| ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py |
| ; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE,FALLBACK0 |
| ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-ONLY,FALLBACK1 |
| ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX1,AVX2,AVX2-ONLY,AVX2-SLOW,FALLBACK2 |
| ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX1,AVX2,AVX2-ONLY,AVX2-FAST,FALLBACK3 |
| ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX1,AVX2,AVX2-ONLY,AVX2-FAST-PERLANE,FALLBACK4 |
| ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-SLOW,AVX512F-ONLY-SLOW,FALLBACK5 |
| ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-FAST,AVX512F-ONLY-FAST,FALLBACK6 |
| ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-SLOW,AVX512DQ-SLOW,FALLBACK7 |
| ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-FAST,AVX512DQ-FAST,FALLBACK8 |
| ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-SLOW,AVX512BW-ONLY-SLOW,FALLBACK9 |
| ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-FAST,AVX512BW-ONLY-FAST,FALLBACK10 |
| ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-SLOW,AVX512DQBW-SLOW,FALLBACK11 |
| ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-FAST,AVX512DQBW-FAST,FALLBACK12 |
| |
| ; These patterns are produced by LoopVectorizer for interleaved stores. |
| |
| define void @store_i32_stride7_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %out.vec) nounwind { |
| ; SSE-LABEL: store_i32_stride7_vf2: |
| ; SSE: # %bb.0: |
| ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero |
| ; SSE-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero |
| ; SSE-NEXT: movsd {{.*#+}} xmm2 = mem[0],zero |
| ; SSE-NEXT: movsd {{.*#+}} xmm3 = mem[0],zero |
| ; SSE-NEXT: movq {{.*#+}} xmm4 = mem[0],zero |
| ; SSE-NEXT: movsd {{.*#+}} xmm5 = mem[0],zero |
| ; SSE-NEXT: movsd {{.*#+}} xmm6 = mem[0],zero |
| ; SSE-NEXT: movaps %xmm6, %xmm7 |
| ; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm5[0] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm4[1,1,1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,1],xmm0[1,3] |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0] |
| ; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[1,1],xmm3[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,3],xmm6[2,0] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm7[3,3,3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm7 = xmm7[0,2],xmm4[0,2] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1] |
| ; SSE-NEXT: movaps %xmm1, 32(%rax) |
| ; SSE-NEXT: movaps %xmm7, 16(%rax) |
| ; SSE-NEXT: movaps %xmm0, (%rax) |
| ; SSE-NEXT: movq %xmm2, 48(%rax) |
| ; SSE-NEXT: retq |
| ; |
| ; AVX1-ONLY-LABEL: store_i32_stride7_vf2: |
| ; AVX1-ONLY: # %bb.0: |
| ; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX1-ONLY-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero |
| ; AVX1-ONLY-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero |
| ; AVX1-ONLY-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero |
| ; AVX1-ONLY-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero |
| ; AVX1-ONLY-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] |
| ; AVX1-ONLY-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero |
| ; AVX1-ONLY-NEXT: vmovq {{.*#+}} xmm5 = mem[0],zero |
| ; AVX1-ONLY-NEXT: vpunpcklqdq {{.*#+}} xmm4 = xmm5[0],xmm4[0] |
| ; AVX1-ONLY-NEXT: vmovsd {{.*#+}} xmm5 = mem[0],zero |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm0[2,3,0,1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm5 = ymm1[3,0],ymm0[1,0],ymm1[7,4],ymm0[5,4] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm5 = ymm5[0,2],ymm0[2,1],ymm5[4,6],ymm0[6,5] |
| ; AVX1-ONLY-NEXT: vpalignr {{.*#+}} xmm6 = xmm2[12,13,14,15],xmm4[0,1,2,3,4,5,6,7,8,9,10,11] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,2],xmm4[3,3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} xmm6 = xmm6[0,1,2],xmm5[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm1[0,2],ymm0[0,3],ymm1[4,6],ymm0[4,7] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm3, %ymm4, %ymm1 |
| ; AVX1-ONLY-NEXT: vpermilps {{.*#+}} ymm1 = ymm1[u,u,0,2,u,u,u,5] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,2],ymm1[2,3],ymm2[4,6],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm5, %xmm0 |
| ; AVX1-ONLY-NEXT: vmovlps %xmm0, 48(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %xmm6, 32(%rax) |
| ; AVX1-ONLY-NEXT: vzeroupper |
| ; AVX1-ONLY-NEXT: retq |
| ; |
| ; AVX2-SLOW-LABEL: store_i32_stride7_vf2: |
| ; AVX2-SLOW: # %bb.0: |
| ; AVX2-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX2-SLOW-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero |
| ; AVX2-SLOW-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero |
| ; AVX2-SLOW-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero |
| ; AVX2-SLOW-NEXT: vmovsd {{.*#+}} xmm3 = mem[0],zero |
| ; AVX2-SLOW-NEXT: vmovlhps {{.*#+}} xmm2 = xmm3[0],xmm2[0] |
| ; AVX2-SLOW-NEXT: vmovsd {{.*#+}} xmm3 = mem[0],zero |
| ; AVX2-SLOW-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero |
| ; AVX2-SLOW-NEXT: vmovlhps {{.*#+}} xmm3 = xmm4[0],xmm3[0] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2 |
| ; AVX2-SLOW-NEXT: vmovsd {{.*#+}} xmm3 = mem[0],zero |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,2,2,1,4,6,6,5] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,2,2,1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm3 = ymm0[0,2,2,3,4,6,6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovaps {{.*#+}} xmm3 = <3,5,7,u> |
| ; AVX2-SLOW-NEXT: vpermps %ymm2, %ymm3, %ymm2 |
| ; AVX2-SLOW-NEXT: vbroadcastf128 {{.*#+}} ymm3 = [3,5,0,1,3,5,0,1] |
| ; AVX2-SLOW-NEXT: # ymm3 = mem[0,1,0,1] |
| ; AVX2-SLOW-NEXT: vpermps %ymm0, %ymm3, %ymm0 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1,2],xmm0[3] |
| ; AVX2-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm0 |
| ; AVX2-SLOW-NEXT: vmovlps %xmm0, 48(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, 32(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm1, (%rax) |
| ; AVX2-SLOW-NEXT: vzeroupper |
| ; AVX2-SLOW-NEXT: retq |
| ; |
| ; AVX2-FAST-LABEL: store_i32_stride7_vf2: |
| ; AVX2-FAST: # %bb.0: |
| ; AVX2-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX2-FAST-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero |
| ; AVX2-FAST-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero |
| ; AVX2-FAST-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero |
| ; AVX2-FAST-NEXT: vmovsd {{.*#+}} xmm3 = mem[0],zero |
| ; AVX2-FAST-NEXT: vmovlhps {{.*#+}} xmm2 = xmm3[0],xmm2[0] |
| ; AVX2-FAST-NEXT: vmovsd {{.*#+}} xmm3 = mem[0],zero |
| ; AVX2-FAST-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero |
| ; AVX2-FAST-NEXT: vmovlhps {{.*#+}} xmm3 = xmm4[0],xmm3[0] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2 |
| ; AVX2-FAST-NEXT: vmovsd {{.*#+}} xmm3 = mem[0],zero |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3] |
| ; AVX2-FAST-NEXT: vmovaps {{.*#+}} xmm1 = <3,5,7,u> |
| ; AVX2-FAST-NEXT: vpermps %ymm2, %ymm1, %ymm1 |
| ; AVX2-FAST-NEXT: vbroadcastf128 {{.*#+}} ymm3 = [3,5,0,1,3,5,0,1] |
| ; AVX2-FAST-NEXT: # ymm3 = mem[0,1,0,1] |
| ; AVX2-FAST-NEXT: vpermps %ymm0, %ymm3, %ymm3 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1,2],xmm3[3] |
| ; AVX2-FAST-NEXT: vmovaps {{.*#+}} ymm4 = <0,2,4,6,u,u,u,1> |
| ; AVX2-FAST-NEXT: vpermps %ymm2, %ymm4, %ymm2 |
| ; AVX2-FAST-NEXT: vbroadcastf128 {{.*#+}} ymm4 = [0,2,4,0,0,2,4,0] |
| ; AVX2-FAST-NEXT: # ymm4 = mem[0,1,0,1] |
| ; AVX2-FAST-NEXT: vpermps %ymm0, %ymm4, %ymm0 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6],ymm2[7] |
| ; AVX2-FAST-NEXT: vextractf128 $1, %ymm3, %xmm2 |
| ; AVX2-FAST-NEXT: vmovlps %xmm2, 48(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, 32(%rax) |
| ; AVX2-FAST-NEXT: vzeroupper |
| ; AVX2-FAST-NEXT: retq |
| ; |
| ; AVX2-FAST-PERLANE-LABEL: store_i32_stride7_vf2: |
| ; AVX2-FAST-PERLANE: # %bb.0: |
| ; AVX2-FAST-PERLANE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-PERLANE-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX2-FAST-PERLANE-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero |
| ; AVX2-FAST-PERLANE-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero |
| ; AVX2-FAST-PERLANE-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero |
| ; AVX2-FAST-PERLANE-NEXT: vmovsd {{.*#+}} xmm3 = mem[0],zero |
| ; AVX2-FAST-PERLANE-NEXT: vmovlhps {{.*#+}} xmm2 = xmm3[0],xmm2[0] |
| ; AVX2-FAST-PERLANE-NEXT: vmovsd {{.*#+}} xmm3 = mem[0],zero |
| ; AVX2-FAST-PERLANE-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero |
| ; AVX2-FAST-PERLANE-NEXT: vmovlhps {{.*#+}} xmm3 = xmm4[0],xmm3[0] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovsd {{.*#+}} xmm3 = mem[0],zero |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,2,2,1,4,6,6,5] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,2,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm3 = ymm0[0,2,2,3,4,6,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,0,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{.*#+}} xmm3 = <3,5,7,u> |
| ; AVX2-FAST-PERLANE-NEXT: vpermps %ymm2, %ymm3, %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastf128 {{.*#+}} ymm3 = [3,5,0,1,3,5,0,1] |
| ; AVX2-FAST-PERLANE-NEXT: # ymm3 = mem[0,1,0,1] |
| ; AVX2-FAST-PERLANE-NEXT: vpermps %ymm0, %ymm3, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1,2],xmm0[3] |
| ; AVX2-FAST-PERLANE-NEXT: vextractf128 $1, %ymm0, %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovlps %xmm0, 48(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, 32(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm1, (%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vzeroupper |
| ; AVX2-FAST-PERLANE-NEXT: retq |
| ; |
| ; AVX512F-SLOW-LABEL: store_i32_stride7_vf2: |
| ; AVX512F-SLOW: # %bb.0: |
| ; AVX512F-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512F-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX512F-SLOW-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero |
| ; AVX512F-SLOW-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero |
| ; AVX512F-SLOW-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero |
| ; AVX512F-SLOW-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero |
| ; AVX512F-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] |
| ; AVX512F-SLOW-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero |
| ; AVX512F-SLOW-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero |
| ; AVX512F-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] |
| ; AVX512F-SLOW-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm2 |
| ; AVX512F-SLOW-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero |
| ; AVX512F-SLOW-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 |
| ; AVX512F-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2] |
| ; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm1 = <0,2,4,6,16,20,18,1,3,5,7,17,21,19,u,u> |
| ; AVX512F-SLOW-NEXT: vpermi2d %zmm0, %zmm2, %zmm1 |
| ; AVX512F-SLOW-NEXT: vextracti32x4 $2, %zmm1, 32(%rax) |
| ; AVX512F-SLOW-NEXT: vextracti32x4 $3, %zmm1, %xmm0 |
| ; AVX512F-SLOW-NEXT: vmovq %xmm0, 48(%rax) |
| ; AVX512F-SLOW-NEXT: vmovdqa %ymm1, (%rax) |
| ; AVX512F-SLOW-NEXT: vzeroupper |
| ; AVX512F-SLOW-NEXT: retq |
| ; |
| ; AVX512F-FAST-LABEL: store_i32_stride7_vf2: |
| ; AVX512F-FAST: # %bb.0: |
| ; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX512F-FAST-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero |
| ; AVX512F-FAST-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero |
| ; AVX512F-FAST-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero |
| ; AVX512F-FAST-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero |
| ; AVX512F-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] |
| ; AVX512F-FAST-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero |
| ; AVX512F-FAST-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero |
| ; AVX512F-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] |
| ; AVX512F-FAST-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm2 |
| ; AVX512F-FAST-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero |
| ; AVX512F-FAST-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 |
| ; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <0,2,4,u> |
| ; AVX512F-FAST-NEXT: vpermi2q %ymm3, %ymm0, %ymm1 |
| ; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm0 |
| ; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = <0,2,4,6,8,10,12,1,3,5,7,9,11,13,u,u> |
| ; AVX512F-FAST-NEXT: vpermd %zmm0, %zmm1, %zmm0 |
| ; AVX512F-FAST-NEXT: vextracti32x4 $2, %zmm0, 32(%rax) |
| ; AVX512F-FAST-NEXT: vextracti32x4 $3, %zmm0, %xmm1 |
| ; AVX512F-FAST-NEXT: vmovq %xmm1, 48(%rax) |
| ; AVX512F-FAST-NEXT: vmovdqa %ymm0, (%rax) |
| ; AVX512F-FAST-NEXT: vzeroupper |
| ; AVX512F-FAST-NEXT: retq |
| ; |
| ; AVX512BW-SLOW-LABEL: store_i32_stride7_vf2: |
| ; AVX512BW-SLOW: # %bb.0: |
| ; AVX512BW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512BW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX512BW-SLOW-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero |
| ; AVX512BW-SLOW-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero |
| ; AVX512BW-SLOW-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero |
| ; AVX512BW-SLOW-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero |
| ; AVX512BW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] |
| ; AVX512BW-SLOW-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero |
| ; AVX512BW-SLOW-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero |
| ; AVX512BW-SLOW-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] |
| ; AVX512BW-SLOW-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm2 |
| ; AVX512BW-SLOW-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero |
| ; AVX512BW-SLOW-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 |
| ; AVX512BW-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[2] |
| ; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm1 = <0,2,4,6,16,20,18,1,3,5,7,17,21,19,u,u> |
| ; AVX512BW-SLOW-NEXT: vpermi2d %zmm0, %zmm2, %zmm1 |
| ; AVX512BW-SLOW-NEXT: vextracti32x4 $2, %zmm1, 32(%rax) |
| ; AVX512BW-SLOW-NEXT: vextracti32x4 $3, %zmm1, %xmm0 |
| ; AVX512BW-SLOW-NEXT: vmovq %xmm0, 48(%rax) |
| ; AVX512BW-SLOW-NEXT: vmovdqa %ymm1, (%rax) |
| ; AVX512BW-SLOW-NEXT: vzeroupper |
| ; AVX512BW-SLOW-NEXT: retq |
| ; |
| ; AVX512BW-FAST-LABEL: store_i32_stride7_vf2: |
| ; AVX512BW-FAST: # %bb.0: |
| ; AVX512BW-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512BW-FAST-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX512BW-FAST-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero |
| ; AVX512BW-FAST-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero |
| ; AVX512BW-FAST-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero |
| ; AVX512BW-FAST-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero |
| ; AVX512BW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] |
| ; AVX512BW-FAST-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero |
| ; AVX512BW-FAST-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero |
| ; AVX512BW-FAST-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] |
| ; AVX512BW-FAST-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm2 |
| ; AVX512BW-FAST-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero |
| ; AVX512BW-FAST-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 |
| ; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <0,2,4,u> |
| ; AVX512BW-FAST-NEXT: vpermi2q %ymm3, %ymm0, %ymm1 |
| ; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm0 |
| ; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = <0,2,4,6,8,10,12,1,3,5,7,9,11,13,u,u> |
| ; AVX512BW-FAST-NEXT: vpermd %zmm0, %zmm1, %zmm0 |
| ; AVX512BW-FAST-NEXT: vextracti32x4 $2, %zmm0, 32(%rax) |
| ; AVX512BW-FAST-NEXT: vextracti32x4 $3, %zmm0, %xmm1 |
| ; AVX512BW-FAST-NEXT: vmovq %xmm1, 48(%rax) |
| ; AVX512BW-FAST-NEXT: vmovdqa %ymm0, (%rax) |
| ; AVX512BW-FAST-NEXT: vzeroupper |
| ; AVX512BW-FAST-NEXT: retq |
| %in.vec0 = load <2 x i32>, ptr %in.vecptr0, align 64 |
| %in.vec1 = load <2 x i32>, ptr %in.vecptr1, align 64 |
| %in.vec2 = load <2 x i32>, ptr %in.vecptr2, align 64 |
| %in.vec3 = load <2 x i32>, ptr %in.vecptr3, align 64 |
| %in.vec4 = load <2 x i32>, ptr %in.vecptr4, align 64 |
| %in.vec5 = load <2 x i32>, ptr %in.vecptr5, align 64 |
| %in.vec6 = load <2 x i32>, ptr %in.vecptr6, align 64 |
| %1 = shufflevector <2 x i32> %in.vec0, <2 x i32> %in.vec1, <4 x i32> <i32 0, i32 1, i32 2, i32 3> |
| %2 = shufflevector <2 x i32> %in.vec2, <2 x i32> %in.vec3, <4 x i32> <i32 0, i32 1, i32 2, i32 3> |
| %3 = shufflevector <2 x i32> %in.vec4, <2 x i32> %in.vec5, <4 x i32> <i32 0, i32 1, i32 2, i32 3> |
| %4 = shufflevector <4 x i32> %1, <4 x i32> %2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7> |
| %5 = shufflevector <2 x i32> %in.vec6, <2 x i32> poison, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef> |
| %6 = shufflevector <4 x i32> %3, <4 x i32> %5, <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5> |
| %7 = shufflevector <6 x i32> %6, <6 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 undef, i32 undef> |
| %8 = shufflevector <8 x i32> %4, <8 x i32> %7, <14 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13> |
| %interleaved.vec = shufflevector <14 x i32> %8, <14 x i32> poison, <14 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13> |
| store <14 x i32> %interleaved.vec, ptr %out.vec, align 64 |
| ret void |
| } |
| |
| define void @store_i32_stride7_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %out.vec) nounwind { |
| ; SSE-LABEL: store_i32_stride7_vf4: |
| ; SSE: # %bb.0: |
| ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; SSE-NEXT: movaps (%rdi), %xmm0 |
| ; SSE-NEXT: movaps (%rsi), %xmm6 |
| ; SSE-NEXT: movaps (%rdx), %xmm5 |
| ; SSE-NEXT: movaps (%rcx), %xmm1 |
| ; SSE-NEXT: movaps (%r8), %xmm4 |
| ; SSE-NEXT: movaps (%r9), %xmm2 |
| ; SSE-NEXT: movaps (%r10), %xmm8 |
| ; SSE-NEXT: movaps %xmm5, %xmm7 |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm7 = xmm7[0],xmm1[0],xmm7[1],xmm1[1] |
| ; SSE-NEXT: movaps %xmm0, %xmm3 |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm7[0] |
| ; SSE-NEXT: movaps %xmm4, %xmm9 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm9 = xmm9[2],xmm2[2],xmm9[3],xmm2[3] |
| ; SSE-NEXT: movaps %xmm5, %xmm7 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm7 = xmm7[2],xmm1[2],xmm7[3],xmm1[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm9[0] |
| ; SSE-NEXT: movaps %xmm8, %xmm9 |
| ; SSE-NEXT: shufps {{.*#+}} xmm9 = xmm9[3,3],xmm2[3,3] |
| ; SSE-NEXT: movaps %xmm4, %xmm10 |
| ; SSE-NEXT: shufps {{.*#+}} xmm10 = xmm10[1,1],xmm1[1,1] |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm1 = xmm1[2],xmm4[2],xmm1[3],xmm4[3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,3],xmm9[2,0] |
| ; SSE-NEXT: movaps %xmm0, %xmm9 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm9 = xmm9[2],xmm6[2],xmm9[3],xmm6[3] |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1] |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm8[0],xmm2[1],xmm8[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[2,3],xmm9[0,1] |
| ; SSE-NEXT: movaps %xmm6, %xmm9 |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm9 = xmm9[0],xmm5[0],xmm9[1],xmm5[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm9 = xmm9[2,3],xmm10[2,0] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm8[2,3,2,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm8 = xmm8[0,1],xmm0[1,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,1],xmm8[0,2] |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm6[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,3],xmm5[2,0] |
| ; SSE-NEXT: movss {{.*#+}} xmm0 = xmm10[0],xmm0[1,2,3] |
| ; SSE-NEXT: movaps %xmm4, 16(%rax) |
| ; SSE-NEXT: movaps %xmm9, 32(%rax) |
| ; SSE-NEXT: movaps %xmm2, 48(%rax) |
| ; SSE-NEXT: movaps %xmm1, 96(%rax) |
| ; SSE-NEXT: movaps %xmm7, 64(%rax) |
| ; SSE-NEXT: movaps %xmm3, (%rax) |
| ; SSE-NEXT: movaps %xmm0, 80(%rax) |
| ; SSE-NEXT: retq |
| ; |
| ; AVX1-ONLY-LABEL: store_i32_stride7_vf4: |
| ; AVX1-ONLY: # %bb.0: |
| ; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX1-ONLY-NEXT: vmovaps (%rdi), %xmm5 |
| ; AVX1-ONLY-NEXT: vmovaps (%rsi), %xmm6 |
| ; AVX1-ONLY-NEXT: vmovaps (%rdx), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps (%rcx), %xmm4 |
| ; AVX1-ONLY-NEXT: vmovaps (%r8), %xmm1 |
| ; AVX1-ONLY-NEXT: vmovaps (%r9), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps (%r10), %xmm0 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm7 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm8 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm9 |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm10 = ymm9[1,0],ymm8[1,0],ymm9[5,4],ymm8[5,4] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm10 = ymm10[2,0],ymm8[2,1],ymm10[6,4],ymm8[6,5] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm5, %ymm6, %ymm5 |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm6 = ymm7[1],ymm5[1],ymm7[3],ymm5[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm6 = ymm5[1,1],ymm6[2,0],ymm5[5,5],ymm6[6,4] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm11 = xmm3[1,1],xmm4[1,1] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm11[1,2],ymm6[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2],ymm10[3,4,5],ymm6[6,7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm10 = ymm5[3,3],ymm7[3,3],ymm5[7,7],ymm7[7,7] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm3, %ymm4, %ymm11 |
| ; AVX1-ONLY-NEXT: vunpckhps {{.*#+}} ymm11 = ymm3[2],ymm11[2],ymm3[3],ymm11[3],ymm3[6],ymm11[6],ymm3[7],ymm11[7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm10 = ymm11[0,1,2,3,4],ymm10[5,6],ymm11[7] |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm11 = xmm2[1],xmm1[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm9 = ymm9[2,1],ymm11[2,0],ymm9[6,5],ymm11[6,4] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1],ymm9[2,3,4],ymm10[5,6,7] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm10 |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} ymm8 = ymm10[0],ymm8[0],ymm10[1],ymm8[1],ymm10[4],ymm8[4],ymm10[5],ymm8[5] |
| ; AVX1-ONLY-NEXT: vbroadcastss (%r10), %ymm10 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm10[6,7] |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} ymm5 = ymm7[0],ymm5[0],ymm7[1],ymm5[1],ymm7[4],ymm5[4],ymm7[5],ymm5[5] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm3 = xmm4[0,0],xmm3[0,0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm3 = xmm3[0,1,2,0] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm5[0,1],ymm3[2,3],ymm5[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm8[4,5,6],ymm3[7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm1[3,3],xmm2[3,3] |
| ; AVX1-ONLY-NEXT: vbroadcastss 12(%rcx), %xmm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0],xmm1[1,2,3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm0, 96(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm3, (%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm9, 64(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm6, 32(%rax) |
| ; AVX1-ONLY-NEXT: vzeroupper |
| ; AVX1-ONLY-NEXT: retq |
| ; |
| ; AVX2-SLOW-LABEL: store_i32_stride7_vf4: |
| ; AVX2-SLOW: # %bb.0: |
| ; AVX2-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX2-SLOW-NEXT: vmovaps (%rdi), %xmm4 |
| ; AVX2-SLOW-NEXT: vmovaps (%rdx), %xmm5 |
| ; AVX2-SLOW-NEXT: vmovaps (%rcx), %xmm0 |
| ; AVX2-SLOW-NEXT: vmovaps (%r8), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps (%r9), %xmm3 |
| ; AVX2-SLOW-NEXT: vmovaps (%r10), %xmm1 |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, (%rsi), %ymm4, %ymm4 |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm0, %ymm5, %ymm6 |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm7 |
| ; AVX2-SLOW-NEXT: vbroadcastf128 {{.*#+}} ymm8 = [2,6,0,3,2,6,0,3] |
| ; AVX2-SLOW-NEXT: # ymm8 = mem[0,1,0,1] |
| ; AVX2-SLOW-NEXT: vpermps %ymm6, %ymm8, %ymm8 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm9 = ymm4[3,3,3,3,7,7,7,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[0,1,0,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4],ymm9[5,6],ymm8[7] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm9 |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm9 = ymm9[2],ymm3[2],ymm9[3],ymm3[3],ymm9[6],ymm3[6],ymm9[7],ymm3[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm9 = ymm9[0,1,0,1,4,5,4,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1],ymm9[2,3,4],ymm8[5,6,7] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm9 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm10 = ymm7[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4],ymm9[5],ymm10[6,7] |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm5 = zero,xmm5[1],xmm0[1],zero |
| ; AVX2-SLOW-NEXT: vbroadcastf128 {{.*#+}} ymm10 = [5,0,2,6,5,0,2,6] |
| ; AVX2-SLOW-NEXT: # ymm10 = mem[0,1,0,1] |
| ; AVX2-SLOW-NEXT: vpermps %ymm4, %ymm10, %ymm10 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm10[0],ymm5[1,2],ymm10[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2],ymm9[3,4,5],ymm5[6,7] |
| ; AVX2-SLOW-NEXT: vmovddup {{.*#+}} xmm9 = [0,4,0,4] |
| ; AVX2-SLOW-NEXT: # xmm9 = mem[0,0] |
| ; AVX2-SLOW-NEXT: vpermps %ymm6, %ymm9, %ymm6 |
| ; AVX2-SLOW-NEXT: vbroadcastf128 {{.*#+}} ymm9 = [0,4,0,1,0,4,0,1] |
| ; AVX2-SLOW-NEXT: # ymm9 = mem[0,1,0,1] |
| ; AVX2-SLOW-NEXT: vpermps %ymm4, %ymm9, %ymm4 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm6[2,3],ymm4[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastsd {{.*#+}} ymm6 = [0,4,0,4,0,4,0,4] |
| ; AVX2-SLOW-NEXT: vpermps %ymm7, %ymm6, %ymm6 |
| ; AVX2-SLOW-NEXT: vbroadcastss (%r10), %ymm7 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm7[6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6],ymm4[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3] |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[3],xmm2[1,2],zero |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[3] |
| ; AVX2-SLOW-NEXT: vmovaps %xmm0, 96(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm4, (%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm5, 32(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm8, 64(%rax) |
| ; AVX2-SLOW-NEXT: vzeroupper |
| ; AVX2-SLOW-NEXT: retq |
| ; |
| ; AVX2-FAST-LABEL: store_i32_stride7_vf4: |
| ; AVX2-FAST: # %bb.0: |
| ; AVX2-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX2-FAST-NEXT: vmovaps (%rdi), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps (%rdx), %xmm3 |
| ; AVX2-FAST-NEXT: vmovaps (%rcx), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps (%r8), %xmm4 |
| ; AVX2-FAST-NEXT: vmovaps (%r9), %xmm5 |
| ; AVX2-FAST-NEXT: vmovaps (%r10), %xmm0 |
| ; AVX2-FAST-NEXT: vinsertf128 $1, (%rsi), %ymm2, %ymm2 |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm3, %ymm3 |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm5, %ymm4, %ymm6 |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm7 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm8 = ymm6[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0,1,2,3,4],ymm7[5],ymm8[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastsd {{.*#+}} ymm8 = [5,1,5,1,5,1,5,1] |
| ; AVX2-FAST-NEXT: vpermps %ymm3, %ymm8, %ymm8 |
| ; AVX2-FAST-NEXT: vbroadcastf128 {{.*#+}} ymm9 = [5,0,2,6,5,0,2,6] |
| ; AVX2-FAST-NEXT: # ymm9 = mem[0,1,0,1] |
| ; AVX2-FAST-NEXT: vpermps %ymm2, %ymm9, %ymm9 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm8 = ymm9[0],ymm8[1,2],ymm9[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0,1,2],ymm7[3,4,5],ymm8[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastsd {{.*#+}} ymm8 = [7,3,7,3,7,3,7,3] |
| ; AVX2-FAST-NEXT: vpermps %ymm2, %ymm8, %ymm8 |
| ; AVX2-FAST-NEXT: vbroadcastf128 {{.*#+}} ymm9 = [2,6,0,3,2,6,0,3] |
| ; AVX2-FAST-NEXT: # ymm9 = mem[0,1,0,1] |
| ; AVX2-FAST-NEXT: vpermps %ymm3, %ymm9, %ymm9 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5,6],ymm9[7] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm4, %ymm4 |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm4 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm8[0,1],ymm4[2,3,4],ymm8[5,6,7] |
| ; AVX2-FAST-NEXT: vmovddup {{.*#+}} xmm5 = [0,4,0,4] |
| ; AVX2-FAST-NEXT: # xmm5 = mem[0,0] |
| ; AVX2-FAST-NEXT: vpermps %ymm3, %ymm5, %ymm3 |
| ; AVX2-FAST-NEXT: vbroadcastf128 {{.*#+}} ymm5 = [0,4,0,1,0,4,0,1] |
| ; AVX2-FAST-NEXT: # ymm5 = mem[0,1,0,1] |
| ; AVX2-FAST-NEXT: vpermps %ymm2, %ymm5, %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastsd {{.*#+}} ymm3 = [0,4,0,4,0,4,0,4] |
| ; AVX2-FAST-NEXT: vpermps %ymm6, %ymm3, %ymm3 |
| ; AVX2-FAST-NEXT: vbroadcastss (%r10), %ymm5 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm5[6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6],ymm2[7] |
| ; AVX2-FAST-NEXT: vbroadcastsd {{.*#+}} ymm3 = [7,3,7,3,7,3,7,3] |
| ; AVX2-FAST-NEXT: vpermps %ymm6, %ymm3, %ymm3 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm1[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm3[1,2,3] |
| ; AVX2-FAST-NEXT: vmovaps %ymm2, (%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm4, 64(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm7, 32(%rax) |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3] |
| ; AVX2-FAST-NEXT: vmovaps %xmm0, 96(%rax) |
| ; AVX2-FAST-NEXT: vzeroupper |
| ; AVX2-FAST-NEXT: retq |
| ; |
| ; AVX2-FAST-PERLANE-LABEL: store_i32_stride7_vf4: |
| ; AVX2-FAST-PERLANE: # %bb.0: |
| ; AVX2-FAST-PERLANE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-PERLANE-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdi), %xmm4 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdx), %xmm5 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rcx), %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r8), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r9), %xmm3 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r10), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, (%rsi), %ymm4, %ymm4 |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm0, %ymm5, %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm7 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastf128 {{.*#+}} ymm8 = [2,6,0,3,2,6,0,3] |
| ; AVX2-FAST-PERLANE-NEXT: # ymm8 = mem[0,1,0,1] |
| ; AVX2-FAST-PERLANE-NEXT: vpermps %ymm6, %ymm8, %ymm8 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm9 = ymm4[3,3,3,3,7,7,7,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[0,1,0,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4],ymm9[5,6],ymm8[7] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm9 |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm9 = ymm9[2],ymm3[2],ymm9[3],ymm3[3],ymm9[6],ymm3[6],ymm9[7],ymm3[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm9 = ymm9[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1],ymm9[2,3,4],ymm8[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm9 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm10 = ymm7[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4],ymm9[5],ymm10[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm5 = zero,xmm5[1],xmm0[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastf128 {{.*#+}} ymm10 = [5,0,2,6,5,0,2,6] |
| ; AVX2-FAST-PERLANE-NEXT: # ymm10 = mem[0,1,0,1] |
| ; AVX2-FAST-PERLANE-NEXT: vpermps %ymm4, %ymm10, %ymm10 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm10[0],ymm5[1,2],ymm10[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2],ymm9[3,4,5],ymm5[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovddup {{.*#+}} xmm9 = [0,4,0,4] |
| ; AVX2-FAST-PERLANE-NEXT: # xmm9 = mem[0,0] |
| ; AVX2-FAST-PERLANE-NEXT: vpermps %ymm6, %ymm9, %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastf128 {{.*#+}} ymm9 = [0,4,0,1,0,4,0,1] |
| ; AVX2-FAST-PERLANE-NEXT: # ymm9 = mem[0,1,0,1] |
| ; AVX2-FAST-PERLANE-NEXT: vpermps %ymm4, %ymm9, %ymm4 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm6[2,3],ymm4[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd {{.*#+}} ymm6 = [0,4,0,4,0,4,0,4] |
| ; AVX2-FAST-PERLANE-NEXT: vpermps %ymm7, %ymm6, %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss (%r10), %ymm7 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm7[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6],ymm4[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[3],xmm2[1,2],zero |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[3] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm0, 96(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm4, (%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm5, 32(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm8, 64(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vzeroupper |
| ; AVX2-FAST-PERLANE-NEXT: retq |
| ; |
| ; AVX512-LABEL: store_i32_stride7_vf4: |
| ; AVX512: # %bb.0: |
| ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX512-NEXT: vmovdqa (%rdi), %xmm0 |
| ; AVX512-NEXT: vmovdqa (%rdx), %xmm1 |
| ; AVX512-NEXT: vmovdqa (%r8), %xmm2 |
| ; AVX512-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 |
| ; AVX512-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 |
| ; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 |
| ; AVX512-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm1 |
| ; AVX512-NEXT: vinserti32x4 $2, (%r10), %zmm1, %zmm1 |
| ; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,4,8,12,16,20,24,1,5,9,13,17,21,25,2,6] |
| ; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm2 |
| ; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm3 = <10,14,18,22,26,3,7,11,15,19,23,27,u,u,u,u> |
| ; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm3 |
| ; AVX512-NEXT: vextracti32x4 $2, %zmm3, 96(%rax) |
| ; AVX512-NEXT: vmovdqa64 %zmm2, (%rax) |
| ; AVX512-NEXT: vmovdqa %ymm3, 64(%rax) |
| ; AVX512-NEXT: vzeroupper |
| ; AVX512-NEXT: retq |
| %in.vec0 = load <4 x i32>, ptr %in.vecptr0, align 64 |
| %in.vec1 = load <4 x i32>, ptr %in.vecptr1, align 64 |
| %in.vec2 = load <4 x i32>, ptr %in.vecptr2, align 64 |
| %in.vec3 = load <4 x i32>, ptr %in.vecptr3, align 64 |
| %in.vec4 = load <4 x i32>, ptr %in.vecptr4, align 64 |
| %in.vec5 = load <4 x i32>, ptr %in.vecptr5, align 64 |
| %in.vec6 = load <4 x i32>, ptr %in.vecptr6, align 64 |
| %1 = shufflevector <4 x i32> %in.vec0, <4 x i32> %in.vec1, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7> |
| %2 = shufflevector <4 x i32> %in.vec2, <4 x i32> %in.vec3, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7> |
| %3 = shufflevector <4 x i32> %in.vec4, <4 x i32> %in.vec5, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7> |
| %4 = shufflevector <8 x i32> %1, <8 x i32> %2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15> |
| %5 = shufflevector <4 x i32> %in.vec6, <4 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef> |
| %6 = shufflevector <8 x i32> %3, <8 x i32> %5, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11> |
| %7 = shufflevector <12 x i32> %6, <12 x i32> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 undef, i32 undef, i32 undef, i32 undef> |
| %8 = shufflevector <16 x i32> %4, <16 x i32> %7, <28 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27> |
| %interleaved.vec = shufflevector <28 x i32> %8, <28 x i32> poison, <28 x i32> <i32 0, i32 4, i32 8, i32 12, i32 16, i32 20, i32 24, i32 1, i32 5, i32 9, i32 13, i32 17, i32 21, i32 25, i32 2, i32 6, i32 10, i32 14, i32 18, i32 22, i32 26, i32 3, i32 7, i32 11, i32 15, i32 19, i32 23, i32 27> |
| store <28 x i32> %interleaved.vec, ptr %out.vec, align 64 |
| ret void |
| } |
| |
| define void @store_i32_stride7_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %out.vec) nounwind { |
| ; SSE-LABEL: store_i32_stride7_vf8: |
| ; SSE: # %bb.0: |
| ; SSE-NEXT: subq $104, %rsp |
| ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; SSE-NEXT: movdqa (%rdi), %xmm8 |
| ; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 16(%rdi), %xmm14 |
| ; SSE-NEXT: movdqa (%rsi), %xmm3 |
| ; SSE-NEXT: movdqa 16(%rsi), %xmm5 |
| ; SSE-NEXT: movdqa 16(%rdx), %xmm13 |
| ; SSE-NEXT: movdqa 16(%rcx), %xmm9 |
| ; SSE-NEXT: movdqa 16(%r8), %xmm11 |
| ; SSE-NEXT: movdqa %xmm11, (%rsp) # 16-byte Spill |
| ; SSE-NEXT: movdqa (%r9), %xmm4 |
| ; SSE-NEXT: movaps 16(%r9), %xmm1 |
| ; SSE-NEXT: movdqa (%rax), %xmm7 |
| ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 16(%rax), %xmm12 |
| ; SSE-NEXT: movaps %xmm12, %xmm0 |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3] |
| ; SSE-NEXT: movaps %xmm1, %xmm6 |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm11[3,3,3,3] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm9[3,3,3,3] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm11, %xmm0 |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm9[1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm13[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm14, %xmm0 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm5[2],xmm0[3],xmm5[3] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm12[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm6[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm4[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa %xmm8, %xmm0 |
| ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps (%rdx), %xmm11 |
| ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm3[1,1,1,1] |
| ; SSE-NEXT: movaps %xmm11, %xmm1 |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm5[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps (%rcx), %xmm3 |
| ; SSE-NEXT: movaps (%r8), %xmm7 |
| ; SSE-NEXT: movaps %xmm7, %xmm15 |
| ; SSE-NEXT: shufps {{.*#+}} xmm15 = xmm15[1,1],xmm3[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm15[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm13, %xmm15 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm15 = xmm15[2],xmm9[2],xmm15[3],xmm9[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm9 = xmm9[0],xmm13[0] |
| ; SSE-NEXT: movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm13 = xmm13[3,3],xmm10[3,3] |
| ; SSE-NEXT: movdqa %xmm14, %xmm1 |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm12[0,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm12 = xmm12[2,1],xmm14[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm12 = xmm12[0,2],xmm13[2,0] |
| ; SSE-NEXT: movaps (%rsp), %xmm8 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm8, %xmm13 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm13 = xmm13[0],xmm6[0],xmm13[1],xmm6[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm13 = xmm13[0,1],xmm1[2,0] |
| ; SSE-NEXT: movaps %xmm7, %xmm1 |
| ; SSE-NEXT: movdqa %xmm4, %xmm9 |
| ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm1 = xmm1[2],xmm4[2],xmm1[3],xmm4[3] |
| ; SSE-NEXT: movaps %xmm11, %xmm2 |
| ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm2 = xmm2[2],xmm3[2],xmm2[3],xmm3[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm4, %xmm1 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm5[1,3] |
| ; SSE-NEXT: movaps %xmm7, %xmm0 |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm9[0],xmm0[1],xmm9[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[0,2] |
