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
6445 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm0 = [16,9,2,19,20,13,6,23,24,0,26,27,28,0,30,31]
6448 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [16,9,2,19,12,5,22,23,24,0,26,27,0,29,30,31]
6450 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [8,1,2,19,12,5,22,15,0,9,26,11,0,29,14,0]
6452 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [8,1,18,11,4,5,22,15,0,25,10,0,12,29,14,0]
6454 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,1,18,11,4,21,14,7,8,25,10,0,28,13,0,15]
6456 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,17,10,3,4,21,14,7,24,9,0,11,28,13,0,31]
6458 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [16,17,10,3,20,13,6,23,24,25,0,27,28,0,30,31]
6462 ; AVX512BW-NEXT: movw $-28382, %r11w # imm = 0x9122
6467 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[0,7,14],zero,zero,xmm1[3,10],zero,zero,zero,xmm1[u,u,u,u,u,u]
6470 ; AVX512BW-NEXT: movw $992, %r11w # imm = 0x3E0
6475 ; AVX512BW-NEXT: movw $8772, %r11w # imm = 0x2244
6488 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm15[0],xmm16[0],xmm15[1],xmm16[1],xmm15[2],xmm16[2],xmm15[3],xmm16[3]
6490 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5,6],ymm15[7]
6491 ; AVX512BW-NEXT: movl $-524288, %edi # imm = 0xFFF80000
6494 ; AVX512BW-NEXT: movw $4644, %di # imm = 0x1224
6501 ; AVX512BW-NEXT: vpshufb {{.*#+}} ymm10 = ymm10[u,u,u,u,u,u,u,u,u,0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
6502 ; AVX512BW-NEXT: movl $511, %edi # imm = 0x1FF
6505 ; AVX512BW-NEXT: movw $9288, %di # imm = 0x2448
6510 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm13 = xmm13[u,u,u],zero,zero,xmm13[4,11],zero,zero,xmm13[0,7,14,u,u,u,u]
6515 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm16[0],xmm15[0],xmm16[1],xmm15[1],xmm16[2],xmm15[2],xmm16[3],xmm15[3]
6517 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5,6],ymm15[7]
6522 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm13 = zero,zero,xmm13[0,7,14],zero,zero,xmm13[3,10,u,u,u,u,u,u,u]
6524 ; AVX512BW-NEXT: movl $261632, %edi # imm = 0x3FE00
6530 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[u,u,0,7,14],zero,zero,xmm12[3,10],zero,zero,zero,xmm12[u,u,u,u]
6535 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
6537 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5,6],ymm14[7]
6538 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm12 = ymm13[0],ymm12[1,2,3,4,5,6,7],ymm13[8],ymm12[9,10,11,12,13,14,15]
6539 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]
6547 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm15 = xmm8[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
6550 ; AVX512BW-NEXT: movl $-134217728, %edi # imm = 0xF8000000
6558 ; AVX512BW-NEXT: vpshufb {{.*#+}} ymm14 {%k4} = ymm11[u,u,u,u,u,u,u,u,u,2,9,0,7,14,5,12,19,26,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
6559 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm11 = ymm14[0],ymm13[1,2,3,4,5,6,7],ymm14[8],ymm13[9,10,11,12,13,14,15]
6560 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm11 = ymm14[0,1,2,3],ymm11[4,5,6,7]
6564 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm13 = xmm13[u,u],zero,zero,xmm13[0,7,14],zero,zero,xmm13[3,10,u,u,u,u,u]
6575 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[4,11],zero,zero,xmm14[0,7,14],zero,zero,xmm14[u,u,u,u,u,u,u]
6578 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm5 = ymm14[0],ymm13[1,2,3,4,5,6,7],ymm14[8],ymm13[9,10,11,12,13,14,15]
6579 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3],ymm5[4,5,6,7]
6587 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm15 = xmm9[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm9[0,7,14]
6596 ; AVX512BW-NEXT: vpshufb {{.*#+}} ymm14 {%k4} = ymm4[u,u,u,u,u,u,u,u,u,4,11,2,9,0,7,14,21,28,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
6597 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm4 = ymm14[0],ymm13[1,2,3,4,5,6,7],ymm14[8],ymm13[9,10,11,12,13,14,15]
6598 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm14[0,1,2,3],ymm4[4,5,6,7]
6602 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u,4,11],zero,zero,xmm6[0,7,14],zero,zero,xmm6[u,u,u,u,u]
6613 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14,u,u,u,u,u,u,u]
6616 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0],ymm6[1,2,3,4,5,6,7],ymm2[8],ymm6[9,10,11,12,13,14,15]
6617 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
6629 ; AVX512BW-FCP: # %bb.0:
6630 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
6631 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
6637 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [8,1,2,19,12,5,22,15,0,9,26,11,4,29,14,7]
6639 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [8,1,18,11,4,5,22,15,0,25,10,3,12,29,14,7]
6641 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,1,18,11,4,21,14,7,8,25,10,3,28,13,6,15]
6643 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,17,10,3,4,21,14,7,24,9,2,11,28,13,6,31]
6649 ; AVX512BW-FCP-NEXT: movw $-28382, %r11w # imm = 0x9122
6654 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[0,7,14],zero,zero,xmm1[3,10],zero,zero,zero,xmm1[u,u,u,u,u,u]
6657 ; AVX512BW-FCP-NEXT: movw $992, %r11w # imm = 0x3E0
6662 ; AVX512BW-FCP-NEXT: movw $8772, %r11w # imm = 0x2244
6670 ; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [0,0,0,0,1,2,4,6]
6674 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5,6],ymm12[7]
6675 ; AVX512BW-FCP-NEXT: movl $-524288, %r11d # imm = 0xFFF80000
6678 ; AVX512BW-FCP-NEXT: movw $4644, %r11w # imm = 0x1224
6685 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[u,u,u,u,u,u,u,u,u,0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
6686 ; AVX512BW-FCP-NEXT: movl $511, %r11d # imm = 0x1FF
6689 ; AVX512BW-FCP-NEXT: movw $9288, %r11w # imm = 0x2448
6694 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[u,u,u],zero,zero,xmm11[4,11],zero,zero,xmm11[0,7,14,u,u,u,u]
6697 ; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [0,0,0,0,1,3,4,6]
6700 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5,6],ymm12[7]
6705 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm11 = zero,zero,xmm11[0,7,14],zero,zero,xmm11[3,10,u,u,u,u,u,u,u]
6707 ; AVX512BW-FCP-NEXT: movl $261632, %r11d # imm = 0x3FE00
6713 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[u,u,0,7,14],zero,zero,xmm10[3,10],zero,zero,zero,xmm10[u,u,u,u]
6716 ; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [0,0,0,0,1,3,5,6]
6719 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5,6],ymm12[7]
6720 ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} ymm10 = ymm11[0],ymm10[1,2,3,4,5,6,7],ymm11[8],ymm10[9,10,11,12,13,14,15]
6721 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
6731 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm12[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
6734 ; AVX512BW-FCP-NEXT: movl $-134217728, %edi # imm = 0xF8000000
6742 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm14 {%k4} = ymm9[u,u,u,u,u,u,u,u,u,2,9,0,7,14,5,12,19,26,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
6743 ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} ymm9 = ymm14[0],ymm13[1,2,3,4,5,6,7],ymm14[8],ymm13[9,10,11,12,13,14,15]
6744 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm14[0,1,2,3],ymm9[4,5,6,7]
6748 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm13 = xmm13[u,u],zero,zero,xmm13[0,7,14],zero,zero,xmm13[3,10,u,u,u,u,u]
6759 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[4,11],zero,zero,xmm14[0,7,14],zero,zero,xmm14[u,u,u,u,u,u,u]
6762 ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm14[0],ymm13[1,2,3,4,5,6,7],ymm14[8],ymm13[9,10,11,12,13,14,15]
6763 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3],ymm5[4,5,6,7]
6771 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm11[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm11[0,7,14]
6780 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm14 {%k4} = ymm4[u,u,u,u,u,u,u,u,u,4,11,2,9,0,7,14,21,28,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
6781 ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm14[0],ymm13[1,2,3,4,5,6,7],ymm14[8],ymm13[9,10,11,12,13,14,15]
6782 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm14[0,1,2,3],ymm4[4,5,6,7]
6786 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u,4,11],zero,zero,xmm6[0,7,14],zero,zero,xmm6[u,u,u,u,u]
6797 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14,u,u,u,u,u,u,u]
6800 ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0],ymm6[1,2,3,4,5,6,7],ymm2[8],ymm6[9,10,11,12,13,14,15]
6801 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
6813 ; AVX512DQ-BW: # %bb.0:
6814 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
6815 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
6816 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm0 = [16,9,2,19,20,13,6,23,24,0,26,27,28,0,30,31]
6819 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [16,9,2,19,12,5,22,23,24,0,26,27,0,29,30,31]
6821 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [8,1,2,19,12,5,22,15,0,9,26,11,0,29,14,0]
6823 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [8,1,18,11,4,5,22,15,0,25,10,0,12,29,14,0]
6825 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,1,18,11,4,21,14,7,8,25,10,0,28,13,0,15]
6827 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,17,10,3,4,21,14,7,24,9,0,11,28,13,0,31]
6829 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [16,17,10,3,20,13,6,23,24,25,0,27,28,0,30,31]
6833 ; AVX512DQ-BW-NEXT: movw $-28382, %r11w # imm = 0x9122
6838 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[0,7,14],zero,zero,xmm1[3,10],zero,zero,zero,xmm1[u,u,u,u,u,u]
6841 ; AVX512DQ-BW-NEXT: movw $992, %r11w # imm = 0x3E0
6846 ; AVX512DQ-BW-NEXT: movw $8772, %r11w # imm = 0x2244
6859 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm15[0],xmm16[0],xmm15[1],xmm16[1],xmm15[2],xmm16[2],xmm15[3],xmm16[3]
6861 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5,6],ymm15[7]
6862 ; AVX512DQ-BW-NEXT: movl $-524288, %edi # imm = 0xFFF80000
6865 ; AVX512DQ-BW-NEXT: movw $4644, %di # imm = 0x1224
6872 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm10 = ymm10[u,u,u,u,u,u,u,u,u,0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
6873 ; AVX512DQ-BW-NEXT: movl $511, %edi # imm = 0x1FF
6876 ; AVX512DQ-BW-NEXT: movw $9288, %di # imm = 0x2448
6881 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm13 = xmm13[u,u,u],zero,zero,xmm13[4,11],zero,zero,xmm13[0,7,14,u,u,u,u]
6886 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm16[0],xmm15[0],xmm16[1],xmm15[1],xmm16[2],xmm15[2],xmm16[3],xmm15[3]
6888 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5,6],ymm15[7]
6893 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm13 = zero,zero,xmm13[0,7,14],zero,zero,xmm13[3,10,u,u,u,u,u,u,u]
6895 ; AVX512DQ-BW-NEXT: movl $261632, %edi # imm = 0x3FE00
6901 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[u,u,0,7,14],zero,zero,xmm12[3,10],zero,zero,zero,xmm12[u,u,u,u]
6906 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
6908 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5,6],ymm14[7]
6909 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm12 = ymm13[0],ymm12[1,2,3,4,5,6,7],ymm13[8],ymm12[9,10,11,12,13,14,15]
6910 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]
6918 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm15 = xmm8[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
6921 ; AVX512DQ-BW-NEXT: movl $-134217728, %edi # imm = 0xF8000000
6929 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm14 {%k4} = ymm11[u,u,u,u,u,u,u,u,u,2,9,0,7,14,5,12,19,26,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
6930 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm11 = ymm14[0],ymm13[1,2,3,4,5,6,7],ymm14[8],ymm13[9,10,11,12,13,14,15]
6931 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm11 = ymm14[0,1,2,3],ymm11[4,5,6,7]
6935 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm13 = xmm13[u,u],zero,zero,xmm13[0,7,14],zero,zero,xmm13[3,10,u,u,u,u,u]
6946 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[4,11],zero,zero,xmm14[0,7,14],zero,zero,xmm14[u,u,u,u,u,u,u]
6949 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm5 = ymm14[0],ymm13[1,2,3,4,5,6,7],ymm14[8],ymm13[9,10,11,12,13,14,15]
6950 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3],ymm5[4,5,6,7]
6958 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm15 = xmm9[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm9[0,7,14]
6967 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm14 {%k4} = ymm4[u,u,u,u,u,u,u,u,u,4,11,2,9,0,7,14,21,28,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
6968 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm4 = ymm14[0],ymm13[1,2,3,4,5,6,7],ymm14[8],ymm13[9,10,11,12,13,14,15]
6969 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm14[0,1,2,3],ymm4[4,5,6,7]
6973 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u,4,11],zero,zero,xmm6[0,7,14],zero,zero,xmm6[u,u,u,u,u]
6984 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14,u,u,u,u,u,u,u]
6987 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0],ymm6[1,2,3,4,5,6,7],ymm2[8],ymm6[9,10,11,12,13,14,15]
6988 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
7000 ; AVX512DQ-BW-FCP: # %bb.0:
7001 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
7002 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
7008 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [8,1,2,19,12,5,22,15,0,9,26,11,4,29,14,7]
7010 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [8,1,18,11,4,5,22,15,0,25,10,3,12,29,14,7]
7012 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,1,18,11,4,21,14,7,8,25,10,3,28,13,6,15]
7014 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,17,10,3,4,21,14,7,24,9,2,11,28,13,6,31]
7020 ; AVX512DQ-BW-FCP-NEXT: movw $-28382, %r11w # imm = 0x9122
7025 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[0,7,14],zero,zero,xmm1[3,10],zero,zero,zero,xmm1[u,u,u,u,u,u]
7028 ; AVX512DQ-BW-FCP-NEXT: movw $992, %r11w # imm = 0x3E0
7033 ; AVX512DQ-BW-FCP-NEXT: movw $8772, %r11w # imm = 0x2244
7041 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [0,0,0,0,1,2,4,6]
7045 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5,6],ymm12[7]
7046 ; AVX512DQ-BW-FCP-NEXT: movl $-524288, %r11d # imm = 0xFFF80000
7049 ; AVX512DQ-BW-FCP-NEXT: movw $4644, %r11w # imm = 0x1224
7056 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[u,u,u,u,u,u,u,u,u,0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
7057 ; AVX512DQ-BW-FCP-NEXT: movl $511, %r11d # imm = 0x1FF
7060 ; AVX512DQ-BW-FCP-NEXT: movw $9288, %r11w # imm = 0x2448
7065 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[u,u,u],zero,zero,xmm11[4,11],zero,zero,xmm11[0,7,14,u,u,u,u]
7068 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [0,0,0,0,1,3,4,6]
7071 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5,6],ymm12[7]
7076 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm11 = zero,zero,xmm11[0,7,14],zero,zero,xmm11[3,10,u,u,u,u,u,u,u]
7078 ; AVX512DQ-BW-FCP-NEXT: movl $261632, %r11d # imm = 0x3FE00
7084 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[u,u,0,7,14],zero,zero,xmm10[3,10],zero,zero,zero,xmm10[u,u,u,u]
7087 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [0,0,0,0,1,3,5,6]
7090 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5,6],ymm12[7]
7091 ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} ymm10 = ymm11[0],ymm10[1,2,3,4,5,6,7],ymm11[8],ymm10[9,10,11,12,13,14,15]
7092 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
7102 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm12[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
7105 ; AVX512DQ-BW-FCP-NEXT: movl $-134217728, %edi # imm = 0xF8000000
7113 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm14 {%k4} = ymm9[u,u,u,u,u,u,u,u,u,2,9,0,7,14,5,12,19,26,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
7114 ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} ymm9 = ymm14[0],ymm13[1,2,3,4,5,6,7],ymm14[8],ymm13[9,10,11,12,13,14,15]
7115 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm14[0,1,2,3],ymm9[4,5,6,7]
7119 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm13 = xmm13[u,u],zero,zero,xmm13[0,7,14],zero,zero,xmm13[3,10,u,u,u,u,u]
7130 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[4,11],zero,zero,xmm14[0,7,14],zero,zero,xmm14[u,u,u,u,u,u,u]
7133 ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm14[0],ymm13[1,2,3,4,5,6,7],ymm14[8],ymm13[9,10,11,12,13,14,15]
7134 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3],ymm5[4,5,6,7]
7142 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm11[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm11[0,7,14]
7151 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm14 {%k4} = ymm4[u,u,u,u,u,u,u,u,u,4,11,2,9,0,7,14,21,28,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
7152 ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm14[0],ymm13[1,2,3,4,5,6,7],ymm14[8],ymm13[9,10,11,12,13,14,15]
7153 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm14[0,1,2,3],ymm4[4,5,6,7]
7157 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u,4,11],zero,zero,xmm6[0,7,14],zero,zero,xmm6[u,u,u,u,u]
7168 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14,u,u,u,u,u,u,u]
7171 ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0],ymm6[1,2,3,4,5,6,7],ymm2[8],ymm6[9,10,11,12,13,14,15]
7172 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
7183 %strided.vec0 = shufflevector <224 x i8> %wide.vec, <224 x i8> poison, <32 x i32> <i32 0, i32 7, i32 14, i32 21, i32 28, i32 35, i32 42, i32 49, i32 56, i32 63, i32 70, i32 77, i32 84, i32 91, i32 98, i32 105, i32 112, i32 119, i32 126, i32 133, i32 140, i32 147, i32 154, i32 161, i32 168, i32 175, i32 182, i32 189, i32 196, i32 203, i32 210, i32 217>
7202 ; SSE: # %bb.0:
7203 ; SSE-NEXT: subq $1528, %rsp # imm = 0x5F8
7206 ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7208 ; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7210 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7212 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7214 ; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7216 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7217 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,0,65535,65535,65535,0,65535,65535]
7227 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
7228 ; SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
7229 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,5,6,7]
7230 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
7231 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm1[0,1,2,3,4,4,5,6]
7232 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,3,2,3]
7233 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,3,2,3,4,5,6,7]
7235 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [255,255,255,255,255,0,0,0,0,0,255,255,255,255,255,255]
7236 ; SSE-NEXT: movdqa {{.*#+}} xmm11 = [65535,65535,0,65535,65535,65535,0,65535]
7244 ; SSE-NEXT: movdqa {{.*#+}} xmm10 = [65535,65535,65535,65535,0,65535,0,65535]
7247 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
7250 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,2,1,3]
7251 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,3,4,5,6,7]
7252 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,3,1,1]
7253 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7]
7259 ; SSE-NEXT: movdqa {{.*#+}} xmm9 = [65535,65535,65535,0,65535,65535,0,65535]
7266 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
7267 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,4,7]
7269 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,2,1]
7270 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,7,7]
7275 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7276 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
7277 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7278 ; SSE-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
7279 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,2,1]
7280 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,7]
7282 ; SSE-NEXT: movdqa {{.*#+}} xmm12 = [65535,65535,65535,65535,65535,65535,65535,0]
7288 ; SSE-NEXT: movdqa {{.*#+}} xmm8 = [65535,65535,65535,65535,65535,0,0,0]
7293 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7295 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7304 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
7305 ; SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
7306 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,6,7]
7307 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
7308 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,5,6]
7309 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,3,2,3]
7310 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,3,4,5,6,7]
7315 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7319 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7326 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
7329 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,2,1,3]
7330 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,3,4,5,6,7]
7331 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,3,1,1]
7332 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7]
7337 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7341 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7345 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
7346 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,4,7]
7348 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,2,1]
7349 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,7,7]
7354 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7355 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
7356 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7357 ; SSE-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
7358 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,2,1]
7359 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,7]
7370 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7372 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7376 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7381 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
7382 ; SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
7383 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,6,7]
7384 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
7385 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,5,6]
7386 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,3,2,3]
7387 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,3,4,5,6,7]
7392 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7396 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7403 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
7406 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,2,1,3]
7407 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,3,4,5,6,7]
7408 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,3,1,1]
7409 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7]
7414 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7419 ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7423 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
7424 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,4,7]
7426 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,2,1]
7427 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,7,7]
7432 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7433 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
7434 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7435 ; SSE-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
7436 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,2,1]
7437 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,7]
7448 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7453 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7458 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
7459 ; SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
7460 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,6,7]
7461 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
7462 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,5,6]
7463 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,3,2,3]
7464 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,3,4,5,6,7]
7467 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7471 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7478 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3],xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
7481 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm4[0,2,1,3]
7482 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,3,4,5,6,7]
7483 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,3,1,1]
7484 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7]
7490 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7494 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7499 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3],xmm0[4],xmm5[4],xmm0[5],xmm5[5],xmm0[6],xmm5[6],xmm0[7],xmm5[7]
7500 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,4,7]
7502 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
7503 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,7,7]
7508 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7509 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3],xmm2[4],xmm5[4],xmm2[5],xmm5[5],xmm2[6],xmm5[6],xmm2[7],xmm5[7]
7510 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7512 ; SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
7513 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
7514 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,7]
7524 ; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7525 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,65535,0,65535,65535,0,65535,65535]
7527 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
7528 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
7534 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,65535,65535,0,65535,65535,65535,65535]
7537 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm7[0],xmm1[1],xmm7[1],xmm1[2],xmm7[2],xmm1[3],xmm7[3],xmm1[4],xmm7[4],xmm1[5],xmm7[5],xmm1[6],xmm7[6],xmm1[7],xmm7[7]
7540 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
7541 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,7,6]
7544 ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [255,255,255,255,255,0,0,0,0,255,255,255,255,255,255,255]
7549 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
7550 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
7554 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm7[0],xmm1[1],xmm7[1],xmm1[2],xmm7[2],xmm1[3],xmm7[3],xmm1[4],xmm7[4],xmm1[5],xmm7[5],xmm1[6],xmm7[6],xmm1[7],xmm7[7]
7555 ; SSE-NEXT: movdqa {{.*#+}} xmm6 = [65535,0,65535,65535,65535,65535,0,65535]
7561 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm4[0,3,2,3]
7562 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[1,0,3,2,4,5,6,7]
7563 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5]
7568 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
7569 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
7576 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm7[0],xmm4[1],xmm7[1],xmm4[2],xmm7[2],xmm4[3],xmm7[3],xmm4[4],xmm7[4],xmm4[5],xmm7[5],xmm4[6],xmm7[6],xmm4[7],xmm7[7]
7579 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
7581 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
7582 ; SSE-NEXT: punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
7586 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm4[0,3,2,3]
7587 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[0,0,3,2,4,5,6,7]
7588 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,4,6,7]
7592 ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0]
7598 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7600 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
7608 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm7[0],xmm2[1],xmm7[1],xmm2[2],xmm7[2],xmm2[3],xmm7[3],xmm2[4],xmm7[4],xmm2[5],xmm7[5],xmm2[6],xmm7[6],xmm2[7],xmm7[7]
7611 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
7612 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,7,6]
7618 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
7619 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
7623 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm7[0],xmm1[1],xmm7[1],xmm1[2],xmm7[2],xmm1[3],xmm7[3],xmm1[4],xmm7[4],xmm1[5],xmm7[5],xmm1[6],xmm7[6],xmm1[7],xmm7[7]
7629 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,3,2,3]
7630 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[1,0,3,2,4,5,6,7]
7631 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5]
7636 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
7637 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
7644 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm7[0],xmm4[1],xmm7[1],xmm4[2],xmm7[2],xmm4[3],xmm7[3],xmm4[4],xmm7[4],xmm4[5],xmm7[5],xmm4[6],xmm7[6],xmm4[7],xmm7[7]
7647 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
7649 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
7650 ; SSE-NEXT: punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
7654 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm4[0,3,2,3]
7655 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[0,0,3,2,4,5,6,7]
7656 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,4,6,7]
7664 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7666 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
7667 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
7674 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm7[0],xmm2[1],xmm7[1],xmm2[2],xmm7[2],xmm2[3],xmm7[3],xmm2[4],xmm7[4],xmm2[5],xmm7[5],xmm2[6],xmm7[6],xmm2[7],xmm7[7]
7677 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,2,1]
7678 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,7,6]
7684 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
7685 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
7689 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm7[0],xmm1[1],xmm7[1],xmm1[2],xmm7[2],xmm1[3],xmm7[3],xmm1[4],xmm7[4],xmm1[5],xmm7[5],xmm1[6],xmm7[6],xmm1[7],xmm7[7]
7695 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,3,2,3]
7696 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[1,0,3,2,4,5,6,7]
7697 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5]
7703 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
7710 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm7[0],xmm4[1],xmm7[1],xmm4[2],xmm7[2],xmm4[3],xmm7[3],xmm4[4],xmm7[4],xmm4[5],xmm7[5],xmm4[6],xmm7[6],xmm4[7],xmm7[7]
7713 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
7715 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
7716 ; SSE-NEXT: punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
7720 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm4[0,3,2,3]
7721 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[0,0,3,2,4,5,6,7]
7722 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,4,6,7]
7730 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7734 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
7741 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
7746 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,2,1]
7747 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,7,6]
7753 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
7754 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
7759 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3],xmm0[4],xmm5[4],xmm0[5],xmm5[5],xmm0[6],xmm5[6],xmm0[7],xmm5[7]
7765 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,3,2,3]
7766 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[1,0,3,2,4,5,6,7]
7767 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,5,5]
7771 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7773 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7775 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
7781 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
7785 ; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7787 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
7789 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
7793 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7797 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
7801 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7803 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
7805 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
7809 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7810 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7816 ; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7817 ; SSE-NEXT: movdqa {{.*#+}} xmm15 = [65535,0,65535,65535,0,65535,65535,65535]
7823 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7824 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm7[0,2,2,3]
7825 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7826 ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7829 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7835 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7836 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm8[0,2,2,3]
7837 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7838 ; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7848 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7849 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm9[0,2,2,3]
7850 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7851 ; SSE-NEXT: movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7854 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7855 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
7857 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
7859 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
7861 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7862 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
7864 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7865 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
7867 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7868 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
7870 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7871 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
7873 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7876 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7880 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7881 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm6[0,2,2,3]
7882 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7883 ; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7885 ; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7886 ; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7887 ; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7888 ; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7889 ; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7890 ; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7891 ; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7892 ; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7895 ; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7896 ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [65535,0,65535,65535,65535,65535,65535,65535]
7898 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
7900 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7904 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
7906 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7909 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
7911 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7914 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
7918 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7919 ; SSE-NEXT: punpcklwd {{.*#+}} xmm10 = xmm10[0],xmm14[0],xmm10[1],xmm14[1],xmm10[2],xmm14[2],xmm10[3],xmm14[3]
7922 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
7923 ; SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm15[0],xmm0[1],xmm15[1],xmm0[2],xmm15[2],xmm0[3],xmm15[3]
7924 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7926 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
7927 ; SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
7928 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7929 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
7932 ; SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
7933 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7935 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7938 ; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7941 ; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7946 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7948 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
7949 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7951 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3],xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
7952 ; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm3[0,1,2,3,7,5,6,7]
7953 ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[0,1,2,0]
7954 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm14[0,1,2,3,6,4,6,5]
7955 ; SSE-NEXT: movdqa {{.*#+}} xmm14 = [65535,65535,65535,65535,65535,65535,0,65535]
7957 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7958 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
7959 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7960 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3],xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
7961 ; SSE-NEXT: pshufhw {{.*#+}} xmm15 = xmm3[0,1,2,3,7,5,6,7]
7962 ; SSE-NEXT: pshufd {{.*#+}} xmm15 = xmm15[0,1,2,0]
7963 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm15[0,1,2,3,6,4,6,5]
7965 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7966 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
7967 ; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7968 ; SSE-NEXT: punpcklbw {{.*#+}} xmm15 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3],xmm15[4],xmm0[4],xmm15[5],xmm0[5],xmm15[6],xmm0[6],xmm15[7],xmm0[7]
7970 ; SSE-NEXT: pshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,7,5,6,7]
7971 ; SSE-NEXT: pshufd {{.*#+}} xmm15 = xmm15[0,1,2,0]
7972 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm15[0,1,2,3,6,4,6,5]
7974 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7975 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
7976 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7977 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
7978 ; SSE-NEXT: pshufhw {{.*#+}} xmm15 = xmm0[0,1,2,3,7,5,6,7]
7979 ; SSE-NEXT: pshufd {{.*#+}} xmm15 = xmm15[0,1,2,0]
7980 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm15[0,1,2,3,6,4,6,5]
7983 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7986 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7988 ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7991 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7993 ; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7997 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7999 ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8002 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8003 ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8004 ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8005 ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8006 ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8008 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8009 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8010 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8012 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3],xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
8013 ; SSE-NEXT: movdqa {{.*#+}} xmm11 = [65535,0,65535,65535,65535,65535,65535,65535]
8015 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8017 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8019 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8020 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3],xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
8024 ; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8025 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8026 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8027 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3],xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
8030 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8032 ; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8033 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
8034 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8035 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3],xmm0[4],xmm8[4],xmm0[5],xmm8[5],xmm0[6],xmm8[6],xmm0[7],xmm8[7]
8037 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8039 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8040 ; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8041 ; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8042 ; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8043 ; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8045 ; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8046 ; SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm13[0],xmm2[1],xmm13[1],xmm2[2],xmm13[2],xmm2[3],xmm13[3]
8048 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,65535,65535,65535,65535,65535,65535,0]
8051 ; SSE-NEXT: pshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
8052 ; SSE-NEXT: # xmm3 = mem[0,3,2,3]
8053 ; SSE-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[0,0,3,2,4,5,6,7]
8054 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,4,6,7]
8059 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0]
8062 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
8066 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8067 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,65535,65535,0,65535,65535,0,65535]
8069 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
8070 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
8075 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm8[0],xmm3[1],xmm8[1],xmm3[2],xmm8[2],xmm3[3],xmm8[3],xmm3[4],xmm8[4],xmm3[5],xmm8[5],xmm3[6],xmm8[6],xmm3[7],xmm8[7]
8076 ; SSE-NEXT: movdqa {{.*#+}} xmm14 = [65535,0,65535,65535,65535,65535,0,65535]
8085 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm15[0,3,2,3]
8086 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,1,0,3,4,5,6,7]
8087 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,5,6,7]
8093 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
8094 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
8100 ; SSE-NEXT: punpcklbw {{.*#+}} xmm9 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3],xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
8103 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
8105 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
8106 ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
8109 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,65535,65535,65535,0,65535,65535,65535]
8112 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3],xmm1[4],xmm8[4],xmm1[5],xmm8[5],xmm1[6],xmm8[6],xmm1[7],xmm8[7]
8115 ; SSE-NEXT: pshufd {{.*#+}} xmm11 = xmm1[0,1,2,1]
8116 ; SSE-NEXT: pshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,5,4,7]
8119 ; SSE-NEXT: movdqa {{.*#+}} xmm6 = [255,255,255,255,255,0,0,0,0,255,255,255,255,255,255,255]
8122 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm9[0,2,1,3]
8123 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,7]
8124 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,3,3]
8125 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[2,1,0,3,4,5,6,7]
8131 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8134 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
8135 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
8139 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3],xmm1[4],xmm8[4],xmm1[5],xmm8[5],xmm1[6],xmm8[6],xmm1[7],xmm8[7]
8145 ; SSE-NEXT: packuswb {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
8148 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,3,2,3]
8149 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,1,0,3,4,5,6,7]
8150 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,5,6,7]
8157 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
8158 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
8164 ; SSE-NEXT: punpcklbw {{.*#+}} xmm12 = xmm12[0],xmm8[0],xmm12[1],xmm8[1],xmm12[2],xmm8[2],xmm12[3],xmm8[3],xmm12[4],xmm8[4],xmm12[5],xmm8[5],xmm12[6],xmm8[6],xmm12[7],xmm8[7]
8167 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
8169 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
8170 ; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
8175 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm8[0],xmm4[1],xmm8[1],xmm4[2],xmm8[2],xmm4[3],xmm8[3],xmm4[4],xmm8[4],xmm4[5],xmm8[5],xmm4[6],xmm8[6],xmm4[7],xmm8[7]
8178 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm4[0,1,2,1]
8179 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,7]
8184 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm12[0,2,1,3]
8185 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,7]
8186 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,3,3]
8187 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[2,1,0,3,4,5,6,7]
8193 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8195 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
8196 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
8200 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3],xmm1[4],xmm8[4],xmm1[5],xmm8[5],xmm1[6],xmm8[6],xmm1[7],xmm8[7]
8206 ; SSE-NEXT: packuswb {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
8209 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,3,2,3]
8210 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,1,0,3,4,5,6,7]
8211 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,5,6,7]
8217 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
8218 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
8219 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
8225 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm8[0],xmm4[1],xmm8[1],xmm4[2],xmm8[2],xmm4[3],xmm8[3],xmm4[4],xmm8[4],xmm4[5],xmm8[5],xmm4[6],xmm8[6],xmm4[7],xmm8[7]
8229 ; SSE-NEXT: pshufd $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
8231 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
8232 ; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
8237 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm8[0],xmm4[1],xmm8[1],xmm4[2],xmm8[2],xmm4[3],xmm8[3],xmm4[4],xmm8[4],xmm4[5],xmm8[5],xmm4[6],xmm8[6],xmm4[7],xmm8[7]
8240 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm4[0,1,2,1]
8241 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,7]
8246 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm5[0,2,1,3]
8247 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,7]
8248 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,3,3]
8249 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[2,1,0,3,4,5,6,7]
8255 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8257 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
8258 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
8262 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3],xmm1[4],xmm8[4],xmm1[5],xmm8[5],xmm1[6],xmm8[6],xmm1[7],xmm8[7]
8268 ; SSE-NEXT: packuswb {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
8271 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,3,2,3]
8272 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,1,0,3,4,5,6,7]
8273 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,5,6,7]
8279 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
8280 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
8281 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
8287 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm8[0],xmm4[1],xmm8[1],xmm4[2],xmm8[2],xmm4[3],xmm8[3],xmm4[4],xmm8[4],xmm4[5],xmm8[5],xmm4[6],xmm8[6],xmm4[7],xmm8[7]
8290 ; SSE-NEXT: pshufd $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
8292 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8293 ; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
8296 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm8[0],xmm3[1],xmm8[1],xmm3[2],xmm8[2],xmm3[3],xmm8[3],xmm3[4],xmm8[4],xmm3[5],xmm8[5],xmm3[6],xmm8[6],xmm3[7],xmm8[7]
8300 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,2,1]
8301 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,4,7]
8306 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,2,1,3]
8307 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,7,6,7]
8308 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,3,3]
8309 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[2,1,0,3,4,5,6,7]
8315 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8319 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
8324 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3],xmm1[4],xmm8[4],xmm1[5],xmm8[5],xmm1[6],xmm8[6],xmm1[7],xmm8[7]
8326 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,3,4,5,6,7]
8327 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
8328 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,7,6]
8333 ; SSE-NEXT: movdqa {{.*#+}} xmm12 = [65535,0,65535,65535,65535,0,65535,65535]
8335 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
8336 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
8340 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3],xmm1[4],xmm8[4],xmm1[5],xmm8[5],xmm1[6],xmm8[6],xmm1[7],xmm8[7]
8341 ; SSE-NEXT: movdqa {{.*#+}} xmm5 = [65535,0,65535,0,65535,65535,65535,65535]
8347 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm2[3,2,1,0,4,5,6,7]
8348 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,7,7,7,7]
8352 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
8353 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
8354 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8360 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm8[0],xmm3[1],xmm8[1],xmm3[2],xmm8[2],xmm3[3],xmm8[3],xmm3[4],xmm8[4],xmm3[5],xmm8[5],xmm3[6],xmm8[6],xmm3[7],xmm8[7]
8363 ; SSE-NEXT: pshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
8365 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
8370 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm3[0,2,1,0,4,5,6,7]
8371 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,6,7]
8380 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8383 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
8388 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3],xmm1[4],xmm8[4],xmm1[5],xmm8[5],xmm1[6],xmm8[6],xmm1[7],xmm8[7]
8390 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,3,4,5,6,7]
8391 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
8392 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,7,6]
8398 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
8400 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
8405 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3],xmm1[4],xmm8[4],xmm1[5],xmm8[5],xmm1[6],xmm8[6],xmm1[7],xmm8[7]
8411 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm2[3,2,1,0,4,5,6,7]
8412 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,7,7,7,7]
8416 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
8417 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
8418 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8424 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm8[0],xmm3[1],xmm8[1],xmm3[2],xmm8[2],xmm3[3],xmm8[3],xmm3[4],xmm8[4],xmm3[5],xmm8[5],xmm3[6],xmm8[6],xmm3[7],xmm8[7]
8427 ; SSE-NEXT: pshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
8429 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
8433 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm3[0,2,1,0,4,5,6,7]
8434 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,6,7]
8442 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8444 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