| ; SSE-NEXT: movaps %xmm11, %xmm1 |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1] |
| ; SSE-NEXT: movaps %xmm5, %xmm3 |
| ; SSE-NEXT: movaps %xmm5, %xmm9 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm9 = xmm9[0],xmm5[0],xmm9[1],xmm5[1] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm9 = xmm9[0],xmm1[0] |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm8 = xmm8[2],xmm6[2],xmm8[3],xmm6[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm15 = xmm15[0],xmm8[0] |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm14 = xmm14[0],xmm10[0],xmm14[1],xmm10[1] |
| ; SSE-NEXT: shufps $36, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm14 = xmm14[0,1],mem[2,0] |
| ; SSE-NEXT: shufps {{.*#+}} xmm11 = xmm11[3,3],xmm5[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,3],xmm11[2,0] |
| ; SSE-NEXT: movaps %xmm3, %xmm6 |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm4[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm6 = xmm1[0],xmm6[1,2,3] |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = xmm4[3,3],mem[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm7 = xmm7[0,3],xmm4[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm1 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm7 = xmm1[0],xmm7[1,2,3] |
| ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; SSE-NEXT: movaps %xmm14, 112(%rax) |
| ; SSE-NEXT: movdqa %xmm15, 176(%rax) |
| ; SSE-NEXT: movaps %xmm9, (%rax) |
| ; SSE-NEXT: movaps %xmm0, 16(%rax) |
| ; SSE-NEXT: movaps %xmm2, 64(%rax) |
| ; SSE-NEXT: movaps %xmm13, 128(%rax) |
| ; SSE-NEXT: movaps %xmm12, 192(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 32(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 48(%rax) |
| ; SSE-NEXT: movaps %xmm7, 96(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 160(%rax) |
| ; SSE-NEXT: movaps %xmm6, 80(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 144(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 208(%rax) |
| ; SSE-NEXT: addq $104, %rsp |
| ; SSE-NEXT: retq |
| ; |
| ; AVX1-ONLY-LABEL: store_i32_stride7_vf8: |
| ; AVX1-ONLY: # %bb.0: |
| ; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX1-ONLY-NEXT: vmovaps (%rdi), %ymm0 |
| ; AVX1-ONLY-NEXT: vmovaps (%rsi), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovaps (%rdx), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovaps (%rcx), %ymm3 |
| ; AVX1-ONLY-NEXT: vmovaps (%r8), %ymm7 |
| ; AVX1-ONLY-NEXT: vmovaps (%r9), %ymm8 |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm4 = ymm2[1,1],ymm3[1,1],ymm2[5,5],ymm3[5,5] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm5 = ymm1[1,1],ymm0[1,1],ymm1[5,5],ymm0[5,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm5 = ymm5[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5,6],ymm5[7] |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm5 = ymm8[0],ymm7[0],ymm8[2],ymm7[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm5 = ymm5[2,0],ymm7[2,1],ymm5[6,4],ymm7[6,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm6 = mem[2,3],ymm5[2,3] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm5, %xmm5 |
| ; AVX1-ONLY-NEXT: vshufpd {{.*#+}} ymm5 = ymm5[0],ymm6[0],ymm5[2],ymm6[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3,4,5,6],ymm5[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%r8), %xmm5 |
| ; AVX1-ONLY-NEXT: vmovaps (%r9), %xmm6 |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm4 = xmm6[1,1],xmm5[1,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm4, %ymm4, %ymm4 |
| ; AVX1-ONLY-NEXT: vmovaps (%rsi), %xmm11 |
| ; AVX1-ONLY-NEXT: vmovaps (%rdi), %xmm13 |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm9 = xmm13[1],xmm11[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm9 = xmm11[1,1],xmm9[0,2] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm9, %ymm9, %ymm9 |
| ; AVX1-ONLY-NEXT: vmovaps (%rcx), %xmm15 |
| ; AVX1-ONLY-NEXT: vmovaps (%rdx), %xmm14 |
| ; AVX1-ONLY-NEXT: vinsertps {{.*#+}} xmm10 = zero,xmm14[1],xmm15[1],zero |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0],ymm10[1,2],ymm9[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovaps (%rax), %xmm10 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm10, %ymm0, %ymm12 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4],ymm12[5],ymm4[6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm4 = ymm9[0,1,2],ymm4[3,4,5],ymm9[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm4 = xmm5[0],xmm6[0],xmm5[1],xmm6[1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm4 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm12[0],ymm4[2],ymm12[2] |
| ; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm12 = xmm11[0],xmm13[0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm12 = xmm12[2,0],xmm13[2,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm12, %ymm12, %ymm12 |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm9 = xmm14[0],xmm15[0],xmm14[1],xmm15[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm9 = xmm9[0,1,0,1] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm9 = ymm12[0,1],ymm9[2,3],ymm12[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm4 = ymm9[0,1,2,3],ymm4[4,5,6],ymm9[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm4 = ymm1[1],ymm0[1],ymm1[3],ymm0[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,1,2,0,4,5,6,4] |
| ; AVX1-ONLY-NEXT: vunpckhps {{.*#+}} ymm9 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],ymm9[2,3] |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm9 = ymm7[1],ymm8[1],ymm7[3],ymm8[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm9 = ymm8[1,1],ymm9[0,2],ymm8[5,5],ymm9[4,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm9[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 16(%rax), %xmm12 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0],ymm12[1],ymm9[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm4 = ymm9[0,1],ymm4[2,3,4,5],ymm9[6,7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm9 = ymm3[3,3],ymm2[3,3],ymm3[7,7],ymm2[7,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm9[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm12 = ymm0[3,3],ymm1[3,3],ymm0[7,7],ymm1[7,7] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm12, %xmm12 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm9 = ymm12[0,1,2],ymm9[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm7 = ymm7[3,3],ymm8[3,3],ymm7[7,7],ymm8[7,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm7 = ymm9[0,1,2,3,4],ymm7[5,6],ymm9[7] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 24(%rax), %ymm8 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0],ymm7[1,2,3,4,5,6],ymm8[7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm8 = xmm13[3,3],xmm11[3,3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm8 |
| ; AVX1-ONLY-NEXT: vunpckhps {{.*#+}} xmm9 = xmm14[2],xmm15[2],xmm14[3],xmm15[3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm14, %ymm9, %ymm9 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5,6],ymm9[7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm9 = xmm6[2,2,2,2] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} xmm9 = xmm5[0,1,2],xmm9[3] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 8(%rax), %ymm11 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm11[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1],ymm9[2,3,4],ymm8[5,6,7] |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5] |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm2[0],ymm3[0],ymm2[2],ymm3[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm3[3,1],ymm1[0,2],ymm3[7,5],ymm1[4,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm5[3,3],xmm6[3,3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1,2],xmm10[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 96(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm8, 64(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm4, 160(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm7, 192(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 128(%rax) |
| ; AVX1-ONLY-NEXT: vzeroupper |
| ; AVX1-ONLY-NEXT: retq |
| ; |
| ; AVX2-SLOW-LABEL: store_i32_stride7_vf8: |
| ; AVX2-SLOW: # %bb.0: |
| ; AVX2-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-SLOW-NEXT: vmovaps (%rdi), %ymm0 |
| ; AVX2-SLOW-NEXT: vmovaps (%rsi), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovaps (%rdx), %ymm8 |
| ; AVX2-SLOW-NEXT: vmovaps (%rcx), %ymm3 |
| ; AVX2-SLOW-NEXT: vmovaps (%r8), %ymm6 |
| ; AVX2-SLOW-NEXT: vmovaps (%r9), %ymm7 |
| ; AVX2-SLOW-NEXT: vmovaps (%rax), %xmm1 |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm9 |
| ; AVX2-SLOW-NEXT: vmovaps (%r8), %xmm4 |
| ; AVX2-SLOW-NEXT: vmovaps (%r9), %xmm5 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm10 = xmm5[1,1,1,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm10 = xmm10[0],xmm4[1],xmm10[2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm10, %ymm10 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4],ymm9[5],ymm10[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps (%rcx), %xmm12 |
| ; AVX2-SLOW-NEXT: vmovaps (%rdx), %xmm13 |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm10 = zero,xmm13[1],xmm12[1],zero |
| ; AVX2-SLOW-NEXT: vmovaps (%rdi), %xmm14 |
| ; AVX2-SLOW-NEXT: vmovaps (%rsi), %xmm15 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm11 = xmm15[1,1,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm11 = xmm11[0,1],xmm14[2],xmm11[3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm10 = ymm11[0],ymm10[1,2],ymm11[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2],ymm9[3,4,5],ymm10[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm10 = ymm8[2],ymm3[2],ymm8[3],ymm3[3],ymm8[6],ymm3[6],ymm8[7],ymm3[7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm11 = ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[6],ymm2[6],ymm0[7],ymm2[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm11 = ymm7[1,1,2,2,5,5,6,6] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1],ymm6[2],ymm11[3,4,5],ymm6[6],ymm11[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vmovaps 16(%rax), %xmm9 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm9 = ymm11[0],ymm9[1],ymm11[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1],ymm10[2,3,4,5],ymm9[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm9 = xmm14[3,3],xmm15[3,3] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm9 |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} xmm11 = xmm13[2],xmm12[2],xmm13[3],xmm12[3] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm11 = xmm11[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm9 = ymm11[0,1,2,3,4],ymm9[5,6],ymm11[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm11 = xmm5[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm11 = xmm4[0,1,2],xmm11[3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 8(%rax), %ymm10 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm11 = ymm9[0,1],ymm10[2,3,4],ymm9[5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm12, %xmm9 |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm13, %xmm10 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm10 = xmm14[0],xmm15[0],xmm14[1],xmm15[1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm10 = xmm10[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1],ymm9[2,3],ymm10[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm10 = xmm4[0],xmm5[0],xmm4[1],xmm5[1] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm10, %ymm0, %ymm10 |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm1, %ymm12 |
| ; AVX2-SLOW-NEXT: vunpcklpd {{.*#+}} ymm10 = ymm10[0],ymm12[0],ymm10[2],ymm12[2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm12 = ymm9[0,1,2,3],ymm10[4,5,6],ymm9[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm9 = ymm8[1,1],ymm3[1,1],ymm8[5,5],ymm3[5,5] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm10 = ymm2[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0],ymm0[1],ymm10[2,3,4],ymm0[5],ymm10[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4],ymm9[5,6],ymm10[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm10 = ymm7[0,0,0,0,4,4,4,4] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm13 = ymm6[0,1,0,1,4,5,4,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm10 = ymm13[0],ymm10[1],ymm13[2,3,4],ymm10[5],ymm13[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 16(%rax), %ymm13 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1],ymm13[2,3],ymm10[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2],ymm9[3,4,5,6],ymm10[7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm8 = ymm3[2],ymm8[2],ymm3[3],ymm8[3],ymm3[6],ymm8[6],ymm3[7],ymm8[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm10 = ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[6],ymm0[6],ymm2[7],ymm0[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm8 = ymm10[0,1,2],ymm8[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm6 = ymm6[3,3],ymm7[3,3],ymm6[7,7],ymm7[7,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm8[0,1,2,3,4],ymm6[5,6],ymm8[7] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 24(%rax), %ymm7 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0],ymm6[1,2,3,4,5,6],ymm7[7] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[1],ymm2[1],ymm0[4],ymm2[4],ymm0[5],ymm2[5] |
| ; AVX2-SLOW-NEXT: vbroadcastss 16(%rdx), %ymm2 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm3 = ymm3[3,1,2,0,7,5,6,4] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6],ymm3[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5],ymm2[6,7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm2 = xmm4[3,3],xmm5[3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5,6,7] |
| ; AVX2-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 96(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm6, 192(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 160(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm9, 128(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm12, (%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm11, 64(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX2-SLOW-NEXT: vzeroupper |
| ; AVX2-SLOW-NEXT: retq |
| ; |
| ; AVX2-FAST-LABEL: store_i32_stride7_vf8: |
| ; AVX2-FAST: # %bb.0: |
| ; AVX2-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-NEXT: vmovaps (%rdi), %ymm1 |
| ; AVX2-FAST-NEXT: vmovaps (%rsi), %ymm2 |
| ; AVX2-FAST-NEXT: vmovaps (%rdx), %ymm10 |
| ; AVX2-FAST-NEXT: vmovaps (%rcx), %ymm3 |
| ; AVX2-FAST-NEXT: vmovaps (%r8), %ymm7 |
| ; AVX2-FAST-NEXT: vmovaps (%r9), %ymm8 |
| ; AVX2-FAST-NEXT: vmovaps (%rax), %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm6 |
| ; AVX2-FAST-NEXT: vmovaps (%r8), %xmm4 |
| ; AVX2-FAST-NEXT: vmovaps (%r9), %xmm5 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm9 = xmm5[1,1,1,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm9 = xmm9[0],xmm4[1],xmm9[2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm9, %ymm9 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm9[0,1,2,3,4],ymm6[5],ymm9[6,7] |
| ; AVX2-FAST-NEXT: vmovaps (%rcx), %xmm11 |
| ; AVX2-FAST-NEXT: vmovaps (%rdx), %xmm12 |
| ; AVX2-FAST-NEXT: vinsertps {{.*#+}} xmm9 = zero,xmm12[1],xmm11[1],zero |
| ; AVX2-FAST-NEXT: vmovaps (%rdi), %xmm13 |
| ; AVX2-FAST-NEXT: vmovaps (%rsi), %xmm14 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm15 = xmm14[1,1,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm15 = xmm15[0,1],xmm13[2],xmm15[3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[0,1,2,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm9 = ymm15[0],ymm9[1,2],ymm15[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2],ymm6[3,4,5],ymm9[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm6 = ymm10[2],ymm3[2],ymm10[3],ymm3[3],ymm10[6],ymm3[6],ymm10[7],ymm3[7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm9 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm9[0,1,2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm9 = ymm8[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1],ymm7[2],ymm9[3,4,5],ymm7[6],ymm9[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[2,1,2,3] |
| ; AVX2-FAST-NEXT: vmovaps 16(%rax), %xmm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0],ymm15[1],ymm9[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1],ymm6[2,3,4,5],ymm9[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm6 = ymm2[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm1[1],ymm6[2,3,4],ymm1[5],ymm6[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[2,2,2,2] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm15 = ymm10[1,1],ymm3[1,1],ymm10[5,5],ymm3[5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4],ymm15[5,6],ymm6[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm15 = ymm8[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm9 = ymm7[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0],ymm15[1],ymm9[2,3,4],ymm15[5],ymm9[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[2,1,2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 16(%rax), %ymm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1],ymm15[2,3],ymm9[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2],ymm6[3,4,5,6],ymm9[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm6 = xmm13[3,3],xmm14[3,3] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm6 |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} xmm9 = xmm12[2],xmm11[2],xmm12[3],xmm11[3] |
| ; AVX2-FAST-NEXT: vbroadcastf128 {{.*#+}} ymm0 = [0,1,2,2,0,1,2,2] |
| ; AVX2-FAST-NEXT: # ymm0 = mem[0,1,0,1] |
| ; AVX2-FAST-NEXT: vpermps %ymm9, %ymm0, %ymm9 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm9[0,1,2,3,4],ymm6[5,6],ymm9[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm9 = xmm5[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm9 = xmm4[0,1,2],xmm9[3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 8(%rax), %ymm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm15[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1],ymm9[2,3,4],ymm6[5,6,7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm9 = ymm3[2],ymm10[2],ymm3[3],ymm10[3],ymm3[6],ymm10[6],ymm3[7],ymm10[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[3,3,3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm10 = ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[6],ymm1[6],ymm2[7],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2],ymm9[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm7 = ymm7[3,3],ymm8[3,3],ymm7[7,7],ymm8[7,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm7 = ymm9[0,1,2,3,4],ymm7[5,6],ymm9[7] |
| ; AVX2-FAST-NEXT: vbroadcastsd 24(%rax), %ymm8 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0],ymm7[1,2,3,4,5,6],ymm8[7] |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm11, %xmm8 |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm12, %xmm9 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1] |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm9 = xmm13[0],xmm14[0],xmm13[1],xmm14[1] |
| ; AVX2-FAST-NEXT: vpermps %ymm9, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm8[2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm8 = xmm4[0],xmm5[0],xmm4[1],xmm5[1] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm8 |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm10, %ymm9 |
| ; AVX2-FAST-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm8[0],ymm9[0],ymm8[2],ymm9[2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm8[4,5,6],ymm0[7] |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[4],ymm2[4],ymm1[5],ymm2[5] |
| ; AVX2-FAST-NEXT: vbroadcastss 16(%rdx), %ymm2 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm3 = ymm3[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6],ymm3[7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5],ymm2[6,7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm2 = xmm4[3,3],xmm5[3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1,2],xmm10[3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-NEXT: vmovaps %ymm1, 96(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm7, 192(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm6, 64(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 160(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 128(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX2-FAST-NEXT: vzeroupper |
| ; AVX2-FAST-NEXT: retq |
| ; |
| ; AVX2-FAST-PERLANE-LABEL: store_i32_stride7_vf8: |
| ; AVX2-FAST-PERLANE: # %bb.0: |
| ; AVX2-FAST-PERLANE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdi), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rsi), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdx), %ymm8 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rcx), %ymm3 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r8), %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r9), %ymm7 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rax), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm9 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r8), %xmm4 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r9), %xmm5 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm10 = xmm5[1,1,1,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm10 = xmm10[0],xmm4[1],xmm10[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm10, %ymm10 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4],ymm9[5],ymm10[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rcx), %xmm12 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdx), %xmm13 |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm10 = zero,xmm13[1],xmm12[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdi), %xmm14 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rsi), %xmm15 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm11 = xmm15[1,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm11 = xmm11[0,1],xmm14[2],xmm11[3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm10 = ymm11[0],ymm10[1,2],ymm11[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2],ymm9[3,4,5],ymm10[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm10 = ymm8[2],ymm3[2],ymm8[3],ymm3[3],ymm8[6],ymm3[6],ymm8[7],ymm3[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm11 = ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[6],ymm2[6],ymm0[7],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm11 = ymm7[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1],ymm6[2],ymm11[3,4,5],ymm6[6],ymm11[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 16(%rax), %xmm9 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm9 = ymm11[0],ymm9[1],ymm11[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1],ymm10[2,3,4,5],ymm9[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm9 = xmm14[3,3],xmm15[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm9 |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} xmm11 = xmm13[2],xmm12[2],xmm13[3],xmm12[3] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm11 = xmm11[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm9 = ymm11[0,1,2,3,4],ymm9[5,6],ymm11[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm11 = xmm5[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm11 = xmm4[0,1,2],xmm11[3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 8(%rax), %ymm10 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm11 = ymm9[0,1],ymm10[2,3,4],ymm9[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm12, %xmm9 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm13, %xmm10 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm10 = xmm14[0],xmm15[0],xmm14[1],xmm15[1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm10 = xmm10[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1],ymm9[2,3],ymm10[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm10 = xmm4[0],xmm5[0],xmm4[1],xmm5[1] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm10, %ymm0, %ymm10 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm1, %ymm12 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklpd {{.*#+}} ymm10 = ymm10[0],ymm12[0],ymm10[2],ymm12[2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm12 = ymm9[0,1,2,3],ymm10[4,5,6],ymm9[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm9 = ymm8[1,1],ymm3[1,1],ymm8[5,5],ymm3[5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm10 = ymm2[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0],ymm0[1],ymm10[2,3,4],ymm0[5],ymm10[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4],ymm9[5,6],ymm10[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm10 = ymm7[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm13 = ymm6[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm10 = ymm13[0],ymm10[1],ymm13[2,3,4],ymm10[5],ymm13[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 16(%rax), %ymm13 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1],ymm13[2,3],ymm10[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2],ymm9[3,4,5,6],ymm10[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm8 = ymm3[2],ymm8[2],ymm3[3],ymm8[3],ymm3[6],ymm8[6],ymm3[7],ymm8[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm10 = ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[6],ymm0[6],ymm2[7],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm8 = ymm10[0,1,2],ymm8[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm6 = ymm6[3,3],ymm7[3,3],ymm6[7,7],ymm7[7,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm8[0,1,2,3,4],ymm6[5,6],ymm8[7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 24(%rax), %ymm7 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0],ymm6[1,2,3,4,5,6],ymm7[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[1],ymm2[1],ymm0[4],ymm2[4],ymm0[5],ymm2[5] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 16(%rdx), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm3 = ymm3[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6],ymm3[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5],ymm2[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm2 = xmm4[3,3],xmm5[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 96(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm6, 192(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 160(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm9, 128(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm12, (%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm11, 64(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vzeroupper |
| ; AVX2-FAST-PERLANE-NEXT: retq |
| ; |
| ; AVX512F-LABEL: store_i32_stride7_vf8: |
| ; AVX512F: # %bb.0: |
| ; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX512F-NEXT: vmovdqa (%rdi), %ymm0 |
| ; AVX512F-NEXT: vmovdqa (%rdx), %ymm1 |
| ; AVX512F-NEXT: vmovdqa (%r8), %ymm2 |
| ; AVX512F-NEXT: vmovdqa (%r10), %ymm3 |
| ; AVX512F-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm4 |
| ; AVX512F-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm5 |
| ; AVX512F-NEXT: vinserti64x4 $1, (%r9), %zmm2, %zmm2 |
| ; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [31,7,15,23,31,7,15,23] |
| ; AVX512F-NEXT: # ymm0 = mem[0,1,0,1] |
| ; AVX512F-NEXT: vpermi2d %zmm5, %zmm4, %zmm0 |
| ; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [6,23,31,7,6,23,31,7] |
| ; AVX512F-NEXT: # ymm1 = mem[0,1,0,1] |
| ; AVX512F-NEXT: vpermi2d %zmm2, %zmm3, %zmm1 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm6 = <u,u,u,u,0,8,16,u,u,u,u,1,9,17,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm3, %zmm2, %zmm6 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm7 = <0,8,16,24,u,u,u,1,9,17,25,u,u,u,2,10> |
| ; AVX512F-NEXT: vpermi2d %zmm5, %zmm4, %zmm7 |
| ; AVX512F-NEXT: movw $14448, %cx # imm = 0x3870 |
| ; AVX512F-NEXT: kmovw %ecx, %k1 |
| ; AVX512F-NEXT: vmovdqa32 %zmm6, %zmm7 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm6 = <u,u,2,10,18,u,u,u,u,3,11,19,u,u,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm3, %zmm2, %zmm6 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm8 = <2,10,u,u,u,19,27,3,11,u,u,u,20,28,4,12> |
| ; AVX512F-NEXT: vpermi2d %zmm4, %zmm5, %zmm8 |
| ; AVX512F-NEXT: movw $3612, %cx # imm = 0xE1C |
| ; AVX512F-NEXT: kmovw %ecx, %k1 |
| ; AVX512F-NEXT: vmovdqa32 %zmm6, %zmm8 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm6 = <u,u,u,5,13,21,29,u,u,u,6,14,22,30,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm5, %zmm4, %zmm6 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm4 = <4,12,20,u,u,u,u,5,13,21,u,u,u,u,6,14> |
| ; AVX512F-NEXT: vpermi2d %zmm3, %zmm2, %zmm4 |
| ; AVX512F-NEXT: movw $15480, %cx # imm = 0x3C78 |
| ; AVX512F-NEXT: kmovw %ecx, %k1 |
| ; AVX512F-NEXT: vmovdqa32 %zmm6, %zmm4 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 %zmm4, 128(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm8, 64(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm7, (%rax) |
| ; AVX512F-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX512F-NEXT: vmovdqa %ymm0, 192(%rax) |
| ; AVX512F-NEXT: vzeroupper |
| ; AVX512F-NEXT: retq |
| ; |
| ; AVX512BW-LABEL: store_i32_stride7_vf8: |
| ; AVX512BW: # %bb.0: |
| ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0 |
| ; AVX512BW-NEXT: vmovdqa (%rdx), %ymm1 |
| ; AVX512BW-NEXT: vmovdqa (%r8), %ymm2 |
| ; AVX512BW-NEXT: vmovdqa (%r10), %ymm3 |
| ; AVX512BW-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm4 |
| ; AVX512BW-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm5 |
| ; AVX512BW-NEXT: vinserti64x4 $1, (%r9), %zmm2, %zmm2 |
| ; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [31,7,15,23,31,7,15,23] |
| ; AVX512BW-NEXT: # ymm0 = mem[0,1,0,1] |
| ; AVX512BW-NEXT: vpermi2d %zmm5, %zmm4, %zmm0 |
| ; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [6,23,31,7,6,23,31,7] |
| ; AVX512BW-NEXT: # ymm1 = mem[0,1,0,1] |
| ; AVX512BW-NEXT: vpermi2d %zmm2, %zmm3, %zmm1 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm6 = <u,u,u,u,0,8,16,u,u,u,u,1,9,17,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm6 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm7 = <0,8,16,24,u,u,u,1,9,17,25,u,u,u,2,10> |
| ; AVX512BW-NEXT: vpermi2d %zmm5, %zmm4, %zmm7 |
| ; AVX512BW-NEXT: movw $14448, %cx # imm = 0x3870 |
| ; AVX512BW-NEXT: kmovd %ecx, %k1 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm6, %zmm7 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm6 = <u,u,2,10,18,u,u,u,u,3,11,19,u,u,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm6 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm8 = <2,10,u,u,u,19,27,3,11,u,u,u,20,28,4,12> |
| ; AVX512BW-NEXT: vpermi2d %zmm4, %zmm5, %zmm8 |
| ; AVX512BW-NEXT: movw $3612, %cx # imm = 0xE1C |
| ; AVX512BW-NEXT: kmovd %ecx, %k1 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm6, %zmm8 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm6 = <u,u,u,5,13,21,29,u,u,u,6,14,22,30,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm5, %zmm4, %zmm6 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm4 = <4,12,20,u,u,u,u,5,13,21,u,u,u,u,6,14> |
| ; AVX512BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm4 |
| ; AVX512BW-NEXT: movw $15480, %cx # imm = 0x3C78 |
| ; AVX512BW-NEXT: kmovd %ecx, %k1 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm6, %zmm4 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 %zmm4, 128(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm8, 64(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm7, (%rax) |
| ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX512BW-NEXT: vmovdqa %ymm0, 192(%rax) |
| ; AVX512BW-NEXT: vzeroupper |
| ; AVX512BW-NEXT: retq |
| %in.vec0 = load <8 x i32>, ptr %in.vecptr0, align 64 |
| %in.vec1 = load <8 x i32>, ptr %in.vecptr1, align 64 |
| %in.vec2 = load <8 x i32>, ptr %in.vecptr2, align 64 |
| %in.vec3 = load <8 x i32>, ptr %in.vecptr3, align 64 |
| %in.vec4 = load <8 x i32>, ptr %in.vecptr4, align 64 |
| %in.vec5 = load <8 x i32>, ptr %in.vecptr5, align 64 |
| %in.vec6 = load <8 x i32>, ptr %in.vecptr6, align 64 |
| %1 = shufflevector <8 x i32> %in.vec0, <8 x i32> %in.vec1, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15> |
| %2 = shufflevector <8 x i32> %in.vec2, <8 x i32> %in.vec3, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15> |
| %3 = shufflevector <8 x i32> %in.vec4, <8 x i32> %in.vec5, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15> |
| %4 = shufflevector <16 x i32> %1, <16 x i32> %2, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31> |
| %5 = shufflevector <8 x i32> %in.vec6, <8 x i32> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef> |
| %6 = shufflevector <16 x i32> %3, <16 x i32> %5, <24 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23> |
| %7 = shufflevector <24 x i32> %6, <24 x i32> poison, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef> |
| %8 = shufflevector <32 x i32> %4, <32 x i32> %7, <56 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55> |
| %interleaved.vec = shufflevector <56 x i32> %8, <56 x i32> poison, <56 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55> |
| store <56 x i32> %interleaved.vec, ptr %out.vec, align 64 |
| ret void |
| } |
| |
| define void @store_i32_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %out.vec) nounwind { |
| ; SSE-LABEL: store_i32_stride7_vf16: |
| ; SSE: # %bb.0: |
| ; SSE-NEXT: subq $536, %rsp # imm = 0x218 |
| ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; SSE-NEXT: movdqa (%rdi), %xmm3 |
| ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa (%rsi), %xmm7 |
| ; SSE-NEXT: movdqa 16(%rsi), %xmm9 |
| ; SSE-NEXT: movaps (%rdx), %xmm5 |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 16(%rdx), %xmm10 |
| ; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps (%rcx), %xmm14 |
| ; SSE-NEXT: movaps 16(%rcx), %xmm2 |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps (%r8), %xmm15 |
| ; SSE-NEXT: movaps 16(%r8), %xmm13 |
| ; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa (%r9), %xmm11 |
| ; SSE-NEXT: movdqa 16(%r9), %xmm12 |
| ; SSE-NEXT: movdqa %xmm12, (%rsp) # 16-byte Spill |
| ; SSE-NEXT: movdqa (%rax), %xmm4 |
| ; SSE-NEXT: movaps %xmm15, %xmm0 |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm14[1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm7[1,1,1,1] |
| ; SSE-NEXT: movss {{.*#+}} xmm5 = xmm1[0],xmm5[1,2,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm3, %xmm0 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm7[2],xmm0[3],xmm7[3] |
| ; SSE-NEXT: movdqa %xmm7, %xmm3 |
| ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm4[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm11[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm5 = xmm5[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm9[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movaps %xmm13, %xmm0 |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm2[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 16(%rax), %xmm10 |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm10[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm12[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 16(%rdi), %xmm13 |
| ; SSE-NEXT: movdqa %xmm13, %xmm0 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm9[2],xmm0[3],xmm9[3] |
| ; SSE-NEXT: movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 32(%rsi), %xmm8 |
| ; SSE-NEXT: movaps 32(%rdx), %xmm5 |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm8[1,1,1,1] |
| ; SSE-NEXT: movaps %xmm5, %xmm1 |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps 32(%rcx), %xmm5 |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 32(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm5[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 32(%r9), %xmm5 |
| ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 32(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 32(%rdi), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm8[2],xmm0[3],xmm8[3] |
| ; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 48(%rsi), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 48(%rdx), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm0[0],xmm5[1],xmm0[1] |
| ; SSE-NEXT: movaps 48(%rcx), %xmm2 |
| ; SSE-NEXT: movaps 48(%r8), %xmm1 |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm2[1,1] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 48(%rdi), %xmm0 |
| ; SSE-NEXT: movaps 48(%rax), %xmm7 |
| ; SSE-NEXT: movaps %xmm0, %xmm5 |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,1],xmm7[0,3] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 48(%r9), %xmm6 |
| ; SSE-NEXT: movaps %xmm6, %xmm5 |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,1],xmm7[1,1] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm7[2,3] |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm7 = xmm7[3,3],xmm6[3,3] |
| ; SSE-NEXT: movaps %xmm1, %xmm5 |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm1[1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[3,3,3,3] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm2[3,3,3,3] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm7[2,0] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm5, %xmm0 |
| ; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movdqa %xmm2, %xmm7 |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm3[0],xmm7[1],xmm3[1] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm7 = xmm7[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,1],xmm2[1,3] |
| ; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm15, %xmm2 |
| ; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm11[0],xmm2[1],xmm11[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm4[0,2] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm15 = xmm15[2],xmm11[2],xmm15[3],xmm11[3] |
| ; SSE-NEXT: movaps %xmm5, %xmm2 |
| ; SSE-NEXT: movaps %xmm5, %xmm7 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm2 = xmm2[2],xmm14[2],xmm2[3],xmm14[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm15[0] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm2, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm4[0] |
| ; SSE-NEXT: movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm13, %xmm15 |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm9[0],xmm15[1],xmm9[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm15 = xmm15[0,1],xmm0[2,0] |
| ; SSE-NEXT: movdqa %xmm13, %xmm0 |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm10[0,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm14, %xmm12 |
| ; SSE-NEXT: movaps (%rsp), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm12 = xmm12[0],xmm3[0],xmm12[1],xmm3[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm12 = xmm12[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm4, %xmm10 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm10 = xmm10[2],xmm2[2],xmm10[3],xmm2[3] |
| ; SSE-NEXT: movaps %xmm14, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm10 = xmm10[0],xmm0[0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; SSE-NEXT: movdqa %xmm5, %xmm0 |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload |
| ; SSE-NEXT: movdqa %xmm13, %xmm11 |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm11 = xmm11[0],xmm8[0],xmm11[1],xmm8[1] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm11 = xmm11[0],xmm0[0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm13[1,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm9, %xmm8 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm8 = xmm8[0],xmm4[0],xmm8[1],xmm4[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm8 = xmm8[0,1],xmm0[0,2] |
| ; SSE-NEXT: movaps %xmm9, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3] |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm5 = xmm5[2],xmm2[2],xmm5[3],xmm2[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm5 = xmm5[0],xmm0[0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm4 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm4 = xmm4[2],xmm3[2],xmm4[3],xmm3[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm1[0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm3[2,0] |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: shufps $36, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm3 = xmm3[0,1],mem[2,0] |
| ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,2],xmm0[0,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,1],xmm6[2,0] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm2[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,0] |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm7 = xmm7[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm7[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm7[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm6[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, %xmm2 |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm7 = xmm7[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm7[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm6 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm6[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm6 = xmm6[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm7 = xmm7[0,3],xmm6[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm1[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm7 = xmm6[0],xmm7[1,2,3] |
| ; SSE-NEXT: shufps $255, (%rsp), %xmm1 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm14 = xmm14[0,3],xmm1[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm6 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm14 = xmm6[0],xmm14[1,2,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm13 = xmm13[0,3],xmm1[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm1[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm13 = xmm6[0],xmm13[1,2,3] |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm9 = xmm9[0,3],xmm1[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm6 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm9 = xmm6[0],xmm9[1,2,3] |
| ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; SSE-NEXT: movaps %xmm0, 416(%rax) |
| ; SSE-NEXT: movaps %xmm4, 400(%rax) |
| ; SSE-NEXT: movaps %xmm3, 384(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 352(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 336(%rax) |
| ; SSE-NEXT: movdqa %xmm5, 288(%rax) |
| ; SSE-NEXT: movaps %xmm8, 240(%rax) |
| ; SSE-NEXT: movdqa %xmm11, 224(%rax) |
| ; SSE-NEXT: movaps %xmm10, 176(%rax) |
| ; SSE-NEXT: movaps %xmm12, 128(%rax) |
| ; SSE-NEXT: movaps %xmm15, 112(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 64(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 16(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, (%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 432(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 368(%rax) |
| ; SSE-NEXT: movaps %xmm9, 320(%rax) |
| ; SSE-NEXT: movaps %xmm13, 304(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 272(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 256(%rax) |
| ; SSE-NEXT: movaps %xmm14, 208(%rax) |
| ; SSE-NEXT: movaps %xmm7, 192(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 160(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 144(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 96(%rax) |
| ; SSE-NEXT: movaps %xmm2, 80(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 48(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 32(%rax) |
| ; SSE-NEXT: addq $536, %rsp # imm = 0x218 |
| ; SSE-NEXT: retq |
| ; |
| ; AVX1-ONLY-LABEL: store_i32_stride7_vf16: |
| ; AVX1-ONLY: # %bb.0: |
| ; AVX1-ONLY-NEXT: subq $488, %rsp # imm = 0x1E8 |
| ; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rdi), %ymm5 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rsi), %ymm6 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rdx), %ymm4 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rcx), %ymm7 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%r8), %ymm0 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%r9), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vunpckhps {{.*#+}} ymm3 = ymm4[2],ymm7[2],ymm4[3],ymm7[3],ymm4[6],ymm7[6],ymm4[7],ymm7[7] |
| ; AVX1-ONLY-NEXT: vmovaps %ymm4, %ymm13 |
| ; AVX1-ONLY-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm4 = ymm6[1],ymm5[1],ymm6[3],ymm5[3] |
| ; AVX1-ONLY-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm4 = ymm5[0,1],ymm4[2,0],ymm5[4,5],ymm4[6,4] |
| ; AVX1-ONLY-NEXT: vmovaps %ymm5, %ymm8 |
| ; AVX1-ONLY-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm4[2,3],ymm3[2,3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm4 = ymm4[1,1,2,2,5,5,6,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1,2,3,4,5,6],ymm4[7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm2[2,3],ymm3[2,3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 32(%r9), %xmm9 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 32(%r8), %xmm10 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm10, (%rsp) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rdi), %xmm4 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rsi), %xmm11 |
| ; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm3 = xmm11[0],xmm4[0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm3 = xmm3[2,0],xmm4[2,1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm4, %xmm12 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm3, %ymm3, %ymm3 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rcx), %xmm5 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rdx), %xmm14 |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm4 = xmm14[0],xmm5[0],xmm14[1],xmm5[1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm14, %xmm15 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps %xmm5, %xmm14 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,0,1] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm4[2,3],ymm3[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm4 = xmm10[0],xmm9[0],xmm10[1],xmm9[1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm4 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rax), %xmm5 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm5[0],ymm4[2],ymm5[2] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6],ymm3[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm3 = xmm9[1,1],xmm10[1,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm3, %ymm3, %ymm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm5[5],ymm3[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm12[1],xmm11[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm4 = xmm11[1,1],xmm4[0,2] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm4, %ymm4, %ymm4 |
| ; AVX1-ONLY-NEXT: vinsertps {{.*#+}} xmm5 = zero,xmm15[1],xmm14[1],zero |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm5[1,2],ymm4[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3,4,5],ymm4[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm3 = ymm6[1,1],ymm8[1,1],ymm6[5,5],ymm8[5,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm3[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm4 = ymm13[1,1],ymm7[1,1],ymm13[5,5],ymm7[5,5] |
| ; AVX1-ONLY-NEXT: vmovaps %ymm7, %ymm13 |
| ; AVX1-ONLY-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm4[5,6],ymm3[7] |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm0[0],ymm1[2],ymm0[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm1[2,0],ymm0[2,1],ymm1[6,4],ymm0[6,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm0[2,3] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm0, %xmm0 |
| ; AVX1-ONLY-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm3[3,4,5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%rdi), %xmm12 |
| ; AVX1-ONLY-NEXT: vmovaps (%rsi), %xmm11 |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm12[1],xmm11[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm11[1,1],xmm0[0,2] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vmovaps (%rcx), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps (%rdx), %xmm9 |
| ; AVX1-ONLY-NEXT: vinsertps {{.*#+}} xmm1 = zero,xmm9[1],xmm2[1],zero |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2],ymm0[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovaps (%r9), %xmm15 |
| ; AVX1-ONLY-NEXT: vmovaps (%r8), %xmm14 |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm15[1,1],xmm14[1,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vmovaps (%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm2[5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm0 = xmm14[0],xmm15[0],xmm14[1],xmm15[1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm1 = xmm11[0],xmm12[0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm1[2,0],xmm12[2,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm2 = xmm9[0],xmm3[0],xmm9[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,0,1] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%rdi), %ymm10 |
| ; AVX1-ONLY-NEXT: vmovaps (%rsi), %ymm8 |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm8[1,1],ymm10[1,1],ymm8[5,5],ymm10[5,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps (%rdx), %ymm7 |
| ; AVX1-ONLY-NEXT: vmovaps (%rcx), %ymm6 |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm7[1,1],ymm6[1,1],ymm7[5,5],ymm6[5,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovaps (%r8), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovaps (%r9), %ymm1 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm4[2,0],ymm2[2,1],ymm4[6,4],ymm2[6,5] |
| ; AVX1-ONLY-NEXT: vmovaps (%rax), %ymm4 |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm4[2,3],ymm0[2,3] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm0, %xmm0 |
| ; AVX1-ONLY-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[2],ymm3[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm5[3,4,5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm8[1],ymm10[1],ymm8[3],ymm10[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,2,0,4,5,6,4] |
| ; AVX1-ONLY-NEXT: vunpckhps {{.*#+}} ymm3 = ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[6],ymm6[6],ymm7[7],ymm6[7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3] |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm3 = ymm2[1],ymm1[1],ymm2[3],ymm1[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm3 = ymm1[1,1],ymm3[0,2],ymm1[5,5],ymm3[4,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm3[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 16(%rax), %xmm5 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm5[1],ymm3[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1],ymm0[2,3,4,5],ymm3[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm0 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm0 = ymm13[3,3],mem[3,3],ymm13[7,7],mem[7,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm3 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm3 = ymm13[3,3],mem[3,3],ymm13[7,7],mem[7,7] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm3, %xmm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 60(%r8), %ymm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 60(%r9), %ymm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 56(%rax), %ymm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0],ymm0[1,2,3,4,5,6],ymm3[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm6[3,3],ymm7[3,3],ymm6[7,7],ymm7[7,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm3 = ymm10[3,3],ymm8[3,3],ymm10[7,7],ymm8[7,7] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm3, %xmm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm2[3,3],ymm1[3,3],ymm2[7,7],ymm1[7,7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm4[2,3],ymm1[1,2],ymm4[6,7],ymm1[5,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,2,3,1,4,6,7,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = xmm2[2],mem[2],xmm2[3],mem[3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm5[2,2,2,2] |
| ; AVX1-ONLY-NEXT: vmovaps (%rsp), %xmm0 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0,1,2],xmm1[3] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 40(%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3,4],ymm3[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm12[3,3],xmm11[3,3] |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm2 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm2 = xmm9[2],mem[2],xmm9[3],mem[3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm9, %ymm2, %ymm2 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5,6],ymm2[7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,2,2,2] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} xmm2 = xmm14[0,1,2],xmm2[3] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 8(%rax), %ymm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3,4],ymm1[5,6,7] |
| ; AVX1-ONLY-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm2 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm2 = ymm13[0],mem[0],ymm13[1],mem[1],ymm13[4],mem[4],ymm13[5],mem[5] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm3 = ymm4[3,1],ymm3[0,2],ymm4[7,5],ymm3[4,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5],ymm3[6,7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm3 = xmm0[3,3],xmm5[3,3] |
| ; AVX1-ONLY-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm3 = xmm3[0,1,2],mem[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm3[1,2,3],ymm2[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} ymm3 = ymm10[0],ymm8[0],ymm10[1],ymm8[1],ymm10[4],ymm8[4],ymm10[5],ymm8[5] |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm7[0],ymm6[0],ymm7[2],ymm6[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm4 = ymm6[3,1],ymm4[0,2],ymm6[7,5],ymm4[4,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5],ymm4[6,7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm4 = xmm14[3,3],xmm15[3,3] |
| ; AVX1-ONLY-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm4 = xmm4[0,1,2],mem[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm4[1,2,3],ymm3[4,5,6,7] |
| ; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX1-ONLY-NEXT: vmovaps %ymm3, 96(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm2, 320(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm1, 64(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 288(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 160(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 192(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 128(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 352(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 256(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 224(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 416(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 384(%rax) |
| ; AVX1-ONLY-NEXT: addq $488, %rsp # imm = 0x1E8 |
| ; AVX1-ONLY-NEXT: vzeroupper |
| ; AVX1-ONLY-NEXT: retq |
| ; |
| ; AVX2-SLOW-LABEL: store_i32_stride7_vf16: |
| ; AVX2-SLOW: # %bb.0: |
| ; AVX2-SLOW-NEXT: subq $504, %rsp # imm = 0x1F8 |
| ; AVX2-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-SLOW-NEXT: vmovaps (%rax), %xmm10 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm10, (%rsp) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rax), %xmm0 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vmovaps (%r8), %xmm13 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%r8), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%r9), %xmm14 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%r9), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps (%rcx), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rcx), %xmm11 |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rdx), %xmm9 |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm5 = zero,xmm9[1],xmm11[1],zero |
| ; AVX2-SLOW-NEXT: vmovaps (%rdi), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rdi), %xmm8 |
| ; AVX2-SLOW-NEXT: vmovaps (%rsi), %xmm12 |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rsi), %xmm7 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm7[1,1,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm8[2],xmm6[3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0],ymm5[1,2],ymm6[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1,2],ymm4[3,4,5],ymm5[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm10, %ymm0, %ymm4 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm5 = xmm14[1,1,1,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0],xmm13[1],xmm5[2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm5, %ymm5 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5],ymm5[6,7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm5 = xmm12[1,1,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm2[2],xmm5[3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm5[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vmovaps (%rdx), %xmm5 |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm10 = zero,xmm5[1],xmm1[1],zero |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm10[1,2],ymm6[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2],ymm4[3,4,5],ymm6[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%rdi), %ymm0 |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%rsi), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm4 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps (%rdx), %ymm0 |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%rcx), %ymm13 |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm6 = ymm0[2],ymm13[2],ymm0[3],ymm13[3],ymm0[6],ymm13[6],ymm0[7],ymm13[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps (%r8), %ymm0 |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%r9), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm6 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1],ymm0[2],ymm6[3,4,5],ymm0[6],ymm6[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vmovaps 16(%rax), %xmm14 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm14[1],ymm6[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1],ymm4[2,3,4,5],ymm6[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rdi), %ymm14 |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rsi), %ymm15 |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm4 = ymm14[2],ymm15[2],ymm14[3],ymm15[3],ymm14[6],ymm15[6],ymm14[7],ymm15[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm4[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rdx), %ymm4 |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rcx), %ymm6 |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm4[2],ymm6[2],ymm4[3],ymm6[3],ymm4[6],ymm6[6],ymm4[7],ymm6[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 32(%r8), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 32(%r9), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm10 = ymm2[1,2,2,3,5,6,6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm10[0],ymm1[1,2,3,4,5,6],ymm10[7] |
| ; AVX2-SLOW-NEXT: vperm2f128 {{.*#+}} ymm0 = mem[2,3],ymm0[2,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm6[2],ymm4[2],ymm6[3],ymm4[3],ymm6[6],ymm4[6],ymm6[7],ymm4[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm15[2],ymm14[2],ymm15[3],ymm14[3],ymm15[6],ymm14[6],ymm15[7],ymm14[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 60(%r8), %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 60(%r9), %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 56(%rax), %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm8[3,3],xmm7[3,3] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} xmm1 = xmm9[2],xmm11[2],xmm9[3],xmm11[3] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm2[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0,1,2],xmm1[3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 40(%rax), %ymm10 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm10[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3,4],ymm3[5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm11, %xmm3 |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm9, %xmm1 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm3 = xmm1[0],xmm3[0],xmm1[1],xmm3[1] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm1 = xmm8[0],xmm7[0],xmm8[1],xmm7[1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1],ymm3[2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm1 = xmm0[0],xmm2[0],xmm0[1],xmm2[1] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm7, %ymm2 |
| ; AVX2-SLOW-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm1[4,5,6],ymm3[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm10[3,3],xmm12[3,3] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} xmm1 = xmm5[2],xmm9[2],xmm5[3],xmm9[3] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm3[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm8[0,1,2],xmm1[3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 8(%rax), %ymm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm9, %xmm0 |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm5, %xmm1 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm1 = xmm10[0],xmm12[0],xmm10[1],xmm12[1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm1 = xmm8[0],xmm3[0],xmm8[1],xmm3[1] |
| ; AVX2-SLOW-NEXT: vmovaps %xmm8, %xmm10 |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-SLOW-NEXT: vmovaps (%rsp), %xmm11 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm11, %ymm2 |
| ; AVX2-SLOW-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm4[1,1],ymm6[1,1],ymm4[5,5],ymm6[5,5] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm2 = ymm15[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm14[1],ymm2[2,3,4],ymm14[5],ymm2[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5,6],ymm2[7] |
| ; AVX2-SLOW-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm2 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-SLOW-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm3 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3,4],ymm2[5],ymm3[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 48(%rax), %ymm3 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm9 = ymm2[0,1,2],ymm1[3,4,5,6],ymm2[7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm2 = ymm12[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm5[1],ymm2[2,3,4],ymm5[5],ymm2[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm2[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm3 = ymm2[1,1],ymm13[1,1],ymm2[5,5],ymm13[5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm8 = ymm0[0,1,2,3,4],ymm3[5,6],ymm0[7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm3 = ymm1[0,0,0,0,4,4,4,4] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm4 = ymm0[0,1,0,1,4,5,4,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3,4],ymm3[5],ymm4[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 16(%rax), %ymm4 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm4[2,3],ymm3[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm8 = ymm3[0,1,2],ymm8[3,4,5,6],ymm3[7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm3 = ymm13[2],ymm2[2],ymm13[3],ymm2[3],ymm13[6],ymm2[6],ymm13[7],ymm2[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm4 = ymm12[2],ymm5[2],ymm12[3],ymm5[3],ymm12[6],ymm5[6],ymm12[7],ymm5[7] |
| ; AVX2-SLOW-NEXT: vmovaps %ymm12, %ymm2 |
| ; AVX2-SLOW-NEXT: vmovaps %ymm5, %ymm12 |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm4 = ymm0[3,3],ymm1[3,3],ymm0[7,7],ymm1[7,7] |
| ; AVX2-SLOW-NEXT: vpermilps {{.*#+}} ymm5 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0],ymm4[1,2],ymm5[3,4],ymm4[5,6],ymm5[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1,2,3,4],ymm4[5,6,7] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} ymm4 = ymm14[0],ymm15[0],ymm14[1],ymm15[1],ymm14[4],ymm15[4],ymm14[5],ymm15[5] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm5 = ymm6[3,1,2,0,7,5,6,4] |
| ; AVX2-SLOW-NEXT: vbroadcastss 48(%rdx), %ymm6 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm6[6],ymm5[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5],ymm5[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm5 = xmm0[3,3],mem[3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0,1,2],xmm7[3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm5[1,2,3],ymm4[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} ymm5 = ymm12[0],ymm2[0],ymm12[1],ymm2[1],ymm12[4],ymm2[4],ymm12[5],ymm2[5] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm6 = ymm13[3,1,2,0,7,5,6,4] |
| ; AVX2-SLOW-NEXT: vbroadcastss 16(%rdx), %ymm7 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm7[6],ymm6[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5],ymm6[6,7] |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm6 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm6 = xmm10[3,3],mem[3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm6 = xmm6[0,1,2],xmm11[3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm6[1,2,3],ymm5[4,5,6,7] |
| ; AVX2-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-SLOW-NEXT: vmovaps %ymm5, 96(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm4, 320(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm3, 192(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm8, 128(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm9, 352(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm1, 160(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 64(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 256(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 224(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 288(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 416(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 384(%rax) |
| ; AVX2-SLOW-NEXT: addq $504, %rsp # imm = 0x1F8 |
| ; AVX2-SLOW-NEXT: vzeroupper |
| ; AVX2-SLOW-NEXT: retq |
| ; |
| ; AVX2-FAST-LABEL: store_i32_stride7_vf16: |
| ; AVX2-FAST: # %bb.0: |
| ; AVX2-FAST-NEXT: subq $536, %rsp # imm = 0x218 |
| ; AVX2-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-NEXT: vmovaps (%rax), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rax), %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vmovaps (%r8), %xmm7 |
| ; AVX2-FAST-NEXT: vmovaps %xmm7, (%rsp) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%r8), %xmm3 |
| ; AVX2-FAST-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%r9), %xmm8 |
| ; AVX2-FAST-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%r9), %xmm4 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm4[1,1,1,1] |
| ; AVX2-FAST-NEXT: vmovaps %xmm4, %xmm14 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm3[1],xmm1[2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovaps (%rcx), %xmm11 |
| ; AVX2-FAST-NEXT: vmovaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rcx), %xmm9 |
| ; AVX2-FAST-NEXT: vmovaps 32(%rdx), %xmm6 |
| ; AVX2-FAST-NEXT: vinsertps {{.*#+}} xmm1 = zero,xmm6[1],xmm9[1],zero |
| ; AVX2-FAST-NEXT: vmovaps (%rdi), %xmm12 |
| ; AVX2-FAST-NEXT: vmovaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rdi), %xmm5 |
| ; AVX2-FAST-NEXT: vmovaps (%rsi), %xmm10 |
| ; AVX2-FAST-NEXT: vmovaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rsi), %xmm3 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm4 = xmm3[1,1,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm5[2],xmm4[3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,2,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm4[0],ymm1[1,2],ymm4[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm8[1,1,1,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm7[1],xmm1[2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm10[1,1,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm12[2],xmm1[3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-NEXT: vmovaps (%rdx), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertps {{.*#+}} xmm4 = zero,xmm2[1],xmm11[1],zero |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm4[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%rdi), %ymm0 |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%rsi), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps (%rdx), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%rcx), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps (%r8), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%r9), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vmovaps 16(%rax), %xmm4 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm4[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rdi), %ymm13 |
| ; AVX2-FAST-NEXT: vmovaps 32(%rsi), %ymm12 |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm13[2],ymm12[2],ymm13[3],ymm12[3],ymm13[6],ymm12[6],ymm13[7],ymm12[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm4 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps 32(%rdx), %ymm2 |
| ; AVX2-FAST-NEXT: vmovaps 32(%rcx), %ymm11 |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm2[2],ymm11[2],ymm2[3],ymm11[3],ymm2[6],ymm11[6],ymm2[7],ymm11[7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps 32(%r8), %ymm10 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm10[6,7] |
| ; AVX2-FAST-NEXT: vmovaps 32(%r9), %ymm4 |
| ; AVX2-FAST-NEXT: vbroadcastsd {{.*#+}} ymm15 = [5,6,5,6,5,6,5,6] |
| ; AVX2-FAST-NEXT: vpermps %ymm4, %ymm15, %ymm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm15[0],ymm1[1,2,3,4,5,6],ymm15[7] |
| ; AVX2-FAST-NEXT: vperm2f128 {{.*#+}} ymm0 = mem[2,3],ymm0[2,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm11[2],ymm2[2],ymm11[3],ymm2[3],ymm11[6],ymm2[6],ymm11[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm12[2],ymm13[2],ymm12[3],ymm13[3],ymm12[6],ymm13[6],ymm12[7],ymm13[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 60(%r8), %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 60(%r9), %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastsd 56(%rax), %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm5[3,3],xmm3[3,3] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} xmm15 = xmm6[2],xmm9[2],xmm6[3],xmm9[3] |
| ; AVX2-FAST-NEXT: vbroadcastf128 {{.*#+}} ymm7 = [0,1,2,2,0,1,2,2] |
| ; AVX2-FAST-NEXT: # ymm7 = mem[0,1,0,1] |
| ; AVX2-FAST-NEXT: vpermps %ymm15, %ymm7, %ymm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm15[0,1,2,3,4],ymm1[5,6],ymm15[7] |
| ; AVX2-FAST-NEXT: vmovaps %xmm14, %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm15 = xmm14[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm15 = xmm8[0,1,2],xmm15[3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 40(%rax), %ymm14 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm14[2,3,4],ymm1[5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm9, %xmm1 |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm6, %xmm6 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm6[0],xmm1[0],xmm6[1],xmm1[1] |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1] |
| ; AVX2-FAST-NEXT: vpermps %ymm3, %ymm7, %ymm3 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5,6,7] |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm3 = xmm8[0],xmm0[0],xmm8[1],xmm0[1] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm3 |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm15, %ymm5 |
| ; AVX2-FAST-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm3[0],ymm5[0],ymm3[2],ymm5[2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm3[4,5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm2[1,1],ymm11[1,1],ymm2[5,5],ymm11[5,5] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm2 = ymm12[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm13[1],ymm2[2,3,4],ymm13[5],ymm2[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5,6],ymm2[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm2 = ymm4[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm3 = ymm10[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3,4],ymm2[5],ymm3[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[2,1,2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 48(%rax), %ymm3 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm1[3,4,5,6],ymm2[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm2[3,3],xmm6[3,3] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} xmm3 = xmm9[2],xmm0[2],xmm9[3],xmm0[3] |
| ; AVX2-FAST-NEXT: vpermps %ymm3, %ymm7, %ymm3 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3,4],ymm1[5,6],ymm3[7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm3 = xmm10[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps (%rsp), %xmm14 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm3 = xmm14[0,1,2],xmm3[3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 8(%rax), %ymm4 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm3[2,3,4],ymm1[5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm0, %xmm3 |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm9, %xmm4 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1] |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm4 = xmm2[0],xmm6[0],xmm2[1],xmm6[1] |
| ; AVX2-FAST-NEXT: vpermps %ymm4, %ymm7, %ymm0 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm3[2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm3 = xmm14[0],xmm10[0],xmm14[1],xmm10[1] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm3 |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm10, %ymm4 |
| ; AVX2-FAST-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6],ymm0[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm3 = ymm0[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm6[1],ymm3[2,3,4],ymm6[5],ymm3[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm2 = ymm3[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm4 = ymm3[1,1],ymm1[1,1],ymm3[5,5],ymm1[5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm9 = ymm2[0,1,2,3,4],ymm4[5,6],ymm2[7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm4 = ymm7[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm5 = ymm2[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3,4],ymm4[5],ymm5[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[2,1,2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 16(%rax), %ymm5 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm5[2,3],ymm4[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm9 = ymm4[0,1,2],ymm9[3,4,5,6],ymm4[7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm4 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7] |
| ; AVX2-FAST-NEXT: vmovaps %ymm1, %ymm3 |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[3,3,3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm5 = ymm0[2],ymm6[2],ymm0[3],ymm6[3],ymm0[6],ymm6[6],ymm0[7],ymm6[7] |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vmovaps %ymm6, %ymm0 |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm5 = ymm2[3,3],ymm7[3,3],ymm2[7,7],ymm7[7,7] |
| ; AVX2-FAST-NEXT: vpermilps {{.*#+}} ymm6 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0],ymm5[1,2],ymm6[3,4],ymm5[5,6],ymm6[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[2,1,2,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0],ymm4[1,2,3,4],ymm5[5,6,7] |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} ymm5 = ymm13[0],ymm12[0],ymm13[1],ymm12[1],ymm13[4],ymm12[4],ymm13[5],ymm12[5] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm6 = ymm11[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-NEXT: vbroadcastss 48(%rdx), %ymm7 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm7[6],ymm6[7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5],ymm6[6,7] |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm6 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm6 = xmm8[3,3],mem[3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm6 = xmm6[0,1,2],xmm15[3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm6[1,2,3],ymm5[4,5,6,7] |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} ymm6 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm7 = ymm3[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-NEXT: vbroadcastss 16(%rdx), %ymm8 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],ymm8[6],ymm7[7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5],ymm7[6,7] |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm7 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm7 = xmm14[3,3],mem[3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0,1,2],xmm10[3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm7[1,2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-NEXT: vmovaps %ymm6, 96(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm5, 320(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm4, 192(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm9, 128(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 64(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 352(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 224(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 288(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 160(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 256(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 416(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 384(%rax) |
| ; AVX2-FAST-NEXT: addq $536, %rsp # imm = 0x218 |
| ; AVX2-FAST-NEXT: vzeroupper |
| ; AVX2-FAST-NEXT: retq |
| ; |
| ; AVX2-FAST-PERLANE-LABEL: store_i32_stride7_vf16: |
| ; AVX2-FAST-PERLANE: # %bb.0: |
| ; AVX2-FAST-PERLANE-NEXT: subq $504, %rsp # imm = 0x1F8 |
| ; AVX2-FAST-PERLANE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rax), %xmm10 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm10, (%rsp) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rax), %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r8), %xmm13 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%r8), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r9), %xmm14 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%r9), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm4 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rcx), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rcx), %xmm11 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rdx), %xmm9 |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm5 = zero,xmm9[1],xmm11[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdi), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rdi), %xmm8 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rsi), %xmm12 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rsi), %xmm7 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm7[1,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm8[2],xmm6[3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0],ymm5[1,2],ymm6[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1,2],ymm4[3,4,5],ymm5[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm10, %ymm0, %ymm4 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm5 = xmm14[1,1,1,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0],xmm13[1],xmm5[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm5, %ymm5 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5],ymm5[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm5 = xmm12[1,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm2[2],xmm5[3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm5[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdx), %xmm5 |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm10 = zero,xmm5[1],xmm1[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm10[1,2],ymm6[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2],ymm4[3,4,5],ymm6[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdi), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rsi), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm4 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdx), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rcx), %ymm13 |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm6 = ymm0[2],ymm13[2],ymm0[3],ymm13[3],ymm0[6],ymm13[6],ymm0[7],ymm13[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r8), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r9), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm6 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1],ymm0[2],ymm6[3,4,5],ymm0[6],ymm6[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 16(%rax), %xmm14 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm14[1],ymm6[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1],ymm4[2,3,4,5],ymm6[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rdi), %ymm14 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rsi), %ymm15 |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm4 = ymm14[2],ymm15[2],ymm14[3],ymm15[3],ymm14[6],ymm15[6],ymm14[7],ymm15[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm4[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rdx), %ymm4 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rcx), %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm4[2],ymm6[2],ymm4[3],ymm6[3],ymm4[6],ymm6[6],ymm4[7],ymm6[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%r8), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%r9), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm10 = ymm2[1,2,2,3,5,6,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm10[0],ymm1[1,2,3,4,5,6],ymm10[7] |
| ; AVX2-FAST-PERLANE-NEXT: vperm2f128 {{.*#+}} ymm0 = mem[2,3],ymm0[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm6[2],ymm4[2],ymm6[3],ymm4[3],ymm6[6],ymm4[6],ymm6[7],ymm4[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm15[2],ymm14[2],ymm15[3],ymm14[3],ymm15[6],ymm14[6],ymm15[7],ymm14[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 60(%r8), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 60(%r9), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 56(%rax), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm8[3,3],xmm7[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} xmm1 = xmm9[2],xmm11[2],xmm9[3],xmm11[3] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm2[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0,1,2],xmm1[3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 40(%rax), %ymm10 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm10[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3,4],ymm3[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm11, %xmm3 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm9, %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm3 = xmm1[0],xmm3[0],xmm1[1],xmm3[1] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm1 = xmm8[0],xmm7[0],xmm8[1],xmm7[1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm1[0,1],ymm3[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm1 = xmm0[0],xmm2[0],xmm0[1],xmm2[1] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm7, %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm1[4,5,6],ymm3[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm10[3,3],xmm12[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} xmm1 = xmm5[2],xmm9[2],xmm5[3],xmm9[3] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm3[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm8[0,1,2],xmm1[3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 8(%rax), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm9, %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm5, %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm1 = xmm10[0],xmm12[0],xmm10[1],xmm12[1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm1 = xmm8[0],xmm3[0],xmm8[1],xmm3[1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm8, %xmm10 |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rsp), %xmm11 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm11, %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm4[1,1],ymm6[1,1],ymm4[5,5],ymm6[5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm2 = ymm15[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm14[1],ymm2[2,3,4],ymm14[5],ymm2[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5,6],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm2 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm3 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3,4],ymm2[5],ymm3[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 48(%rax), %ymm3 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm9 = ymm2[0,1,2],ymm1[3,4,5,6],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm2 = ymm12[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm5[1],ymm2[2,3,4],ymm5[5],ymm2[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm2[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm3 = ymm2[1,1],ymm13[1,1],ymm2[5,5],ymm13[5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm8 = ymm0[0,1,2,3,4],ymm3[5,6],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm3 = ymm1[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm4 = ymm0[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3,4],ymm3[5],ymm4[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 16(%rax), %ymm4 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm4[2,3],ymm3[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm8 = ymm3[0,1,2],ymm8[3,4,5,6],ymm3[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm3 = ymm13[2],ymm2[2],ymm13[3],ymm2[3],ymm13[6],ymm2[6],ymm13[7],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm4 = ymm12[2],ymm5[2],ymm12[3],ymm5[3],ymm12[6],ymm5[6],ymm12[7],ymm5[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm12, %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm5, %ymm12 |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm4 = ymm0[3,3],ymm1[3,3],ymm0[7,7],ymm1[7,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps {{.*#+}} ymm5 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0],ymm4[1,2],ymm5[3,4],ymm4[5,6],ymm5[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1,2,3,4],ymm4[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} ymm4 = ymm14[0],ymm15[0],ymm14[1],ymm15[1],ymm14[4],ymm15[4],ymm14[5],ymm15[5] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm5 = ymm6[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 48(%rdx), %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm6[6],ymm5[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5],ymm5[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm5 = xmm0[3,3],mem[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0,1,2],xmm7[3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm5[1,2,3],ymm4[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} ymm5 = ymm12[0],ymm2[0],ymm12[1],ymm2[1],ymm12[4],ymm2[4],ymm12[5],ymm2[5] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm6 = ymm13[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 16(%rdx), %ymm7 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm7[6],ymm6[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5],ymm6[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm6 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm6 = xmm10[3,3],mem[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm6 = xmm6[0,1,2],xmm11[3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm6[1,2,3],ymm5[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm5, 96(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm4, 320(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm3, 192(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm8, 128(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm9, 352(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm1, 160(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 64(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 256(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 224(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 288(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 416(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 384(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: addq $504, %rsp # imm = 0x1F8 |
| ; AVX2-FAST-PERLANE-NEXT: vzeroupper |
| ; AVX2-FAST-PERLANE-NEXT: retq |
| ; |
| ; AVX512F-LABEL: store_i32_stride7_vf16: |
| ; AVX512F: # %bb.0: |
| ; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX512F-NEXT: vmovdqa64 (%rdi), %zmm4 |
| ; AVX512F-NEXT: vmovdqa64 (%rsi), %zmm6 |
| ; AVX512F-NEXT: vmovdqa64 (%rdx), %zmm3 |
| ; AVX512F-NEXT: vmovdqa64 (%rcx), %zmm5 |
| ; AVX512F-NEXT: vmovdqa64 (%r8), %zmm1 |
| ; AVX512F-NEXT: vmovdqa64 (%r9), %zmm2 |
| ; AVX512F-NEXT: vmovdqa64 (%r10), %zmm0 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm7 = <u,u,u,u,14,30,u,u,u,u,u,15,31,u,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm5, %zmm3, %zmm7 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm8 = <u,u,14,30,u,u,u,u,u,15,31,u,u,u,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm6, %zmm4, %zmm8 |
| ; AVX512F-NEXT: movw $6192, %cx # imm = 0x1830 |
| ; AVX512F-NEXT: kmovw %ecx, %k1 |
| ; AVX512F-NEXT: vmovdqa32 %zmm7, %zmm8 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm7 = <13,u,u,u,u,u,30,14,u,u,u,u,u,31,15,u> |
| ; AVX512F-NEXT: vpermi2d %zmm1, %zmm2, %zmm7 |
| ; AVX512F-NEXT: movw $24769, %cx # imm = 0x60C1 |
| ; AVX512F-NEXT: kmovw %ecx, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm7, %zmm8 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm7 = [0,29,2,3,4,5,6,7,30,9,10,11,12,13,14,31] |
| ; AVX512F-NEXT: vpermi2d %zmm0, %zmm8, %zmm7 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm9 = <u,u,0,16,u,u,u,u,u,1,17,u,u,u,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm5, %zmm3, %zmm9 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm8 = <0,16,u,u,u,u,u,1,17,u,u,u,u,u,2,18> |
| ; AVX512F-NEXT: vpermi2d %zmm6, %zmm4, %zmm8 |
| ; AVX512F-NEXT: movw $1548, %cx # imm = 0x60C |
| ; AVX512F-NEXT: kmovw %ecx, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm9, %zmm8 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm9 = <u,u,u,u,0,16,u,u,u,u,u,1,17,u,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm2, %zmm1, %zmm9 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm10 = <u,u,u,u,4,5,16,u,u,u,u,11,12,17,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm0, %zmm9, %zmm10 |
| ; AVX512F-NEXT: movw $14448, %cx # imm = 0x3870 |
| ; AVX512F-NEXT: kmovw %ecx, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm10, %zmm8 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm10 = <u,u,u,u,u,3,19,u,u,u,u,u,4,20,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm6, %zmm4, %zmm10 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm9 = <2,18,u,u,u,u,u,3,19,u,u,u,u,u,4,20> |
| ; AVX512F-NEXT: vpermi2d %zmm5, %zmm3, %zmm9 |
| ; AVX512F-NEXT: movw $12384, %cx # imm = 0x3060 |
| ; AVX512F-NEXT: kmovw %ecx, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm10, %zmm9 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm10 = <u,u,2,18,u,u,u,u,u,3,19,u,u,u,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm2, %zmm1, %zmm10 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,u,2,3,18,u,u,u,u,9,10,19,u,u,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm0, %zmm10, %zmm11 |
| ; AVX512F-NEXT: movw $3612, %cx # imm = 0xE1C |
| ; AVX512F-NEXT: kmovw %ecx, %k3 |
| ; AVX512F-NEXT: vmovdqa32 %zmm11, %zmm9 {%k3} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm10 = <u,u,u,u,u,5,21,u,u,u,u,u,6,22,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm5, %zmm3, %zmm10 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,u,u,5,21,u,u,u,u,u,6,22,u,u,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm6, %zmm4, %zmm11 |
| ; AVX512F-NEXT: vmovdqa32 %zmm10, %zmm11 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm12 = <4,20,u,u,u,u,u,5,21,u,u,u,u,u,6,22> |
| ; AVX512F-NEXT: vpermi2d %zmm2, %zmm1, %zmm12 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm10 = <0,1,20,u,u,u,u,7,8,21,u,u,u,u,14,15> |
| ; AVX512F-NEXT: vpermi2d %zmm0, %zmm12, %zmm10 |
| ; AVX512F-NEXT: movw $15480, %cx # imm = 0x3C78 |
| ; AVX512F-NEXT: kmovw %ecx, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm11, %zmm10 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,u,u,7,23,u,u,u,u,u,8,24,u,u,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm5, %zmm3, %zmm11 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm12 = <u,7,23,u,u,u,u,u,8,24,u,u,u,u,u,9> |
| ; AVX512F-NEXT: vpermi2d %zmm6, %zmm4, %zmm12 |
| ; AVX512F-NEXT: movw $3096, %cx # imm = 0xC18 |
| ; AVX512F-NEXT: kmovw %ecx, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm11, %zmm12 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,u,u,u,u,7,23,u,u,u,u,u,8,24,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm2, %zmm1, %zmm11 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm13 = <22,u,u,u,u,5,6,23,u,u,u,u,12,13,24,u> |
| ; AVX512F-NEXT: vpermi2d %zmm0, %zmm11, %zmm13 |
| ; AVX512F-NEXT: movw $28897, %cx # imm = 0x70E1 |
| ; AVX512F-NEXT: kmovw %ecx, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm13, %zmm12 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,9,25,u,u,u,u,u,10,26,u,u,u,u,u,11> |
| ; AVX512F-NEXT: vpermi2d %zmm5, %zmm3, %zmm11 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm13 = <9,u,u,u,u,u,26,10,u,u,u,u,u,27,11,u> |
| ; AVX512F-NEXT: vpermi2d %zmm4, %zmm6, %zmm13 |
| ; AVX512F-NEXT: movw $-31994, %cx # imm = 0x8306 |
| ; AVX512F-NEXT: kmovw %ecx, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm11, %zmm13 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,u,u,9,25,u,u,u,u,u,10,26,u,u,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm2, %zmm1, %zmm11 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm14 = <u,u,u,3,4,25,u,u,u,u,10,11,26,u,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm0, %zmm11, %zmm14 |
| ; AVX512F-NEXT: movw $7224, %cx # imm = 0x1C38 |
| ; AVX512F-NEXT: kmovw %ecx, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm14, %zmm13 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,u,u,u,12,28,u,u,u,u,u,13,29,u,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm6, %zmm4, %zmm11 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm4 = <11,u,u,u,u,u,28,12,u,u,u,u,u,29,13,u> |
| ; AVX512F-NEXT: vpermi2d %zmm3, %zmm5, %zmm4 |
| ; AVX512F-NEXT: vmovdqa32 %zmm11, %zmm4 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm3 = <u,11,27,u,u,u,u,u,12,28,u,u,u,u,u,13> |
| ; AVX512F-NEXT: vpermi2d %zmm2, %zmm1, %zmm3 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm1 = <u,1,2,27,u,u,u,u,8,9,28,u,u,u,u,15> |
| ; AVX512F-NEXT: vpermi2d %zmm0, %zmm3, %zmm1 |
| ; AVX512F-NEXT: movw $-30962, %cx # imm = 0x870E |
| ; AVX512F-NEXT: kmovw %ecx, %k1 |
| ; AVX512F-NEXT: vmovdqa32 %zmm1, %zmm4 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 %zmm4, 320(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm13, 256(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm12, 192(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm10, 128(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm9, 64(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm8, (%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm7, 384(%rax) |
| ; AVX512F-NEXT: vzeroupper |
| ; AVX512F-NEXT: retq |
| ; |
| ; AVX512BW-LABEL: store_i32_stride7_vf16: |
| ; AVX512BW: # %bb.0: |
| ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10 |
| ; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm4 |
| ; AVX512BW-NEXT: vmovdqa64 (%rsi), %zmm6 |
| ; AVX512BW-NEXT: vmovdqa64 (%rdx), %zmm3 |
| ; AVX512BW-NEXT: vmovdqa64 (%rcx), %zmm5 |
| ; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm1 |
| ; AVX512BW-NEXT: vmovdqa64 (%r9), %zmm2 |
| ; AVX512BW-NEXT: vmovdqa64 (%r10), %zmm0 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm7 = <u,u,u,u,14,30,u,u,u,u,u,15,31,u,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm5, %zmm3, %zmm7 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm8 = <u,u,14,30,u,u,u,u,u,15,31,u,u,u,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm8 |
| ; AVX512BW-NEXT: movw $6192, %cx # imm = 0x1830 |
| ; AVX512BW-NEXT: kmovd %ecx, %k1 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm8 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm7 = <13,u,u,u,u,u,30,14,u,u,u,u,u,31,15,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm1, %zmm2, %zmm7 |
| ; AVX512BW-NEXT: movw $24769, %cx # imm = 0x60C1 |
| ; AVX512BW-NEXT: kmovd %ecx, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm8 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm7 = [0,29,2,3,4,5,6,7,30,9,10,11,12,13,14,31] |
| ; AVX512BW-NEXT: vpermi2d %zmm0, %zmm8, %zmm7 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm9 = <u,u,0,16,u,u,u,u,u,1,17,u,u,u,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm5, %zmm3, %zmm9 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm8 = <0,16,u,u,u,u,u,1,17,u,u,u,u,u,2,18> |
| ; AVX512BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm8 |
| ; AVX512BW-NEXT: movw $1548, %cx # imm = 0x60C |
| ; AVX512BW-NEXT: kmovd %ecx, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm9, %zmm8 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm9 = <u,u,u,u,0,16,u,u,u,u,u,1,17,u,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm9 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm10 = <u,u,u,u,4,5,16,u,u,u,u,11,12,17,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm0, %zmm9, %zmm10 |
| ; AVX512BW-NEXT: movw $14448, %cx # imm = 0x3870 |
| ; AVX512BW-NEXT: kmovd %ecx, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm10, %zmm8 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm10 = <u,u,u,u,u,3,19,u,u,u,u,u,4,20,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm10 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm9 = <2,18,u,u,u,u,u,3,19,u,u,u,u,u,4,20> |
| ; AVX512BW-NEXT: vpermi2d %zmm5, %zmm3, %zmm9 |
| ; AVX512BW-NEXT: movw $12384, %cx # imm = 0x3060 |
| ; AVX512BW-NEXT: kmovd %ecx, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm10, %zmm9 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm10 = <u,u,2,18,u,u,u,u,u,3,19,u,u,u,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm10 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,u,2,3,18,u,u,u,u,9,10,19,u,u,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm0, %zmm10, %zmm11 |
| ; AVX512BW-NEXT: movw $3612, %cx # imm = 0xE1C |
| ; AVX512BW-NEXT: kmovd %ecx, %k3 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm11, %zmm9 {%k3} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm10 = <u,u,u,u,u,5,21,u,u,u,u,u,6,22,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm5, %zmm3, %zmm10 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,u,u,5,21,u,u,u,u,u,6,22,u,u,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm11 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm10, %zmm11 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm12 = <4,20,u,u,u,u,u,5,21,u,u,u,u,u,6,22> |
| ; AVX512BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm12 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm10 = <0,1,20,u,u,u,u,7,8,21,u,u,u,u,14,15> |
| ; AVX512BW-NEXT: vpermi2d %zmm0, %zmm12, %zmm10 |
| ; AVX512BW-NEXT: movw $15480, %cx # imm = 0x3C78 |
| ; AVX512BW-NEXT: kmovd %ecx, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm11, %zmm10 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,u,u,7,23,u,u,u,u,u,8,24,u,u,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm5, %zmm3, %zmm11 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm12 = <u,7,23,u,u,u,u,u,8,24,u,u,u,u,u,9> |
| ; AVX512BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm12 |
| ; AVX512BW-NEXT: movw $3096, %cx # imm = 0xC18 |
| ; AVX512BW-NEXT: kmovd %ecx, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm11, %zmm12 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,u,u,u,u,7,23,u,u,u,u,u,8,24,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm11 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm13 = <22,u,u,u,u,5,6,23,u,u,u,u,12,13,24,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm0, %zmm11, %zmm13 |
| ; AVX512BW-NEXT: movw $28897, %cx # imm = 0x70E1 |
| ; AVX512BW-NEXT: kmovd %ecx, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm13, %zmm12 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,9,25,u,u,u,u,u,10,26,u,u,u,u,u,11> |
| ; AVX512BW-NEXT: vpermi2d %zmm5, %zmm3, %zmm11 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm13 = <9,u,u,u,u,u,26,10,u,u,u,u,u,27,11,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm4, %zmm6, %zmm13 |
| ; AVX512BW-NEXT: movw $-31994, %cx # imm = 0x8306 |
| ; AVX512BW-NEXT: kmovd %ecx, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm11, %zmm13 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,u,u,9,25,u,u,u,u,u,10,26,u,u,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm11 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm14 = <u,u,u,3,4,25,u,u,u,u,10,11,26,u,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm0, %zmm11, %zmm14 |
| ; AVX512BW-NEXT: movw $7224, %cx # imm = 0x1C38 |
| ; AVX512BW-NEXT: kmovd %ecx, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm14, %zmm13 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,u,u,u,12,28,u,u,u,u,u,13,29,u,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm11 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm4 = <11,u,u,u,u,u,28,12,u,u,u,u,u,29,13,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm3, %zmm5, %zmm4 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm11, %zmm4 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm3 = <u,11,27,u,u,u,u,u,12,28,u,u,u,u,u,13> |
| ; AVX512BW-NEXT: vpermi2d %zmm2, %zmm1, %zmm3 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm1 = <u,1,2,27,u,u,u,u,8,9,28,u,u,u,u,15> |
| ; AVX512BW-NEXT: vpermi2d %zmm0, %zmm3, %zmm1 |
| ; AVX512BW-NEXT: movw $-30962, %cx # imm = 0x870E |
| ; AVX512BW-NEXT: kmovd %ecx, %k1 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm1, %zmm4 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 %zmm4, 320(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm13, 256(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm12, 192(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm10, 128(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm9, 64(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm8, (%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm7, 384(%rax) |
| ; AVX512BW-NEXT: vzeroupper |
| ; AVX512BW-NEXT: retq |
| %in.vec0 = load <16 x i32>, ptr %in.vecptr0, align 64 |
| %in.vec1 = load <16 x i32>, ptr %in.vecptr1, align 64 |
| %in.vec2 = load <16 x i32>, ptr %in.vecptr2, align 64 |
| %in.vec3 = load <16 x i32>, ptr %in.vecptr3, align 64 |
| %in.vec4 = load <16 x i32>, ptr %in.vecptr4, align 64 |
| %in.vec5 = load <16 x i32>, ptr %in.vecptr5, align 64 |
| %in.vec6 = load <16 x i32>, ptr %in.vecptr6, align 64 |
| %1 = shufflevector <16 x i32> %in.vec0, <16 x i32> %in.vec1, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31> |
| %2 = shufflevector <16 x i32> %in.vec2, <16 x i32> %in.vec3, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31> |
| %3 = shufflevector <16 x i32> %in.vec4, <16 x i32> %in.vec5, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31> |
| %4 = shufflevector <32 x i32> %1, <32 x i32> %2, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63> |
| %5 = shufflevector <16 x i32> %in.vec6, <16 x i32> poison, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef> |
| %6 = shufflevector <32 x i32> %3, <32 x i32> %5, <48 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47> |
| %7 = shufflevector <48 x i32> %6, <48 x i32> poison, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef> |
| %8 = shufflevector <64 x i32> %4, <64 x i32> %7, <112 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111> |
| %interleaved.vec = shufflevector <112 x i32> %8, <112 x i32> poison, <112 x i32> <i32 0, i32 16, i32 32, i32 48, i32 64, i32 80, i32 96, i32 1, i32 17, i32 33, i32 49, i32 65, i32 81, i32 97, i32 2, i32 18, i32 34, i32 50, i32 66, i32 82, i32 98, i32 3, i32 19, i32 35, i32 51, i32 67, i32 83, i32 99, i32 4, i32 20, i32 36, i32 52, i32 68, i32 84, i32 100, i32 5, i32 21, i32 37, i32 53, i32 69, i32 85, i32 101, i32 6, i32 22, i32 38, i32 54, i32 70, i32 86, i32 102, i32 7, i32 23, i32 39, i32 55, i32 71, i32 87, i32 103, i32 8, i32 24, i32 40, i32 56, i32 72, i32 88, i32 104, i32 9, i32 25, i32 41, i32 57, i32 73, i32 89, i32 105, i32 10, i32 26, i32 42, i32 58, i32 74, i32 90, i32 106, i32 11, i32 27, i32 43, i32 59, i32 75, i32 91, i32 107, i32 12, i32 28, i32 44, i32 60, i32 76, i32 92, i32 108, i32 13, i32 29, i32 45, i32 61, i32 77, i32 93, i32 109, i32 14, i32 30, i32 46, i32 62, i32 78, i32 94, i32 110, i32 15, i32 31, i32 47, i32 63, i32 79, i32 95, i32 111> |
| store <112 x i32> %interleaved.vec, ptr %out.vec, align 64 |
| ret void |
| } |
| |
| define void @store_i32_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %out.vec) nounwind { |
| ; SSE-LABEL: store_i32_stride7_vf32: |
| ; SSE: # %bb.0: |
| ; SSE-NEXT: subq $1256, %rsp # imm = 0x4E8 |
| ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; SSE-NEXT: movdqa (%rdi), %xmm4 |
| ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa (%rsi), %xmm7 |
| ; SSE-NEXT: movdqa 16(%rsi), %xmm5 |
| ; SSE-NEXT: movaps (%rdx), %xmm9 |
| ; SSE-NEXT: movdqa 16(%rdx), %xmm12 |
| ; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps (%rcx), %xmm10 |
| ; SSE-NEXT: movaps 16(%rcx), %xmm13 |
| ; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps (%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 16(%r8), %xmm14 |
| ; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa (%r9), %xmm2 |
| ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 16(%r9), %xmm15 |
| ; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa (%rax), %xmm11 |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm10[1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm7[1,1,1,1] |
| ; SSE-NEXT: movaps %xmm9, %xmm3 |
| ; SSE-NEXT: movss {{.*#+}} xmm3 = xmm1[0],xmm3[1,2,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm4, %xmm0 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm7[2],xmm0[3],xmm7[3] |
| ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm11[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm2[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm12[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movaps %xmm14, %xmm0 |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm13[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 16(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm15[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 16(%rdi), %xmm14 |
| ; SSE-NEXT: movdqa %xmm14, %xmm0 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm5[2],xmm0[3],xmm5[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 32(%rsi), %xmm1 |
| ; SSE-NEXT: movaps 32(%rdx), %xmm2 |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm1, %xmm8 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm2, %xmm1 |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps 32(%rcx), %xmm2 |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 32(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm2[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 32(%r9), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 32(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 32(%rdi), %xmm6 |
| ; SSE-NEXT: movdqa %xmm6, %xmm0 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm8[2],xmm0[3],xmm8[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 48(%rsi), %xmm8 |
| ; SSE-NEXT: movdqa 48(%rdx), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm8[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm8, %xmm12 |
| ; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movaps 48(%rcx), %xmm13 |
| ; SSE-NEXT: movaps 48(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm13[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 48(%r9), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 48(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 48(%rdi), %xmm8 |
| ; SSE-NEXT: movdqa %xmm8, %xmm0 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm12[2],xmm0[3],xmm12[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 64(%rsi), %xmm1 |
| ; SSE-NEXT: movaps 64(%rdx), %xmm2 |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm1, %xmm12 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm2, %xmm1 |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps 64(%rcx), %xmm2 |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 64(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm2[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 64(%r9), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 64(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 64(%rdi), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm12[2],xmm0[3],xmm12[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 80(%rsi), %xmm12 |
| ; SSE-NEXT: movdqa 80(%rdx), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm12[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movaps 80(%rcx), %xmm2 |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 80(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm2[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 80(%r9), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 80(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 80(%rdi), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm12[2],xmm0[3],xmm12[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 96(%rsi), %xmm4 |
| ; SSE-NEXT: movaps 96(%rdx), %xmm12 |
| ; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm12, %xmm1 |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps 96(%rcx), %xmm3 |
| ; SSE-NEXT: movaps 96(%r8), %xmm15 |
| ; SSE-NEXT: movaps %xmm15, %xmm0 |
| ; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm3[1,1] |
| ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 96(%r9), %xmm2 |
| ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 96(%rax), %xmm12 |
| ; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm12[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 96(%rdi), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm15[3,3,3,3] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm3[3,3,3,3] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa %xmm12, %xmm0 |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm2[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 112(%rsi), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 112(%rdx), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1] |
| ; SSE-NEXT: movaps 112(%rcx), %xmm3 |
| ; SSE-NEXT: movaps 112(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, %xmm1 |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm3[1,1] |
| ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm1[2,0] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm0, %xmm1 |
| ; SSE-NEXT: movaps 112(%r9), %xmm15 |
| ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm15[0] |
| ; SSE-NEXT: movaps %xmm1, (%rsp) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm15, %xmm1 |
| ; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: unpckhpd {{.*#+}} xmm15 = xmm15[1],xmm0[1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[3,3,3,3] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1] |
| ; SSE-NEXT: movaps 112(%rax), %xmm12 |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm12[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm1[0,2] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm9, %xmm0 |
| ; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movdqa %xmm2, %xmm1 |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm7[0],xmm1[1],xmm7[1] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm11 = xmm11[0,1],xmm2[1,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm3, %xmm1 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm11[0,2] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm3, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3] |
| ; SSE-NEXT: movaps %xmm9, %xmm1 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm1 = xmm1[2],xmm10[2],xmm1[3],xmm10[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm5[0] |
| ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm14, %xmm2 |
| ; SSE-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm2 = xmm2[0],mem[0],xmm2[1],mem[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $197, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm14 = xmm14[1,1],mem[0,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm2, %xmm3 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,1],xmm14[2,0] |
| ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm5, %xmm3 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3] |
| ; SSE-NEXT: movaps %xmm2, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movdqa %xmm2, %xmm0 |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| ; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm6, %xmm3 |
| ; SSE-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm6[1,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm5, %xmm4 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,1],xmm0[0,2] |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm5, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3] |
| ; SSE-NEXT: movdqa %xmm2, %xmm3 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm13, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| ; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm8, %xmm3 |
| ; SSE-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $197, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm8 = xmm8[1,1],mem[0,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm5, %xmm4 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,1],xmm8[2,0] |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm1, %xmm4 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm4 = xmm4[2],xmm13[2],xmm4[3],xmm13[3] |
| ; SSE-NEXT: movaps %xmm5, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm3, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm4 |
| ; SSE-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = xmm4[0],mem[0],xmm4[1],mem[1] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[1,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm5 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm0[0,2] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3] |
| ; SSE-NEXT: movaps %xmm3, %xmm1 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm1 = xmm1[2],xmm2[2],xmm1[3],xmm2[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm3, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload |
| ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm10[0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm13 |
| ; SSE-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm13 = xmm13[0],mem[0],xmm13[1],mem[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm13 = xmm13[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: shufps $197, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[1,1],mem[0,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm11 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm11 = xmm11[0],xmm4[0],xmm11[1],xmm4[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm11 = xmm11[0,1],xmm0[2,0] |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm10 = xmm10[2],xmm3[2],xmm10[3],xmm3[3] |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm10 = xmm10[0],xmm0[0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm7, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm14, %xmm9 |
| ; SSE-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm9 = xmm9[0],mem[0],xmm9[1],mem[1] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm9 = xmm9[0],xmm0[0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm14[1,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm8 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm8 = xmm8[0],xmm4[0],xmm8[1],xmm4[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm8 = xmm8[0,1],xmm0[0,2] |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm1 = xmm1[2],xmm4[2],xmm1[3],xmm4[3] |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm7 = xmm7[2],xmm3[2],xmm7[3],xmm3[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm1[0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm2, %xmm5 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm5 = xmm5[2],xmm6[2],xmm5[3],xmm6[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm6 = xmm6[0],xmm2[0] |
| ; SSE-NEXT: movaps 112(%rdi), %xmm3 |
| ; SSE-NEXT: movaps %xmm3, %xmm4 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,1],xmm6[2,0] |
| ; SSE-NEXT: movaps %xmm3, %xmm0 |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm12[0,3] |
| ; SSE-NEXT: movaps (%rsp), %xmm6 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,2],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm6, (%rsp) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm3, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[1,1],xmm12[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,2],xmm0[0,1] |
| ; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm15[2,0] |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[3,3],xmm1[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[3,1],xmm12[2,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[2,0],xmm2[2,0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm15[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm15 = xmm15[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm15[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,3],xmm1[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,3],xmm1[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm15 = xmm15[0,3],xmm1[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm15 = xmm0[0],xmm15[1,2,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm12 = xmm12[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm12 = xmm0[0],xmm12[1,2,3] |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,3],xmm1[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm6 = xmm0[0],xmm6[1,2,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3] |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm14 = xmm14[0,3],xmm0[2,0] |