8445 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
8451 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3],xmm1[4],xmm8[4],xmm1[5],xmm8[5],xmm1[6],xmm8[6],xmm1[7],xmm8[7]
8453 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,3,4,5,6,7]
8454 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
8455 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,7,6]
8461 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
8462 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
8467 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3],xmm1[4],xmm8[4],xmm1[5],xmm8[5],xmm1[6],xmm8[6],xmm1[7],xmm8[7]
8473 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm2[3,2,1,0,4,5,6,7]
8474 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,7,7,7,7]
8478 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
8479 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
8481 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8487 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm8[0],xmm3[1],xmm8[1],xmm3[2],xmm8[2],xmm3[3],xmm8[3],xmm3[4],xmm8[4],xmm3[5],xmm8[5],xmm3[6],xmm8[6],xmm3[7],xmm8[7]
8490 ; SSE-NEXT: pshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
8492 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
8496 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm3[0,2,1,0,4,5,6,7]
8497 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,6,7]
8505 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8507 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
8508 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
8513 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm8[0],xmm2[1],xmm8[1],xmm2[2],xmm8[2],xmm2[3],xmm8[3],xmm2[4],xmm8[4],xmm2[5],xmm8[5],xmm2[6],xmm8[6],xmm2[7],xmm8[7]
8515 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm2[0,3,2,3,4,5,6,7]
8516 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
8517 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,7,6]
8521 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
8523 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
8527 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3],xmm0[4],xmm8[4],xmm0[5],xmm8[5],xmm0[6],xmm8[6],xmm0[7],xmm8[7]
8533 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm2[3,2,1,0,4,5,6,7]
8534 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,7,7,7]
8539 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
8540 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
8541 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8545 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm8[0],xmm3[1],xmm8[1],xmm3[2],xmm8[2],xmm3[3],xmm8[3],xmm3[4],xmm8[4],xmm3[5],xmm8[5],xmm3[6],xmm8[6],xmm3[7],xmm8[7]
8549 ; SSE-NEXT: pshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
8551 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,0,0,0]
8555 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm14[0,2,1,0,4,5,6,7]
8556 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,7,6,6,7]
8564 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8565 ; SSE-NEXT: movdqa {{.*#+}} xmm8 = [65535,65535,0,65535,65535,0,65535,65535]
8567 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
8568 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
8575 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
8577 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
8578 ; SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
8579 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
8582 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8584 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
8588 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,1,3,3]
8589 ; SSE-NEXT: movss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
8591 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
8592 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
8597 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
8603 ; SSE-NEXT: pshufd $100, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
8604 ; SSE-NEXT: # xmm1 = mem[0,1,2,1]
8605 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,7]
8609 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,1,0,3]
8610 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7]
8611 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,7]
8619 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8628 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3],xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
8630 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[0,2,2,3,4,5,6,7]
8631 ; SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
8632 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
8634 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
8636 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8640 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[0,1,3,3]
8641 ; SSE-NEXT: movss {{.*#+}} xmm0 = xmm2[0],xmm0[1,2,3]
8643 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
8644 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
8649 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
8655 ; SSE-NEXT: pshufd $100, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
8656 ; SSE-NEXT: # xmm2 = mem[0,1,2,1]
8657 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,7]
8661 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[0,1,0,3]
8662 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[0,3,2,1,4,5,6,7]
8663 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,7]
8671 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8673 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
8680 ; SSE-NEXT: punpcklbw {{.*#+}} xmm9 = xmm9[0],xmm1[0],xmm9[1],xmm1[1],xmm9[2],xmm1[2],xmm9[3],xmm1[3],xmm9[4],xmm1[4],xmm9[5],xmm1[5],xmm9[6],xmm1[6],xmm9[7],xmm1[7]
8682 ; SSE-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[0,2,2,3,4,5,6,7]
8683 ; SSE-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
8684 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
8686 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
8688 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
8692 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,1,3,3]
8693 ; SSE-NEXT: movss {{.*#+}} xmm0 = xmm3[0],xmm0[1,2,3]
8696 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
8701 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
8707 ; SSE-NEXT: pshufd $100, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
8708 ; SSE-NEXT: # xmm3 = mem[0,1,2,1]
8709 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,7]
8713 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm4[0,1,0,3]
8714 ; SSE-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[0,3,2,1,4,5,6,7]
8715 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,7]
8723 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8726 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
8733 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3],xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
8735 ; SSE-NEXT: pshuflw {{.*#+}} xmm4 = xmm4[0,2,2,3,4,5,6,7]
8736 ; SSE-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
8737 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
8739 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
8741 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8745 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[0,1,3,3]
8746 ; SSE-NEXT: movss {{.*#+}} xmm0 = xmm4[0],xmm0[1,2,3]
8747 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
8750 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
8753 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3],xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
8759 ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm5[0,1,0,3]
8760 ; SSE-NEXT: pshuflw {{.*#+}} xmm4 = xmm4[0,3,2,1,4,5,6,7]
8761 ; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,7,6,7]
8764 ; SSE-NEXT: pshufd $100, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
8765 ; SSE-NEXT: # xmm5 = mem[0,1,2,1]
8766 ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,4,7]
8774 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8775 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
8776 ; SSE-NEXT: movdqa {{.*#+}} xmm13 = [65535,65535,65535,0,65535,65535,0,65535]
8778 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
8781 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [0,65535,65535,65535,65535,65535,65535,0]
8784 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm12[0],xmm4[1],xmm12[1],xmm4[2],xmm12[2],xmm4[3],xmm12[3],xmm4[4],xmm12[4],xmm4[5],xmm12[5],xmm4[6],xmm12[6],xmm4[7],xmm12[7]
8787 ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,1,0,3]
8788 ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm4[0,1,2,3,5,4,7,6]
8792 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
8794 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8796 ; SSE-NEXT: pshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload
8797 ; SSE-NEXT: # xmm6 = mem[0,2,2,3]
8798 ; SSE-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm4[0],xmm6[1],xmm4[1]
8802 ; SSE-NEXT: pshuflw {{.*#+}} xmm4 = xmm4[0,2,2,3,4,5,6,7]
8803 ; SSE-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm12[0],xmm6[1],xmm12[1],xmm6[2],xmm12[2],xmm6[3],xmm12[3],xmm6[4],xmm12[4],xmm6[5],xmm12[5],xmm6[6],xmm12[6],xmm6[7],xmm12[7]
8806 ; SSE-NEXT: punpcklwd {{.*#+}} xmm6 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
8808 ; SSE-NEXT: movss {{.*#+}} xmm7 = xmm6[0],xmm7[1,2,3]
8811 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
8817 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm12[0],xmm4[1],xmm12[1],xmm4[2],xmm12[2],xmm4[3],xmm12[3],xmm4[4],xmm12[4],xmm4[5],xmm12[5],xmm4[6],xmm12[6],xmm4[7],xmm12[7]
8818 ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [65535,65535,0,65535,0,65535,65535,65535]
8824 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8825 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
8826 ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,1,0,3]
8827 ; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,4,7,6]
8829 ; SSE-NEXT: movdqa {{.*#+}} xmm6 = [255,255,255,255,255,255,255,255,255,255,255,255,255,0,0,0]
8832 ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm8[2,1,0,3]
8833 ; SSE-NEXT: pshuflw {{.*#+}} xmm4 = xmm4[0,0,3,2,4,5,6,7]
8834 ; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,5,5,5,5]
8843 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8844 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
8847 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload
8852 ; SSE-NEXT: punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm12[0],xmm7[1],xmm12[1],xmm7[2],xmm12[2],xmm7[3],xmm12[3],xmm7[4],xmm12[4],xmm7[5],xmm12[5],xmm7[6],xmm12[6],xmm7[7],xmm12[7]
8855 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,1,0,3]
8856 ; SSE-NEXT: pshufhw {{.*#+}} xmm9 = xmm7[0,1,2,3,5,4,7,6]
8860 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8862 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8864 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
8865 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm4[0,2,2,3]
8866 ; SSE-NEXT: punpckldq {{.*#+}} xmm8 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]
8870 ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm7[0,2,2,3,4,5,6,7]
8871 ; SSE-NEXT: punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm12[0],xmm8[1],xmm12[1],xmm8[2],xmm12[2],xmm8[3],xmm12[3],xmm8[4],xmm12[4],xmm8[5],xmm12[5],xmm8[6],xmm12[6],xmm8[7],xmm12[7]
8874 ; SSE-NEXT: punpcklwd {{.*#+}} xmm8 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
8876 ; SSE-NEXT: movss {{.*#+}} xmm9 = xmm8[0],xmm9[1,2,3]
8878 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload
8883 ; SSE-NEXT: punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm12[0],xmm7[1],xmm12[1],xmm7[2],xmm12[2],xmm7[3],xmm12[3],xmm7[4],xmm12[4],xmm7[5],xmm12[5],xmm7[6],xmm12[6],xmm7[7],xmm12[7]
8889 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8890 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
8891 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm3[0,1,0,3]
8892 ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,7,6]
8896 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm8[2,1,0,3]
8897 ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm7[0,0,3,2,4,5,6,7]
8898 ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5]
8906 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
8908 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload
8913 ; SSE-NEXT: punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm12[0],xmm8[1],xmm12[1],xmm8[2],xmm12[2],xmm8[3],xmm12[3],xmm8[4],xmm12[4],xmm8[5],xmm12[5],xmm8[6],xmm12[6],xmm8[7],xmm12[7]
8916 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,1,0,3]
8917 ; SSE-NEXT: pshufhw {{.*#+}} xmm15 = xmm8[0,1,2,3,5,4,7,6]
8921 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8924 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8925 ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm3[0,2,2,3]
8926 ; SSE-NEXT: punpckldq {{.*#+}} xmm9 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]
8930 ; SSE-NEXT: pshuflw {{.*#+}} xmm8 = xmm8[0,2,2,3,4,5,6,7]
8931 ; SSE-NEXT: punpcklbw {{.*#+}} xmm9 = xmm9[0],xmm12[0],xmm9[1],xmm12[1],xmm9[2],xmm12[2],xmm9[3],xmm12[3],xmm9[4],xmm12[4],xmm9[5],xmm12[5],xmm9[6],xmm12[6],xmm9[7],xmm12[7]
8934 ; SSE-NEXT: punpcklwd {{.*#+}} xmm9 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
8936 ; SSE-NEXT: movss {{.*#+}} xmm15 = xmm9[0],xmm15[1,2,3]
8943 ; SSE-NEXT: punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm12[0],xmm8[1],xmm12[1],xmm8[2],xmm12[2],xmm8[3],xmm12[3],xmm8[4],xmm12[4],xmm8[5],xmm12[5],xmm8[6],xmm12[6],xmm8[7],xmm12[7]
8949 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
8950 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
8951 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm2[0,1,0,3]
8952 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,4,7,6]
8956 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm9[2,1,0,3]
8957 ; SSE-NEXT: pshuflw {{.*#+}} xmm8 = xmm8[0,0,3,2,4,5,6,7]
8958 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,5,5,5,5]
8966 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
8968 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload
8971 ; SSE-NEXT: punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm12[0],xmm8[1],xmm12[1],xmm8[2],xmm12[2],xmm8[3],xmm12[3],xmm8[4],xmm12[4],xmm8[5],xmm12[5],xmm8[6],xmm12[6],xmm8[7],xmm12[7]
8975 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,3]
8976 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,4,7,6]
8980 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
8984 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8985 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8986 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm3[0,2,2,3]
8988 ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8989 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
8992 ; SSE-NEXT: punpckldq {{.*#+}} xmm8 = xmm8[0],xmm11[0],xmm8[1],xmm11[1]
8996 ; SSE-NEXT: pshuflw {{.*#+}} xmm11 = xmm11[0,2,2,3,4,5,6,7]
8997 ; SSE-NEXT: punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm12[0],xmm8[1],xmm12[1],xmm8[2],xmm12[2],xmm8[3],xmm12[3],xmm8[4],xmm12[4],xmm8[5],xmm12[5],xmm8[6],xmm12[6],xmm8[7],xmm12[7]
9000 ; SSE-NEXT: punpcklwd {{.*#+}} xmm8 = xmm8[0],xmm11[0],xmm8[1],xmm11[1],xmm8[2],xmm11[2],xmm8[3],xmm11[3]
9002 ; SSE-NEXT: movss {{.*#+}} xmm0 = xmm8[0],xmm0[1,2,3]
9003 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
9006 ; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
9009 ; SSE-NEXT: punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm12[0],xmm8[1],xmm12[1],xmm8[2],xmm12[2],xmm8[3],xmm12[3],xmm8[4],xmm12[4],xmm8[5],xmm12[5],xmm8[6],xmm12[6],xmm8[7],xmm12[7]
9014 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
9015 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
9016 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm3[0,1,0,3]
9017 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,4,7,6]
9021 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,1,0,3]
9022 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,0,3,2,4,5,6,7]
9023 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5]
9032 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
9035 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
9036 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,4,6,7]
9038 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
9039 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,7,6,7]
9041 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
9043 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
9046 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
9048 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,1,3]
9049 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,7]
9050 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
9051 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
9055 ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
9056 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9057 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
9058 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,3]
9059 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,4,7]
9063 ; SSE-NEXT: pshufd $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
9065 ; SSE-NEXT: pshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload
9066 ; SSE-NEXT: # xmm14 = mem[0,2,2,3]
9067 ; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm0[0],xmm14[1],xmm0[1]
9069 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
9070 ; SSE-NEXT: movdqa {{.*#+}} xmm10 = [65535,65535,65535,0,65535,0,65535,65535]
9077 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,1,0,3,4,5,6,7]
9085 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
9088 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
9089 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,4,6,7]
9091 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm4[0,1,2,1]
9092 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,5,7,6,7]
9094 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
9096 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
9100 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,2,1,3]
9101 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,7,6,7]
9102 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,1,2,0]
9103 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,7,6,5]
9107 ; SSE-NEXT: movss {{.*#+}} xmm8 = xmm1[0],xmm8[1,2,3]
9108 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
9109 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
9110 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,0,3]
9111 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,7]
9115 ; SSE-NEXT: pshufd $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
9117 ; SSE-NEXT: pshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload
9118 ; SSE-NEXT: # xmm14 = mem[0,2,2,3]
9119 ; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm1[0],xmm14[1],xmm1[1]
9121 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
9128 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,1,0,3,4,5,6,7]
9137 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9141 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm13[0],xmm1[1],xmm13[1],xmm1[2],xmm13[2],xmm1[3],xmm13[3],xmm1[4],xmm13[4],xmm1[5],xmm13[5],xmm1[6],xmm13[6],xmm1[7],xmm13[7]
9142 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,4,6,7]
9144 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm0[0,1,2,1]
9145 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,5,7,6,7]
9147 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
9150 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
9152 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
9154 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,2,1,3]
9155 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,7,6,7]
9156 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,1,2,0]
9157 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,7,6,5]
9161 ; SSE-NEXT: movss {{.*#+}} xmm8 = xmm1[0],xmm8[1,2,3]
9162 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
9163 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
9164 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,0,3]
9165 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,7]
9169 ; SSE-NEXT: pshufd $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
9171 ; SSE-NEXT: pshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload
9172 ; SSE-NEXT: # xmm14 = mem[0,2,2,3]
9173 ; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm1[0],xmm14[1],xmm1[1]
9175 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
9182 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,1,0,3,4,5,6,7]
9191 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
9194 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm14[0],xmm1[1],xmm14[1],xmm1[2],xmm14[2],xmm1[3],xmm14[3],xmm1[4],xmm14[4],xmm1[5],xmm14[5],xmm1[6],xmm14[6],xmm1[7],xmm14[7]
9195 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,4,6,7]
9197 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm2[0,1,2,1]
9198 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,5,7,6,7]
9200 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
9203 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
9205 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
9207 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm4[0,2,1,3]
9208 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,7,6,7]
9209 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,1,2,0]
9210 ; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm8[0,1,2,3,4,7,6,5]
9214 ; SSE-NEXT: movss {{.*#+}} xmm14 = xmm1[0],xmm14[1,2,3]
9215 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
9216 ; SSE-NEXT: por {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
9217 ; SSE-NEXT: pshufd $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
9219 ; SSE-NEXT: pshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload
9220 ; SSE-NEXT: # xmm8 = mem[0,2,2,3]
9221 ; SSE-NEXT: punpckldq {{.*#+}} xmm8 = xmm8[0],xmm1[0],xmm8[1],xmm1[1]
9223 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1],xmm1[2],xmm15[2],xmm1[3],xmm15[3],xmm1[4],xmm15[4],xmm1[5],xmm15[5],xmm1[6],xmm15[6],xmm1[7],xmm15[7]
9229 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,1,0,3,4,5,6,7]
9232 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm2[0,1,0,3]
9233 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,5,4,7]
9241 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9243 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9245 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9247 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9249 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9251 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9253 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9255 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9257 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9259 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9261 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9263 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9265 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9267 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9269 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9271 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9273 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9275 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9277 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9279 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9281 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
9285 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9287 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
9292 ; SSE-NEXT: addq $1528, %rsp # imm = 0x5F8
9296 ; AVX: # %bb.0:
9297 ; AVX-NEXT: subq $744, %rsp # imm = 0x2E8
9298 ; AVX-NEXT: vmovddup {{.*#+}} xmm0 = [128,128,6,13,0,0,0,128,128,128,6,13,0,0,0,128]
9299 ; AVX-NEXT: # xmm0 = mem[0,0]
9301 ; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9305 ; AVX-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9306 ; AVX-NEXT: vmovq {{.*#+}} xmm2 = [128,128,128,5,12,0,0,0,0,0,0,0,0,0,0,0]
9308 ; AVX-NEXT: vmovq {{.*#+}} xmm3 = [0,7,14,128,128,0,0,0,0,0,0,0,0,0,0,0]
9310 ; AVX-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9313 ; AVX-NEXT: vmovddup {{.*#+}} xmm4 = [128,128,0,0,0,3,10,128,128,128,0,0,0,3,10,128]
9314 ; AVX-NEXT: # xmm4 = mem[0,0]
9316 ; AVX-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9318 ; AVX-NEXT: vmovddup {{.*#+}} xmm5 = [8,15,0,0,0,128,128,1,8,15,0,0,0,128,128,1]
9319 ; AVX-NEXT: # xmm5 = mem[0,0]
9322 ; AVX-NEXT: vmovq {{.*#+}} xmm6 = [255,255,255,255,255,0,0,0,0,0,0,0,0,0,0,0]
9324 ; AVX-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9326 ; AVX-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9329 ; AVX-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9333 ; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9336 ; AVX-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9339 ; AVX-NEXT: vmovddup {{.*#+}} xmm4 = [8,15,128,128,0,0,0,1,8,15,128,128,0,0,0,1]
9340 ; AVX-NEXT: # xmm4 = mem[0,0]
9342 ; AVX-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9353 ; AVX-NEXT: vmovdqa {{.*#+}} xmm15 = [u,u,u,0,0,0,0,255,255,255,255,255,u,u,u,u]
9355 ; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9367 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9368 ; AVX-NEXT: vmovddup {{.*#+}} xmm0 = [9,128,128,128,0,0,0,2,9,128,128,128,0,0,0,2]
9369 ; AVX-NEXT: # xmm0 = mem[0,0]
9372 ; AVX-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9373 ; AVX-NEXT: vmovddup {{.*#+}} xmm2 = [128,0,7,14,0,0,0,128,128,0,7,14,0,0,0,128]
9374 ; AVX-NEXT: # xmm2 = mem[0,0]