| ; SSE-NEXT: pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = mem[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm14 = xmm0[0],xmm14[1,2,3] |
| ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; SSE-NEXT: movaps %xmm3, 864(%rax) |
| ; SSE-NEXT: movaps %xmm5, 848(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 832(%rax) |
| ; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 800(%rax) |
| ; SSE-NEXT: movaps %xmm4, 784(%rax) |
| ; SSE-NEXT: movaps %xmm7, 736(%rax) |
| ; SSE-NEXT: movaps %xmm8, 688(%rax) |
| ; SSE-NEXT: movaps %xmm9, 672(%rax) |
| ; SSE-NEXT: movaps %xmm10, 624(%rax) |
| ; SSE-NEXT: movaps %xmm11, 576(%rax) |
| ; SSE-NEXT: movaps %xmm13, 560(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 512(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 464(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 448(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 400(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 352(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 336(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 288(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 240(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 224(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 176(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 128(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 112(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 64(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 16(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, (%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 880(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 816(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 768(%rax) |
| ; SSE-NEXT: movaps %xmm14, 752(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 720(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 704(%rax) |
| ; SSE-NEXT: movaps %xmm1, 656(%rax) |
| ; SSE-NEXT: movaps %xmm2, 640(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 608(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 592(%rax) |
| ; SSE-NEXT: movaps %xmm6, 544(%rax) |
| ; SSE-NEXT: movaps %xmm12, 528(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 496(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 480(%rax) |
| ; SSE-NEXT: movaps %xmm15, 432(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 416(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 384(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 368(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 320(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 304(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 272(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 256(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 208(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 192(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 160(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 144(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 96(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 80(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 48(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 32(%rax) |
| ; SSE-NEXT: addq $1256, %rsp # imm = 0x4E8 |
| ; SSE-NEXT: retq |
| ; |
| ; AVX1-ONLY-LABEL: store_i32_stride7_vf32: |
| ; AVX1-ONLY: # %bb.0: |
| ; AVX1-ONLY-NEXT: subq $1624, %rsp # imm = 0x658 |
| ; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rdi), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rsi), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rdx), %ymm4 |
| ; AVX1-ONLY-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rcx), %ymm0 |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rax), %ymm3 |
| ; AVX1-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhps {{.*#+}} ymm0 = ymm4[2],ymm0[2],ymm4[3],ymm0[3],ymm4[6],ymm0[6],ymm4[7],ymm0[7] |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,0],ymm2[4,5],ymm1[6,4] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = mem[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm3[2,3],ymm0[2,3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%rax), %xmm0 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vmovaps (%r9), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%r8), %xmm4 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%rdi), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps (%rsi), %xmm5 |
| ; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm2[0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm1[2,0],xmm2[2,1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, %xmm6 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vmovaps (%rcx), %xmm7 |
| ; AVX1-ONLY-NEXT: vmovaps (%rdx), %xmm8 |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm2 = xmm8[0],xmm7[0],xmm8[1],xmm7[1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,0,1] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm2[0],ymm0[0],ymm2[2],ymm0[2] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm3[1,1],xmm4[1,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm6[1],xmm5[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm5[1,1],xmm1[0,2] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm8[1],xmm7[1],zero |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%rdi), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%rsi), %ymm0 |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,1],ymm1[1,1],ymm0[5,5],ymm1[5,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps (%rdx), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%rcx), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1],ymm2[1,1],ymm1[5,5],ymm2[5,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovaps (%r8), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%r9), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm2[2,1],ymm1[6,4],ymm2[6,5] |
| ; AVX1-ONLY-NEXT: vmovaps (%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rdi), %xmm1 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rsi), %xmm8 |
| ; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm8[0],xmm1[0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, %xmm6 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rcx), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rdx), %xmm3 |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, %xmm9 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, %xmm7 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,0,1] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovaps 32(%r9), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 32(%r8), %xmm4 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm3[1,1],xmm4[1,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm6[1],xmm8[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm8[1,1],xmm1[0,2] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm9[1],xmm7[1],zero |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rdi), %ymm8 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rsi), %ymm0 |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,1],ymm8[1,1],ymm0[5,5],ymm8[5,5] |
| ; AVX1-ONLY-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rdx), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rcx), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1],ymm2[1,1],ymm1[5,5],ymm2[5,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovaps 32(%r8), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 32(%r9), %ymm11 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm11[0],ymm2[0],ymm11[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm2[2,1],ymm1[6,4],ymm2[6,5] |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rdi), %xmm1 |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rsi), %xmm7 |
| ; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm7[0],xmm1[0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, %xmm6 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rcx), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rdx), %xmm3 |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, %xmm9 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, %xmm10 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,0,1] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovaps 64(%r9), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 64(%r8), %xmm4 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm3[1,1],xmm4[1,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm6[1],xmm7[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm7[1,1],xmm1[0,2] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm9[1],xmm10[1],zero |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rdi), %ymm15 |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rsi), %ymm0 |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,1],ymm15[1,1],ymm0[5,5],ymm15[5,5] |
| ; AVX1-ONLY-NEXT: vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rdx), %ymm10 |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rcx), %ymm9 |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm10[1,1],ymm9[1,1],ymm10[5,5],ymm9[5,5] |
| ; AVX1-ONLY-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovaps 64(%r8), %ymm12 |
| ; AVX1-ONLY-NEXT: vmovaps 64(%r9), %ymm13 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm13[0],ymm12[0],ymm13[2],ymm12[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm12[2,1],ymm1[6,4],ymm12[6,5] |
| ; AVX1-ONLY-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rdi), %xmm6 |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rsi), %xmm7 |
| ; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm7[0],xmm6[0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm6[2,1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rcx), %xmm5 |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rdx), %xmm3 |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm14 = xmm3[0],xmm5[0],xmm3[1],xmm5[1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, (%rsp) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm14 = xmm14[0,1,0,1] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1],ymm14[2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovaps 96(%r9), %xmm0 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 96(%r8), %xmm4 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm14 = xmm4[0],xmm0[0],xmm4[1],xmm0[1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm14, %ymm0, %ymm14 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, 96(%rax), %ymm0, %ymm2 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm14[0],ymm2[0],ymm14[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm14[4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm0[1,1],xmm4[1,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm2[5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm6[1],xmm7[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm7[1,1],xmm1[0,2] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertps {{.*#+}} xmm14 = zero,xmm3[1],xmm5[1],zero |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm14[1,2],ymm1[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,2,0,4,5,6,4] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{.*#+}} ymm1 = ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[6],ymm3[6],ymm4[7],ymm3[7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm2[1,1],ymm1[0,2],ymm2[5,5],ymm1[4,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 16(%rax), %xmm14 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm14[1],ymm1[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm8[1],ymm0[3],ymm8[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,2,0,4,5,6,4] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{.*#+}} ymm1 = ymm8[2],ymm5[2],ymm8[3],ymm5[3],ymm8[6],ymm5[6],ymm8[7],ymm5[7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm11[1],ymm1[3],ymm11[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm11[1,1],ymm1[0,2],ymm11[5,5],ymm1[4,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 48(%rax), %xmm14 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm14[1],ymm1[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm15[1],ymm0[3],ymm15[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,2,0,4,5,6,4] |
| ; AVX1-ONLY-NEXT: vunpckhps {{.*#+}} ymm1 = ymm10[2],ymm9[2],ymm10[3],ymm9[3],ymm10[6],ymm9[6],ymm10[7],ymm9[7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm12[1],ymm13[1],ymm12[3],ymm13[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,1],ymm1[0,2],ymm13[5,5],ymm1[4,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 80(%rax), %xmm14 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm14[1],ymm1[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm7[3,3],ymm2[3,3],ymm7[7,7],ymm2[7,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm6[3,3],ymm14[3,3],ymm6[7,7],ymm14[7,7] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 124(%r8), %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 124(%r9), %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 120(%rax), %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm2[0],ymm7[0],ymm2[2],ymm7[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm7[3,1],ymm0[0,2],ymm7[7,5],ymm0[4,6] |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} ymm1 = ymm6[0],ymm14[0],ymm6[1],ymm14[1],ymm6[4],ymm14[4],ymm6[5],ymm14[5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 108(%r8), %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 108(%r9), %xmm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1],ymm7[1,1],ymm2[5,5],ymm7[5,5] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm14[1,1],ymm6[1,1],ymm14[5,5],ymm6[5,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 112(%r8), %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 112(%r9), %xmm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 112(%rax), %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = xmm6[2],mem[2],xmm6[3],mem[3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm6, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = mem[2,2,2,2] |
| ; AVX1-ONLY-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = mem[0,1,2],xmm1[3] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 8(%rax), %ymm14 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm14[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm3[3,3],ymm4[3,3],ymm3[7,7],ymm4[7,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,3],ymm1[1,2],ymm2[6,7],ymm1[5,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,2,3,1,4,6,7,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = xmm2[2],mem[2],xmm2[3],mem[3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm15[2,2,2,2] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} xmm1 = xmm14[0,1,2],xmm1[3] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 40(%rax), %ymm13 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm13[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm5[3,3],ymm8[3,3],ymm5[7,7],ymm8[7,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm13 = ymm7[3,3],ymm0[3,3],ymm7[7,7],ymm0[7,7] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm13, %xmm13 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm13[0,1,2],ymm1[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm11 = ymm2[3,3],ymm11[3,3],ymm2[7,7],ymm11[7,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm11 = ymm2[2,3],ymm11[1,2],ymm2[6,7],ymm11[5,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm11 = ymm11[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm11 = ymm11[0,2,3,1,4,6,7,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm11[0],ymm1[1,2,3,4],ymm11[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm9 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm9 = xmm1[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm10 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm10 = xmm1[2],mem[2],xmm1[3],mem[3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm10, %ymm8 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm9 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4],ymm9[5,6],ymm8[7] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm9 = xmm12[2,2,2,2] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} xmm9 = xmm13[0,1,2],xmm9[3] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 72(%rax), %ymm10 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm8[0,1],ymm9[2,3,4],ymm8[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm9 = ymm8[3,3],ymm11[3,3],ymm8[7,7],ymm11[7,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm9[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm10 = ymm6[3,3],ymm1[3,3],ymm6[7,7],ymm1[7,7] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm10, %xmm10 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2],ymm9[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm4 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm4 = ymm2[3,3],mem[3,3],ymm2[7,7],mem[7,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm4 = ymm2[2,3],ymm4[1,2],ymm2[6,7],ymm4[5,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,2,3,1,4,6,7,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm9[1,2,3,4],ymm4[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm3 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm3 = xmm2[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vmovaps (%rsp), %xmm2 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm5 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm5 = xmm2[2],mem[2],xmm2[3],mem[3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm5, %ymm2 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm3[5,6],ymm2[7] |
| ; AVX1-ONLY-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm3 = mem[2,2,2,2] |
| ; AVX1-ONLY-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm3 = mem[0,1,2],xmm3[3] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 104(%rax), %ymm5 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3,4],ymm2[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm3 = ymm3[0],mem[0],ymm3[1],mem[1],ymm3[4],mem[4],ymm3[5],mem[5] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm5 = ymm5[0],ymm9[0],ymm5[2],ymm9[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm5 = ymm9[3,1],ymm5[0,2],ymm9[7,5],ymm5[4,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm3[4,5],ymm5[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm5 = xmm5[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm5 = xmm5[0,1,2],mem[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm5[1,2,3],ymm3[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} ymm5 = ymm7[0],ymm0[0],ymm7[1],ymm0[1],ymm7[4],ymm0[4],ymm7[5],ymm0[5] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm7[0],ymm0[0],ymm7[2],ymm0[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm7 = ymm0[3,1],ymm7[0,2],ymm0[7,5],ymm7[4,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3],ymm5[4,5],ymm7[6,7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm7 = xmm14[3,3],xmm15[3,3] |
| ; AVX1-ONLY-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm7 = xmm7[0,1,2],mem[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm7[1,2,3],ymm5[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} ymm7 = ymm6[0],ymm1[0],ymm6[1],ymm1[1],ymm6[4],ymm1[4],ymm6[5],ymm1[5] |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm11[0],ymm8[0],ymm11[2],ymm8[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm6 = ymm8[3,1],ymm9[0,2],ymm8[7,5],ymm9[4,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5],ymm6[6,7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm7 = xmm13[3,3],xmm12[3,3] |
| ; AVX1-ONLY-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm7 = xmm7[0,1,2],mem[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm7[1,2,3],ymm6[4,5,6,7] |
| ; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX1-ONLY-NEXT: vmovaps %ymm6, 544(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm5, 320(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm3, 96(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm2, 736(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm4, 640(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 608(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 512(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 416(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 384(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 288(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 192(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 160(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 64(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 800(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 768(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 704(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 672(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 576(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 480(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 448(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 352(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 256(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 224(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 128(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 864(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 832(%rax) |
| ; AVX1-ONLY-NEXT: addq $1624, %rsp # imm = 0x658 |
| ; AVX1-ONLY-NEXT: vzeroupper |
| ; AVX1-ONLY-NEXT: retq |
| ; |
| ; AVX2-SLOW-LABEL: store_i32_stride7_vf32: |
| ; AVX2-SLOW: # %bb.0: |
| ; AVX2-SLOW-NEXT: subq $1320, %rsp # imm = 0x528 |
| ; AVX2-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-SLOW-NEXT: vmovaps (%rax), %xmm0 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rax), %xmm3 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vmovaps (%r8), %xmm14 |
| ; AVX2-SLOW-NEXT: vmovaps 32(%r8), %xmm6 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%r9), %xmm8 |
| ; AVX2-SLOW-NEXT: vmovaps 32(%r9), %xmm9 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm8[1,1,1,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm14[1],xmm1[2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps (%rcx), %xmm10 |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rcx), %xmm11 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%rdx), %xmm7 |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm1 = zero,xmm7[1],xmm10[1],zero |
| ; AVX2-SLOW-NEXT: vmovaps (%rdi), %xmm5 |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rdi), %xmm13 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%rsi), %xmm4 |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rsi), %xmm12 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm2 = xmm4[1,1,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm5[2],xmm2[3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm9[1,1,1,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm6[1],xmm1[2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm12[1,1,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm13[2],xmm1[3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rdx), %xmm6 |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm6[1],xmm11[1],zero |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 64(%r8), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 64(%r9), %xmm0 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rax), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rdi), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rsi), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rcx), %xmm13 |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rdx), %xmm12 |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm12[1],xmm13[1],zero |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 96(%r8), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 96(%r9), %xmm11 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm11[1,1,1,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rax), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rdi), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rsi), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rcx), %xmm3 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm3, (%rsp) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rdx), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%rdi), %ymm0 |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%rsi), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps (%rdx), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%rcx), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[6],ymm1[6],ymm2[7],ymm1[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps (%r8), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%r9), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vmovaps 16(%rax), %xmm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rdi), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rsi), %ymm0 |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rdx), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rcx), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 32(%r8), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%r9), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vmovaps 48(%rax), %xmm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rdi), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rsi), %ymm0 |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rdx), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rcx), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 64(%r8), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 64(%r9), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vmovaps 80(%rax), %xmm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rdi), %ymm3 |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rsi), %ymm2 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rdx), %ymm9 |
| ; AVX2-SLOW-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rcx), %ymm1 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm15 = ymm9[1,1],ymm1[1,1],ymm9[5,5],ymm1[5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm15[5,6],ymm0[7] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 112(%r8), %ymm15 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0],ymm0[1,2,3,4,5,6],ymm15[7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 112(%r9), %xmm15 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm15[1],ymm0[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 112(%rax), %ymm15 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm15[2],ymm0[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm10, %xmm0 |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm7, %xmm15 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm15 = xmm5[0],xmm4[0],xmm5[1],xmm4[1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm15 = xmm15[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1],ymm0[2,3],ymm15[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps %xmm14, %xmm9 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm15 = xmm14[0],xmm8[0],xmm14[1],xmm8[1] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm15 |
| ; AVX2-SLOW-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm15[0],ymm14[0],ymm15[2],ymm14[2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm14[4,5,6],ymm0[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm5[3,3],xmm4[3,3] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} xmm4 = xmm7[2],xmm10[2],xmm7[3],xmm10[3] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3,4],ymm0[5,6],ymm4[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm4 = xmm8[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm4 = xmm9[0,1,2],xmm4[3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 8(%rax), %ymm5 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm4[2,3,4],ymm0[5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm8, %xmm0 |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm6, %xmm4 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm4 = xmm10[0],xmm14[0],xmm10[1],xmm14[1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1],ymm0[2,3],ymm4[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm4 = xmm7[0],xmm9[0],xmm7[1],xmm9[1] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm4 |
| ; AVX2-SLOW-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm5[0],ymm4[2],ymm5[2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6],ymm0[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm10[3,3],xmm14[3,3] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} xmm4 = xmm6[2],xmm8[2],xmm6[3],xmm8[3] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3,4],ymm0[5,6],ymm4[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm4 = xmm9[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm4 = xmm7[0,1,2],xmm4[3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 40(%rax), %ymm5 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm4[2,3,4],ymm0[5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm13, %xmm0 |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm12, %xmm4 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm4 = xmm9[0],xmm8[0],xmm9[1],xmm8[1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1],ymm0[2,3],ymm4[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm4 = xmm7[0],xmm6[0],xmm7[1],xmm6[1] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm4 |
| ; AVX2-SLOW-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm5[0],ymm4[2],ymm5[2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6],ymm0[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm9[3,3],xmm8[3,3] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} xmm4 = xmm12[2],xmm13[2],xmm12[3],xmm13[3] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3,4],ymm0[5,6],ymm4[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm4 = xmm6[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm4 = xmm7[0,1,2],xmm4[3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 72(%rax), %ymm5 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm4[2,3,4],ymm0[5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%rsp), %xmm9 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm9, %xmm0 |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm10, %xmm4 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm4 = xmm8[0],xmm7[0],xmm8[1],xmm7[1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1],ymm0[2,3],ymm4[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm4 = xmm6[0],xmm11[0],xmm6[1],xmm11[1] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm4 |
| ; AVX2-SLOW-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm5[0],ymm4[2],ymm5[2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6],ymm0[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm8[3,3],xmm7[3,3] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} xmm4 = xmm10[2],xmm9[2],xmm10[3],xmm9[3] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3,4],ymm0[5,6],ymm4[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm4 = xmm11[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm4 = xmm6[0,1,2],xmm4[3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 104(%rax), %ymm5 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm4[2,3,4],ymm0[5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vbroadcastss 112(%rdx), %ymm0 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm5 = ymm1[3,1,2,0,7,5,6,4] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5],ymm0[6],ymm5[7] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} ymm5 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 108(%r8), %ymm5 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm5 = xmm11[2,2,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm5[2,3],ymm0[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm5 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm6 = ymm4[2],ymm1[2],ymm4[3],ymm1[3],ymm4[6],ymm1[6],ymm4[7],ymm1[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],mem[6,7] |
| ; AVX2-SLOW-NEXT: vpermilps {{.*#+}} ymm7 = mem[1,2,2,3,5,6,6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0],ymm5[1,2,3,4,5,6],ymm7[7] |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rax), %ymm7 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm7[3],ymm0[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm6[2,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2,3,4],ymm0[5],ymm5[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm4[2],ymm1[3],ymm4[3],ymm1[6],ymm4[6],ymm1[7],ymm4[7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 124(%r8), %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 124(%r9), %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 120(%rax), %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm8[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm5[1],ymm1[2,3,4],ymm5[5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm2 = ymm9[1,1],ymm4[1,1],ymm9[5,5],ymm4[5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm2[5,6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm2 = ymm13[0,0,0,0,4,4,4,4] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm3 = ymm11[0,1,0,1,4,5,4,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3,4],ymm2[5],ymm3[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 16(%rax), %ymm3 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm1[3,4,5,6],ymm2[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm2 = ymm14[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm12[1],ymm2[2,3,4],ymm12[5],ymm2[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm3 = ymm0[1,1],ymm10[1,1],ymm0[5,5],ymm10[5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm3[5,6],ymm2[7] |
| ; AVX2-SLOW-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm3 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-SLOW-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm6 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm6[0],ymm3[1],ymm6[2,3,4],ymm3[5],ymm6[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 48(%rax), %ymm6 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm6[2,3],ymm3[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm2[3,4,5,6],ymm3[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm3 = mem[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm3 = ymm3[0],mem[1],ymm3[2,3,4],mem[5],ymm3[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm6 = ymm0[1,1],ymm15[1,1],ymm0[5,5],ymm15[5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm6[5,6],ymm3[7] |
| ; AVX2-SLOW-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm6 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-SLOW-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm7 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0],ymm6[1],ymm7[2,3,4],ymm6[5],ymm7[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 80(%rax), %ymm7 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1],ymm7[2,3],ymm6[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2],ymm3[3,4,5,6],ymm6[7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 16(%rdx), %ymm3 |
| ; AVX2-SLOW-NEXT: vmovaps %ymm4, %ymm1 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm7 = ymm4[3,1,2,0,7,5,6,4] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm7[0,1,2,3,4,5],ymm3[6],ymm7[7] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} ymm7 = ymm5[0],ymm8[0],ymm5[1],ymm8[1],ymm5[4],ymm8[4],ymm5[5],ymm8[5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm7[4,5],ymm3[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm7 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm7 = xmm4[3,3],mem[3,3] |
| ; AVX2-SLOW-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm7 = xmm7[0,1,2],mem[3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm7 = ymm3[0],ymm7[1,2,3],ymm3[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm3 = ymm1[2],ymm9[2],ymm1[3],ymm9[3],ymm1[6],ymm9[6],ymm1[7],ymm9[7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm8 = ymm8[2],ymm5[2],ymm8[3],ymm5[3],ymm8[6],ymm5[6],ymm8[7],ymm5[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm8[0,1,2],ymm3[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm8 = ymm11[3,3],ymm13[3,3],ymm11[7,7],ymm13[7,7] |
| ; AVX2-SLOW-NEXT: vpermilps {{.*#+}} ymm9 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm8 = ymm9[0],ymm8[1,2],ymm9[3,4],ymm8[5,6],ymm9[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0],ymm3[1,2,3,4],ymm8[5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 48(%rdx), %ymm3 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm9 = ymm10[3,1,2,0,7,5,6,4] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm9[0,1,2,3,4,5],ymm3[6],ymm9[7] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} ymm9 = ymm12[0],ymm14[0],ymm12[1],ymm14[1],ymm12[4],ymm14[4],ymm12[5],ymm14[5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm9[4,5],ymm3[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm9 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm9 = xmm2[3,3],mem[3,3] |
| ; AVX2-SLOW-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm9 = xmm9[0,1,2],mem[3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm9 = ymm3[0],ymm9[1,2,3],ymm3[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm3 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm3 = ymm10[2],mem[2],ymm10[3],mem[3],ymm10[6],mem[6],ymm10[7],mem[7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm11 = ymm14[2],ymm12[2],ymm14[3],ymm12[3],ymm14[6],ymm12[6],ymm14[7],ymm12[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm11[0,1,2],ymm3[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm11 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm11 = ymm0[3,3],mem[3,3],ymm0[7,7],mem[7,7] |
| ; AVX2-SLOW-NEXT: vpermilps {{.*#+}} ymm12 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm11 = ymm12[0],ymm11[1,2],ymm12[3,4],ymm11[5,6],ymm12[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm11[0],ymm3[1,2,3,4],ymm11[5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 80(%rdx), %ymm11 |
| ; AVX2-SLOW-NEXT: vmovaps %ymm15, %ymm1 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm12 = ymm15[3,1,2,0,7,5,6,4] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm11 = ymm12[0,1,2,3,4,5],ymm11[6],ymm12[7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} ymm12 = ymm15[0],ymm0[0],ymm15[1],ymm0[1],ymm15[4],ymm0[4],ymm15[5],ymm0[5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5],ymm11[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm12 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm12 = xmm2[3,3],mem[3,3] |
| ; AVX2-SLOW-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm12 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm12 = xmm12[0,1,2],mem[3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0],ymm12[1,2,3],ymm11[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm12 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm12 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm13 = ymm0[2],ymm15[2],ymm0[3],ymm15[3],ymm0[6],ymm15[6],ymm0[7],ymm15[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm12 = ymm12[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2],ymm12[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm13 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm13 = ymm0[3,3],mem[3,3],ymm0[7,7],mem[7,7] |
| ; AVX2-SLOW-NEXT: vpermilps {{.*#+}} ymm14 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm13 = ymm14[0],ymm13[1,2],ymm14[3,4],ymm13[5,6],ymm14[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0],ymm12[1,2,3,4],ymm13[5,6,7] |
| ; AVX2-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-SLOW-NEXT: vmovaps %ymm12, 640(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm11, 544(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm3, 416(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm9, 320(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm8, 192(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm7, 96(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm3, 608(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm6, 576(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm3, 384(%rax) |
| ; AVX2-SLOW-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 352(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm2, 160(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 128(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 864(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 832(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 768(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 736(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 704(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 672(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 512(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 480(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 448(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 288(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 256(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 224(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 64(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 800(%rax) |
| ; AVX2-SLOW-NEXT: addq $1320, %rsp # imm = 0x528 |
| ; AVX2-SLOW-NEXT: vzeroupper |
| ; AVX2-SLOW-NEXT: retq |
| ; |
| ; AVX2-FAST-LABEL: store_i32_stride7_vf32: |
| ; AVX2-FAST: # %bb.0: |
| ; AVX2-FAST-NEXT: subq $1400, %rsp # imm = 0x578 |
| ; AVX2-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-NEXT: vmovaps (%rax), %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps %xmm0, (%rsp) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rax), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vmovaps (%r8), %xmm3 |
| ; AVX2-FAST-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%r8), %xmm4 |
| ; AVX2-FAST-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%r9), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%r9), %xmm5 |
| ; AVX2-FAST-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm3[1],xmm1[2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovaps (%rcx), %xmm6 |
| ; AVX2-FAST-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rcx), %xmm3 |
| ; AVX2-FAST-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%rdx), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertps {{.*#+}} xmm1 = zero,xmm1[1],xmm6[1],zero |
| ; AVX2-FAST-NEXT: vmovaps (%rdi), %xmm9 |
| ; AVX2-FAST-NEXT: vmovaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rdi), %xmm7 |
| ; AVX2-FAST-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%rsi), %xmm13 |
| ; AVX2-FAST-NEXT: vmovaps 32(%rsi), %xmm8 |
| ; AVX2-FAST-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm6 = xmm13[1,1,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm9[2],xmm6[3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm6[0],ymm1[1,2],ymm6[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm5[1,1,1,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm4[1],xmm1[2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm8[1,1,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm7[2],xmm1[3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-NEXT: vmovaps 32(%rdx), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertps {{.*#+}} xmm7 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm7[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%r8), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%r9), %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-FAST-NEXT: vmovaps 64(%rax), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vmovaps 64(%rdi), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%rsi), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-NEXT: vmovaps 64(%rcx), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%rdx), %xmm3 |
| ; AVX2-FAST-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertps {{.*#+}} xmm7 = zero,xmm3[1],xmm2[1],zero |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm7[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 96(%r8), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 96(%r9), %xmm2 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm2[1,1,1,1] |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, %xmm12 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-FAST-NEXT: vmovaps 96(%rax), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vmovaps 96(%rdi), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 96(%rsi), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-NEXT: vmovaps 96(%rcx), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 96(%rdx), %xmm3 |
| ; AVX2-FAST-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertps {{.*#+}} xmm7 = zero,xmm3[1],xmm2[1],zero |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm7[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%rdi), %ymm0 |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%rsi), %ymm10 |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm10[2],ymm0[3],ymm10[3],ymm0[6],ymm10[6],ymm0[7],ymm10[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps (%rdx), %ymm9 |
| ; AVX2-FAST-NEXT: vmovaps (%rcx), %ymm8 |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[6],ymm8[6],ymm9[7],ymm8[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps (%r8), %ymm6 |
| ; AVX2-FAST-NEXT: vmovaps (%r9), %ymm5 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm5[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm6[2],ymm1[3,4,5],ymm6[6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vmovaps 16(%rax), %xmm11 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm11[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rdi), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rsi), %ymm0 |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps 32(%rdx), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rcx), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps 32(%r8), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%r9), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vmovaps 48(%rax), %xmm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm15[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%rdi), %ymm0 |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%rsi), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps 64(%rdx), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%rcx), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps 64(%r8), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%r9), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vmovaps 80(%rax), %xmm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm15[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 96(%rdi), %ymm3 |
| ; AVX2-FAST-NEXT: vmovaps 96(%rsi), %ymm2 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm15 = ymm2[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0],ymm3[1],ymm15[2,3,4],ymm3[5],ymm15[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps 96(%rdx), %ymm4 |
| ; AVX2-FAST-NEXT: vmovaps 96(%rcx), %ymm1 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm14 = ymm4[1,1],ymm1[1,1],ymm4[5,5],ymm1[5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5,6],ymm15[7] |
| ; AVX2-FAST-NEXT: vbroadcastsd 112(%r8), %ymm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0],ymm14[1,2,3,4,5,6],ymm15[7] |
| ; AVX2-FAST-NEXT: vbroadcastss 112(%r9), %xmm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0],ymm15[1],ymm14[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 112(%rax), %ymm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm14[0,1],ymm15[2],ymm14[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vbroadcastss 112(%rdx), %ymm14 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm15 = ymm1[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3,4,5],ymm14[6],ymm15[7] |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} ymm15 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm15[4,5],ymm14[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 108(%r8), %ymm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0],ymm15[1],ymm14[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm15 = xmm12[2,2,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1],ymm15[2,3],ymm14[4,5,6,7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm15 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[2,2,2,2] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm4[2],ymm1[2],ymm4[3],ymm1[3],ymm4[6],ymm1[6],ymm4[7],ymm1[7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3,4,5],mem[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastsd {{.*#+}} ymm11 = [5,6,5,6,5,6,5,6] |
| ; AVX2-FAST-NEXT: vpermps 96(%r9), %ymm11, %ymm11 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0],ymm15[1,2,3,4,5,6],ymm11[7] |
| ; AVX2-FAST-NEXT: vmovaps 96(%rax), %ymm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm12 = ymm14[0,1,2],ymm15[3],ymm14[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm15[2,3],ymm0[2,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm11[0],ymm0[1],ymm11[2,3,4],ymm0[5],ymm11[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm4[2],ymm1[3],ymm4[3],ymm1[6],ymm4[6],ymm1[7],ymm4[7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 124(%r8), %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 124(%r9), %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastsd 120(%rax), %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm14, %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm12, %xmm1 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm2 = xmm15[0],xmm13[0],xmm15[1],xmm13[1] |
| ; AVX2-FAST-NEXT: vbroadcastf128 {{.*#+}} ymm7 = [0,1,2,2,0,1,2,2] |
| ; AVX2-FAST-NEXT: # ymm7 = mem[0,1,0,1] |
| ; AVX2-FAST-NEXT: vpermps %ymm2, %ymm7, %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm2 = xmm4[0],xmm11[0],xmm4[1],xmm11[1] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2 |
| ; AVX2-FAST-NEXT: vbroadcastsd (%rsp), %ymm3 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[2],ymm3[2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm2[4,5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm15[3,3],xmm13[3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} xmm2 = xmm12[2],xmm14[2],xmm12[3],xmm14[3] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vpermps %ymm2, %ymm7, %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5,6],ymm2[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm2 = xmm11[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm2 = xmm4[0,1,2],xmm2[3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 8(%rax), %ymm3 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm2[2,3,4],ymm1[5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm10[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm10[1],ymm1[2,3,4],ymm10[5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm2 = ymm9[1,1],ymm8[1,1],ymm9[5,5],ymm8[5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm2[5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm2 = ymm5[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm3 = ymm6[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3,4],ymm2[5],ymm3[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[2,1,2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 16(%rax), %ymm3 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm1[3,4,5,6],ymm2[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm0, %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm6, %xmm2 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm2 = xmm4[0],xmm5[0],xmm4[1],xmm5[1] |
| ; AVX2-FAST-NEXT: vpermps %ymm2, %ymm7, %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm2 = xmm9[0],xmm8[0],xmm9[1],xmm8[1] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2 |
| ; AVX2-FAST-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[2],ymm3[2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm4[3,3],xmm5[3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} xmm2 = xmm6[2],xmm0[2],xmm6[3],xmm0[3] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vpermps %ymm2, %ymm7, %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5,6],ymm2[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm2 = xmm8[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm2 = xmm9[0,1,2],xmm2[3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 40(%rax), %ymm3 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm2[2,3,4],ymm1[5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm1 = mem[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm12[1],ymm1[2,3,4],ymm12[5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm3 = ymm3[1,1],ymm11[1,1],ymm3[5,5],ymm11[5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm3[5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm3 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm4 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3,4],ymm3[5],ymm4[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[2,1,2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 48(%rax), %ymm4 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm4[2,3],ymm3[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm1[3,4,5,6],ymm3[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm8, %xmm3 |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm9, %xmm4 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm4 = xmm2[0],xmm0[0],xmm2[1],xmm0[1] |
| ; AVX2-FAST-NEXT: vpermps %ymm4, %ymm7, %ymm4 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1],ymm3[2,3],ymm4[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm4 = xmm6[0],xmm1[0],xmm6[1],xmm1[1] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm4 |
| ; AVX2-FAST-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm5[0],ymm4[2],ymm5[2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6],ymm3[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm4 = xmm2[3,3],xmm0[3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm4 |
| ; AVX2-FAST-NEXT: vpermps %ymm5, %ymm7, %ymm5 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5,6],ymm5[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm5 = xmm1[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm5 = xmm6[0,1,2],xmm5[3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 72(%rax), %ymm6 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1],ymm5[2,3,4],ymm4[5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm5 = ymm15[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm14[1],ymm5[2,3,4],ymm14[5],ymm5[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm6 = ymm0[1,1],ymm13[1,1],ymm0[5,5],ymm13[5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4],ymm6[5,6],ymm5[7] |
| ; AVX2-FAST-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm6 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm8 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2,3,4],ymm6[5],ymm8[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[2,1,2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 80(%rax), %ymm8 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1],ymm8[2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2],ymm5[3,4,5,6],ymm6[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm4, %xmm6 |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm3, %xmm8 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm6 = xmm8[0],xmm6[0],xmm8[1],xmm6[1] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm8 = xmm0[0],xmm2[0],xmm0[1],xmm2[1] |
| ; AVX2-FAST-NEXT: vpermps %ymm8, %ymm7, %ymm8 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm8[0,1],ymm6[2,3],ymm8[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm8 = xmm1[0],xmm5[0],xmm1[1],xmm5[1] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm8 |
| ; AVX2-FAST-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: vunpcklpd {{.*#+}} ymm8 = ymm8[0],ymm9[0],ymm8[2],ymm9[2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm8 = ymm6[0,1,2,3],ymm8[4,5,6],ymm6[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm6 = xmm0[3,3],xmm2[3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} xmm9 = xmm3[2],xmm4[2],xmm3[3],xmm4[3] |
| ; AVX2-FAST-NEXT: vpermps %ymm9, %ymm7, %ymm0 |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm6 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm6[5,6],ymm0[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm6 = xmm5[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm6 = xmm1[0,1,2],xmm6[3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 104(%rax), %ymm9 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm9[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0,1],ymm6[2,3,4],ymm0[5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 16(%rdx), %ymm6 |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm9 = ymm0[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm9[0,1,2,3,4,5],ymm6[6],ymm9[7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} ymm9 = ymm10[0],ymm1[0],ymm10[1],ymm1[1],ymm10[4],ymm1[4],ymm10[5],ymm1[5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm9[4,5],ymm6[6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm9 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm9 = xmm3[3,3],mem[3,3] |
| ; AVX2-FAST-NEXT: vblendps $8, (%rsp), %xmm9, %xmm9 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm9 = xmm9[0,1,2],mem[3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm9 = ymm6[0],ymm9[1,2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm6 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm7 = ymm1[2],ymm10[2],ymm1[3],ymm10[3],ymm1[6],ymm10[6],ymm1[7],ymm10[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[3,3,3,3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2],ymm6[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm7 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm7 = ymm0[3,3],mem[3,3],ymm0[7,7],mem[7,7] |
| ; AVX2-FAST-NEXT: vpermilps {{.*#+}} ymm10 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm7 = ymm10[0],ymm7[1,2],ymm10[3,4],ymm7[5,6],ymm10[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[2,1,2,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0],ymm6[1,2,3,4],ymm7[5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 48(%rdx), %ymm7 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm10 = ymm11[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6],ymm10[7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} ymm10 = ymm12[0],ymm0[0],ymm12[1],ymm0[1],ymm12[4],ymm0[4],ymm12[5],ymm0[5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm10[4,5],ymm7[6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm10 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm10 = xmm3[3,3],mem[3,3] |
| ; AVX2-FAST-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm10 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm10 = xmm10[0,1,2],mem[3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm10[1,2,3],ymm7[4,5,6,7] |
| ; AVX2-FAST-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm10 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm10 = ymm11[2],mem[2],ymm11[3],mem[3],ymm11[6],mem[6],ymm11[7],mem[7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm11 = ymm0[2],ymm12[2],ymm0[3],ymm12[3],ymm0[6],ymm12[6],ymm0[7],ymm12[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[3,3,3,3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm10 = ymm11[0,1,2],ymm10[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm11 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm11 = ymm0[3,3],mem[3,3],ymm0[7,7],mem[7,7] |
| ; AVX2-FAST-NEXT: vpermilps {{.*#+}} ymm12 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm11 = ymm12[0],ymm11[1,2],ymm12[3,4],ymm11[5,6],ymm12[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[2,1,2,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm10 = ymm11[0],ymm10[1,2,3,4],ymm11[5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 80(%rdx), %ymm11 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm12 = ymm13[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm11 = ymm12[0,1,2,3,4,5],ymm11[6],ymm12[7] |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} ymm12 = ymm14[0],ymm15[0],ymm14[1],ymm15[1],ymm14[4],ymm15[4],ymm14[5],ymm15[5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5],ymm11[6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm12 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm12 = xmm1[3,3],mem[3,3] |
| ; AVX2-FAST-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm12 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm12 = xmm12[0,1,2],mem[3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0],ymm12[1,2,3],ymm11[4,5,6,7] |
| ; AVX2-FAST-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm12 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm12 = ymm13[2],mem[2],ymm13[3],mem[3],ymm13[6],mem[6],ymm13[7],mem[7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm13 = ymm15[2],ymm14[2],ymm15[3],ymm14[3],ymm15[6],ymm14[6],ymm15[7],ymm14[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm12 = ymm12[3,3,3,3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2],ymm12[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm13 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm13 = ymm0[3,3],mem[3,3],ymm0[7,7],mem[7,7] |
| ; AVX2-FAST-NEXT: vpermilps {{.*#+}} ymm14 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm13 = ymm14[0],ymm13[1,2],ymm14[3,4],ymm13[5,6],ymm14[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[2,1,2,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0],ymm12[1,2,3,4],ymm13[5,6,7] |
| ; AVX2-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-NEXT: vmovaps %ymm12, 640(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm11, 544(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm10, 416(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm7, 320(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm6, 192(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm9, 96(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm5, 736(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm8, 672(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 608(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 576(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 512(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 448(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 384(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 352(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 288(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 224(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 160(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 128(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 64(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 864(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 832(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 768(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 704(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 480(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 256(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 800(%rax) |
| ; AVX2-FAST-NEXT: addq $1400, %rsp # imm = 0x578 |
| ; AVX2-FAST-NEXT: vzeroupper |
| ; AVX2-FAST-NEXT: retq |
| ; |
| ; AVX2-FAST-PERLANE-LABEL: store_i32_stride7_vf32: |
| ; AVX2-FAST-PERLANE: # %bb.0: |
| ; AVX2-FAST-PERLANE-NEXT: subq $1320, %rsp # imm = 0x528 |
| ; AVX2-FAST-PERLANE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rax), %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rax), %xmm3 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r8), %xmm14 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%r8), %xmm6 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r9), %xmm8 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%r9), %xmm9 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm8[1,1,1,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm14[1],xmm1[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rcx), %xmm10 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rcx), %xmm11 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdx), %xmm7 |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm1 = zero,xmm7[1],xmm10[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdi), %xmm5 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rdi), %xmm13 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rsi), %xmm4 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rsi), %xmm12 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm2 = xmm4[1,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm5[2],xmm2[3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm9[1,1,1,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm6[1],xmm1[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm12[1,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm13[2],xmm1[3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rdx), %xmm6 |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm6[1],xmm11[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%r8), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%r9), %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rax), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rdi), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rsi), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rcx), %xmm13 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rdx), %xmm12 |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm12[1],xmm13[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%r8), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%r9), %xmm11 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm11[1,1,1,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rax), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rdi), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rsi), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rcx), %xmm3 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm3, (%rsp) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rdx), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdi), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rsi), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdx), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rcx), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[6],ymm1[6],ymm2[7],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r8), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r9), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 16(%rax), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rdi), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rsi), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rdx), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rcx), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%r8), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%r9), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 48(%rax), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rdi), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rsi), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rdx), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rcx), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%r8), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%r9), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 80(%rax), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rdi), %ymm3 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rsi), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rdx), %ymm9 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rcx), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm15 = ymm9[1,1],ymm1[1,1],ymm9[5,5],ymm1[5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm15[5,6],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 112(%r8), %ymm15 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0],ymm0[1,2,3,4,5,6],ymm15[7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 112(%r9), %xmm15 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm15[1],ymm0[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 112(%rax), %ymm15 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm15[2],ymm0[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm10, %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm7, %xmm15 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm15 = xmm5[0],xmm4[0],xmm5[1],xmm4[1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm15 = xmm15[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1],ymm0[2,3],ymm15[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm14, %xmm9 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm15 = xmm14[0],xmm8[0],xmm14[1],xmm8[1] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm15 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm15[0],ymm14[0],ymm15[2],ymm14[2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm14[4,5,6],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm5[3,3],xmm4[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} xmm4 = xmm7[2],xmm10[2],xmm7[3],xmm10[3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3,4],ymm0[5,6],ymm4[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm4 = xmm8[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm4 = xmm9[0,1,2],xmm4[3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 8(%rax), %ymm5 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm4[2,3,4],ymm0[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm8, %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm6, %xmm4 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm4 = xmm10[0],xmm14[0],xmm10[1],xmm14[1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1],ymm0[2,3],ymm4[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm4 = xmm7[0],xmm9[0],xmm7[1],xmm9[1] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm4 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm5[0],ymm4[2],ymm5[2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm10[3,3],xmm14[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} xmm4 = xmm6[2],xmm8[2],xmm6[3],xmm8[3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3,4],ymm0[5,6],ymm4[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm4 = xmm9[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm4 = xmm7[0,1,2],xmm4[3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 40(%rax), %ymm5 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm4[2,3,4],ymm0[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm13, %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm12, %xmm4 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm4 = xmm9[0],xmm8[0],xmm9[1],xmm8[1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1],ymm0[2,3],ymm4[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm4 = xmm7[0],xmm6[0],xmm7[1],xmm6[1] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm4 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm5[0],ymm4[2],ymm5[2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm9[3,3],xmm8[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} xmm4 = xmm12[2],xmm13[2],xmm12[3],xmm13[3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3,4],ymm0[5,6],ymm4[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm4 = xmm6[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm4 = xmm7[0,1,2],xmm4[3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 72(%rax), %ymm5 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm4[2,3,4],ymm0[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rsp), %xmm9 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm9, %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm10, %xmm4 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm4 = xmm8[0],xmm7[0],xmm8[1],xmm7[1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1],ymm0[2,3],ymm4[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm4 = xmm6[0],xmm11[0],xmm6[1],xmm11[1] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm4 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm5[0],ymm4[2],ymm5[2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm8[3,3],xmm7[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} xmm4 = xmm10[2],xmm9[2],xmm10[3],xmm9[3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3,4],ymm0[5,6],ymm4[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm4 = xmm11[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm4 = xmm6[0,1,2],xmm4[3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 104(%rax), %ymm5 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm4[2,3,4],ymm0[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 112(%rdx), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm5 = ymm1[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5],ymm0[6],ymm5[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} ymm5 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 108(%r8), %ymm5 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm5 = xmm11[2,2,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm5[2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm5 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm6 = ymm4[2],ymm1[2],ymm4[3],ymm1[3],ymm4[6],ymm1[6],ymm4[7],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],mem[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps {{.*#+}} ymm7 = mem[1,2,2,3,5,6,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0],ymm5[1,2,3,4,5,6],ymm7[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rax), %ymm7 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm7[3],ymm0[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm7[2,3],ymm6[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2,3,4],ymm0[5],ymm5[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm4[2],ymm1[3],ymm4[3],ymm1[6],ymm4[6],ymm1[7],ymm4[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 124(%r8), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 124(%r9), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 120(%rax), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm8[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm5[1],ymm1[2,3,4],ymm5[5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm2 = ymm9[1,1],ymm4[1,1],ymm9[5,5],ymm4[5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm2[5,6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm2 = ymm13[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm3 = ymm11[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3,4],ymm2[5],ymm3[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 16(%rax), %ymm3 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm1[3,4,5,6],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm2 = ymm14[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm12[1],ymm2[2,3,4],ymm12[5],ymm2[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm3 = ymm0[1,1],ymm10[1,1],ymm0[5,5],ymm10[5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm3[5,6],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm3 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm6 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm6[0],ymm3[1],ymm6[2,3,4],ymm3[5],ymm6[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 48(%rax), %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm6[2,3],ymm3[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm2[3,4,5,6],ymm3[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm3 = mem[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm3 = ymm3[0],mem[1],ymm3[2,3,4],mem[5],ymm3[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm6 = ymm0[1,1],ymm15[1,1],ymm0[5,5],ymm15[5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm6[5,6],ymm3[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm6 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm7 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0],ymm6[1],ymm7[2,3,4],ymm6[5],ymm7[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 80(%rax), %ymm7 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1],ymm7[2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2],ymm3[3,4,5,6],ymm6[7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 16(%rdx), %ymm3 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm4, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm7 = ymm4[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm7[0,1,2,3,4,5],ymm3[6],ymm7[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} ymm7 = ymm5[0],ymm8[0],ymm5[1],ymm8[1],ymm5[4],ymm8[4],ymm5[5],ymm8[5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm7[4,5],ymm3[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm7 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm7 = xmm4[3,3],mem[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm7 = xmm7[0,1,2],mem[3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm7 = ymm3[0],ymm7[1,2,3],ymm3[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm3 = ymm1[2],ymm9[2],ymm1[3],ymm9[3],ymm1[6],ymm9[6],ymm1[7],ymm9[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm8 = ymm8[2],ymm5[2],ymm8[3],ymm5[3],ymm8[6],ymm5[6],ymm8[7],ymm5[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm8[0,1,2],ymm3[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm8 = ymm11[3,3],ymm13[3,3],ymm11[7,7],ymm13[7,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps {{.*#+}} ymm9 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm8 = ymm9[0],ymm8[1,2],ymm9[3,4],ymm8[5,6],ymm9[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0],ymm3[1,2,3,4],ymm8[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 48(%rdx), %ymm3 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm9 = ymm10[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm9[0,1,2,3,4,5],ymm3[6],ymm9[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} ymm9 = ymm12[0],ymm14[0],ymm12[1],ymm14[1],ymm12[4],ymm14[4],ymm12[5],ymm14[5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm9[4,5],ymm3[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm9 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm9 = xmm2[3,3],mem[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm9 = xmm9[0,1,2],mem[3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm9 = ymm3[0],ymm9[1,2,3],ymm3[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm3 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm3 = ymm10[2],mem[2],ymm10[3],mem[3],ymm10[6],mem[6],ymm10[7],mem[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm11 = ymm14[2],ymm12[2],ymm14[3],ymm12[3],ymm14[6],ymm12[6],ymm14[7],ymm12[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm11[0,1,2],ymm3[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm11 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm11 = ymm0[3,3],mem[3,3],ymm0[7,7],mem[7,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps {{.*#+}} ymm12 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm11 = ymm12[0],ymm11[1,2],ymm12[3,4],ymm11[5,6],ymm12[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm11[0],ymm3[1,2,3,4],ymm11[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 80(%rdx), %ymm11 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm15, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm12 = ymm15[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm11 = ymm12[0,1,2,3,4,5],ymm11[6],ymm12[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} ymm12 = ymm15[0],ymm0[0],ymm15[1],ymm0[1],ymm15[4],ymm0[4],ymm15[5],ymm0[5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4,5],ymm11[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm12 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm12 = xmm2[3,3],mem[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm12 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm12 = xmm12[0,1,2],mem[3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0],ymm12[1,2,3],ymm11[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm12 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm12 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm13 = ymm0[2],ymm15[2],ymm0[3],ymm15[3],ymm0[6],ymm15[6],ymm0[7],ymm15[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm12 = ymm12[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2],ymm12[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm13 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm13 = ymm0[3,3],mem[3,3],ymm0[7,7],mem[7,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps {{.*#+}} ymm14 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm13 = ymm14[0],ymm13[1,2],ymm14[3,4],ymm13[5,6],ymm14[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0],ymm12[1,2,3,4],ymm13[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm12, 640(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm11, 544(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm3, 416(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm9, 320(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm8, 192(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm7, 96(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm3, 608(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm6, 576(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm3, 384(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 352(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm2, 160(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 128(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 864(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 832(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 768(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 736(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 704(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 672(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 512(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 480(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 448(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 288(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 256(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 224(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 64(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 800(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: addq $1320, %rsp # imm = 0x528 |