9375 ; AVX-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9380 ; AVX-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9383 ; AVX-NEXT: vmovdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9387 ; AVX-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9390 ; AVX-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9393 ; AVX-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9395 ; AVX-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9399 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9400 ; AVX-NEXT: vmovq {{.*#+}} xmm0 = [0,0,128,128,128,5,12,0,0,0,0,0,0,0,0,0]
9402 ; AVX-NEXT: vmovq {{.*#+}} xmm12 = [0,0,0,7,14,128,128,0,0,0,0,0,0,0,0,0]
9405 ; AVX-NEXT: vmovddup {{.*#+}} xmm2 = [10,128,128,128,0,0,0,3,10,128,128,128,0,0,0,3]
9406 ; AVX-NEXT: # xmm2 = mem[0,0]
9408 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = [128,1,8,15,0,0,0,128,128,1,8,15,0,0,0,128]
9409 ; AVX-NEXT: # xmm3 = mem[0,0]
9412 ; AVX-NEXT: vmovq {{.*#+}} xmm4 = [0,0,255,255,255,255,255,0,0,0,0,0,0,0,0,0]
9414 ; AVX-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9422 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9423 ; AVX-NEXT: vmovq {{.*#+}} xmm14 = [128,128,128,6,13,0,0,0,0,0,0,0,0,0,0,0]
9424 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
9426 ; AVX-NEXT: vmovq {{.*#+}} xmm2 = [1,8,15,128,128,0,0,0,0,0,0,0,0,0,0,0]
9427 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
9431 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
9434 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
9437 ; AVX-NEXT: vmovq {{.*#+}} xmm12 = [255,255,255,255,255,0,0,0,0,0,0,0,0,0,0,0]
9439 ; AVX-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9440 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
9442 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
9445 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
9447 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9451 ; AVX-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9452 ; AVX-NEXT: vmovq {{.*#+}} xmm9 = [2,9,128,128,128,0,0,0,0,0,0,0,0,0,0,0]
9454 ; AVX-NEXT: vmovq {{.*#+}} xmm3 = [128,128,0,7,14,0,0,0,0,0,0,0,0,0,0,0]
9464 ; AVX-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9475 ; AVX-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9476 ; AVX-NEXT: vmovq {{.*#+}} xmm1 = [3,10,128,128,128,0,0,0,0,0,0,0,0,0,0,0]
9478 ; AVX-NEXT: vmovq {{.*#+}} xmm2 = [128,128,1,8,15,0,0,0,0,0,0,0,0,0,0,0]
9487 ; AVX-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9495 ; AVX-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9496 ; AVX-NEXT: vmovq {{.*#+}} xmm0 = [0,0,128,128,128,6,13,0,0,0,0,0,0,0,0,0]
9497 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
9499 ; AVX-NEXT: vmovq {{.*#+}} xmm3 = [0,0,1,8,15,128,128,0,0,0,0,0,0,0,0,0]
9500 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
9504 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
9507 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
9510 ; AVX-NEXT: vmovq {{.*#+}} xmm2 = [0,0,255,255,255,255,255,0,0,0,0,0,0,0,0,0]
9512 ; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9513 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
9515 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
9518 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
9524 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9525 ; AVX-NEXT: vmovq {{.*#+}} xmm1 = [0,0,2,9,128,128,128,0,0,0,0,0,0,0,0,0]
9527 ; AVX-NEXT: vmovq {{.*#+}} xmm3 = [0,0,128,128,0,7,14,0,0,0,0,0,0,0,0,0]
9538 ; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9548 ; AVX-NEXT: vmovq {{.*#+}} xmm1 = [0,0,3,10,128,128,128,0,0,0,0,0,0,0,0,0]
9550 ; AVX-NEXT: vmovq {{.*#+}} xmm3 = [0,0,128,128,1,8,15,0,0,0,0,0,0,0,0,0]
9551 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
9571 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = [0,0,0,128,128,128,5,12,0,0,0,128,128,128,5,12]
9572 ; AVX-NEXT: # xmm3 = mem[0,0]
9574 ; AVX-NEXT: vmovddup {{.*#+}} xmm11 = [0,0,0,0,7,14,128,128,0,0,0,0,7,14,128,128]
9575 ; AVX-NEXT: # xmm11 = mem[0,0]
9579 ; AVX-NEXT: vpblendvb %xmm13, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm0 # 16-byte Folded Reload
9580 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9586 ; AVX-NEXT: vpblendvb %xmm13, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
9587 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9588 ; AVX-NEXT: vmovddup {{.*#+}} xmm1 = [0,0,0,128,128,128,6,13,0,0,0,128,128,128,6,13]
9589 ; AVX-NEXT: # xmm1 = mem[0,0]
9591 ; AVX-NEXT: vmovddup {{.*#+}} xmm0 = [0,0,0,1,8,15,128,128,0,0,0,1,8,15,128,128]
9592 ; AVX-NEXT: # xmm0 = mem[0,0]
9595 ; AVX-NEXT: vpblendvb %xmm13, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm10 # 16-byte Folded Reload
9596 ; AVX-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9601 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9602 ; AVX-NEXT: vmovddup {{.*#+}} xmm0 = [0,0,0,2,9,128,128,128,0,0,0,2,9,128,128,128]
9603 ; AVX-NEXT: # xmm0 = mem[0,0]
9605 ; AVX-NEXT: vmovddup {{.*#+}} xmm11 = [0,0,0,128,128,0,7,14,0,0,0,128,128,0,7,14]
9606 ; AVX-NEXT: # xmm11 = mem[0,0]
9610 ; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9615 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9616 ; AVX-NEXT: vmovd {{.*#+}} xmm14 = [2,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
9618 ; AVX-NEXT: vmovd {{.*#+}} xmm0 = [0,0,4,11,0,0,0,0,0,0,0,0,0,0,0,0]
9619 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
9621 ; AVX-NEXT: vpunpckldq {{.*#+}} xmm7 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
9622 ; AVX-NEXT: vmovddup {{.*#+}} xmm1 = [128,5,12,0,0,0,128,128,128,5,12,0,0,0,128,128]
9623 ; AVX-NEXT: # xmm1 = mem[0,0]
9624 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
9626 ; AVX-NEXT: vmovddup {{.*#+}} xmm2 = [14,128,128,0,0,0,0,7,14,128,128,0,0,0,0,7]
9627 ; AVX-NEXT: # xmm2 = mem[0,0]
9630 ; AVX-NEXT: vpblendw {{.*#+}} xmm8 = xmm8[0],xmm7[1,2],xmm8[3,4,5,6,7]
9631 ; AVX-NEXT: vmovddup {{.*#+}} xmm7 = [0,0,0,3,10,128,128,128,0,0,0,3,10,128,128,128]
9632 ; AVX-NEXT: # xmm7 = mem[0,0]
9634 ; AVX-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9635 ; AVX-NEXT: vmovddup {{.*#+}} xmm11 = [0,0,0,128,128,1,8,15,0,0,0,128,128,1,8,15]
9636 ; AVX-NEXT: # xmm11 = mem[0,0]
9640 ; AVX-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9641 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
9643 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
9645 ; AVX-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1]
9646 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
9650 ; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2],xmm1[3,4,5,6,7]
9653 ; AVX-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9658 ; AVX-NEXT: vbroadcastss {{.*#+}} xmm14 = [0,0,2,9,0,0,2,9,0,0,2,9,0,0,2,9]
9660 ; AVX-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9662 ; AVX-NEXT: vbroadcastss {{.*#+}} xmm0 = [0,0,4,11,0,0,4,11,0,0,4,11,0,0,4,11]
9665 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
9666 ; AVX-NEXT: vpblendw $63, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
9667 ; AVX-NEXT: # xmm1 = mem[0,1,2,3,4,5],xmm1[6,7]
9668 ; AVX-NEXT: vbroadcastss {{.*#+}} xmm15 = [4,11,0,0,4,11,0,0,4,11,0,0,4,11,0,0]
9671 ; AVX-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9672 ; AVX-NEXT: vbroadcastss {{.*#+}} xmm6 = [0,0,6,13,0,0,6,13,0,0,6,13,0,0,6,13]
9674 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9676 ; AVX-NEXT: vpunpckldq {{.*#+}} xmm7 = xmm10[0],xmm7[0],xmm10[1],xmm7[1]
9678 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9679 ; AVX-NEXT: vpalignr {{.*#+}} xmm7 = xmm7[10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7,8,9]
9680 ; AVX-NEXT: vmovddup {{.*#+}} xmm11 = [0,0,0,1,2,3,8,15,0,0,0,1,2,3,8,15]
9681 ; AVX-NEXT: # xmm11 = mem[0,0]
9683 ; AVX-NEXT: vmovd {{.*#+}} xmm0 = [0,7,14,0,0,0,0,0,0,0,0,0,0,0,0,0]
9685 ; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9688 ; AVX-NEXT: vmovaps {{.*#+}} ymm2 = [65535,65535,65535,65535,65535,0,0,0,0,0,65535,65535,65535,65535,65535,65535]
9689 ; AVX-NEXT: vandps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm12 # 32-byte Folded Reload
9693 ; AVX-NEXT: vmovaps {{.*#+}} ymm12 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0,0,0,0,0]
9697 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9701 ; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9702 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm10[0],xmm1[0],xmm10[1],xmm1[1],xmm10[2],xmm1[2],xmm10[3],xmm1[3]
9703 ; AVX-NEXT: vpblendw $63, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
9704 ; AVX-NEXT: # xmm1 = mem[0,1,2,3,4,5],xmm1[6,7]
9706 ; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9709 ; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9711 ; AVX-NEXT: vpunpckldq {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1]
9713 ; AVX-NEXT: vpalignr {{.*#+}} xmm14 = xmm14[10,11,12,13,14,15],xmm9[0,1,2,3,4,5,6,7,8,9]
9714 ; AVX-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9718 ; AVX-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
9720 ; AVX-NEXT: vandps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm11 # 32-byte Folded Reload
9727 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9729 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
9731 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm11[0],xmm1[0],xmm11[1],xmm1[1],xmm11[2],xmm1[2],xmm11[3],xmm1[3]
9732 ; AVX-NEXT: vpblendw $63, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
9733 ; AVX-NEXT: # xmm1 = mem[0,1,2,3,4,5],xmm1[6,7]
9735 ; AVX-NEXT: vbroadcastss {{.*#+}} xmm7 = [0,0,7,14,0,0,7,14,0,0,7,14,0,0,7,14]
9736 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
9738 ; AVX-NEXT: vpunpckldq {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1]
9740 ; AVX-NEXT: vpblendw {{.*#+}} xmm10 = xmm10[0,1,2,3,4,5,6],xmm0[7]
9741 ; AVX-NEXT: vmovddup {{.*#+}} xmm2 = [0,128,128,128,128,128,3,10,0,128,128,128,128,128,3,10]
9742 ; AVX-NEXT: # xmm2 = mem[0,0]
9743 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
9746 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
9749 ; AVX-NEXT: vmovaps {{.*#+}} ymm14 = [0,0,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
9750 ; AVX-NEXT: vandnps {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm0 # 32-byte Folded Reload
9757 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9758 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
9761 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
9762 ; AVX-NEXT: vpblendw $63, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
9763 ; AVX-NEXT: # xmm0 = mem[0,1,2,3,4,5],xmm0[6,7]
9764 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
9766 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
9768 ; AVX-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm7[0],xmm1[0],xmm7[1],xmm1[1]
9770 ; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,6],xmm7[7]
9775 ; AVX-NEXT: vandnps {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm2 # 32-byte Folded Reload
9782 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9783 ; AVX-NEXT: vbroadcastss {{.*#+}} xmm2 = [0,0,4,11,0,0,4,11,0,0,4,11,0,0,4,11]
9784 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9786 ; AVX-NEXT: vbroadcastss {{.*#+}} xmm1 = [0,0,6,13,0,0,6,13,0,0,6,13,0,0,6,13]
9788 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
9789 ; AVX-NEXT: vpblendw $63, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
9790 ; AVX-NEXT: # xmm0 = mem[0,1,2,3,4,5],xmm0[6,7]
9791 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
9793 ; AVX-NEXT: vbroadcastss {{.*#+}} xmm14 = [0,1,8,15,0,1,8,15,0,1,8,15,0,1,8,15]
9796 ; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,6],xmm7[7]
9797 ; AVX-NEXT: vmovddup {{.*#+}} xmm6 = [0,128,128,128,128,128,4,11,0,128,128,128,128,128,4,11]
9798 ; AVX-NEXT: # xmm6 = mem[0,0]
9802 ; AVX-NEXT: vmovd {{.*#+}} xmm9 = [2,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
9805 ; AVX-NEXT: vmovaps {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
9806 ; AVX-NEXT: vandps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm8 # 32-byte Folded Reload
9810 ; AVX-NEXT: vmovaps {{.*#+}} ymm15 = [65535,65535,65535,65535,65535,65535,65535,65535,65535,0,0,0,0,0,0,0]
9814 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9816 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9818 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
9819 ; AVX-NEXT: vpblendw $63, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
9820 ; AVX-NEXT: # xmm3 = mem[0,1,2,3,4,5],xmm3[6,7]
9827 ; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,6],xmm7[7]
9828 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
9831 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
9834 ; AVX-NEXT: vmovaps {{.*#+}} ymm5 = [255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
9835 ; AVX-NEXT: vandps {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm4 # 32-byte Folded Reload
9842 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9843 ; AVX-NEXT: vmovddup {{.*#+}} xmm3 = [0,2,9,128,128,128,0,0,0,2,9,128,128,128,0,0]
9844 ; AVX-NEXT: # xmm3 = mem[0,0]
9845 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
9847 ; AVX-NEXT: vmovddup {{.*#+}} xmm4 = [0,128,128,0,7,14,0,0,0,128,128,0,7,14,0,0]
9848 ; AVX-NEXT: # xmm4 = mem[0,0]
9851 ; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,6],xmm7[7]
9852 ; AVX-NEXT: vmovddup {{.*#+}} xmm5 = [0,128,128,128,128,128,5,12,0,128,128,128,128,128,5,12]
9853 ; AVX-NEXT: # xmm5 = mem[0,0]
9857 ; AVX-NEXT: vmovd {{.*#+}} xmm8 = [3,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
9858 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
9861 ; AVX-NEXT: vmovaps {{.*#+}} ymm12 = [255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
9862 ; AVX-NEXT: vandps {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm7 # 32-byte Folded Reload
9865 ; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm7 # 16-byte Folded Reload
9869 ; AVX-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9873 ; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,6],mem[7]
9878 ; AVX-NEXT: vandps {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm4 # 32-byte Folded Reload
9881 ; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 16-byte Folded Reload
9885 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9886 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
9887 ; AVX-NEXT: vmovd {{.*#+}} xmm14 = [2,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
9889 ; AVX-NEXT: vmovd {{.*#+}} xmm0 = [4,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
9890 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
9892 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
9893 ; AVX-NEXT: vmovddup {{.*#+}} xmm4 = [12,0,0,0,128,128,128,5,12,0,0,0,128,128,128,5]
9894 ; AVX-NEXT: # xmm4 = mem[0,0]
9895 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
9897 ; AVX-NEXT: vmovddup {{.*#+}} xmm6 = [128,0,0,0,0,7,14,128,128,0,0,0,0,7,14,128]
9898 ; AVX-NEXT: # xmm6 = mem[0,0]
9899 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
9902 ; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm5[2,3,4,5,6,7]
9904 ; AVX-NEXT: vmovddup {{.*#+}} xmm5 = [0,128,128,1,8,15,0,0,0,128,128,1,8,15,0,0]
9905 ; AVX-NEXT: # xmm5 = mem[0,0]
9908 ; AVX-NEXT: vpblendw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,5,6],mem[7]
9909 ; AVX-NEXT: vmovddup {{.*#+}} xmm9 = [0,128,128,128,128,128,6,13,0,128,128,128,128,128,6,13]
9910 ; AVX-NEXT: # xmm9 = mem[0,0]
9916 ; AVX-NEXT: vmovaps {{.*#+}} ymm0 = [255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
9920 ; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm7 # 16-byte Folded Reload
9924 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9925 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
9927 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
9930 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3]
9931 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
9933 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
9936 ; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3,4,5,6,7]
9937 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9939 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
9942 ; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,5,6],mem[7]
9943 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
9946 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
9949 ; AVX-NEXT: vmovaps {{.*#+}} ymm2 = [255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
9953 ; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 16-byte Folded Reload
9957 ; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9958 ; AVX-NEXT: vmovd {{.*#+}} xmm2 = [3,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
9960 ; AVX-NEXT: vmovd {{.*#+}} xmm8 = [5,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
9962 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
9963 ; AVX-NEXT: vmovddup {{.*#+}} xmm4 = [13,0,0,0,128,128,128,6,13,0,0,0,128,128,128,6]
9964 ; AVX-NEXT: # xmm4 = mem[0,0]
9966 ; AVX-NEXT: vmovddup {{.*#+}} xmm6 = [128,0,0,0,1,8,15,128,128,0,0,0,1,8,15,128]
9967 ; AVX-NEXT: # xmm6 = mem[0,0]
9970 ; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm3[0,1],xmm5[2,3,4,5,6,7]
9971 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
9973 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
9977 ; AVX-NEXT: vmovddup {{.*#+}} xmm7 = [0,128,128,128,128,0,7,14,0,128,128,128,128,0,7,14]
9978 ; AVX-NEXT: # xmm7 = mem[0,0]
9979 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
9982 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
9985 ; AVX-NEXT: vmovaps {{.*#+}} ymm13 = [255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
9989 ; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm9 # 16-byte Folded Reload
9993 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
9997 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm9[0],xmm2[0],xmm9[1],xmm2[1],xmm9[2],xmm2[2],xmm9[3],xmm2[3]
9999 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
10002 ; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm4[2,3,4,5,6,7]
10004 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
10007 ; AVX-NEXT: vmovddup {{.*#+}} xmm11 = [0,9,10,11,12,128,128,128,0,9,10,11,12,128,128,128]
10008 ; AVX-NEXT: # xmm11 = mem[0,0]
10010 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
10013 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
10019 ; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 16-byte Folded Reload
10023 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10024 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10026 ; AVX-NEXT: vmovd {{.*#+}} xmm2 = [6,13,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
10027 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10029 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
10030 ; AVX-NEXT: vmovddup {{.*#+}} xmm6 = [128,0,0,0,2,9,128,128,128,0,0,0,2,9,128,128]
10031 ; AVX-NEXT: # xmm6 = mem[0,0]
10032 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10034 ; AVX-NEXT: vmovddup {{.*#+}} xmm0 = [14,0,0,0,128,128,0,7,14,0,0,0,128,128,0,7]
10035 ; AVX-NEXT: # xmm0 = mem[0,0]
10036 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
10039 ; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0,1],xmm7[2,3,4,5,6,7]
10041 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
10046 ; AVX-NEXT: vmovddup {{.*#+}} xmm13 = [0,128,128,128,128,1,8,15,0,128,128,128,128,1,8,15]
10047 ; AVX-NEXT: # xmm13 = mem[0,0]
10048 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
10053 ; AVX-NEXT: vmovaps {{.*#+}} ymm11 = [255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
10057 ; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm9 # 16-byte Folded Reload
10061 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
10063 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
10065 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm12[0],xmm9[0],xmm12[1],xmm9[1],xmm12[2],xmm9[2],xmm12[3],xmm9[3]
10067 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
10070 ; AVX-NEXT: vpblendw {{.*#+}} xmm6 = xmm9[0,1],xmm6[2,3,4,5,6,7]
10071 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
10087 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10089 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10091 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10093 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10095 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10097 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10099 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10101 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10103 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10105 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10107 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
10108 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10110 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10112 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
10115 ; AVX-NEXT: addq $744, %rsp # imm = 0x2E8
10120 ; AVX2: # %bb.0:
10121 ; AVX2-NEXT: subq $760, %rsp # imm = 0x2F8
10129 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm13 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
10131 ; AVX2-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10132 ; AVX2-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10136 ; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = [0,7,14,128,128,3,10,128,128,128,u,u,u,u,u,u]
10139 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm14 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
10142 ; AVX2-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10143 ; AVX2-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10145 ; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
10146 ; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [0,7,14,0,0,0,0,0,0,0,6,13,4,11,2,9,0,7,14,0,0,0,0,0,0,0,6,13,4,11,2,9]
10147 ; AVX2-NEXT: # ymm3 = mem[0,1,0,1]
10149 ; AVX2-NEXT: vpmovsxbw {{.*#+}} xmm4 = [65535,65535,65535,65535,65535,0,0,0]
10151 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10153 ; AVX2-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10155 ; AVX2-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10162 ; AVX2-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10163 ; AVX2-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10165 ; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm5[2],ymm2[3,4],ymm5[5],ymm2[6,7,8,9],ymm5[10],ymm2[11,12],ymm5[13],ymm2[14,15]
10168 ; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10171 ; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = ymm2[0,1],ymm3[2],ymm2[3,4,5],ymm3[6],ymm2[7,8,9],ymm3[10],ymm2[11,12,13],ymm3[14],ymm2[15]
10172 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm1 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
10180 ; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [1,8,15,0,0,0,0,0,0,0,7,14,5,12,3,10,1,8,15,0,0,0,0,0,0,0,7,14,5,12,3,10]
10181 ; AVX2-NEXT: # ymm7 = mem[0,1,0,1]
10183 ; AVX2-NEXT: vpmovsxdq {{.*#+}} ymm9 = [0,18446744073709551360,16777215,0]
10185 ; AVX2-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10188 ; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4,5],ymm5[6],ymm4[7,8,9],ymm5[10],ymm4[11,12,13],ymm5[14],ymm4[15]
10196 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10199 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm4 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
10201 ; AVX2-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10214 ; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10215 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm10[0],xmm3[0],xmm10[1],xmm3[1],xmm10[2],xmm3[2],xmm10[3],xmm3[3]
10218 ; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6],ymm3[7]
10219 ; AVX2-NEXT: vpmovsxdq {{.*#+}} ymm14 = [18446744073709551615,18446744073709551615,16777215,0]
10220 ; AVX2-NEXT: vpblendvb %ymm14, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm0 # 32-byte Folded Reload
10221 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10234 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
10237 ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm6[7]
10238 ; AVX2-NEXT: vpblendvb %ymm14, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
10239 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10240 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm7 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
10245 ; AVX2-NEXT: vmovdqa {{.*#+}} xmm8 = [u,u,u,128,128,4,11,128,128,0,7,14,u,u,u,u]
10252 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
10255 ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm9[7]
10256 ; AVX2-NEXT: vpblendvb %ymm14, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
10257 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10267 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3]
10270 ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm1[7]
10271 ; AVX2-NEXT: vpblendvb %ymm14, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
10272 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10273 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
10275 ; AVX2-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10282 ; AVX2-NEXT: vpbroadcastq {{.*#+}} xmm1 = [0,0,0,128,128,128,5,12,0,0,0,128,128,128,5,12]
10283 ; AVX2-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10285 ; AVX2-NEXT: vpbroadcastq {{.*#+}} xmm9 = [0,0,0,0,7,14,128,128,0,0,0,0,7,14,128,128]
10288 ; AVX2-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10292 ; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [0,0,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,0,0,0,0,0]
10293 ; AVX2-NEXT: # ymm0 = mem[0,1,0,1]
10295 ; AVX2-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10296 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
10303 ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
10309 ; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10310 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
10315 ; AVX2-NEXT: vmovdqa {{.*#+}} xmm8 = [u,u,128,128,0,7,14,128,128,3,10,u,u,u,u,u]
10318 ; AVX2-NEXT: vpbroadcastq {{.*#+}} xmm7 = [0,0,0,128,128,128,6,13,0,0,0,128,128,128,6,13]
10320 ; AVX2-NEXT: vpbroadcastq {{.*#+}} xmm10 = [0,0,0,1,8,15,128,128,0,0,0,1,8,15,128,128]
10326 ; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10327 ; AVX2-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10341 ; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10342 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm6 = [0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535]
10343 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
10351 ; AVX2-NEXT: vpbroadcastq {{.*#+}} xmm7 = [0,0,0,2,9,128,128,128,0,0,0,2,9,128,128,128]
10352 ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
10354 ; AVX2-NEXT: vpbroadcastq {{.*#+}} xmm10 = [0,0,0,128,128,0,7,14,0,0,0,128,128,0,7,14]
10355 ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
10361 ; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10367 ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
10375 ; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10376 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm10 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
10378 ; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10379 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm15 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
10384 ; AVX2-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,4,11,128,128,0,7,14,128,128,u,u,u,u,u]
10387 ; AVX2-NEXT: vpbroadcastq {{.*#+}} xmm2 = [0,0,0,3,10,128,128,128,0,0,0,3,10,128,128,128]
10389 ; AVX2-NEXT: vpbroadcastq {{.*#+}} xmm8 = [0,0,0,128,128,1,8,15,0,0,0,128,128,1,8,15]
10395 ; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10396 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10397 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
10399 ; AVX2-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10411 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10412 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10413 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
10414 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
10416 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm5 = [0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535]
10419 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm7 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
10421 ; AVX2-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10422 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm3 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
10424 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10425 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10426 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
10428 ; AVX2-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10432 ; AVX2-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10434 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10435 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
10436 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
10437 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm10 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
10442 ; AVX2-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10444 ; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10445 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
10446 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
10451 ; AVX2-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10452 ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm5 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
10457 ; AVX2-NEXT: vmovdqa {{.*#+}} xmm7 = [128,128,0,7,14,128,128,3,10,u,u,u,u,u,u,u]
10461 ; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm15[3],ymm0[4,5],ymm15[6],ymm0[7,8,9,10],ymm15[11],ymm0[12,13],ymm15[14],ymm0[15]
10462 ; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [2,9,0,0,0,0,0,0,0,1,8,15,6,13,4,11,2,9,0,0,0,0,0,0,0,1,8,15,6,13,4,11]
10463 ; AVX2-NEXT: # ymm5 = mem[0,1,0,1]
10467 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10473 ; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
10476 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10484 ; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm4[1,2],ymm2[3],ymm4[4,5,6],ymm2[7,8],ymm4[9,10],ymm2[11],ymm4[12,13,14],ymm2[15]
10485 ; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [3,10,0,0,0,0,0,0,0,2,9,0,7,14,5,12,3,10,0,0,0,0,0,0,0,2,9,0,7,14,5,12]
10486 ; AVX2-NEXT: # ymm4 = mem[0,1,0,1]
10494 ; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm6[1,2],ymm1[3],ymm6[4,5,6],ymm1[7,8],ymm6[9,10],ymm1[11],ymm6[12,13,14],ymm1[15]
10500 ; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = [4,11,128,128,0,7,14,128,128,u,u,u,u,u,u,u]
10504 ; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm9[1,2,3],ymm3[4],ymm9[5,6],ymm3[7,8],ymm9[9,10,11],ymm3[12],ymm9[13,14],ymm3[15]
10505 ; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [4,11,0,0,0,0,0,0,0,3,10,1,8,15,6,13,4,11,0,0,0,0,0,0,0,3,10,1,8,15,6,13]
10506 ; AVX2-NEXT: # ymm4 = mem[0,1,0,1]
10509 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
10515 ; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm10[1,2,3],ymm1[4],ymm10[5,6],ymm1[7,8],ymm10[9,10,11],ymm1[12],ymm10[13,14],ymm1[15]
10518 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
10525 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
10527 ; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0],ymm2[1],ymm5[2,3],ymm2[4],ymm5[5,6,7,8],ymm2[9],ymm5[10,11],ymm2[12],ymm5[13,14,15]
10528 ; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [5,12,0,0,0,0,0,0,0,4,11,2,9,0,7,14,5,12,0,0,0,0,0,0,0,4,11,2,9,0,7,14]
10529 ; AVX2-NEXT: # ymm5 = mem[0,1,0,1]
10532 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
10537 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
10539 ; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6,7,8],ymm3[9],ymm4[10,11],ymm3[12],ymm4[13,14,15]
10543 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
10546 ; AVX2-NEXT: vmovdqa {{.*#+}} xmm7 = [128,128,4,11,128,128,0,7,14,u,u,u,u,u,u,u]
10549 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
10551 ; AVX2-NEXT: vpblendw {{.*#+}} ymm5 = ymm8[0],ymm5[1],ymm8[2,3,4],ymm5[5],ymm8[6,7,8],ymm5[9],ymm8[10,11,12],ymm5[13],ymm8[14,15]
10552 ; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [6,13,0,0,0,0,0,0,0,5,12,3,10,1,8,15,6,13,0,0,0,0,0,0,0,5,12,3,10,1,8,15]
10553 ; AVX2-NEXT: # ymm8 = mem[0,1,0,1]
10556 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
10562 ; AVX2-NEXT: vpblendw {{.*#+}} ymm5 = ymm11[0],ymm5[1],ymm11[2,3,4],ymm5[5],ymm11[6,7,8],ymm5[9],ymm11[10,11,12],ymm5[13],ymm11[14,15]
10565 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
10569 ; AVX2-NEXT: vmovdqa {{.*#+}} xmm8 = [u,u,0,7,14,128,128,3,10,128,128,128,u,u,u,u]
10573 ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
10576 ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
10578 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
10581 ; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5,6],ymm9[7]
10582 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
10583 ; AVX2-NEXT: vpblendw {{.*#+}} ymm5 = ymm9[0],ymm5[1,2,3,4,5,6,7],ymm9[8],ymm5[9,10,11,12,13,14,15]
10584 ; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm9[0,1,2,3],ymm5[4,5,6,7]
10585 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
10592 ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
10594 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
10597 ; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5,6],ymm8[7]
10598 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
10599 ; AVX2-NEXT: vpblendw {{.*#+}} ymm7 = ymm8[0],ymm7[1,2,3,4,5,6,7],ymm8[8],ymm7[9,10,11,12,13,14,15]
10600 ; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]
10601 ; AVX2-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm8 # 32-byte Folded Reload
10602 ; AVX2-NEXT: # ymm8 = ymm12[0],mem[1,2,3,4,5,6,7],ymm12[8],mem[9,10,11,12,13,14,15]
10603 ; AVX2-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
10604 ; AVX2-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm9 # 32-byte Folded Reload
10605 ; AVX2-NEXT: # ymm9 = ymm14[0],mem[1,2,3,4,5,6,7],ymm14[8],mem[9,10,11,12,13,14,15]
10606 ; AVX2-NEXT: vpblendd {{.*#+}} ymm9 = ymm14[0,1,2,3],ymm9[4,5,6,7]
10607 ; AVX2-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm10 # 32-byte Folded Reload
10608 ; AVX2-NEXT: # ymm10 = ymm6[0],mem[1,2,3,4,5,6,7],ymm6[8],mem[9,10,11,12,13,14,15]
10609 ; AVX2-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm10[4,5,6,7]
10610 ; AVX2-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm10 # 32-byte Folded Reload
10611 ; AVX2-NEXT: # ymm10 = ymm0[0],mem[1,2,3,4,5,6,7],ymm0[8],mem[9,10,11,12,13,14,15]
10612 ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm10[4,5,6,7]
10613 ; AVX2-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm10 # 32-byte Folded Reload
10614 ; AVX2-NEXT: # ymm10 = ymm2[0],mem[1,2,3,4,5,6,7],ymm2[8],mem[9,10,11,12,13,14,15]
10615 ; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm10[4,5,6,7]
10616 ; AVX2-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm10 # 32-byte Folded Reload
10617 ; AVX2-NEXT: # ymm10 = ymm1[0],mem[1,2,3,4,5,6,7],ymm1[8],mem[9,10,11,12,13,14,15]
10618 ; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm10[4,5,6,7]
10619 ; AVX2-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm10 # 32-byte Folded Reload
10620 ; AVX2-NEXT: # ymm10 = ymm3[0],mem[1,2,3,4,5,6,7],ymm3[8],mem[9,10,11,12,13,14,15]
10621 ; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm10[4,5,6,7]
10622 ; AVX2-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm10 # 32-byte Folded Reload
10623 ; AVX2-NEXT: # ymm10 = ymm4[0],mem[1,2,3,4,5,6,7],ymm4[8],mem[9,10,11,12,13,14,15]
10624 ; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]
10625 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
10627 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
10629 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
10631 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
10639 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
10642 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
10645 ; AVX2-NEXT: addq $760, %rsp # imm = 0x2F8
10650 ; AVX2-FP: # %bb.0:
10651 ; AVX2-FP-NEXT: subq $760, %rsp # imm = 0x2F8
10659 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm13 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
10661 ; AVX2-FP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10662 ; AVX2-FP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10666 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} xmm2 = [0,7,14,128,128,3,10,128,128,128,u,u,u,u,u,u]
10669 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm14 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
10672 ; AVX2-FP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10673 ; AVX2-FP-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10675 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
10676 ; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [0,7,14,0,0,0,0,0,0,0,6,13,4,11,2,9,0,7,14,0,0,0,0,0,0,0,6,13,4,11,2,9]
10677 ; AVX2-FP-NEXT: # ymm3 = mem[0,1,0,1]
10679 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} xmm4 = [65535,65535,65535,65535,65535,0,0,0]
10681 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10683 ; AVX2-FP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10685 ; AVX2-FP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10692 ; AVX2-FP-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10693 ; AVX2-FP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10695 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm5[2],ymm2[3,4],ymm5[5],ymm2[6,7,8,9],ymm5[10],ymm2[11,12],ymm5[13],ymm2[14,15]
10698 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10701 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = ymm2[0,1],ymm3[2],ymm2[3,4,5],ymm3[6],ymm2[7,8,9],ymm3[10],ymm2[11,12,13],ymm3[14],ymm2[15]
10702 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
10710 ; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [1,8,15,0,0,0,0,0,0,0,7,14,5,12,3,10,1,8,15,0,0,0,0,0,0,0,7,14,5,12,3,10]
10711 ; AVX2-FP-NEXT: # ymm7 = mem[0,1,0,1]
10713 ; AVX2-FP-NEXT: vpmovsxdq {{.*#+}} ymm9 = [0,18446744073709551360,16777215,0]
10715 ; AVX2-FP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10718 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4,5],ymm5[6],ymm4[7,8,9],ymm5[10],ymm4[11,12,13],ymm5[14],ymm4[15]
10726 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10729 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm4 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
10731 ; AVX2-FP-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10744 ; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10745 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm10[0],xmm3[0],xmm10[1],xmm3[1],xmm10[2],xmm3[2],xmm10[3],xmm3[3]
10748 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6],ymm3[7]
10749 ; AVX2-FP-NEXT: vpmovsxdq {{.*#+}} ymm14 = [18446744073709551615,18446744073709551615,16777215,0]
10750 ; AVX2-FP-NEXT: vpblendvb %ymm14, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm0 # 32-byte Folded Reload
10751 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10764 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
10767 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm6[7]
10768 ; AVX2-FP-NEXT: vpblendvb %ymm14, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
10769 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10770 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm7 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
10775 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} xmm8 = [u,u,u,128,128,4,11,128,128,0,7,14,u,u,u,u]
10782 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
10785 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm9[7]
10786 ; AVX2-FP-NEXT: vpblendvb %ymm14, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
10787 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10797 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3]
10800 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm1[7]
10801 ; AVX2-FP-NEXT: vpblendvb %ymm14, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
10802 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10803 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
10805 ; AVX2-FP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10812 ; AVX2-FP-NEXT: vpbroadcastq {{.*#+}} xmm1 = [0,0,0,128,128,128,5,12,0,0,0,128,128,128,5,12]
10813 ; AVX2-FP-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10815 ; AVX2-FP-NEXT: vpbroadcastq {{.*#+}} xmm9 = [0,0,0,0,7,14,128,128,0,0,0,0,7,14,128,128]
10818 ; AVX2-FP-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
10822 ; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [0,0,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,0,0,0,0,0]
10823 ; AVX2-FP-NEXT: # ymm0 = mem[0,1,0,1]
10825 ; AVX2-FP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10826 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
10833 ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
10839 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10840 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
10845 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} xmm8 = [u,u,128,128,0,7,14,128,128,3,10,u,u,u,u,u]
10848 ; AVX2-FP-NEXT: vpbroadcastq {{.*#+}} xmm7 = [0,0,0,128,128,128,6,13,0,0,0,128,128,128,6,13]
10850 ; AVX2-FP-NEXT: vpbroadcastq {{.*#+}} xmm10 = [0,0,0,1,8,15,128,128,0,0,0,1,8,15,128,128]
10856 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10857 ; AVX2-FP-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10871 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10872 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm6 = [0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535]
10873 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
10881 ; AVX2-FP-NEXT: vpbroadcastq {{.*#+}} xmm7 = [0,0,0,2,9,128,128,128,0,0,0,2,9,128,128,128]
10882 ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
10884 ; AVX2-FP-NEXT: vpbroadcastq {{.*#+}} xmm10 = [0,0,0,128,128,0,7,14,0,0,0,128,128,0,7,14]
10885 ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
10891 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10897 ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
10905 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10906 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm10 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
10908 ; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10909 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm15 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
10914 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,4,11,128,128,0,7,14,128,128,u,u,u,u,u]
10917 ; AVX2-FP-NEXT: vpbroadcastq {{.*#+}} xmm2 = [0,0,0,3,10,128,128,128,0,0,0,3,10,128,128,128]
10919 ; AVX2-FP-NEXT: vpbroadcastq {{.*#+}} xmm8 = [0,0,0,128,128,1,8,15,0,0,0,128,128,1,8,15]
10925 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10926 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
10927 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
10929 ; AVX2-FP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10941 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10942 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10943 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
10944 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
10946 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm5 = [0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535]
10949 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm7 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
10951 ; AVX2-FP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10952 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm3 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
10954 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10955 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
10956 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
10958 ; AVX2-FP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10962 ; AVX2-FP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10964 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10965 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
10966 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
10967 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm10 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
10972 ; AVX2-FP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10974 ; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10975 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
10976 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
10981 ; AVX2-FP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
10982 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm5 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
10987 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} xmm7 = [128,128,0,7,14,128,128,3,10,u,u,u,u,u,u,u]
10991 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm15[3],ymm0[4,5],ymm15[6],ymm0[7,8,9,10],ymm15[11],ymm0[12,13],ymm15[14],ymm0[15]
10992 ; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [2,9,0,0,0,0,0,0,0,1,8,15,6,13,4,11,2,9,0,0,0,0,0,0,0,1,8,15,6,13,4,11]
10993 ; AVX2-FP-NEXT: # ymm5 = mem[0,1,0,1]
10997 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11003 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
11006 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11014 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm4[1,2],ymm2[3],ymm4[4,5,6],ymm2[7,8],ymm4[9,10],ymm2[11],ymm4[12,13,14],ymm2[15]
11015 ; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [3,10,0,0,0,0,0,0,0,2,9,0,7,14,5,12,3,10,0,0,0,0,0,0,0,2,9,0,7,14,5,12]
11016 ; AVX2-FP-NEXT: # ymm4 = mem[0,1,0,1]
11024 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm6[1,2],ymm1[3],ymm6[4,5,6],ymm1[7,8],ymm6[9,10],ymm1[11],ymm6[12,13,14],ymm1[15]
11030 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} xmm2 = [4,11,128,128,0,7,14,128,128,u,u,u,u,u,u,u]
11034 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm9[1,2,3],ymm3[4],ymm9[5,6],ymm3[7,8],ymm9[9,10,11],ymm3[12],ymm9[13,14],ymm3[15]
11035 ; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [4,11,0,0,0,0,0,0,0,3,10,1,8,15,6,13,4,11,0,0,0,0,0,0,0,3,10,1,8,15,6,13]
11036 ; AVX2-FP-NEXT: # ymm4 = mem[0,1,0,1]
11039 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
11045 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm10[1,2,3],ymm1[4],ymm10[5,6],ymm1[7,8],ymm10[9,10,11],ymm1[12],ymm10[13,14],ymm1[15]
11048 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
11055 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
11057 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0],ymm2[1],ymm5[2,3],ymm2[4],ymm5[5,6,7,8],ymm2[9],ymm5[10,11],ymm2[12],ymm5[13,14,15]
11058 ; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [5,12,0,0,0,0,0,0,0,4,11,2,9,0,7,14,5,12,0,0,0,0,0,0,0,4,11,2,9,0,7,14]
11059 ; AVX2-FP-NEXT: # ymm5 = mem[0,1,0,1]
11062 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
11067 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
11069 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6,7,8],ymm3[9],ymm4[10,11],ymm3[12],ymm4[13,14,15]
11073 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
11076 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} xmm7 = [128,128,4,11,128,128,0,7,14,u,u,u,u,u,u,u]
11079 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
11081 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm5 = ymm8[0],ymm5[1],ymm8[2,3,4],ymm5[5],ymm8[6,7,8],ymm5[9],ymm8[10,11,12],ymm5[13],ymm8[14,15]
11082 ; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [6,13,0,0,0,0,0,0,0,5,12,3,10,1,8,15,6,13,0,0,0,0,0,0,0,5,12,3,10,1,8,15]
11083 ; AVX2-FP-NEXT: # ymm8 = mem[0,1,0,1]
11086 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
11092 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm5 = ymm11[0],ymm5[1],ymm11[2,3,4],ymm5[5],ymm11[6,7,8],ymm5[9],ymm11[10,11,12],ymm5[13],ymm11[14,15]
11095 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
11099 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} xmm8 = [u,u,0,7,14,128,128,3,10,128,128,128,u,u,u,u]
11103 ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
11106 ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
11108 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
11111 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5,6],ymm9[7]
11112 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
11113 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm5 = ymm9[0],ymm5[1,2,3,4,5,6,7],ymm9[8],ymm5[9,10,11,12,13,14,15]
11114 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm9[0,1,2,3],ymm5[4,5,6,7]
11115 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
11122 ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
11124 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
11127 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5,6],ymm8[7]
11128 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
11129 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm7 = ymm8[0],ymm7[1,2,3,4,5,6,7],ymm8[8],ymm7[9,10,11,12,13,14,15]
11130 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]
11131 ; AVX2-FP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm8 # 32-byte Folded Reload
11132 ; AVX2-FP-NEXT: # ymm8 = ymm12[0],mem[1,2,3,4,5,6,7],ymm12[8],mem[9,10,11,12,13,14,15]
11133 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
11134 ; AVX2-FP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm9 # 32-byte Folded Reload
11135 ; AVX2-FP-NEXT: # ymm9 = ymm14[0],mem[1,2,3,4,5,6,7],ymm14[8],mem[9,10,11,12,13,14,15]
11136 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm9 = ymm14[0,1,2,3],ymm9[4,5,6,7]
11137 ; AVX2-FP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm10 # 32-byte Folded Reload
11138 ; AVX2-FP-NEXT: # ymm10 = ymm6[0],mem[1,2,3,4,5,6,7],ymm6[8],mem[9,10,11,12,13,14,15]
11139 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm10[4,5,6,7]
11140 ; AVX2-FP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm10 # 32-byte Folded Reload
11141 ; AVX2-FP-NEXT: # ymm10 = ymm0[0],mem[1,2,3,4,5,6,7],ymm0[8],mem[9,10,11,12,13,14,15]
11142 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm10[4,5,6,7]
11143 ; AVX2-FP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm10 # 32-byte Folded Reload
11144 ; AVX2-FP-NEXT: # ymm10 = ymm2[0],mem[1,2,3,4,5,6,7],ymm2[8],mem[9,10,11,12,13,14,15]
11145 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm10[4,5,6,7]
11146 ; AVX2-FP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm10 # 32-byte Folded Reload
11147 ; AVX2-FP-NEXT: # ymm10 = ymm1[0],mem[1,2,3,4,5,6,7],ymm1[8],mem[9,10,11,12,13,14,15]
11148 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm10[4,5,6,7]
11149 ; AVX2-FP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm10 # 32-byte Folded Reload
11150 ; AVX2-FP-NEXT: # ymm10 = ymm3[0],mem[1,2,3,4,5,6,7],ymm3[8],mem[9,10,11,12,13,14,15]
11151 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm10[4,5,6,7]
11152 ; AVX2-FP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm10 # 32-byte Folded Reload
11153 ; AVX2-FP-NEXT: # ymm10 = ymm4[0],mem[1,2,3,4,5,6,7],ymm4[8],mem[9,10,11,12,13,14,15]
11154 ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm10[4,5,6,7]
11155 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
11157 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
11159 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
11161 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
11169 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
11172 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
11175 ; AVX2-FP-NEXT: addq $760, %rsp # imm = 0x2F8
11180 ; AVX2-FCP: # %bb.0:
11181 ; AVX2-FCP-NEXT: subq $776, %rsp # imm = 0x308
11189 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm7 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
11191 ; AVX2-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11192 ; AVX2-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11196 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [0,7,14,128,128,3,10,128,128,128,u,u,u,u,u,u]
11199 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm14 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
11201 ; AVX2-FCP-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11202 ; AVX2-FCP-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11204 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
11205 ; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [0,7,14,0,0,0,0,0,0,0,6,13,4,11,2,9,0,7,14,0,0,0,0,0,0,0,6,13,4,11,2,9]
11206 ; AVX2-FCP-NEXT: # ymm3 = mem[0,1,0,1]
11208 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} xmm4 = [65535,65535,65535,65535,65535,0,0,0]
11210 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11212 ; AVX2-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11214 ; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11221 ; AVX2-FCP-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11223 ; AVX2-FCP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11225 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm5[2],ymm2[3,4],ymm5[5],ymm2[6,7,8,9],ymm5[10],ymm2[11,12],ymm5[13],ymm2[14,15]
11228 ; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11229 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm15 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
11232 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm2[0,1],ymm3[2],ymm2[3,4,5],ymm3[6],ymm2[7,8,9],ymm3[10],ymm2[11,12,13],ymm3[14],ymm2[15]
11233 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm9 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
11241 ; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [1,8,15,0,0,0,0,0,0,0,7,14,5,12,3,10,1,8,15,0,0,0,0,0,0,0,7,14,5,12,3,10]
11242 ; AVX2-FCP-NEXT: # ymm5 = mem[0,1,0,1]
11244 ; AVX2-FCP-NEXT: vpmovsxdq {{.*#+}} ymm1 = [0,18446744073709551360,16777215,0]
11246 ; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11249 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm8[2],ymm3[3,4,5],ymm8[6],ymm3[7,8,9],ymm8[10],ymm3[11,12,13],ymm8[14],ymm3[15]
11259 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm9 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
11261 ; AVX2-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11262 ; AVX2-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11271 ; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [1,2,0,2,1,2,4,6]
11273 ; AVX2-FCP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11274 ; AVX2-FCP-NEXT: vpbroadcastd {{.*#+}} ymm1 = [0,7,10,13,0,7,10,13,0,7,10,13,0,7,10,13,0,7,10,13,0,7,10,13,0,7,10,13,0,7,10,13]
11276 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6],ymm3[7]
11277 ; AVX2-FCP-NEXT: vpmovsxdq {{.*#+}} ymm6 = [18446744073709551615,18446744073709551615,16777215,0]
11278 ; AVX2-FCP-NEXT: vpblendvb %ymm6, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm0 # 32-byte Folded Reload
11279 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11290 ; AVX2-FCP-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11293 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm1[7]
11294 ; AVX2-FCP-NEXT: vpblendvb %ymm6, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11295 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11296 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm13 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
11301 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [u,u,u,128,128,4,11,128,128,0,7,14,u,u,u,u]
11305 ; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [1,3,0,2,1,3,4,6]
11309 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm11[7]
11310 ; AVX2-FCP-NEXT: vpblendvb %ymm6, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11311 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11320 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm1[7]
11322 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11323 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm7 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
11325 ; AVX2-FCP-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11332 ; AVX2-FCP-NEXT: vpbroadcastq {{.*#+}} xmm10 = [0,0,0,128,128,128,5,12,0,0,0,128,128,128,5,12]
11335 ; AVX2-FCP-NEXT: vpbroadcastq {{.*#+}} xmm11 = [0,0,0,0,7,14,128,128,0,0,0,0,7,14,128,128]
11341 ; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [0,0,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,0,0,0,0,0]
11342 ; AVX2-FCP-NEXT: # ymm9 = mem[0,1,0,1]
11344 ; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11345 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11346 ; AVX2-FCP-NEXT: vpblendvb %ymm7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
11353 ; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11360 ; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11361 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm0 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
11368 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [u,u,128,128,0,7,14,128,128,3,10,u,u,u,u,u]
11371 ; AVX2-FCP-NEXT: vpbroadcastq {{.*#+}} xmm11 = [0,0,0,128,128,128,6,13,0,0,0,128,128,128,6,13]
11375 ; AVX2-FCP-NEXT: vpbroadcastq {{.*#+}} xmm1 = [0,0,0,1,8,15,128,128,0,0,0,1,8,15,128,128]
11382 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11383 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11384 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
11396 ; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11397 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm6 = [0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535]
11399 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
11407 ; AVX2-FCP-NEXT: vpbroadcastq {{.*#+}} xmm10 = [0,0,0,2,9,128,128,128,0,0,0,2,9,128,128,128]
11409 ; AVX2-FCP-NEXT: vpbroadcastq {{.*#+}} xmm15 = [0,0,0,128,128,0,7,14,0,0,0,128,128,0,7,14]
11415 ; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11424 ; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
11430 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11431 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm12 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
11433 ; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11434 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm6 = [65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0]
11439 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [u,u,4,11,128,128,0,7,14,128,128,u,u,u,u,u]
11442 ; AVX2-FCP-NEXT: vpbroadcastq {{.*#+}} xmm2 = [0,0,0,3,10,128,128,128,0,0,0,3,10,128,128,128]
11444 ; AVX2-FCP-NEXT: vpbroadcastq {{.*#+}} xmm11 = [0,0,0,128,128,1,8,15,0,0,0,128,128,1,8,15]
11454 ; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11461 ; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
11467 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11468 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11469 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11470 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm12 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
11472 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm15 = [0,65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535]
11475 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm5 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
11477 ; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11478 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,65535,0,0,65535,0,0,0,65535,0,0,65535,0,0,0,65535]
11480 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11481 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11482 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
11484 ; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11488 ; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11490 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11491 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11492 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
11493 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm7 = [65535,0,0,0,65535,0,0,65535,0,0,0,65535,0,0,65535,0]
11498 ; AVX2-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11500 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11502 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
11503 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
11508 ; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11510 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11514 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [128,128,0,7,14,128,128,3,10,u,u,u,u,u,u,u]
11518 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm11[3],ymm1[4,5],ymm11[6],ymm1[7,8,9,10],ymm11[11],ymm1[12,13],ymm11[14],ymm1[15]
11519 ; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [2,9,0,0,0,0,0,0,0,1,8,15,6,13,4,11,2,9,0,0,0,0,0,0,0,1,8,15,6,13,4,11]
11520 ; AVX2-FCP-NEXT: # ymm5 = mem[0,1,0,1]
11524 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11530 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
11533 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11541 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm4[1,2],ymm2[3],ymm4[4,5,6],ymm2[7,8],ymm4[9,10],ymm2[11],ymm4[12,13,14],ymm2[15]
11542 ; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [3,10,0,0,0,0,0,0,0,2,9,0,7,14,5,12,3,10,0,0,0,0,0,0,0,2,9,0,7,14,5,12]
11543 ; AVX2-FCP-NEXT: # ymm4 = mem[0,1,0,1]
11551 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm12[1,2],ymm1[3],ymm12[4,5,6],ymm1[7,8],ymm12[9,10],ymm1[11],ymm12[12,13,14],ymm1[15]
11557 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [4,11,128,128,0,7,14,128,128,u,u,u,u,u,u,u]
11561 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm14[1,2,3],ymm3[4],ymm14[5,6],ymm3[7,8],ymm14[9,10,11],ymm3[12],ymm14[13,14],ymm3[15]
11562 ; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [4,11,0,0,0,0,0,0,0,3,10,1,8,15,6,13,4,11,0,0,0,0,0,0,0,3,10,1,8,15,6,13]
11563 ; AVX2-FCP-NEXT: # ymm4 = mem[0,1,0,1]
11566 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
11572 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm15[1,2,3],ymm1[4],ymm15[5,6],ymm1[7,8],ymm15[9,10,11],ymm1[12],ymm15[13,14],ymm1[15]
11575 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
11582 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
11584 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6,7,8],ymm3[9],ymm4[10,11],ymm3[12],ymm4[13,14,15]
11585 ; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [5,12,0,0,0,0,0,0,0,4,11,2,9,0,7,14,5,12,0,0,0,0,0,0,0,4,11,2,9,0,7,14]
11586 ; AVX2-FCP-NEXT: # ymm4 = mem[0,1,0,1]
11589 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
11594 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
11596 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6,7,8],ymm2[9],ymm3[10,11],ymm2[12],ymm3[13,14,15]
11600 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
11603 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [128,128,4,11,128,128,0,7,14,u,u,u,u,u,u,u]
11606 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
11608 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm6[0],ymm4[1],ymm6[2,3,4],ymm4[5],ymm6[6,7,8],ymm4[9],ymm6[10,11,12],ymm4[13],ymm6[14,15]
11609 ; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [6,13,0,0,0,0,0,0,0,5,12,3,10,1,8,15,6,13,0,0,0,0,0,0,0,5,12,3,10,1,8,15]
11610 ; AVX2-FCP-NEXT: # ymm6 = mem[0,1,0,1]
11613 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
11618 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
11620 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3,4],ymm4[5],ymm5[6,7,8],ymm4[9],ymm5[10,11,12],ymm4[13],ymm5[14,15]
11623 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
11627 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,0,7,14,128,128,3,10,128,128,128,u,u,u,u]
11632 ; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm8 # 32-byte Folded Reload
11635 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6],ymm8[7]
11636 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
11637 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm8[0],ymm4[1,2,3,4,5,6,7],ymm8[8],ymm4[9,10,11,12,13,14,15]
11638 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]
11639 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
11644 ; AVX2-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm6 # 32-byte Folded Reload
11647 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5,6],ymm6[7]
11648 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
11649 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm6[0],ymm5[1,2,3,4,5,6,7],ymm6[8],ymm5[9,10,11,12,13,14,15]
11650 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
11651 ; AVX2-FCP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm6 # 32-byte Folded Reload
11652 ; AVX2-FCP-NEXT: # ymm6 = ymm9[0],mem[1,2,3,4,5,6,7],ymm9[8],mem[9,10,11,12,13,14,15]
11653 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm9[0,1,2,3],ymm6[4,5,6,7]
11654 ; AVX2-FCP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm7 # 32-byte Folded Reload
11655 ; AVX2-FCP-NEXT: # ymm7 = ymm12[0],mem[1,2,3,4,5,6,7],ymm12[8],mem[9,10,11,12,13,14,15]
11656 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm7 = ymm12[0,1,2,3],ymm7[4,5,6,7]
11657 ; AVX2-FCP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm8 # 32-byte Folded Reload
11658 ; AVX2-FCP-NEXT: # ymm8 = ymm14[0],mem[1,2,3,4,5,6,7],ymm14[8],mem[9,10,11,12,13,14,15]
11659 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm8[4,5,6,7]
11660 ; AVX2-FCP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm9 # 32-byte Folded Reload
11661 ; AVX2-FCP-NEXT: # ymm9 = ymm15[0],mem[1,2,3,4,5,6,7],ymm15[8],mem[9,10,11,12,13,14,15]
11662 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm15[0,1,2,3],ymm9[4,5,6,7]
11663 ; AVX2-FCP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm10 # 32-byte Folded Reload
11664 ; AVX2-FCP-NEXT: # ymm10 = ymm0[0],mem[1,2,3,4,5,6,7],ymm0[8],mem[9,10,11,12,13,14,15]
11665 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm10[4,5,6,7]
11666 ; AVX2-FCP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm10 # 32-byte Folded Reload
11667 ; AVX2-FCP-NEXT: # ymm10 = ymm1[0],mem[1,2,3,4,5,6,7],ymm1[8],mem[9,10,11,12,13,14,15]
11668 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm10[4,5,6,7]
11670 ; AVX2-FCP-NEXT: # ymm10 = ymm2[0],mem[1,2,3,4,5,6,7],ymm2[8],mem[9,10,11,12,13,14,15]
11671 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm10[4,5,6,7]
11672 ; AVX2-FCP-NEXT: vpblendw $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm10 # 32-byte Folded Reload
11673 ; AVX2-FCP-NEXT: # ymm10 = ymm3[0],mem[1,2,3,4,5,6,7],ymm3[8],mem[9,10,11,12,13,14,15]
11674 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm10[4,5,6,7]
11675 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
11677 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
11679 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
11681 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
11689 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
11692 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
11695 ; AVX2-FCP-NEXT: addq $776, %rsp # imm = 0x308
11700 ; AVX512: # %bb.0:
11701 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm0 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0]
11710 ; AVX512-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[0,7,14],zero,zero,xmm1[3,10],zero,zero,zero,xmm1[u,u,u,u,u,u]
11712 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535]
11717 ; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm11[2],ymm2[3,4],ymm11[5],ymm2[6,7,8,9],ymm11[10],ymm2[11,12],ymm11[13],ymm2[14,15]
11720 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm14 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535]
11731 ; AVX512-NEXT: vpbroadcastd {{.*#+}} xmm4 = [0,0,4,11,0,0,4,11,0,0,4,11,0,0,4,11]
11737 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
11739 ; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6],ymm3[7]
11743 ; AVX512-NEXT: vpshufb {{.*#+}} xmm7 = xmm6[0,7,14],zero,zero,xmm6[u,u,u,u,u,u,u,u,u,u,u]
11746 ; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm4 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
11760 ; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm7[2,3,0,1]
11761 ; AVX512-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm8[2],ymm7[3,4,5],ymm8[6],ymm7[7,8,9],ymm8[10],ymm7[11,12,13],ymm8[14],ymm7[15]
11763 ; AVX512-NEXT: vmovdqa64 {{.*#+}} ymm23 = [65535,65535,65535,65535,65535,65535,65535,0,0,0,0,0,65535,65535,65535,65535]
11765 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm7 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
11775 ; AVX512-NEXT: vpblendw {{.*#+}} ymm15 = ymm15[0,1],ymm11[2],ymm15[3,4,5],ymm11[6],ymm15[7,8,9],ymm11[10],ymm15[11,12,13],ymm11[14],ymm15[15]
11776 ; AVX512-NEXT: vpshufb {{.*#+}} ymm15 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm15[0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
11782 ; AVX512-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u],zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14,u,u,u,u]
11792 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3]
11795 ; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm3[7]
11802 ; AVX512-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[0,7,14],zero,zero,xmm0[3,10,u,u,u,u,u,u,u]
11806 ; AVX512-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
11807 ; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1,2],ymm11[3],ymm2[4,5],ymm11[6],ymm2[7,8,9,10],ymm11[11],ymm2[12,13],ymm11[14],ymm2[15]
11815 ; AVX512-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,0,7,14],zero,zero,xmm0[3,10],zero,zero,zero,xmm0[u,u,u,u]
11818 ; AVX512-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11823 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3]
11825 ; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm3[7]
11827 ; AVX512-NEXT: vmovdqa64 %xmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11830 ; AVX512-NEXT: vmovdqa64 %xmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
11831 ; AVX512-NEXT: vpshufb {{.*#+}} xmm6 = zero,zero,xmm4[0,7,14,u,u,u,u,u,u,u,u,u,u,u]
11834 ; AVX512-NEXT: vpmovsxdq {{.*#+}} zmm20 = [0,0,18446744073709486080,18446744073709551615,18446744073709551615,18446744073709551615,18446744073709551615,18446744073709551615]
11844 ; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm11[0],ymm2[1,2],ymm11[3],ymm2[4,5,6],ymm11[7,8],ymm2[9,10],ymm11[11],ymm2[12,13,14],ymm11[15]
11845 ; AVX512-NEXT: vpshufb {{.*#+}} ymm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[2,9,0,7,14,5,12,19,26],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
11857 ; AVX512-NEXT: vpshufb {{.*#+}} xmm15 = xmm5[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
11876 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0]
11879 ; AVX512-NEXT: vpmovsxwd {{.*#+}} zmm8 = [4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,255,0,0,0,0,0,0]
11885 ; AVX512-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,u,u,u,4,11],zero,zero,xmm0[0,7,14],zero,zero,xmm0[u,u]
11889 ; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm3[2,3,0,1]
11890 ; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm6[3],ymm3[4,5],ymm6[6],ymm3[7,8,9,10],ymm6[11],ymm3[12,13],ymm6[14],ymm3[15]
11897 ; AVX512-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,u,u,u,u,u],zero,zero,xmm0[0,7,14],zero,zero,xmm0[3,10]
11911 ; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm3[2,3,0,1]
11912 ; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm6[0],ymm3[1,2],ymm6[3],ymm3[4,5,6],ymm6[7,8],ymm3[9,10],ymm6[11],ymm3[12,13,14],ymm6[15]
11927 ; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm0[2,3,0,1]
11928 ; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0],ymm0[1,2,3],ymm3[4],ymm0[5,6],ymm3[7,8],ymm0[9,10,11],ymm3[12],ymm0[13,14],ymm3[15]
11934 ; AVX512-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,u],zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14,u,u]
11941 ; AVX512-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,u,u,u,u,u,4,11],zero,zero,xmm0[0,7,14],zero,zero
11949 ; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm0[2,3,0,1]
11950 ; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6,7,8],ymm2[9],ymm0[10,11],ymm2[12],ymm0[13,14,15]
11951 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm4 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0]
11956 ; AVX512-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,0,7,14],zero,zero,xmm2[3,10],zero,zero,zero,xmm2[u,u]
11970 ; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm0[2,3,0,1]
11971 ; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3,4],ymm2[5],ymm0[6,7,8],ymm2[9],ymm0[10,11,12],ymm2[13],ymm0[14,15]
11978 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm5 = [u,u,u,u,128,128,128,128,128,128,128,128,128,0,7,14,21,28,19,26,17,24,31,u,u,u,u,u,u,u,u,u]
11985 ; AVX512-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,u,u,u,u,u],zero,zero,xmm0[4,11],zero,zero,xmm0[0,7,14]
11993 ; AVX512-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[4,11],zero,zero,xmm0[0,7,14],zero,zero,xmm0[u,u,u,u,u,u,u]
12008 ; AVX512-NEXT: vpshufb {{.*#+}} xmm9 = zero,zero,xmm9[4,11],zero,zero,xmm9[0,7,14,u,u,u,u,u,u,u]
12015 ; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
12016 ; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm8[0],ymm3[1,2,3],ymm8[4],ymm3[5,6],ymm8[7,8],ymm3[9,10,11],ymm8[12],ymm3[13,14],ymm8[15]
12020 ; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm9[0],ymm8[1],ymm9[2,3],ymm8[4],ymm9[5,6,7,8],ymm8[9],ymm9[10,11],ymm8[12],ymm9[13,14,15]
12021 ; AVX512-NEXT: vpshufb {{.*#+}} ymm11 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm0[4,11,2,9,0,7,14,21,28],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
12023 ; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm14[0],ymm8[1],ymm14[2,3,4],ymm8[5],ymm14[6,7,8],ymm8[9],ymm14[10,11,12],ymm8[13],ymm14[14,15]
12029 ; AVX512-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u],zero,zero,xmm2[0,7,14],zero,zero,xmm2[3,10,u,u,u,u,u]
12031 ; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
12042 ; AVX512-NEXT: vmovd {{.*#+}} xmm10 = [4,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
12045 ; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
12047 ; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
12049 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3]
12051 ; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,65535,65535,65535,65535,65535,0,0,0,0,0,0,0,0,0,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535]
12066 ; AVX512-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,u,4,11],zero,zero,xmm7[0,7,14],zero,zero,xmm7[u,u,u,u,u]
12077 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm13[0],xmm3[0],xmm13[1],xmm3[1],xmm13[2],xmm3[2],xmm13[3],xmm3[3]
12082 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm10[0],xmm6[0],xmm10[1],xmm6[1],xmm10[2],xmm6[2],xmm10[3],xmm6[3]
12086 ; AVX512-NEXT: movw $-512, %ax # imm = 0xFE00
12092 ; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm28[2,3,0,1]
12093 ; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
12098 ; AVX512-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,u,u],zero,zero,xmm1[0,7,14],zero,zero,xmm1[3,10,u,u,u]
12103 ; AVX512-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,u,u,u,u,u,0,7,14],zero,zero,xmm4[3,10],zero,zero,zero
12106 ; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7],ymm1[8,9,10],ymm0[11,12,13,14,15]
12107 ; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
12115 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
12117 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
12123 ; AVX512-FCP: # %bb.0:
12124 ; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} ymm19 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0]
12132 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,7,14],zero,zero,xmm0[3,10],zero,zero,zero,xmm0[u,u,u,u,u,u]
12134 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535]
12139 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm6[2],ymm1[3,4],ymm6[5],ymm1[6,7,8,9],ymm6[10],ymm1[11,12],ymm6[13],ymm1[14,15]
12142 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535]
12152 ; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [0,0,0,0,1,2,4,6]
12156 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm2[7]
12160 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm5[0,7,14],zero,zero,xmm5[u,u,u,u,u,u,u,u,u,u,u]
12163 ; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm21 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
12177 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm7[2,3,0,1]
12178 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm13[2],ymm7[3,4,5],ymm13[6],ymm7[7,8,9],ymm13[10],ymm7[11,12,13],ymm13[14],ymm7[15]
12180 ; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} ymm26 = [65535,65535,65535,65535,65535,65535,65535,0,0,0,0,0,65535,65535,65535,65535]
12182 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
12192 ; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} ymm23 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0]
12195 ; AVX512-FCP-NEXT: vpmovsxwd {{.*#+}} zmm29 = [4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,255,0,0,0,0,0,0]
12197 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12206 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0,1],ymm6[2],ymm13[3,4,5],ymm6[6],ymm13[7,8,9],ymm6[10],ymm13[11,12,13],ymm6[14],ymm13[15]
12207 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm13 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm13[0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
12213 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[u,u,u],zero,zero,xmm8[4,11],zero,zero,xmm8[0,7,14,u,u,u,u]
12216 ; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm15 = [0,0,0,0,1,3,4,6]
12219 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5,6],ymm15[7]
12229 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[u,u,u,u,u,4,11],zero,zero,xmm8[0,7,14],zero,zero,xmm8[u,u]
12233 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm13[2,3,0,1]
12234 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0,1,2],ymm15[3],ymm13[4,5],ymm15[6],ymm13[7,8,9,10],ymm15[11],ymm13[12,13],ymm15[14],ymm13[15]
12241 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[u,u,u,u,u,u,u],zero,zero,xmm8[0,7,14],zero,zero,xmm8[3,10]
12251 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,xmm7[0,7,14],zero,zero,xmm7[3,10,u,u,u,u,u,u,u]
12255 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0,1,2],ymm6[3],ymm8[4,5],ymm6[6],ymm8[7,8,9,10],ymm6[11],ymm8[12,13],ymm6[14],ymm8[15]
12263 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,u,0,7,14],zero,zero,xmm7[3,10],zero,zero,zero,xmm7[u,u,u,u]
12266 ; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm13 = [0,0,0,0,1,3,5,6]
12269 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm7[0,1,2,3,4,5,6],ymm3[7]
12271 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm13 = zero,zero,xmm4[0,7,14,u,u,u,u,u,u,u,u,u,u,u]
12274 ; AVX512-FCP-NEXT: vpmovsxdq {{.*#+}} zmm22 = [0,0,18446744073709486080,18446744073709551615,18446744073709551615,18446744073709551615,18446744073709551615,18446744073709551615]
12284 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm8[2,3,0,1]
12285 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm13[0],ymm8[1,2],ymm13[3],ymm8[4,5,6],ymm13[7,8],ymm8[9,10],ymm13[11],ymm8[12,13,14],ymm13[15]
12306 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm6[0],ymm7[1,2],ymm6[3],ymm7[4,5,6],ymm6[7,8],ymm7[9,10],ymm6[11],ymm7[12,13,14],ymm6[15]
12307 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm7[2,9,0,7,14,5,12,19,26],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
12317 ; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12320 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm15[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
12332 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm3[2,3,0,1]
12333 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1,2,3],ymm4[4],ymm3[5,6],ymm4[7,8],ymm3[9,10,11],ymm4[12],ymm3[13,14],ymm4[15]
12339 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,u,u,u,u],zero,zero,xmm4[4,11],zero,zero,xmm4[0,7,14,u,u]
12346 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u,u,u,u,u,u,4,11],zero,zero,xmm3[0,7,14],zero,zero
12354 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm0[2,3,0,1]
12355 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3],ymm3[4],ymm0[5,6,7,8],ymm3[9],ymm0[10,11],ymm3[12],ymm0[13,14,15]
12361 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u,u,u,0,7,14],zero,zero,xmm3[3,10],zero,zero,zero,xmm3[u,u]
12372 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [u,u,u,u,128,128,128,128,128,128,128,128,128,0,7,14,21,28,19,26,17,24,31,u,u,u,u,u,u,u,u,u]
12375 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm0[2,3,0,1]
12376 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7,8],ymm3[9],ymm0[10,11,12],ymm3[13],ymm0[14,15]
12389 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,u,u,u,u,u],zero,zero,xmm0[4,11],zero,zero,xmm0[0,7,14]
12403 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm12 = xmm13[4,11],zero,zero,xmm13[0,7,14],zero,zero,xmm13[u,u,u,u,u,u,u]
12417 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm10 = zero,zero,xmm10[4,11],zero,zero,xmm10[0,7,14,u,u,u,u,u,u,u]
12419 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm6[0],ymm7[1,2,3],ymm6[4],ymm7[5,6],ymm6[7,8],ymm7[9,10,11],ymm6[12],ymm7[13,14],ymm6[15]
12422 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm13[0],ymm6[1],ymm13[2,3],ymm6[4],ymm13[5,6,7,8],ymm6[9],ymm13[10,11],ymm6[12],ymm13[13,14,15]
12423 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm3[4,11,2,9,0,7,14,21,28],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
12425 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm9[0],ymm6[1],ymm9[2,3,4],ymm6[5],ymm9[6,7,8],ymm6[9],ymm9[10,11,12],ymm6[13],ymm9[14,15]
12434 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u],zero,zero,xmm3[0,7,14],zero,zero,xmm3[3,10,u,u,u,u,u]
12436 ; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
12457 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,4,11],zero,zero,xmm1[0,7,14],zero,zero,xmm1[u,u,u,u,u]
12466 ; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [1,2,4,6,0,0,0,0]
12469 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[0,7,10,13,u,u,u,u,u,u,u,u,u,u,u,u]
12471 ; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,65535,65535,65535,65535,65535,65535,0,0,0,0,0,0,0,0,0,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535]
12473 ; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [1,3,4,6,0,0,0,0]
12478 ; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [1,3,5,6,0,0,0,0]
12484 ; AVX512-FCP-NEXT: movw $-512, %ax # imm = 0xFE00
12489 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm2[2,3,0,1]
12490 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7,8,9],ymm3[10],ymm2[11,12],ymm3[13],ymm2[14,15]
12494 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u,u,u],zero,zero,xmm3[0,7,14],zero,zero,xmm3[3,10,u,u,u]
12500 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u,u,u,u,u,0,7,14],zero,zero,xmm3[3,10],zero,zero,zero
12503 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7],ymm2[8,9,10],ymm0[11,12,13,14,15]
12504 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
12507 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12513 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
12515 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
12521 ; AVX512DQ: # %bb.0:
12523 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm0 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0]
12532 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[0,7,14],zero,zero,xmm1[3,10],zero,zero,zero,xmm1[u,u,u,u,u,u]
12534 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535]
12539 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm11[2],ymm2[3,4],ymm11[5],ymm2[6,7,8,9],ymm11[10],ymm2[11,12],ymm11[13],ymm2[14,15]
12542 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm14 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535]
12553 ; AVX512DQ-NEXT: vpbroadcastd {{.*#+}} xmm5 = [0,0,4,11,0,0,4,11,0,0,4,11,0,0,4,11]
12559 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3]
12561 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6],ymm3[7]
12565 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm7 = xmm5[0,7,14],zero,zero,xmm5[u,u,u,u,u,u,u,u,u,u,u]
12568 ; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm19 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
12582 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm7[2,3,0,1]
12583 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm8[2],ymm7[3,4,5],ymm8[6],ymm7[7,8,9],ymm8[10],ymm7[11,12,13],ymm8[14],ymm7[15]
12585 ; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} ymm24 = [65535,65535,65535,65535,65535,65535,65535,0,0,0,0,0,65535,65535,65535,65535]
12587 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm7 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
12597 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm15 = ymm15[0,1],ymm11[2],ymm15[3,4,5],ymm11[6],ymm15[7,8,9],ymm11[10],ymm15[11,12,13],ymm11[14],ymm15[15]
12598 ; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm15 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm15[0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
12605 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u],zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14,u,u,u,u]
12615 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
12618 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm3[7]
12625 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[0,7,14],zero,zero,xmm0[3,10,u,u,u,u,u,u,u]
12629 ; AVX512DQ-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12630 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1,2],ymm11[3],ymm2[4,5],ymm11[6],ymm2[7,8,9,10],ymm11[11],ymm2[12,13],ymm11[14],ymm2[15]
12636 ; AVX512DQ-NEXT: vmovdqu64 %ymm25, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12639 ; AVX512DQ-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
12642 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,0,7,14],zero,zero,xmm0[3,10],zero,zero,zero,xmm0[u,u,u,u]
12649 ; AVX512DQ-NEXT: vmovdqa64 %xmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12650 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
12652 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm3[7]
12654 ; AVX512DQ-NEXT: vmovdqa64 %xmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12656 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,xmm4[0,7,14,u,u,u,u,u,u,u,u,u,u,u]
12657 ; AVX512DQ-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12660 ; AVX512DQ-NEXT: vpmovsxdq {{.*#+}} zmm20 = [0,0,18446744073709486080,18446744073709551615,18446744073709551615,18446744073709551615,18446744073709551615,18446744073709551615]
12670 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm2 = ymm11[0],ymm2[1,2],ymm11[3],ymm2[4,5,6],ymm11[7,8],ymm2[9,10],ymm11[11],ymm2[12,13,14],ymm11[15]
12671 ; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[2,9,0,7,14,5,12,19,26],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
12683 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm15 = xmm6[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
12701 ; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} ymm29 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0]
12704 ; AVX512DQ-NEXT: vpmovsxwd {{.*#+}} zmm8 = [4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,255,0,0,0,0,0,0]
12710 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,u,u,u,4,11],zero,zero,xmm0[0,7,14],zero,zero,xmm0[u,u]
12714 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm3[2,3,0,1]
12715 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm5[3],ymm3[4,5],ymm5[6],ymm3[7,8,9,10],ymm5[11],ymm3[12,13],ymm5[14],ymm3[15]
12722 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,u,u,u,u,u],zero,zero,xmm0[0,7,14],zero,zero,xmm0[3,10]
12736 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm3[2,3,0,1]
12737 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1,2],ymm5[3],ymm3[4,5,6],ymm5[7,8],ymm3[9,10],ymm5[11],ymm3[12,13,14],ymm5[15]
12752 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm0[2,3,0,1]
12753 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0],ymm0[1,2,3],ymm3[4],ymm0[5,6],ymm3[7,8],ymm0[9,10,11],ymm3[12],ymm0[13,14],ymm3[15]
12759 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,u],zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14,u,u]
12766 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,u,u,u,u,u,4,11],zero,zero,xmm0[0,7,14],zero,zero
12774 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm0[2,3,0,1]
12775 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6,7,8],ymm2[9],ymm0[10,11],ymm2[12],ymm0[13,14,15]
12776 ; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} ymm19 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0]
12781 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,0,7,14],zero,zero,xmm2[3,10],zero,zero,zero,xmm2[u,u]
12795 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm0[2,3,0,1]
12796 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3,4],ymm2[5],ymm0[6,7,8],ymm2[9],ymm0[10,11,12],ymm2[13],ymm0[14,15]
12803 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm8 = [u,u,u,u,128,128,128,128,128,128,128,128,128,0,7,14,21,28,19,26,17,24,31,u,u,u,u,u,u,u,u,u]
12810 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,u,u,u,u,u],zero,zero,xmm0[4,11],zero,zero,xmm0[0,7,14]
12818 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[4,11],zero,zero,xmm0[0,7,14],zero,zero,xmm0[u,u,u,u,u,u,u]
12833 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,xmm5[4,11],zero,zero,xmm5[0,7,14,u,u,u,u,u,u,u]
12840 ; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
12841 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0],ymm2[1,2,3],ymm4[4],ymm2[5,6],ymm4[7,8],ymm2[9,10,11],ymm4[12],ymm2[13,14],ymm4[15]
12845 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm2 = ymm9[0],ymm4[1],ymm9[2,3],ymm4[4],ymm9[5,6,7,8],ymm4[9],ymm9[10,11],ymm4[12],ymm9[13,14,15]
12846 ; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm11 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[4,11,2,9,0,7,14,21,28],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
12848 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm14[0],ymm4[1],ymm14[2,3,4],ymm4[5],ymm14[6,7,8],ymm4[9],ymm14[10,11,12],ymm4[13],ymm14[14,15]
12851 ; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
12852 ; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
12856 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u],zero,zero,xmm2[0,7,14],zero,zero,xmm2[3,10,u,u,u,u,u]
12861 ; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
12869 ; AVX512DQ-NEXT: vmovd {{.*#+}} xmm10 = [4,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
12872 ; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
12874 ; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
12876 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
12878 ; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,65535,65535,65535,65535,65535,0,0,0,0,0,0,0,0,0,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535]
12892 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,u,4,11],zero,zero,xmm7[0,7,14],zero,zero,xmm7[u,u,u,u,u]
12903 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm12[0],xmm3[0],xmm12[1],xmm3[1],xmm12[2],xmm3[2],xmm12[3],xmm3[3]
12908 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm10[0],xmm5[0],xmm10[1],xmm5[1],xmm10[2],xmm5[2],xmm10[3],xmm5[3]
12911 ; AVX512DQ-NEXT: movw $-512, %ax # imm = 0xFE00
12916 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm16[2,3,0,1]
12917 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
12922 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,u,u],zero,zero,xmm1[0,7,14],zero,zero,xmm1[3,10,u,u,u]
12928 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,u,u,u,u,u,0,7,14],zero,zero,xmm4[3,10],zero,zero,zero
12931 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7],ymm1[8,9,10],ymm0[11,12,13,14,15]
12932 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
12939 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
12941 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
12948 ; AVX512DQ-FCP: # %bb.0:
12950 ; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} ymm19 = [65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0]
12958 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,7,14],zero,zero,xmm0[3,10],zero,zero,zero,xmm0[u,u,u,u,u,u]
12960 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535]
12965 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7,8,9],ymm3[10],ymm2[11,12],ymm3[13],ymm2[14,15]
12969 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535]
12979 ; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,1,2,4,6]
12983 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6],ymm4[7]
12987 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm5[0,7,14],zero,zero,xmm5[u,u,u,u,u,u,u,u,u,u,u]
12990 ; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
13004 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm13[2,3,0,1]
13005 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm13[0,1],ymm8[2],ymm13[3,4,5],ymm8[6],ymm13[7,8,9],ymm8[10],ymm13[11,12,13],ymm8[14],ymm13[15]
13007 ; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} ymm26 = [65535,65535,65535,65535,65535,65535,65535,0,0,0,0,0,65535,65535,65535,65535]
13009 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,65535]
13019 ; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} ymm21 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0,0]
13022 ; AVX512DQ-FCP-NEXT: vpmovsxwd {{.*#+}} zmm29 = [4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,4294967295,255,0,0,0,0,0,0]
13024 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
13034 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0,1],ymm3[2],ymm8[3,4,5],ymm3[6],ymm8[7,8,9],ymm3[10],ymm8[11,12,13],ymm3[14],ymm8[15]
13035 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm8[0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
13041 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,u,u],zero,zero,xmm7[4,11],zero,zero,xmm7[0,7,14,u,u,u,u]
13044 ; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm13 = [0,0,0,0,1,3,4,6]
13047 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5,6],ymm13[7]
13057 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[u,u,u,u,u,4,11],zero,zero,xmm8[0,7,14],zero,zero,xmm8[u,u]
13061 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm13[2,3,0,1]
13062 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0,1,2],ymm15[3],ymm13[4,5],ymm15[6],ymm13[7,8,9,10],ymm15[11],ymm13[12,13],ymm15[14],ymm13[15]
13069 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[u,u,u,u,u,u,u],zero,zero,xmm8[0,7,14],zero,zero,xmm8[3,10]
13079 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,xmm7[0,7,14],zero,zero,xmm7[3,10,u,u,u,u,u,u,u]
13083 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0,1,2],ymm3[3],ymm8[4,5],ymm3[6],ymm8[7,8,9,10],ymm3[11],ymm8[12,13],ymm3[14],ymm8[15]
13085 ; AVX512DQ-FCP-NEXT: vmovdqu64 %ymm20, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13093 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,u,0,7,14],zero,zero,xmm7[3,10],zero,zero,zero,xmm7[u,u,u,u]
13096 ; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,0,0,0,1,3,5,6]
13099 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm7[0,1,2,3,4,5,6],ymm3[7]
13101 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm8 = zero,zero,xmm4[0,7,14,u,u,u,u,u,u,u,u,u,u,u]
13104 ; AVX512DQ-FCP-NEXT: vpmovsxdq {{.*#+}} zmm22 = [0,0,18446744073709486080,18446744073709551615,18446744073709551615,18446744073709551615,18446744073709551615,18446744073709551615]
13114 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm13[2,3,0,1]
13115 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm13 = ymm15[0],ymm13[1,2],ymm15[3],ymm13[4,5,6],ymm15[7,8],ymm13[9,10],ymm15[11],ymm13[12,13,14],ymm15[15]
13136 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm0[0],ymm7[1,2],ymm0[3],ymm7[4,5,6],ymm0[7,8],ymm7[9,10],ymm0[11],ymm7[12,13,14],ymm0[15]
13137 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm7[2,9,0,7,14,5,12,19,26],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
13147 ; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13150 ; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13151 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm0[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
13164 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm4[2,3,0,1]
13165 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1,2,3],ymm5[4],ymm4[5,6],ymm5[7,8],ymm4[9,10,11],ymm5[12],ymm4[13,14],ymm5[15]
13171 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[u,u,u,u,u],zero,zero,xmm5[4,11],zero,zero,xmm5[0,7,14,u,u]
13178 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,u,u,u,u,u,u,4,11],zero,zero,xmm4[0,7,14],zero,zero
13186 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm3[2,3,0,1]
13187 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3],ymm4[4],ymm3[5,6,7,8],ymm4[9],ymm3[10,11],ymm4[12],ymm3[13,14,15]
13193 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,u,u,u,0,7,14],zero,zero,xmm4[3,10],zero,zero,zero,xmm4[u,u]
13206 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm3[2,3,0,1]
13207 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7,8],ymm4[9],ymm3[10,11,12],ymm4[13],ymm3[14,15]
13208 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [u,u,u,u,128,128,128,128,128,128,128,128,128,0,7,14,21,28,19,26,17,24,31,u,u,u,u,u,u,u,u,u]
13221 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u,u,u,u,u,u],zero,zero,xmm3[4,11],zero,zero,xmm3[0,7,14]
13235 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm12 = xmm13[4,11],zero,zero,xmm13[0,7,14],zero,zero,xmm13[u,u,u,u,u,u,u]
13249 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[4,11],zero,zero,xmm0[0,7,14,u,u,u,u,u,u,u]
13251 ; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
13252 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm4[0],ymm8[1,2,3],ymm4[4],ymm8[5,6],ymm4[7,8],ymm8[9,10,11],ymm4[12],ymm8[13,14],ymm4[15]
13255 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm15[0],ymm4[1],ymm15[2,3],ymm4[4],ymm15[5,6,7,8],ymm4[9],ymm15[10,11],ymm4[12],ymm15[13,14,15]
13256 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm0[4,11,2,9,0,7,14,21,28],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
13258 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm9[0],ymm4[1],ymm9[2,3,4],ymm4[5],ymm9[6,7,8],ymm4[9],ymm9[10,11,12],ymm4[13],ymm9[14,15]
13267 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[u,u],zero,zero,xmm9[0,7,14],zero,zero,xmm9[3,10,u,u,u,u,u]
13269 ; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
13272 ; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
13290 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm10[u,u,4,11],zero,zero,xmm10[0,7,14],zero,zero,xmm10[u,u,u,u,u]
13299 ; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [1,2,4,6,0,0,0,0]
13302 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0,7,10,13,u,u,u,u,u,u,u,u,u,u,u,u]
13304 ; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,65535,65535,65535,65535,65535,65535,65535,65535,0,0,0,0,0,0,0,0,0,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535,65535]
13306 ; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [1,3,4,6,0,0,0,0]
13311 ; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [1,3,5,6,0,0,0,0]
13316 ; AVX512DQ-FCP-NEXT: movw $-512, %ax # imm = 0xFE00
13320 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm2[2,3,0,1]
13321 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0,1],ymm0[2],ymm2[3,4],ymm0[5],ymm2[6,7,8,9],ymm0[10],ymm2[11,12],ymm0[13],ymm2[14,15]
13325 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,u,u],zero,zero,xmm1[0,7,14],zero,zero,xmm1[3,10,u,u,u]
13331 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,u,u,0,7,14],zero,zero,xmm2[3,10],zero,zero,zero
13334 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7],ymm1[8,9,10],ymm0[11,12,13,14,15]
13335 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
13337 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
13343 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
13345 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
13352 ; AVX512BW: # %bb.0:
13354 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm0 = [8,1,18,11,4,5,22,15,0,25,10,0,12,29,14,0]
13356 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm0 = [0,1,18,11,4,21,14,7,8,25,10,0,28,13,0,15]
13358 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm0 = [0,17,10,3,4,21,14,7,24,9,0,11,28,13,0,31]
13360 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm0 = [16,17,10,3,20,13,6,23,24,25,0,27,28,0,30,31]
13364 ; AVX512BW-NEXT: movw $-28382, %ax # imm = 0x9122
13368 ; AVX512BW-NEXT: kmovw %k1, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
13371 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0,7,14],zero,zero,xmm3[3,10],zero,zero,zero,xmm3[u,u,u,u,u,u]
13374 ; AVX512BW-NEXT: movw $992, %ax # imm = 0x3E0
13379 ; AVX512BW-NEXT: movw $8772, %ax # imm = 0x2244
13392 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
13394 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm3[7]
13398 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm12 = xmm4[0,7,14],zero,zero,xmm4[u,u,u,u,u,u,u,u,u,u,u]
13401 ; AVX512BW-NEXT: movabsq $137438429184, %rax # imm = 0x1FFFF80000
13406 ; AVX512BW-NEXT: movw $9288, %ax # imm = 0x2448
13416 ; AVX512BW-NEXT: vpermq {{.*#+}} ymm15 = ymm5[2,3,0,1]
13417 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1],ymm15[2],ymm5[3,4,5],ymm15[6],ymm5[7,8,9],ymm15[10],ymm5[11,12,13],ymm15[14],ymm5[15]
13419 ; AVX512BW-NEXT: movw $3968, %ax # imm = 0xF80
13424 ; AVX512BW-NEXT: movw $4644, %ax # imm = 0x1224
13432 ; AVX512BW-NEXT: movl $-8388608, %eax # imm = 0xFF800000
13438 ; AVX512BW-NEXT: vpshufb {{.*#+}} ymm9 = ymm9[u,u,u,u,u,u,u,u,u,0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
13439 ; AVX512BW-NEXT: movl $511, %edi # imm = 0x1FF
13445 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm22 = xmm22[u,u,u],zero,zero,xmm22[4,11],zero,zero,xmm22[0,7,14,u,u,u,u]
13450 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm22 = xmm23[0],xmm22[0],xmm23[1],xmm22[1],xmm23[2],xmm22[2],xmm23[3],xmm22[3]
13452 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm14[0,1,2,3,4,5,6],ymm2[7]
13461 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm2[0,7,14],zero,zero,xmm2[3,10,u,u,u,u,u,u,u]
13463 ; AVX512BW-NEXT: movl $261632, %edi # imm = 0x3FE00
13469 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,0,7,14],zero,zero,xmm14[3,10],zero,zero,zero,xmm14[u,u,u,u]
13474 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm21 = xmm22[0],xmm21[0],xmm22[1],xmm21[1],xmm22[2],xmm21[2],xmm22[3],xmm21[3]
13476 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm14[0,1,2,3,4,5,6],ymm3[7]
13478 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm21 = zero,zero,xmm26[0,7,14,u,u,u,u,u,u,u,u,u,u,u]
13487 ; AVX512BW-NEXT: vpshufb {{.*#+}} ymm2 {%k5} = ymm18[u,u,u,u,u,u,u,u,u,2,9,0,7,14,5,12,19,26,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
13495 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm21 = xmm7[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
13498 ; AVX512BW-NEXT: movl $-134217728, %edi # imm = 0xF8000000
13500 ; AVX512BW-NEXT: kmovd %k2, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
13510 ; AVX512BW-NEXT: movabsq $-137438953472, %rax # imm = 0xFFFFFFE000000000
13516 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,u,4,11],zero,zero,xmm2[0,7,14],zero,zero,xmm2[u,u]
13519 ; AVX512BW-NEXT: vpermq {{.*#+}} ymm14 = ymm3[2,3,0,1]
13520 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm14[3],ymm3[4,5],ymm14[6],ymm3[7,8,9,10],ymm14[11],ymm3[12,13],ymm14[14],ymm3[15]
13526 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u,u,u,u,u,u],zero,zero,xmm3[0,7,14],zero,zero,xmm3[3,10]
13538 ; AVX512BW-NEXT: vpermq {{.*#+}} ymm14 = ymm3[2,3,0,1]
13539 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm3 = ymm14[0],ymm3[1,2],ymm14[3],ymm3[4,5,6],ymm14[7,8],ymm3[9,10],ymm14[11],ymm3[12,13,14],ymm14[15]
13554 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,u],zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14,u,u]
13557 ; AVX512BW-NEXT: vpermq {{.*#+}} ymm14 = ymm3[2,3,0,1]
13558 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm3 = ymm14[0],ymm3[1,2,3],ymm14[4],ymm3[5,6],ymm14[7,8],ymm3[9,10,11],ymm14[12],ymm3[13,14],ymm14[15]
13564 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u,u,u,u,u,u,4,11],zero,zero,xmm3[0,7,14],zero,zero
13570 ; AVX512BW-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 # 2-byte Reload
13574 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,0,7,14],zero,zero,xmm2[3,10],zero,zero,zero,xmm2[u,u]
13577 ; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm2[2,3,0,1]
13578 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2,3],ymm3[4],ymm2[5,6,7,8],ymm3[9],ymm2[10,11],ymm3[12],ymm2[13,14,15]
13588 ; AVX512BW-NEXT: vpermq {{.*#+}} ymm14 = ymm3[2,3,0,1]
13589 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm14[1],ymm3[2,3,4],ymm14[5],ymm3[6,7,8],ymm14[9],ymm3[10,11,12],ymm14[13],ymm3[14,15]
13590 ; AVX512BW-NEXT: vpshufb {{.*#+}} ymm20 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,0,7,14,21,28,19,26,17,24,31,u,u,u,u,u,u,u,u,u]
13591 ; AVX512BW-NEXT: movl $8176, %eax # imm = 0x1FF0
13604 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,u,u,u],zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14]
13611 ; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm0[2,3,0,1]
13612 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm3 = ymm0[0,1],ymm3[2],ymm0[3,4],ymm3[5],ymm0[6,7,8,9],ymm3[10],ymm0[11,12],ymm3[13],ymm0[14,15]
13616 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[u,u,u,u],zero,zero,xmm12[0,7,14],zero,zero,xmm12[3,10,u,u,u]
13618 ; AVX512BW-NEXT: movl $4186112, %eax # imm = 0x3FE000
13628 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u],zero,zero,xmm2[0,7,14],zero,zero,xmm2[3,10,u,u,u,u,u]
13635 ; AVX512BW-NEXT: kmovd {{[-0-9]+}}(%r{{[sb]}}p), %k1 # 4-byte Reload
13643 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm11 = xmm8[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm8[0,7,14]
13649 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u,4,11],zero,zero,xmm6[0,7,14],zero,zero,xmm6[u,u,u,u,u]
13657 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm7 = [16,9,2,19,20,13,6,23,24,0,26,27,28,0,30,31]
13659 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm8 = [16,9,2,19,12,5,22,23,24,0,26,27,0,29,30,31]
13661 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm10 = [8,1,2,19,12,5,22,15,0,9,26,11,0,29,14,0]
13665 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[4,11],zero,zero,xmm12[0,7,14],zero,zero,xmm12[u,u,u,u,u,u,u]
13671 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
13675 ; AVX512BW-NEXT: movw $-512, %ax # imm = 0xFE00
13680 ; AVX512BW-NEXT: vpshufb {{.*#+}} ymm12 {%k5} = ymm8[u,u,u,u,u,u,u,u,u,4,11,2,9,0,7,14,21,28,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
13683 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm13[0],xmm8[0],xmm13[1],xmm8[1],xmm13[2],xmm8[2],xmm13[3],xmm8[3]
13688 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,xmm1[4,11],zero,zero,xmm1[0,7,14,u,u,u,u,u,u,u]
13693 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
13702 ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm3 = xmm5[u,u,u,u,u,u,0,7,14],zero,zero,xmm5[3,10],zero,zero,zero
13705 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm2 = ymm0[0,1,2],ymm2[3,4,5,6,7],ymm0[8,9,10],ymm2[11,12,13,14,15]
13706 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
13709 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
13710 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rdi
13722 ; AVX512BW-FCP: # %bb.0:
13727 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [8,1,18,11,4,5,22,15,0,25,10,3,12,29,14,7]
13731 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [0,1,18,11,4,21,14,7,8,25,10,3,28,13,6,15]
13733 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [0,1,2,11,4,5,14,7,8,9,26,19,12,29,22,15]
13735 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [0,17,10,3,4,21,14,7,24,9,2,11,28,13,6,31]
13737 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [0,1,10,3,4,5,14,7,8,25,18,11,12,29,22,15]
13743 ; AVX512BW-FCP-NEXT: movw $-28382, %ax # imm = 0x9122
13747 ; AVX512BW-FCP-NEXT: kmovw %k1, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
13750 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[0,7,14],zero,zero,xmm1[3,10],zero,zero,zero,xmm1[u,u,u,u,u,u]
13753 ; AVX512BW-FCP-NEXT: movw $992, %ax # imm = 0x3E0
13758 ; AVX512BW-FCP-NEXT: movw $8772, %ax # imm = 0x2244
13762 ; AVX512BW-FCP-NEXT: kmovw %k1, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
13768 ; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,0,0,0,1,2,4,6]
13772 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5,6],ymm6[7]
13776 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm20[0,7,14],zero,zero,xmm20[u,u,u,u,u,u,u,u,u,u,u]
13779 ; AVX512BW-FCP-NEXT: movabsq $137438429184, %rax # imm = 0x1FFFF80000
13784 ; AVX512BW-FCP-NEXT: movw $9288, %ax # imm = 0x2448
13792 ; AVX512BW-FCP-NEXT: movw $3968, %ax # imm = 0xF80
13797 ; AVX512BW-FCP-NEXT: movw $4644, %ax # imm = 0x1224
13805 ; AVX512BW-FCP-NEXT: movl $-8388608, %eax # imm = 0xFF800000
13811 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[u,u,u,u,u,u,u,u,u,0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
13812 ; AVX512BW-FCP-NEXT: movl $511, %r10d # imm = 0x1FF
13818 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm18 = xmm18[u,u,u],zero,zero,xmm18[4,11],zero,zero,xmm18[0,7,14,u,u,u,u]
13821 ; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm18 = [0,0,0,0,1,3,4,6]
13824 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5,6],ymm15[7]
13833 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm14 = zero,zero,xmm14[0,7,14],zero,zero,xmm14[3,10,u,u,u,u,u,u,u]
13835 ; AVX512BW-FCP-NEXT: movl $261632, %r10d # imm = 0x3FE00
13841 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[u,u,0,7,14],zero,zero,xmm12[3,10],zero,zero,zero,xmm12[u,u,u,u]
13844 ; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm15 = [0,0,0,0,1,3,5,6]
13847 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5,6],ymm15[7]
13849 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm17 = zero,zero,xmm19[0,7,14,u,u,u,u,u,u,u,u,u,u,u]
13858 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm14 {%k5} = ymm13[u,u,u,u,u,u,u,u,u,2,9,0,7,14,5,12,19,26,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
13868 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm23 = xmm18[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
13871 ; AVX512BW-FCP-NEXT: movl $-134217728, %r10d # imm = 0xF8000000
13873 ; AVX512BW-FCP-NEXT: kmovd %k2, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
13883 ; AVX512BW-FCP-NEXT: movabsq $-137438953472, %rax # imm = 0xFFFFFFE000000000
13889 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,u,u,u,4,11],zero,zero,xmm14[0,7,14],zero,zero,xmm14[u,u]
13893 ; AVX512BW-FCP-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 # 2-byte Reload
13897 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[u,u,u,u,u,u,u],zero,zero,xmm15[0,7,14],zero,zero,xmm15[3,10]
13922 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,u,u,u],zero,zero,xmm14[4,11],zero,zero,xmm14[0,7,14,u,u]
13929 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[u,u,u,u,u,u,u,4,11],zero,zero,xmm15[0,7,14],zero,zero
13935 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm14 = [0,9,2,3,4,13,6,7,24,17,10,11,28,21,14,31]
13939 ; AVX512BW-FCP-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k2 # 2-byte Reload
13943 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,u,u,0,7,14],zero,zero,xmm14[3,10],zero,zero,zero,xmm14[u,u]
13952 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,0,7,14,21,28,19,26,17,24,31,u,u,u,u,u,u,u,u,u]
13953 ; AVX512BW-FCP-NEXT: movl $8176, %eax # imm = 0x1FF0
13956 ; AVX512BW-FCP-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 # 2-byte Reload
13967 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm16 = xmm16[u,u,u,u,u,u,u],zero,zero,xmm16[4,11],zero,zero,xmm16[0,7,14]
13979 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[u,u],zero,zero,xmm10[0,7,14],zero,zero,xmm10[3,10,u,u,u,u,u]
13986 ; AVX512BW-FCP-NEXT: kmovd {{[-0-9]+}}(%r{{[sb]}}p), %k3 # 4-byte Reload
13994 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm21 = xmm17[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm17[0,7,14]
14000 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[u,u,4,11],zero,zero,xmm9[0,7,14],zero,zero,xmm9[u,u,u,u,u]
14012 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm20 = [8,1,2,19,12,5,22,15,0,9,26,11,4,29,14,7]
14016 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm19 = xmm19[4,11],zero,zero,xmm19[0,7,14],zero,zero,xmm19[u,u,u,u,u,u,u]
14019 ; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm19 = [1,2,4,6,0,0,0,0]
14022 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm19 = xmm19[0,7,10,13,u,u,u,u,u,u,u,u,u,u,u,u]
14029 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm10 {%k5} = ymm18[u,u,u,u,u,u,u,u,u,4,11,2,9,0,7,14,21,28,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
14030 ; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm16 = [1,3,4,6,0,0,0,0]
14037 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm3 = zero,zero,xmm3[4,11],zero,zero,xmm3[0,7,14,u,u,u,u,u,u,u]
14040 ; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [1,3,5,6,0,0,0,0]
14045 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm9 = [0,1,10,3,4,13,6,7,8,25,18,11,28,21,14,15]
14050 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[u,u,u,u],zero,zero,xmm5[0,7,14],zero,zero,xmm5[3,10,u,u,u]
14052 ; AVX512BW-FCP-NEXT: movl $4186112, %eax # imm = 0x3FE000
14056 ; AVX512BW-FCP-NEXT: movw $-512, %ax # imm = 0xFE00
14064 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,u,u,u,u,u,0,7,14],zero,zero,xmm4[3,10],zero,zero,zero
14067 ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm5[0,1,2],ymm0[3,4,5,6,7],ymm5[8,9,10],ymm0[11,12,13,14,15]
14068 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm0[4,5,6,7]
14071 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
14072 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rdi
14084 ; AVX512DQ-BW: # %bb.0:
14086 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm0 = [8,1,18,11,4,5,22,15,0,25,10,0,12,29,14,0]
14088 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm0 = [0,1,18,11,4,21,14,7,8,25,10,0,28,13,0,15]
14090 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm0 = [0,17,10,3,4,21,14,7,24,9,0,11,28,13,0,31]
14092 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm0 = [16,17,10,3,20,13,6,23,24,25,0,27,28,0,30,31]
14096 ; AVX512DQ-BW-NEXT: movw $-28382, %ax # imm = 0x9122
14100 ; AVX512DQ-BW-NEXT: kmovw %k1, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
14103 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0,7,14],zero,zero,xmm3[3,10],zero,zero,zero,xmm3[u,u,u,u,u,u]
14106 ; AVX512DQ-BW-NEXT: movw $992, %ax # imm = 0x3E0
14111 ; AVX512DQ-BW-NEXT: movw $8772, %ax # imm = 0x2244
14124 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
14126 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm3[7]
14130 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm12 = xmm4[0,7,14],zero,zero,xmm4[u,u,u,u,u,u,u,u,u,u,u]
14133 ; AVX512DQ-BW-NEXT: movabsq $137438429184, %rax # imm = 0x1FFFF80000
14138 ; AVX512DQ-BW-NEXT: movw $9288, %ax # imm = 0x2448
14148 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm15 = ymm5[2,3,0,1]
14149 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1],ymm15[2],ymm5[3,4,5],ymm15[6],ymm5[7,8,9],ymm15[10],ymm5[11,12,13],ymm15[14],ymm5[15]
14151 ; AVX512DQ-BW-NEXT: movw $3968, %ax # imm = 0xF80
14156 ; AVX512DQ-BW-NEXT: movw $4644, %ax # imm = 0x1224
14164 ; AVX512DQ-BW-NEXT: movl $-8388608, %eax # imm = 0xFF800000
14170 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm10 = ymm10[u,u,u,u,u,u,u,u,u,0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
14171 ; AVX512DQ-BW-NEXT: movl $511, %edi # imm = 0x1FF
14177 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm22 = xmm22[u,u,u],zero,zero,xmm22[4,11],zero,zero,xmm22[0,7,14,u,u,u,u]
14182 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm22 = xmm23[0],xmm22[0],xmm23[1],xmm22[1],xmm23[2],xmm22[2],xmm23[3],xmm22[3]
14184 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm14[0,1,2,3,4,5,6],ymm2[7]
14193 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm2[0,7,14],zero,zero,xmm2[3,10,u,u,u,u,u,u,u]
14195 ; AVX512DQ-BW-NEXT: movl $261632, %edi # imm = 0x3FE00
14201 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,0,7,14],zero,zero,xmm14[3,10],zero,zero,zero,xmm14[u,u,u,u]
14206 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm21 = xmm22[0],xmm21[0],xmm22[1],xmm21[1],xmm22[2],xmm21[2],xmm22[3],xmm21[3]
14208 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm14[0,1,2,3,4,5,6],ymm3[7]
14210 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm21 = zero,zero,xmm26[0,7,14,u,u,u,u,u,u,u,u,u,u,u]
14219 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm2 {%k5} = ymm18[u,u,u,u,u,u,u,u,u,2,9,0,7,14,5,12,19,26,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
14227 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm21 = xmm7[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
14230 ; AVX512DQ-BW-NEXT: movl $-134217728, %edi # imm = 0xF8000000
14232 ; AVX512DQ-BW-NEXT: kmovd %k2, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
14242 ; AVX512DQ-BW-NEXT: movabsq $-137438953472, %rax # imm = 0xFFFFFFE000000000
14248 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,u,4,11],zero,zero,xmm2[0,7,14],zero,zero,xmm2[u,u]
14251 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm14 = ymm3[2,3,0,1]
14252 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm14[3],ymm3[4,5],ymm14[6],ymm3[7,8,9,10],ymm14[11],ymm3[12,13],ymm14[14],ymm3[15]
14258 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u,u,u,u,u,u],zero,zero,xmm3[0,7,14],zero,zero,xmm3[3,10]
14270 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm14 = ymm3[2,3,0,1]
14271 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm3 = ymm14[0],ymm3[1,2],ymm14[3],ymm3[4,5,6],ymm14[7,8],ymm3[9,10],ymm14[11],ymm3[12,13,14],ymm14[15]
14286 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,u],zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14,u,u]
14289 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm14 = ymm3[2,3,0,1]
14290 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm3 = ymm14[0],ymm3[1,2,3],ymm14[4],ymm3[5,6],ymm14[7,8],ymm3[9,10,11],ymm14[12],ymm3[13,14],ymm14[15]
14296 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u,u,u,u,u,u,4,11],zero,zero,xmm3[0,7,14],zero,zero
14302 ; AVX512DQ-BW-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 # 2-byte Reload
14306 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,0,7,14],zero,zero,xmm2[3,10],zero,zero,zero,xmm2[u,u]
14309 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm3 = ymm2[2,3,0,1]
14310 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2,3],ymm3[4],ymm2[5,6,7,8],ymm3[9],ymm2[10,11],ymm3[12],ymm2[13,14,15]
14320 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm14 = ymm3[2,3,0,1]
14321 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm14[1],ymm3[2,3,4],ymm14[5],ymm3[6,7,8],ymm14[9],ymm3[10,11,12],ymm14[13],ymm3[14,15]
14322 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm20 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,0,7,14,21,28,19,26,17,24,31,u,u,u,u,u,u,u,u,u]
14323 ; AVX512DQ-BW-NEXT: movl $8176, %eax # imm = 0x1FF0
14336 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,u,u,u],zero,zero,xmm2[4,11],zero,zero,xmm2[0,7,14]
14343 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm2 = ymm0[2,3,0,1]
14344 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm2 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7,8,9],ymm2[10],ymm0[11,12],ymm2[13],ymm0[14,15]
14348 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u,u,u],zero,zero,xmm3[0,7,14],zero,zero,xmm3[3,10,u,u,u]
14350 ; AVX512DQ-BW-NEXT: movl $4186112, %eax # imm = 0x3FE000
14360 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[u,u],zero,zero,xmm11[0,7,14],zero,zero,xmm11[3,10,u,u,u,u,u]
14367 ; AVX512DQ-BW-NEXT: kmovd {{[-0-9]+}}(%r{{[sb]}}p), %k1 # 4-byte Reload
14375 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm12 = xmm8[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm8[0,7,14]
14381 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u,4,11],zero,zero,xmm6[0,7,14],zero,zero,xmm6[u,u,u,u,u]
14389 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm7 = [16,9,2,19,20,13,6,23,24,0,26,27,28,0,30,31]
14391 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm8 = [16,9,2,19,12,5,22,23,24,0,26,27,0,29,30,31]
14393 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm11 = [8,1,2,19,12,5,22,15,0,9,26,11,0,29,14,0]
14397 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[4,11],zero,zero,xmm2[0,7,14],zero,zero,xmm2[u,u,u,u,u,u,u]
14403 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
14406 ; AVX512DQ-BW-NEXT: movw $-512, %ax # imm = 0xFE00
14411 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm9 {%k5} = ymm8[u,u,u,u,u,u,u,u,u,4,11,2,9,0,7,14,21,28,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
14414 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm12[0],xmm8[0],xmm12[1],xmm8[1],xmm12[2],xmm8[2],xmm12[3],xmm8[3]
14419 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,xmm1[4,11],zero,zero,xmm1[0,7,14,u,u,u,u,u,u,u]
14424 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
14432 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm4 = xmm5[u,u,u,u,u,u,0,7,14],zero,zero,xmm5[3,10],zero,zero,zero
14435 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm3 = ymm0[0,1,2],ymm3[3,4,5,6,7],ymm0[8,9,10],ymm3[11,12,13,14,15]
14436 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
14438 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
14439 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rdi
14451 ; AVX512DQ-BW-FCP: # %bb.0:
14456 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [8,1,18,11,4,5,22,15,0,25,10,3,12,29,14,7]
14460 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [0,1,18,11,4,21,14,7,8,25,10,3,28,13,6,15]
14462 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [0,1,2,11,4,5,14,7,8,9,26,19,12,29,22,15]
14464 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [0,17,10,3,4,21,14,7,24,9,2,11,28,13,6,31]
14466 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [0,1,10,3,4,5,14,7,8,25,18,11,12,29,22,15]
14472 ; AVX512DQ-BW-FCP-NEXT: movw $-28382, %ax # imm = 0x9122
14476 ; AVX512DQ-BW-FCP-NEXT: kmovw %k1, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
14479 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[0,7,14],zero,zero,xmm1[3,10],zero,zero,zero,xmm1[u,u,u,u,u,u]
14482 ; AVX512DQ-BW-FCP-NEXT: movw $992, %ax # imm = 0x3E0
14487 ; AVX512DQ-BW-FCP-NEXT: movw $8772, %ax # imm = 0x2244
14491 ; AVX512DQ-BW-FCP-NEXT: kmovw %k1, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
14497 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,0,0,0,1,2,4,6]
14501 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5,6],ymm6[7]
14505 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm20[0,7,14],zero,zero,xmm20[u,u,u,u,u,u,u,u,u,u,u]
14508 ; AVX512DQ-BW-FCP-NEXT: movabsq $137438429184, %rax # imm = 0x1FFFF80000
14513 ; AVX512DQ-BW-FCP-NEXT: movw $9288, %ax # imm = 0x2448
14521 ; AVX512DQ-BW-FCP-NEXT: movw $3968, %ax # imm = 0xF80
14526 ; AVX512DQ-BW-FCP-NEXT: movw $4644, %ax # imm = 0x1224
14534 ; AVX512DQ-BW-FCP-NEXT: movl $-8388608, %eax # imm = 0xFF800000
14540 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[u,u,u,u,u,u,u,u,u,0,7,14,5,12,3,10,17,24,31,u,u,u,u,u,u,u,u,u,u,u,u,u]
14541 ; AVX512DQ-BW-FCP-NEXT: movl $511, %r10d # imm = 0x1FF
14547 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm18 = xmm18[u,u,u],zero,zero,xmm18[4,11],zero,zero,xmm18[0,7,14,u,u,u,u]
14550 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm18 = [0,0,0,0,1,3,4,6]
14553 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5,6],ymm15[7]
14562 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm14 = zero,zero,xmm14[0,7,14],zero,zero,xmm14[3,10,u,u,u,u,u,u,u]
14564 ; AVX512DQ-BW-FCP-NEXT: movl $261632, %r10d # imm = 0x3FE00
14570 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[u,u,0,7,14],zero,zero,xmm12[3,10],zero,zero,zero,xmm12[u,u,u,u]
14573 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm15 = [0,0,0,0,1,3,5,6]
14576 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5,6],ymm15[7]
14578 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm17 = zero,zero,xmm19[0,7,14,u,u,u,u,u,u,u,u,u,u,u]
14587 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm14 {%k5} = ymm13[u,u,u,u,u,u,u,u,u,2,9,0,7,14,5,12,19,26,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
14597 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm23 = xmm18[u,u,u,u,u,u,u,u,u,u,u,0,7,14],zero,zero
14600 ; AVX512DQ-BW-FCP-NEXT: movl $-134217728, %r10d # imm = 0xF8000000
14602 ; AVX512DQ-BW-FCP-NEXT: kmovd %k2, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
14612 ; AVX512DQ-BW-FCP-NEXT: movabsq $-137438953472, %rax # imm = 0xFFFFFFE000000000
14618 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,u,u,u,4,11],zero,zero,xmm14[0,7,14],zero,zero,xmm14[u,u]
14622 ; AVX512DQ-BW-FCP-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 # 2-byte Reload
14626 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[u,u,u,u,u,u,u],zero,zero,xmm15[0,7,14],zero,zero,xmm15[3,10]
14651 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,u,u,u],zero,zero,xmm14[4,11],zero,zero,xmm14[0,7,14,u,u]
14658 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[u,u,u,u,u,u,u,4,11],zero,zero,xmm15[0,7,14],zero,zero
14664 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm14 = [0,9,2,3,4,13,6,7,24,17,10,11,28,21,14,31]
14668 ; AVX512DQ-BW-FCP-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 # 2-byte Reload
14672 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,u,u,0,7,14],zero,zero,xmm14[3,10],zero,zero,zero,xmm14[u,u]
14681 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,0,7,14,21,28,19,26,17,24,31,u,u,u,u,u,u,u,u,u]
14682 ; AVX512DQ-BW-FCP-NEXT: movl $8176, %eax # imm = 0x1FF0
14685 ; AVX512DQ-BW-FCP-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 # 2-byte Reload
14696 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm16 = xmm16[u,u,u,u,u,u,u],zero,zero,xmm16[4,11],zero,zero,xmm16[0,7,14]
14708 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[u,u],zero,zero,xmm10[0,7,14],zero,zero,xmm10[3,10,u,u,u,u,u]
14715 ; AVX512DQ-BW-FCP-NEXT: kmovd {{[-0-9]+}}(%r{{[sb]}}p), %k2 # 4-byte Reload
14723 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm21 = xmm17[u,u,u,u,u,u,u,u,u,u,u],zero,zero,xmm17[0,7,14]
14729 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[u,u,4,11],zero,zero,xmm9[0,7,14],zero,zero,xmm9[u,u,u,u,u]
14741 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm20 = [8,1,2,19,12,5,22,15,0,9,26,11,4,29,14,7]
14745 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm19 = xmm19[4,11],zero,zero,xmm19[0,7,14],zero,zero,xmm19[u,u,u,u,u,u,u]
14748 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm19 = [1,2,4,6,0,0,0,0]
14751 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm19 = xmm19[0,7,10,13,u,u,u,u,u,u,u,u,u,u,u,u]
14758 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm10 {%k5} = ymm18[u,u,u,u,u,u,u,u,u,4,11,2,9,0,7,14,21,28,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
14759 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm16 = [1,3,4,6,0,0,0,0]
14766 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm3 = zero,zero,xmm3[4,11],zero,zero,xmm3[0,7,14,u,u,u,u,u,u,u]
14769 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [1,3,5,6,0,0,0,0]
14774 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm9 = [0,1,10,3,4,13,6,7,8,25,18,11,28,21,14,15]
14779 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,u,u,u],zero,zero,xmm4[0,7,14],zero,zero,xmm4[3,10,u,u,u]
14781 ; AVX512DQ-BW-FCP-NEXT: movl $4186112, %eax # imm = 0x3FE000
14784 ; AVX512DQ-BW-FCP-NEXT: movw $-512, %ax # imm = 0xFE00
14785 ; AVX512DQ-BW-FCP-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 # 2-byte Reload
14792 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[u,u,u,u,u,u,0,7,14],zero,zero,xmm5[3,10],zero,zero,zero
14795 ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm4[0,1,2],ymm0[3,4,5,6,7],ymm4[8,9,10],ymm0[11,12,13,14,15]
14796 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
14798 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
14799 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rdi
14810 %strided.vec0 = shufflevector <448 x i8> %wide.vec, <448 x i8> poison, <64 x i32> <i32 0, i32 7, i32 14, i32 21, i32 28, i32 35, i32 42, i32 49, i32 56, i32 63, i32 70, i32 77, i32 84, i32 91, i32 98, i32 105, i32 112, i32 119, i32 126, i32 133, i32 140, i32 147, i32 154, i32 161, i32 168, i32 175, i32 182, i32 189, i32 196, i32 203, i32 210, i32 217, i32 224, i32 231, i32 238, i32 245, i32 252, i32 259, i32 266, i32 273, i32 280, i32 287, i32 294, i32 301, i32 308, i32 315, i32 322, i32 329, i32 336, i32 343, i32 350, i32 357, i32 364, i32 371, i32 378, i32 385, i32 392, i32 399, i32 406, i32 413, i32 420, i32 427, i32 434, i32 441>