| ; AVX2-FAST-PERLANE-NEXT: vzeroupper |
| ; AVX2-FAST-PERLANE-NEXT: retq |
| ; |
| ; AVX512F-LABEL: store_i32_stride7_vf32: |
| ; AVX512F: # %bb.0: |
| ; AVX512F-NEXT: pushq %rax |
| ; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512F-NEXT: vmovdqa64 (%rdi), %zmm15 |
| ; AVX512F-NEXT: vmovdqa64 64(%rdi), %zmm18 |
| ; AVX512F-NEXT: vmovdqa64 (%rsi), %zmm14 |
| ; AVX512F-NEXT: vmovdqa64 64(%rsi), %zmm10 |
| ; AVX512F-NEXT: vmovdqa64 (%rdx), %zmm1 |
| ; AVX512F-NEXT: vmovdqa64 64(%rdx), %zmm21 |
| ; AVX512F-NEXT: vmovdqa64 (%rcx), %zmm12 |
| ; AVX512F-NEXT: vmovdqa64 64(%rcx), %zmm11 |
| ; AVX512F-NEXT: vmovdqa64 (%r8), %zmm3 |
| ; AVX512F-NEXT: vmovdqa64 64(%r8), %zmm28 |
| ; AVX512F-NEXT: vmovdqa64 (%r9), %zmm8 |
| ; AVX512F-NEXT: vmovdqa64 64(%r9), %zmm20 |
| ; AVX512F-NEXT: vmovdqa64 (%rax), %zmm9 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm30 = <u,u,u,u,14,30,u,u,u,u,u,15,31,u,u,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm21, %zmm2 |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm30, %zmm2 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm7 = <u,u,14,30,u,u,u,u,u,15,31,u,u,u,u,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm18, %zmm4 |
| ; AVX512F-NEXT: vpermt2d %zmm10, %zmm7, %zmm4 |
| ; AVX512F-NEXT: movw $6192, %cx # imm = 0x1830 |
| ; AVX512F-NEXT: kmovw %ecx, %k1 |
| ; AVX512F-NEXT: vmovdqa32 %zmm2, %zmm4 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm0 = <u,u,2,3,4,5,30,u,u,9,10,11,12,31,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm28, %zmm4, %zmm0 |
| ; AVX512F-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm22 = <u,u,u,u,12,28,u,u,u,u,u,13,29,u,u,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm18, %zmm2 |
| ; AVX512F-NEXT: vpermt2d %zmm10, %zmm22, %zmm2 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm13 = <11,u,u,u,u,u,28,12,u,u,u,u,u,29,13,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm11, %zmm4 |
| ; AVX512F-NEXT: vpermt2d %zmm21, %zmm13, %zmm4 |
| ; AVX512F-NEXT: vmovdqa32 %zmm2, %zmm4 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm0 = <0,27,u,u,4,5,6,7,28,u,u,11,12,13,14,29> |
| ; AVX512F-NEXT: vpermi2d %zmm28, %zmm4, %zmm0 |
| ; AVX512F-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,u,0,16,u,u,u,u,u,1,17,u,u,u,u,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm16 |
| ; AVX512F-NEXT: vpermt2d %zmm12, %zmm2, %zmm16 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm0 = <0,16,u,u,u,u,u,1,17,u,u,u,u,u,2,18> |
| ; AVX512F-NEXT: vmovdqa64 %zmm15, %zmm6 |
| ; AVX512F-NEXT: vpermt2d %zmm14, %zmm0, %zmm6 |
| ; AVX512F-NEXT: movw $1548, %cx # imm = 0x60C |
| ; AVX512F-NEXT: kmovw %ecx, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm16, %zmm6 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm27 = <u,u,u,u,0,16,u,u,u,u,u,1,17,u,u,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm3, %zmm31 |
| ; AVX512F-NEXT: vpermt2d %zmm8, %zmm27, %zmm31 |
| ; AVX512F-NEXT: vmovdqa64 %zmm21, %zmm25 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm17 = <9,u,u,u,u,u,26,10,u,u,u,u,u,27,11,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm10, %zmm16 |
| ; AVX512F-NEXT: vpermt2d %zmm18, %zmm17, %zmm16 |
| ; AVX512F-NEXT: vmovdqa64 %zmm21, %zmm24 |
| ; AVX512F-NEXT: vmovdqa64 %zmm18, %zmm4 |
| ; AVX512F-NEXT: vmovdqa64 %zmm21, %zmm29 |
| ; AVX512F-NEXT: vmovdqa64 %zmm18, %zmm26 |
| ; AVX512F-NEXT: vmovdqa64 %zmm21, %zmm23 |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm2, %zmm21 |
| ; AVX512F-NEXT: vmovdqa64 %zmm18, %zmm5 |
| ; AVX512F-NEXT: vpermt2d %zmm10, %zmm0, %zmm18 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,u,u,u,4,5,16,u,u,u,u,11,12,17,u,u> |
| ; AVX512F-NEXT: vpermt2d %zmm9, %zmm2, %zmm31 |
| ; AVX512F-NEXT: movw $14448, %cx # imm = 0x3870 |
| ; AVX512F-NEXT: vmovdqa32 %zmm21, %zmm18 {%k2} |
| ; AVX512F-NEXT: kmovw %ecx, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm31, %zmm6 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 %zmm28, %zmm31 |
| ; AVX512F-NEXT: vmovdqa64 %zmm28, %zmm19 |
| ; AVX512F-NEXT: vmovdqa64 %zmm28, %zmm21 |
| ; AVX512F-NEXT: vmovdqa64 %zmm28, %zmm0 |
| ; AVX512F-NEXT: vpermt2d %zmm20, %zmm27, %zmm28 |
| ; AVX512F-NEXT: vmovdqa64 64(%rax), %zmm27 |
| ; AVX512F-NEXT: vpermt2d %zmm27, %zmm2, %zmm28 |
| ; AVX512F-NEXT: vmovdqa32 %zmm28, %zmm18 {%k2} |
| ; AVX512F-NEXT: vpermi2d %zmm12, %zmm1, %zmm30 |
| ; AVX512F-NEXT: vpermi2d %zmm14, %zmm15, %zmm7 |
| ; AVX512F-NEXT: vmovdqa32 %zmm30, %zmm7 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm2 = <13,u,u,u,u,u,30,14,u,u,u,u,u,31,15,u> |
| ; AVX512F-NEXT: vpermi2d %zmm3, %zmm8, %zmm2 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm28 = <0,29,u,u,u,u,6,7,30,u,u,u,u,13,14,31> |
| ; AVX512F-NEXT: vpermi2d %zmm9, %zmm2, %zmm28 |
| ; AVX512F-NEXT: movw $-7741, %ax # imm = 0xE1C3 |
| ; AVX512F-NEXT: kmovw %eax, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm28, %zmm7 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,9,25,u,u,u,u,u,10,26,u,u,u,u,u,11> |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm2, %zmm25 |
| ; AVX512F-NEXT: movw $-31994, %ax # imm = 0x8306 |
| ; AVX512F-NEXT: kmovw %eax, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm25, %zmm16 {%k2} |
| ; AVX512F-NEXT: vpermi2d %zmm14, %zmm15, %zmm22 |
| ; AVX512F-NEXT: vpermi2d %zmm1, %zmm12, %zmm13 |
| ; AVX512F-NEXT: vmovdqa32 %zmm22, %zmm13 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm22 = <u,11,27,u,u,u,u,u,12,28,u,u,u,u,u,13> |
| ; AVX512F-NEXT: vpermi2d %zmm8, %zmm3, %zmm22 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm25 = <u,1,2,27,u,u,u,u,8,9,28,u,u,u,u,15> |
| ; AVX512F-NEXT: vpermi2d %zmm9, %zmm22, %zmm25 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm22 = <u,u,u,9,25,u,u,u,u,u,10,26,u,u,u,u> |
| ; AVX512F-NEXT: vpermt2d %zmm20, %zmm22, %zmm31 |
| ; AVX512F-NEXT: movw $-30962, %ax # imm = 0x870E |
| ; AVX512F-NEXT: kmovw %eax, %k1 |
| ; AVX512F-NEXT: vmovdqa32 %zmm25, %zmm13 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm25 = <u,u,u,3,4,25,u,u,u,u,10,11,26,u,u,u> |
| ; AVX512F-NEXT: vpermt2d %zmm27, %zmm25, %zmm31 |
| ; AVX512F-NEXT: movw $7224, %ax # imm = 0x1C38 |
| ; AVX512F-NEXT: kmovw %eax, %k1 |
| ; AVX512F-NEXT: vmovdqa32 %zmm31, %zmm16 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm28 = <u,u,u,7,23,u,u,u,u,u,8,24,u,u,u,u> |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm28, %zmm24 |
| ; AVX512F-NEXT: vpermi2d %zmm12, %zmm1, %zmm2 |
| ; AVX512F-NEXT: vpermi2d %zmm15, %zmm14, %zmm17 |
| ; AVX512F-NEXT: vmovdqa32 %zmm2, %zmm17 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,7,23,u,u,u,u,u,8,24,u,u,u,u,u,9> |
| ; AVX512F-NEXT: vpermt2d %zmm10, %zmm2, %zmm4 |
| ; AVX512F-NEXT: movw $3096, %ax # imm = 0xC18 |
| ; AVX512F-NEXT: kmovw %eax, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm24, %zmm4 {%k2} |
| ; AVX512F-NEXT: vpermi2d %zmm8, %zmm3, %zmm22 |
| ; AVX512F-NEXT: vpermt2d %zmm9, %zmm25, %zmm22 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm24 = <u,u,u,u,u,7,23,u,u,u,u,u,8,24,u,u> |
| ; AVX512F-NEXT: vpermt2d %zmm20, %zmm24, %zmm19 |
| ; AVX512F-NEXT: vmovdqa32 %zmm22, %zmm17 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm22 = <22,u,u,u,u,5,6,23,u,u,u,u,12,13,24,u> |
| ; AVX512F-NEXT: vpermt2d %zmm27, %zmm22, %zmm19 |
| ; AVX512F-NEXT: movw $28897, %ax # imm = 0x70E1 |
| ; AVX512F-NEXT: kmovw %eax, %k3 |
| ; AVX512F-NEXT: vmovdqa32 %zmm19, %zmm4 {%k3} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm19 = <u,u,u,u,u,5,21,u,u,u,u,u,6,22,u,u> |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm19, %zmm29 |
| ; AVX512F-NEXT: vpermi2d %zmm12, %zmm1, %zmm28 |
| ; AVX512F-NEXT: vpermi2d %zmm14, %zmm15, %zmm2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm28, %zmm2 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm25 = <u,u,u,5,21,u,u,u,u,u,6,22,u,u,u,u> |
| ; AVX512F-NEXT: vpermt2d %zmm10, %zmm25, %zmm26 |
| ; AVX512F-NEXT: movw $12384, %ax # imm = 0x3060 |
| ; AVX512F-NEXT: kmovw %eax, %k1 |
| ; AVX512F-NEXT: vmovdqa32 %zmm29, %zmm26 {%k1} |
| ; AVX512F-NEXT: vpermi2d %zmm8, %zmm3, %zmm24 |
| ; AVX512F-NEXT: vpermt2d %zmm9, %zmm22, %zmm24 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm22 = <4,20,u,u,u,u,u,5,21,u,u,u,u,u,6,22> |
| ; AVX512F-NEXT: vpermt2d %zmm20, %zmm22, %zmm21 |
| ; AVX512F-NEXT: vmovdqa32 %zmm24, %zmm2 {%k3} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm24 = <0,1,20,u,u,u,u,7,8,21,u,u,u,u,14,15> |
| ; AVX512F-NEXT: vpermt2d %zmm27, %zmm24, %zmm21 |
| ; AVX512F-NEXT: movw $15480, %ax # imm = 0x3C78 |
| ; AVX512F-NEXT: kmovw %eax, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm26, %zmm21 {%k2} |
| ; AVX512F-NEXT: vpermi2d %zmm12, %zmm1, %zmm19 |
| ; AVX512F-NEXT: vpermi2d %zmm14, %zmm15, %zmm25 |
| ; AVX512F-NEXT: vmovdqa32 %zmm19, %zmm25 {%k1} |
| ; AVX512F-NEXT: vpermi2d %zmm8, %zmm3, %zmm22 |
| ; AVX512F-NEXT: vpermt2d %zmm9, %zmm24, %zmm22 |
| ; AVX512F-NEXT: vmovdqa32 %zmm25, %zmm22 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm19 = <u,u,u,u,u,3,19,u,u,u,u,u,4,20,u,u> |
| ; AVX512F-NEXT: vpermt2d %zmm10, %zmm19, %zmm5 |
| ; AVX512F-NEXT: vpermt2d %zmm14, %zmm19, %zmm15 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm10 = <2,18,u,u,u,u,u,3,19,u,u,u,u,u,4,20> |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm10, %zmm23 |
| ; AVX512F-NEXT: vpermt2d %zmm12, %zmm10, %zmm1 |
| ; AVX512F-NEXT: vmovdqa32 %zmm5, %zmm23 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm5 = <u,u,2,18,u,u,u,u,u,3,19,u,u,u,u,u> |
| ; AVX512F-NEXT: vpermt2d %zmm8, %zmm5, %zmm3 |
| ; AVX512F-NEXT: vpermt2d %zmm20, %zmm5, %zmm0 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm5 = <u,u,2,3,18,u,u,u,u,9,10,19,u,u,u,u> |
| ; AVX512F-NEXT: vpermt2d %zmm9, %zmm5, %zmm3 |
| ; AVX512F-NEXT: vpermt2d %zmm27, %zmm5, %zmm0 |
| ; AVX512F-NEXT: vmovdqa32 %zmm15, %zmm1 {%k1} |
| ; AVX512F-NEXT: movw $3612, %ax # imm = 0xE1C |
| ; AVX512F-NEXT: kmovw %eax, %k1 |
| ; AVX512F-NEXT: vmovdqa32 %zmm0, %zmm23 {%k1} |
| ; AVX512F-NEXT: vmovdqa32 %zmm3, %zmm1 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm0 = <29,u,2,3,4,5,6,30,u,9,10,11,12,13,31,u> |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512F-NEXT: vpermi2d %zmm20, %zmm3, %zmm0 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm3 = <0,1,27,u,4,5,6,7,8,28,u,11,12,13,14,15> |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload |
| ; AVX512F-NEXT: vpermi2d %zmm20, %zmm5, %zmm3 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,29,2,3,4,5,6,7,30,9,10,11,12,13,14,31] |
| ; AVX512F-NEXT: vpermi2d %zmm27, %zmm0, %zmm5 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,1,2,27,4,5,6,7,8,9,28,11,12,13,14,15] |
| ; AVX512F-NEXT: vpermi2d %zmm27, %zmm3, %zmm0 |
| ; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512F-NEXT: vmovdqa64 %zmm1, 64(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm22, 128(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm2, 192(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm17, 256(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm13, 320(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm7, 384(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm18, 448(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm23, 512(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm21, 576(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm4, 640(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm16, 704(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm6, (%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm0, 768(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm5, 832(%rax) |
| ; AVX512F-NEXT: popq %rax |
| ; AVX512F-NEXT: vzeroupper |
| ; AVX512F-NEXT: retq |
| ; |
| ; AVX512BW-LABEL: store_i32_stride7_vf32: |
| ; AVX512BW: # %bb.0: |
| ; AVX512BW-NEXT: pushq %rax |
| ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm15 |
| ; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm18 |
| ; AVX512BW-NEXT: vmovdqa64 (%rsi), %zmm14 |
| ; AVX512BW-NEXT: vmovdqa64 64(%rsi), %zmm10 |
| ; AVX512BW-NEXT: vmovdqa64 (%rdx), %zmm1 |
| ; AVX512BW-NEXT: vmovdqa64 64(%rdx), %zmm21 |
| ; AVX512BW-NEXT: vmovdqa64 (%rcx), %zmm12 |
| ; AVX512BW-NEXT: vmovdqa64 64(%rcx), %zmm11 |
| ; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm3 |
| ; AVX512BW-NEXT: vmovdqa64 64(%r8), %zmm28 |
| ; AVX512BW-NEXT: vmovdqa64 (%r9), %zmm8 |
| ; AVX512BW-NEXT: vmovdqa64 64(%r9), %zmm20 |
| ; AVX512BW-NEXT: vmovdqa64 (%rax), %zmm9 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm30 = <u,u,u,u,14,30,u,u,u,u,u,15,31,u,u,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm21, %zmm2 |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm30, %zmm2 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm7 = <u,u,14,30,u,u,u,u,u,15,31,u,u,u,u,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm4 |
| ; AVX512BW-NEXT: vpermt2d %zmm10, %zmm7, %zmm4 |
| ; AVX512BW-NEXT: movw $6192, %cx # imm = 0x1830 |
| ; AVX512BW-NEXT: kmovd %ecx, %k1 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm4 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm0 = <u,u,2,3,4,5,30,u,u,9,10,11,12,31,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm28, %zmm4, %zmm0 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm22 = <u,u,u,u,12,28,u,u,u,u,u,13,29,u,u,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm2 |
| ; AVX512BW-NEXT: vpermt2d %zmm10, %zmm22, %zmm2 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm13 = <11,u,u,u,u,u,28,12,u,u,u,u,u,29,13,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm4 |
| ; AVX512BW-NEXT: vpermt2d %zmm21, %zmm13, %zmm4 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm4 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm0 = <0,27,u,u,4,5,6,7,28,u,u,11,12,13,14,29> |
| ; AVX512BW-NEXT: vpermi2d %zmm28, %zmm4, %zmm0 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,u,0,16,u,u,u,u,u,1,17,u,u,u,u,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm16 |
| ; AVX512BW-NEXT: vpermt2d %zmm12, %zmm2, %zmm16 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm0 = <0,16,u,u,u,u,u,1,17,u,u,u,u,u,2,18> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm6 |
| ; AVX512BW-NEXT: vpermt2d %zmm14, %zmm0, %zmm6 |
| ; AVX512BW-NEXT: movw $1548, %cx # imm = 0x60C |
| ; AVX512BW-NEXT: kmovd %ecx, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm16, %zmm6 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm27 = <u,u,u,u,0,16,u,u,u,u,u,1,17,u,u,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm31 |
| ; AVX512BW-NEXT: vpermt2d %zmm8, %zmm27, %zmm31 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm21, %zmm25 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm17 = <9,u,u,u,u,u,26,10,u,u,u,u,u,27,11,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm16 |
| ; AVX512BW-NEXT: vpermt2d %zmm18, %zmm17, %zmm16 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm21, %zmm24 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm4 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm21, %zmm29 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm26 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm21, %zmm23 |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm2, %zmm21 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm5 |
| ; AVX512BW-NEXT: vpermt2d %zmm10, %zmm0, %zmm18 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,u,u,u,4,5,16,u,u,u,u,11,12,17,u,u> |
| ; AVX512BW-NEXT: vpermt2d %zmm9, %zmm2, %zmm31 |
| ; AVX512BW-NEXT: movw $14448, %cx # imm = 0x3870 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm21, %zmm18 {%k2} |
| ; AVX512BW-NEXT: kmovd %ecx, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm31, %zmm6 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 %zmm28, %zmm31 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm28, %zmm19 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm28, %zmm21 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm28, %zmm0 |
| ; AVX512BW-NEXT: vpermt2d %zmm20, %zmm27, %zmm28 |
| ; AVX512BW-NEXT: vmovdqa64 64(%rax), %zmm27 |
| ; AVX512BW-NEXT: vpermt2d %zmm27, %zmm2, %zmm28 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm28, %zmm18 {%k2} |
| ; AVX512BW-NEXT: vpermi2d %zmm12, %zmm1, %zmm30 |
| ; AVX512BW-NEXT: vpermi2d %zmm14, %zmm15, %zmm7 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm30, %zmm7 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = <13,u,u,u,u,u,30,14,u,u,u,u,u,31,15,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm3, %zmm8, %zmm2 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm28 = <0,29,u,u,u,u,6,7,30,u,u,u,u,13,14,31> |
| ; AVX512BW-NEXT: vpermi2d %zmm9, %zmm2, %zmm28 |
| ; AVX512BW-NEXT: movw $-7741, %ax # imm = 0xE1C3 |
| ; AVX512BW-NEXT: kmovd %eax, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm28, %zmm7 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,9,25,u,u,u,u,u,10,26,u,u,u,u,u,11> |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm2, %zmm25 |
| ; AVX512BW-NEXT: movw $-31994, %ax # imm = 0x8306 |
| ; AVX512BW-NEXT: kmovd %eax, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm25, %zmm16 {%k2} |
| ; AVX512BW-NEXT: vpermi2d %zmm14, %zmm15, %zmm22 |
| ; AVX512BW-NEXT: vpermi2d %zmm1, %zmm12, %zmm13 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm22, %zmm13 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm22 = <u,11,27,u,u,u,u,u,12,28,u,u,u,u,u,13> |
| ; AVX512BW-NEXT: vpermi2d %zmm8, %zmm3, %zmm22 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm25 = <u,1,2,27,u,u,u,u,8,9,28,u,u,u,u,15> |
| ; AVX512BW-NEXT: vpermi2d %zmm9, %zmm22, %zmm25 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm22 = <u,u,u,9,25,u,u,u,u,u,10,26,u,u,u,u> |
| ; AVX512BW-NEXT: vpermt2d %zmm20, %zmm22, %zmm31 |
| ; AVX512BW-NEXT: movw $-30962, %ax # imm = 0x870E |
| ; AVX512BW-NEXT: kmovd %eax, %k1 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm25, %zmm13 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm25 = <u,u,u,3,4,25,u,u,u,u,10,11,26,u,u,u> |
| ; AVX512BW-NEXT: vpermt2d %zmm27, %zmm25, %zmm31 |
| ; AVX512BW-NEXT: movw $7224, %ax # imm = 0x1C38 |
| ; AVX512BW-NEXT: kmovd %eax, %k1 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm31, %zmm16 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm28 = <u,u,u,7,23,u,u,u,u,u,8,24,u,u,u,u> |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm28, %zmm24 |
| ; AVX512BW-NEXT: vpermi2d %zmm12, %zmm1, %zmm2 |
| ; AVX512BW-NEXT: vpermi2d %zmm15, %zmm14, %zmm17 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm17 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,7,23,u,u,u,u,u,8,24,u,u,u,u,u,9> |
| ; AVX512BW-NEXT: vpermt2d %zmm10, %zmm2, %zmm4 |
| ; AVX512BW-NEXT: movw $3096, %ax # imm = 0xC18 |
| ; AVX512BW-NEXT: kmovd %eax, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm24, %zmm4 {%k2} |
| ; AVX512BW-NEXT: vpermi2d %zmm8, %zmm3, %zmm22 |
| ; AVX512BW-NEXT: vpermt2d %zmm9, %zmm25, %zmm22 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm24 = <u,u,u,u,u,7,23,u,u,u,u,u,8,24,u,u> |
| ; AVX512BW-NEXT: vpermt2d %zmm20, %zmm24, %zmm19 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm22, %zmm17 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm22 = <22,u,u,u,u,5,6,23,u,u,u,u,12,13,24,u> |
| ; AVX512BW-NEXT: vpermt2d %zmm27, %zmm22, %zmm19 |
| ; AVX512BW-NEXT: movw $28897, %ax # imm = 0x70E1 |
| ; AVX512BW-NEXT: kmovd %eax, %k3 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm19, %zmm4 {%k3} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm19 = <u,u,u,u,u,5,21,u,u,u,u,u,6,22,u,u> |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm19, %zmm29 |
| ; AVX512BW-NEXT: vpermi2d %zmm12, %zmm1, %zmm28 |
| ; AVX512BW-NEXT: vpermi2d %zmm14, %zmm15, %zmm2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm28, %zmm2 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm25 = <u,u,u,5,21,u,u,u,u,u,6,22,u,u,u,u> |
| ; AVX512BW-NEXT: vpermt2d %zmm10, %zmm25, %zmm26 |
| ; AVX512BW-NEXT: movw $12384, %ax # imm = 0x3060 |
| ; AVX512BW-NEXT: kmovd %eax, %k1 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm29, %zmm26 {%k1} |
| ; AVX512BW-NEXT: vpermi2d %zmm8, %zmm3, %zmm24 |
| ; AVX512BW-NEXT: vpermt2d %zmm9, %zmm22, %zmm24 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm22 = <4,20,u,u,u,u,u,5,21,u,u,u,u,u,6,22> |
| ; AVX512BW-NEXT: vpermt2d %zmm20, %zmm22, %zmm21 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm24, %zmm2 {%k3} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm24 = <0,1,20,u,u,u,u,7,8,21,u,u,u,u,14,15> |
| ; AVX512BW-NEXT: vpermt2d %zmm27, %zmm24, %zmm21 |
| ; AVX512BW-NEXT: movw $15480, %ax # imm = 0x3C78 |
| ; AVX512BW-NEXT: kmovd %eax, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm26, %zmm21 {%k2} |
| ; AVX512BW-NEXT: vpermi2d %zmm12, %zmm1, %zmm19 |
| ; AVX512BW-NEXT: vpermi2d %zmm14, %zmm15, %zmm25 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm19, %zmm25 {%k1} |
| ; AVX512BW-NEXT: vpermi2d %zmm8, %zmm3, %zmm22 |
| ; AVX512BW-NEXT: vpermt2d %zmm9, %zmm24, %zmm22 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm25, %zmm22 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm19 = <u,u,u,u,u,3,19,u,u,u,u,u,4,20,u,u> |
| ; AVX512BW-NEXT: vpermt2d %zmm10, %zmm19, %zmm5 |
| ; AVX512BW-NEXT: vpermt2d %zmm14, %zmm19, %zmm15 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm10 = <2,18,u,u,u,u,u,3,19,u,u,u,u,u,4,20> |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm10, %zmm23 |
| ; AVX512BW-NEXT: vpermt2d %zmm12, %zmm10, %zmm1 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm5, %zmm23 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm5 = <u,u,2,18,u,u,u,u,u,3,19,u,u,u,u,u> |
| ; AVX512BW-NEXT: vpermt2d %zmm8, %zmm5, %zmm3 |
| ; AVX512BW-NEXT: vpermt2d %zmm20, %zmm5, %zmm0 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm5 = <u,u,2,3,18,u,u,u,u,9,10,19,u,u,u,u> |
| ; AVX512BW-NEXT: vpermt2d %zmm9, %zmm5, %zmm3 |
| ; AVX512BW-NEXT: vpermt2d %zmm27, %zmm5, %zmm0 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm15, %zmm1 {%k1} |
| ; AVX512BW-NEXT: movw $3612, %ax # imm = 0xE1C |
| ; AVX512BW-NEXT: kmovd %eax, %k1 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm23 {%k1} |
| ; AVX512BW-NEXT: vmovdqa32 %zmm3, %zmm1 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm0 = <29,u,2,3,4,5,6,30,u,9,10,11,12,13,31,u> |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermi2d %zmm20, %zmm3, %zmm0 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm3 = <0,1,27,u,4,5,6,7,8,28,u,11,12,13,14,15> |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermi2d %zmm20, %zmm5, %zmm3 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,29,2,3,4,5,6,7,30,9,10,11,12,13,14,31] |
| ; AVX512BW-NEXT: vpermi2d %zmm27, %zmm0, %zmm5 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,1,2,27,4,5,6,7,8,9,28,11,12,13,14,15] |
| ; AVX512BW-NEXT: vpermi2d %zmm27, %zmm3, %zmm0 |
| ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512BW-NEXT: vmovdqa64 %zmm1, 64(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm22, 128(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm2, 192(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm17, 256(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm13, 320(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm7, 384(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm18, 448(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm23, 512(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm21, 576(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm4, 640(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm16, 704(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm6, (%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm0, 768(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm5, 832(%rax) |
| ; AVX512BW-NEXT: popq %rax |
| ; AVX512BW-NEXT: vzeroupper |
| ; AVX512BW-NEXT: retq |
| %in.vec0 = load <32 x i32>, ptr %in.vecptr0, align 64 |
| %in.vec1 = load <32 x i32>, ptr %in.vecptr1, align 64 |
| %in.vec2 = load <32 x i32>, ptr %in.vecptr2, align 64 |
| %in.vec3 = load <32 x i32>, ptr %in.vecptr3, align 64 |
| %in.vec4 = load <32 x i32>, ptr %in.vecptr4, align 64 |
| %in.vec5 = load <32 x i32>, ptr %in.vecptr5, align 64 |
| %in.vec6 = load <32 x i32>, ptr %in.vecptr6, align 64 |
| %1 = shufflevector <32 x i32> %in.vec0, <32 x i32> %in.vec1, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63> |
| %2 = shufflevector <32 x i32> %in.vec2, <32 x i32> %in.vec3, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63> |
| %3 = shufflevector <32 x i32> %in.vec4, <32 x i32> %in.vec5, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63> |
| %4 = shufflevector <64 x i32> %1, <64 x i32> %2, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127> |
| %5 = shufflevector <32 x i32> %in.vec6, <32 x i32> poison, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef> |
| %6 = shufflevector <64 x i32> %3, <64 x i32> %5, <96 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95> |
| %7 = shufflevector <96 x i32> %6, <96 x i32> poison, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef> |
| %8 = shufflevector <128 x i32> %4, <128 x i32> %7, <224 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223> |
| %interleaved.vec = shufflevector <224 x i32> %8, <224 x i32> poison, <224 x i32> <i32 0, i32 32, i32 64, i32 96, i32 128, i32 160, i32 192, i32 1, i32 33, i32 65, i32 97, i32 129, i32 161, i32 193, i32 2, i32 34, i32 66, i32 98, i32 130, i32 162, i32 194, i32 3, i32 35, i32 67, i32 99, i32 131, i32 163, i32 195, i32 4, i32 36, i32 68, i32 100, i32 132, i32 164, i32 196, i32 5, i32 37, i32 69, i32 101, i32 133, i32 165, i32 197, i32 6, i32 38, i32 70, i32 102, i32 134, i32 166, i32 198, i32 7, i32 39, i32 71, i32 103, i32 135, i32 167, i32 199, i32 8, i32 40, i32 72, i32 104, i32 136, i32 168, i32 200, i32 9, i32 41, i32 73, i32 105, i32 137, i32 169, i32 201, i32 10, i32 42, i32 74, i32 106, i32 138, i32 170, i32 202, i32 11, i32 43, i32 75, i32 107, i32 139, i32 171, i32 203, i32 12, i32 44, i32 76, i32 108, i32 140, i32 172, i32 204, i32 13, i32 45, i32 77, i32 109, i32 141, i32 173, i32 205, i32 14, i32 46, i32 78, i32 110, i32 142, i32 174, i32 206, i32 15, i32 47, i32 79, i32 111, i32 143, i32 175, i32 207, i32 16, i32 48, i32 80, i32 112, i32 144, i32 176, i32 208, i32 17, i32 49, i32 81, i32 113, i32 145, i32 177, i32 209, i32 18, i32 50, i32 82, i32 114, i32 146, i32 178, i32 210, i32 19, i32 51, i32 83, i32 115, i32 147, i32 179, i32 211, i32 20, i32 52, i32 84, i32 116, i32 148, i32 180, i32 212, i32 21, i32 53, i32 85, i32 117, i32 149, i32 181, i32 213, i32 22, i32 54, i32 86, i32 118, i32 150, i32 182, i32 214, i32 23, i32 55, i32 87, i32 119, i32 151, i32 183, i32 215, i32 24, i32 56, i32 88, i32 120, i32 152, i32 184, i32 216, i32 25, i32 57, i32 89, i32 121, i32 153, i32 185, i32 217, i32 26, i32 58, i32 90, i32 122, i32 154, i32 186, i32 218, i32 27, i32 59, i32 91, i32 123, i32 155, i32 187, i32 219, i32 28, i32 60, i32 92, i32 124, i32 156, i32 188, i32 220, i32 29, i32 61, i32 93, i32 125, i32 157, i32 189, i32 221, i32 30, i32 62, i32 94, i32 126, i32 158, i32 190, i32 222, i32 31, i32 63, i32 95, i32 127, i32 159, i32 191, i32 223> |
| store <224 x i32> %interleaved.vec, ptr %out.vec, align 64 |
| ret void |
| } |
| |
| define void @store_i32_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %out.vec) nounwind { |
| ; SSE-LABEL: store_i32_stride7_vf64: |
| ; SSE: # %bb.0: |
| ; SSE-NEXT: subq $2760, %rsp # imm = 0xAC8 |
| ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; SSE-NEXT: movdqa (%rdi), %xmm6 |
| ; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa (%rsi), %xmm5 |
| ; SSE-NEXT: movdqa 16(%rsi), %xmm4 |
| ; SSE-NEXT: movaps (%rdx), %xmm3 |
| ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 16(%rdx), %xmm7 |
| ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps (%rcx), %xmm9 |
| ; SSE-NEXT: movaps 16(%rcx), %xmm10 |
| ; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps (%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 16(%r8), %xmm8 |
| ; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa (%r9), %xmm11 |
| ; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 16(%r9), %xmm12 |
| ; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa (%rax), %xmm15 |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm9[1,1] |
| ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[1,1,1,1] |
| ; SSE-NEXT: movss {{.*#+}} xmm3 = xmm1[0],xmm3[1,2,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm6, %xmm0 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm5[2],xmm0[3],xmm5[3] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm15[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm11[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm4[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movaps %xmm8, %xmm0 |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm10[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 16(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm12[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 16(%rdi), %xmm5 |
| ; SSE-NEXT: movdqa %xmm5, %xmm0 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 32(%rsi), %xmm1 |
| ; SSE-NEXT: movaps 32(%rdx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm1, %xmm2 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm4, %xmm1 |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps 32(%rcx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 32(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 32(%r9), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 32(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 32(%rdi), %xmm6 |
| ; SSE-NEXT: movdqa %xmm6, %xmm0 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 48(%rsi), %xmm2 |
| ; SSE-NEXT: movdqa 48(%rdx), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movaps 48(%rcx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 48(%r8), %xmm7 |
| ; SSE-NEXT: movaps %xmm7, %xmm0 |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 48(%r9), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 48(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 48(%rdi), %xmm10 |
| ; SSE-NEXT: movdqa %xmm10, %xmm0 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 64(%rsi), %xmm1 |
| ; SSE-NEXT: movaps 64(%rdx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm1, %xmm2 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm4, %xmm1 |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps 64(%rcx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 64(%r8), %xmm13 |
| ; SSE-NEXT: movaps %xmm13, %xmm0 |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 64(%r9), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 64(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 64(%rdi), %xmm14 |
| ; SSE-NEXT: movdqa %xmm14, %xmm0 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 80(%rsi), %xmm2 |
| ; SSE-NEXT: movdqa 80(%rdx), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movaps 80(%rcx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 80(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 80(%r9), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 80(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 80(%rdi), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 96(%rsi), %xmm1 |
| ; SSE-NEXT: movaps 96(%rdx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm1, %xmm2 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm4, %xmm1 |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps 96(%rcx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 96(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 96(%r9), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 96(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 96(%rdi), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 112(%rsi), %xmm2 |
| ; SSE-NEXT: movdqa 112(%rdx), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movaps 112(%rcx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 112(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 112(%r9), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 112(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 112(%rdi), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 128(%rsi), %xmm1 |
| ; SSE-NEXT: movaps 128(%rdx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm1, %xmm2 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm4, %xmm1 |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps 128(%rcx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 128(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, (%rsp) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 128(%r9), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 128(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 128(%rdi), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 144(%rsi), %xmm2 |
| ; SSE-NEXT: movdqa 144(%rdx), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movaps 144(%rcx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 144(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 144(%r9), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 144(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 144(%rdi), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 160(%rsi), %xmm1 |
| ; SSE-NEXT: movaps 160(%rdx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm1, %xmm2 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm4, %xmm1 |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps 160(%rcx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 160(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 160(%r9), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 160(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 160(%rdi), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 176(%rsi), %xmm2 |
| ; SSE-NEXT: movdqa 176(%rdx), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movaps 176(%rcx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 176(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 176(%r9), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 176(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 176(%rdi), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 192(%rsi), %xmm1 |
| ; SSE-NEXT: movaps 192(%rdx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm1, %xmm2 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm4, %xmm1 |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps 192(%rcx), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 192(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 192(%r9), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 192(%rax), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 192(%rdi), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 208(%rsi), %xmm3 |
| ; SSE-NEXT: movdqa 208(%rdx), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm3[1,1,1,1] |
| ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movaps 208(%rcx), %xmm8 |
| ; SSE-NEXT: movaps 208(%r8), %xmm4 |
| ; SSE-NEXT: movaps %xmm4, %xmm0 |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm8[1,1] |
| ; SSE-NEXT: movaps %xmm8, %xmm11 |
| ; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 208(%r9), %xmm12 |
| ; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 208(%rax), %xmm8 |
| ; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm8[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm12[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa 208(%rdi), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[3,3,3,3] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm11[3,3,3,3] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; SSE-NEXT: movdqa %xmm8, %xmm0 |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm12[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 224(%rsi), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps 224(%rdx), %xmm2 |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3] |
| ; SSE-NEXT: movaps 224(%rcx), %xmm3 |
| ; SSE-NEXT: movaps 224(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, %xmm1 |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm3[1,1] |
| ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm1[2,0] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm0, %xmm1 |
| ; SSE-NEXT: movaps 224(%r9), %xmm11 |
| ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm11[0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm11, %xmm1 |
| ; SSE-NEXT: movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm0[1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[3,3,3,3] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1] |
| ; SSE-NEXT: movaps 224(%rax), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm0[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm1[0,2] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 240(%rsi), %xmm1 |
| ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa 240(%rdx), %xmm0 |
| ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,1,1,1] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1] |
| ; SSE-NEXT: movaps 240(%rcx), %xmm8 |
| ; SSE-NEXT: movaps 240(%r8), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, %xmm1 |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm8[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm1[2,0] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm0, %xmm1 |
| ; SSE-NEXT: movaps 240(%r9), %xmm12 |
| ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm12[0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm12, %xmm1 |
| ; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm0[1] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3] |
| ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm8[3,3,3,3] |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1] |
| ; SSE-NEXT: movaps 240(%rax), %xmm0 |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm0[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm1[0,2] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: movaps %xmm9, %xmm2 |
| ; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm9[0],xmm0[1],xmm9[1] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm4, %xmm3 |
| ; SSE-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm15 = xmm15[0,1],xmm4[1,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm4, %xmm3 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,1],xmm15[0,2] |
| ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm4, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm9[2],xmm0[3],xmm9[3] |
| ; SSE-NEXT: movaps %xmm1, %xmm3 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm3 = xmm3[2],xmm2[2],xmm3[3],xmm2[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm4[0] |
| ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm5, %xmm2 |
| ; SSE-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm2 = xmm2[0],mem[0],xmm2[1],mem[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm5, %xmm0 |
| ; SSE-NEXT: shufps $197, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[1,1],mem[0,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm2, %xmm5 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm4, %xmm5 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm5 = xmm5[2],xmm1[2],xmm5[3],xmm1[3] |
| ; SSE-NEXT: movaps %xmm2, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movdqa %xmm2, %xmm0 |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| ; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm6, %xmm3 |
| ; SSE-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm6[1,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm4, %xmm6 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm6 = xmm6[0],xmm3[0],xmm6[1],xmm3[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,1],xmm0[0,2] |
| ; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm4, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3] |
| ; SSE-NEXT: movdqa %xmm2, %xmm3 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm2, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0] |
| ; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm10, %xmm3 |
| ; SSE-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm10, %xmm0 |
| ; SSE-NEXT: shufps $197, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[1,1],mem[0,3] |
| ; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm7, %xmm5 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm1, %xmm5 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm5 = xmm5[2],xmm2[2],xmm5[3],xmm2[3] |
| ; SSE-NEXT: movaps %xmm7, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movdqa %xmm2, %xmm0 |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movdqa %xmm14, %xmm3 |
| ; SSE-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm14[1,3] |
| ; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm13, %xmm5 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm0[0,2] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm13, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3] |
| ; SSE-NEXT: movdqa %xmm2, %xmm3 |
| ; SSE-NEXT: punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3] |
| ; SSE-NEXT: punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm0[0] |
| ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm3, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm5 |
| ; SSE-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm5 = xmm5[0],mem[0],xmm5[1],mem[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: shufps $197, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[1,1],mem[0,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm6 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm6 = xmm6[0],xmm4[0],xmm6[1],xmm4[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm2, %xmm6 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm6 = xmm6[2],xmm3[2],xmm6[3],xmm3[3] |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm6 = xmm6[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm3, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm5 |
| ; SSE-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm5 = xmm5[0],mem[0],xmm5[1],mem[1] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[1,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm6 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm6 = xmm6[0],xmm4[0],xmm6[1],xmm4[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,1],xmm0[0,2] |
| ; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3] |
| ; SSE-NEXT: movaps %xmm3, %xmm1 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm1 = xmm1[2],xmm2[2],xmm1[3],xmm2[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm3, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm5 |
| ; SSE-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm5 = xmm5[0],mem[0],xmm5[1],mem[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: shufps $197, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[1,1],mem[0,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm6 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm6 = xmm6[0],xmm4[0],xmm6[1],xmm4[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm2, %xmm6 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm6 = xmm6[2],xmm3[2],xmm6[3],xmm3[3] |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm6 = xmm6[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm3, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm5 |
| ; SSE-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm5 = xmm5[0],mem[0],xmm5[1],mem[1] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[1,3] |
| ; SSE-NEXT: movaps (%rsp), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm6 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm6 = xmm6[0],xmm4[0],xmm6[1],xmm4[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,1],xmm0[0,2] |
| ; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3] |
| ; SSE-NEXT: movaps %xmm3, %xmm1 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm1 = xmm1[2],xmm2[2],xmm1[3],xmm2[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm3, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm5 |
| ; SSE-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm5 = xmm5[0],mem[0],xmm5[1],mem[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: shufps $197, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[1,1],mem[0,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm6 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm6 = xmm6[0],xmm4[0],xmm6[1],xmm4[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm2, %xmm6 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm6 = xmm6[2],xmm3[2],xmm6[3],xmm3[3] |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm6 = xmm6[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm3, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm5 |
| ; SSE-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm5 = xmm5[0],mem[0],xmm5[1],mem[1] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[1,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm6 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm6 = xmm6[0],xmm4[0],xmm6[1],xmm4[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,1],xmm0[0,2] |
| ; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3] |
| ; SSE-NEXT: movaps %xmm3, %xmm1 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm1 = xmm1[2],xmm2[2],xmm1[3],xmm2[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm3, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm4 |
| ; SSE-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = xmm4[0],mem[0],xmm4[1],mem[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: shufps $197, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[1,1],mem[0,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm4 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,1],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm2, %xmm4 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm4 = xmm4[2],xmm3[2],xmm4[3],xmm3[3] |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm5[2],xmm0[3],xmm5[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm15, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm4 |
| ; SSE-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = xmm4[0],mem[0],xmm4[1],mem[1] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[1,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, %xmm5 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm0[0,2] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm1, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3] |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm15 = xmm15[2],xmm2[2],xmm15[3],xmm2[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm15 = xmm15[0],xmm0[0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm14, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm1[0,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,1],xmm14[3,3] |
| ; SSE-NEXT: movaps %xmm1, %xmm4 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm2, %xmm13 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm13 = xmm13[2],xmm3[2],xmm13[3],xmm3[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm2[0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm14 = xmm14[0],xmm1[0],xmm14[1],xmm1[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm14 = xmm14[0,1],xmm3[2,0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm3, %xmm10 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm10 = xmm10[0],xmm5[0],xmm10[1],xmm5[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm10 = xmm10[0,1],xmm0[2,0] |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm3 = xmm3[2],xmm5[2],xmm3[3],xmm5[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm13 = xmm13[0],xmm3[0] |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[3,3],xmm1[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,2],xmm2[2,0] |
| ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm2, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1] |
| ; SSE-NEXT: movaps 224(%rdi), %xmm6 |
| ; SSE-NEXT: movaps %xmm6, %xmm9 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm9 = xmm9[0],xmm3[0],xmm9[1],xmm3[1] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm9 = xmm9[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm6, %xmm0 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm1[0,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,2],xmm0[2,0] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm6, %xmm0 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,1],xmm1[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,2],xmm0[0,1] |
| ; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm2, %xmm7 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm7 = xmm7[0,1],xmm11[2,0] |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[3,1],xmm1[2,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[2,0],xmm2[2,0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, %xmm11 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm11 = xmm11[2],xmm8[2],xmm11[3],xmm8[3] |
| ; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0] |
| ; SSE-NEXT: movaps %xmm0, %xmm4 |
| ; SSE-NEXT: movaps 240(%rdi), %xmm3 |
| ; SSE-NEXT: movaps %xmm3, %xmm5 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: unpcklps {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm8[2,0] |
| ; SSE-NEXT: movaps %xmm3, %xmm8 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm8 = xmm8[1,1],xmm2[0,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm8[2,0] |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps %xmm3, %xmm8 |
| ; SSE-NEXT: unpckhps {{.*#+}} xmm8 = xmm8[2],xmm1[2],xmm8[3],xmm1[3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm2[1,1] |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm8[0,1] |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps {{.*#+}} xmm11 = xmm11[0,1],xmm12[2,0] |
| ; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm1[3,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[3,1],xmm2[2,3] |
| ; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[2,0],xmm4[2,0] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,3],xmm1[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm12[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm0 = xmm4[0],xmm0[1,2,3] |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm12 = xmm12[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,3],xmm12[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm0 = xmm4[0],xmm0[1,2,3] |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm2 = xmm4[0],xmm2[1,2,3] |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps (%rsp), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, (%rsp) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps %xmm0, %xmm4 |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,3],xmm4[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm0 = xmm4[0],xmm0[1,2,3] |
| ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3] |
| ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm12 = xmm12[0,3],xmm0[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm12 = xmm4[0],xmm12[1,2,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm8 = xmm8[0,3],xmm0[2,0] |
| ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm1[2,3,2,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm8 = xmm4[0],xmm8[1,2,3] |
| ; SSE-NEXT: shufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,3],xmm1[2,0] |
| ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload |
| ; SSE-NEXT: # xmm4 = mem[3,3,3,3] |
| ; SSE-NEXT: movss {{.*#+}} xmm0 = xmm4[0],xmm0[1,2,3] |
| ; SSE-NEXT: movaps %xmm0, %xmm4 |
| ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; SSE-NEXT: movaps %xmm3, 1760(%rax) |
| ; SSE-NEXT: movaps %xmm11, 1744(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1728(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1696(%rax) |
| ; SSE-NEXT: movaps %xmm5, 1680(%rax) |
| ; SSE-NEXT: movaps %xmm6, 1648(%rax) |
| ; SSE-NEXT: movaps %xmm7, 1632(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1616(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1584(%rax) |
| ; SSE-NEXT: movaps %xmm9, 1568(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1536(%rax) |
| ; SSE-NEXT: movaps %xmm13, 1520(%rax) |
| ; SSE-NEXT: movaps %xmm10, 1472(%rax) |
| ; SSE-NEXT: movaps %xmm14, 1456(%rax) |
| ; SSE-NEXT: movaps %xmm15, 1408(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1360(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1344(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1296(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1248(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1232(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1184(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1136(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1120(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1072(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1024(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1008(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 960(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 912(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 896(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 848(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 800(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 784(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 736(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 688(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 672(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 624(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 576(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 560(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 512(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 464(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 448(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 400(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 352(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 336(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 288(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 240(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 224(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 176(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 128(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 112(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 64(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 16(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, (%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1776(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1712(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1664(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1600(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1552(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1504(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, 1488(%rax) |
| ; SSE-NEXT: movaps %xmm4, 1440(%rax) |
| ; SSE-NEXT: movaps %xmm8, 1424(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, 1392(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, 1376(%rax) |
| ; SSE-NEXT: movaps %xmm12, 1328(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1312(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, 1280(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, 1264(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1216(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1200(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, 1168(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm1, 1152(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1104(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1088(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1056(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 1040(%rax) |
| ; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 992(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 976(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 944(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 928(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 880(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 864(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 832(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 816(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 768(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 752(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 720(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 704(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 656(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 640(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 608(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 592(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 544(%rax) |
| ; SSE-NEXT: movaps %xmm2, 528(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 496(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 480(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 432(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 416(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 384(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 368(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 320(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 304(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 272(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 256(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 208(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 192(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 160(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 144(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 96(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 80(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 48(%rax) |
| ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; SSE-NEXT: movaps %xmm0, 32(%rax) |
| ; SSE-NEXT: addq $2760, %rsp # imm = 0xAC8 |
| ; SSE-NEXT: retq |
| ; |
| ; AVX1-ONLY-LABEL: store_i32_stride7_vf64: |
| ; AVX1-ONLY: # %bb.0: |
| ; AVX1-ONLY-NEXT: subq $3432, %rsp # imm = 0xD68 |
| ; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX1-ONLY-NEXT: vmovaps 224(%rdi), %ymm5 |
| ; AVX1-ONLY-NEXT: vmovaps 224(%rsi), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 224(%rdx), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 224(%rcx), %ymm0 |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 224(%r8), %ymm3 |
| ; AVX1-ONLY-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 224(%rax), %ymm4 |
| ; AVX1-ONLY-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm5[1],ymm2[3],ymm5[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm5[0,1],ymm1[2,0],ymm5[4,5],ymm1[6,4] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = mem[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm4[2,3],ymm0[2,3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%r9), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%r8), %xmm5 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%rdi), %xmm1 |
| ; AVX1-ONLY-NEXT: vmovaps (%rsi), %xmm4 |
| ; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm4[0],xmm1[0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, %xmm6 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vmovaps (%rcx), %xmm7 |
| ; AVX1-ONLY-NEXT: vmovaps (%rdx), %xmm8 |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm8[0],xmm7[0],xmm8[1],xmm7[1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,0,1] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm5[0],xmm3[0],xmm5[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm3[1,1],xmm5[1,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm6[1],xmm4[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm4[1,1],xmm1[0,2] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm8[1],xmm7[1],zero |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%rdi), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%rsi), %ymm0 |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,1],ymm1[1,1],ymm0[5,5],ymm1[5,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps (%rdx), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%rcx), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1],ymm2[1,1],ymm1[5,5],ymm2[5,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovaps (%r8), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps (%r9), %ymm12 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm12[0],ymm2[0],ymm12[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm2[2,1],ymm1[6,4],ymm2[6,5] |
| ; AVX1-ONLY-NEXT: vmovaps (%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rdi), %xmm1 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rsi), %xmm4 |
| ; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm4[0],xmm1[0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, %xmm7 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rcx), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rdx), %xmm2 |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm3[0],xmm2[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, %xmm8 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, %xmm9 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,0,1] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovaps 32(%r9), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, (%rsp) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 32(%r8), %xmm6 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm6[0],xmm3[0],xmm6[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm3[1,1],xmm6[1,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm7[1],xmm4[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm4[1,1],xmm1[0,2] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm9[1],xmm8[1],zero |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rdi), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rsi), %ymm0 |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,1],ymm1[1,1],ymm0[5,5],ymm1[5,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rdx), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rcx), %ymm8 |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1],ymm8[1,1],ymm1[5,5],ymm8[5,5] |
| ; AVX1-ONLY-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovaps 32(%r8), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 32(%r9), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm2[2,1],ymm1[6,4],ymm2[6,5] |
| ; AVX1-ONLY-NEXT: vmovaps 32(%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rdi), %xmm1 |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rsi), %xmm6 |
| ; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm6[0],xmm1[0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, %xmm5 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rcx), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rdx), %xmm2 |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm3[0],xmm2[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, %xmm7 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, %xmm9 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,0,1] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovaps 64(%r9), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 64(%r8), %xmm4 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm3[1,1],xmm4[1,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm5[1],xmm6[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm6[1,1],xmm1[0,2] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm9[1],xmm7[1],zero |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rdi), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rsi), %ymm0 |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,1],ymm1[1,1],ymm0[5,5],ymm1[5,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rdx), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rcx), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1],ymm2[1,1],ymm1[5,5],ymm2[5,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovaps 64(%r8), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 64(%r9), %ymm15 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm15[0],ymm2[0],ymm15[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm2[2,1],ymm1[6,4],ymm2[6,5] |
| ; AVX1-ONLY-NEXT: vmovaps 64(%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rdi), %xmm1 |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rsi), %xmm5 |
| ; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm5[0],xmm1[0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, %xmm6 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rcx), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rdx), %xmm2 |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm3[0],xmm2[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, %xmm7 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, %xmm9 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,0,1] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovaps 96(%r9), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 96(%r8), %xmm4 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm3[1,1],xmm4[1,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm6[1],xmm5[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm5[1,1],xmm1[0,2] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm9[1],xmm7[1],zero |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rdi), %ymm0 |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rsi), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm1[1,1],ymm0[1,1],ymm1[5,5],ymm0[5,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rdx), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rcx), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm2[1,1],ymm1[1,1],ymm2[5,5],ymm1[5,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovaps 96(%r8), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 96(%r9), %ymm10 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm10[0],ymm2[0],ymm10[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm2[2,1],ymm1[6,4],ymm2[6,5] |
| ; AVX1-ONLY-NEXT: vmovaps 96(%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 128(%rdi), %xmm1 |
| ; AVX1-ONLY-NEXT: vmovaps 128(%rsi), %xmm4 |
| ; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm4[0],xmm1[0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, %xmm5 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vmovaps 128(%rcx), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps 128(%rdx), %xmm2 |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm3[0],xmm2[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, %xmm6 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, %xmm9 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,0,1] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovaps 128(%r9), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 128(%r8), %xmm7 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm7[0],xmm3[0],xmm7[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX1-ONLY-NEXT: vmovaps 128(%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm3[1,1],xmm7[1,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm5[1],xmm4[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm4[1,1],xmm1[0,2] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm9[1],xmm6[1],zero |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 128(%rdi), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 128(%rsi), %ymm0 |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,1],ymm1[1,1],ymm0[5,5],ymm1[5,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 128(%rdx), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 128(%rcx), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm2[1,1],ymm1[1,1],ymm2[5,5],ymm1[5,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovaps 128(%r8), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 128(%r9), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm2[2,1],ymm1[6,4],ymm2[6,5] |
| ; AVX1-ONLY-NEXT: vmovaps 128(%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 160(%rdi), %xmm1 |
| ; AVX1-ONLY-NEXT: vmovaps 160(%rsi), %xmm4 |
| ; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm4[0],xmm1[0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, %xmm5 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vmovaps 160(%rcx), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps 160(%rdx), %xmm2 |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm3[0],xmm2[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, %xmm6 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, %xmm9 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,0,1] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovaps 160(%r9), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 160(%r8), %xmm7 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm7[0],xmm3[0],xmm7[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX1-ONLY-NEXT: vmovaps 160(%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm3[1,1],xmm7[1,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm5[1],xmm4[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm4[1,1],xmm1[0,2] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm9[1],xmm6[1],zero |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 160(%rdi), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 160(%rsi), %ymm0 |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,1],ymm1[1,1],ymm0[5,5],ymm1[5,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 160(%rdx), %ymm6 |
| ; AVX1-ONLY-NEXT: vmovaps 160(%rcx), %ymm1 |
| ; AVX1-ONLY-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm6[1,1],ymm1[1,1],ymm6[5,5],ymm1[5,5] |
| ; AVX1-ONLY-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovaps 160(%r8), %ymm9 |
| ; AVX1-ONLY-NEXT: vmovaps 160(%r9), %ymm13 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm13[0],ymm9[0],ymm13[2],ymm9[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,0],ymm9[2,1],ymm1[6,4],ymm9[6,5] |
| ; AVX1-ONLY-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 160(%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 192(%rdi), %xmm1 |
| ; AVX1-ONLY-NEXT: vmovaps 192(%rsi), %xmm11 |
| ; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm0 = xmm11[0],xmm1[0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, %xmm14 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vmovaps 192(%rcx), %xmm5 |
| ; AVX1-ONLY-NEXT: vmovaps 192(%rdx), %xmm4 |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm4[0],xmm5[0],xmm4[1],xmm5[1] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,1,0,1] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovaps 192(%r9), %xmm3 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 192(%r8), %xmm7 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm1 = xmm7[0],xmm3[0],xmm7[1],xmm3[1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX1-ONLY-NEXT: vmovaps 192(%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm3[1,1],xmm7[1,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm14[1],xmm11[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm11[1,1],xmm1[0,2] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm4[1],xmm5[1],zero |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 192(%rdi), %ymm4 |
| ; AVX1-ONLY-NEXT: vmovaps 192(%rsi), %ymm5 |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm5[1,1],ymm4[1,1],ymm5[5,5],ymm4[5,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 192(%rdx), %ymm7 |
| ; AVX1-ONLY-NEXT: vmovaps 192(%rcx), %ymm14 |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm7[1,1],ymm14[1,1],ymm7[5,5],ymm14[5,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovaps 192(%r8), %ymm11 |
| ; AVX1-ONLY-NEXT: vmovaps 192(%r9), %ymm3 |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm3[0],ymm11[0],ymm3[2],ymm11[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,0],ymm11[2,1],ymm2[6,4],ymm11[6,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = mem[2,3],ymm2[2,3] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm2, %xmm2 |
| ; AVX1-ONLY-NEXT: vshufpd {{.*#+}} ymm0 = ymm2[0],ymm0[0],ymm2[2],ymm0[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4,5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,2,0,4,5,6,4] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm12[1],ymm1[3],ymm12[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm12[1,1],ymm1[0,2],ymm12[5,5],ymm1[4,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 16(%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,2,0,4,5,6,4] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm8[2],ymm1[3],ymm8[3],ymm1[6],ymm8[6],ymm1[7],ymm8[7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm2[1,1],ymm1[0,2],ymm2[5,5],ymm1[4,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 48(%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,2,0,4,5,6,4] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm1 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm1 = ymm8[2],mem[2],ymm8[3],mem[3],ymm8[6],mem[6],ymm8[7],mem[7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm15[1],ymm1[3],ymm15[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm15[1,1],ymm1[0,2],ymm15[5,5],ymm1[4,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 80(%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm0 = ymm15[1],mem[1],ymm15[3],mem[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,2,0,4,5,6,4] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm1 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm1 = ymm12[2],mem[2],ymm12[3],mem[3],ymm12[6],mem[6],ymm12[7],mem[7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm10[1],ymm1[3],ymm10[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm10[1,1],ymm1[0,2],ymm10[5,5],ymm1[4,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 112(%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,2,0,4,5,6,4] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm1 = ymm1[2],mem[2],ymm1[3],mem[3],ymm1[6],mem[6],ymm1[7],mem[7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm2[1,1],ymm1[0,2],ymm2[5,5],ymm1[4,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 144(%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhpd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm0 = ymm0[1],mem[1],ymm0[3],mem[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,2,0,4,5,6,4] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{.*#+}} ymm1 = ymm6[2],ymm2[2],ymm6[3],ymm2[3],ymm6[6],ymm2[6],ymm6[7],ymm2[7] |
| ; AVX1-ONLY-NEXT: vmovaps %ymm2, %ymm10 |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm9[1],ymm13[1],ymm9[3],ymm13[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,1],ymm1[0,2],ymm13[5,5],ymm1[4,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 176(%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps %ymm4, %ymm6 |
| ; AVX1-ONLY-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm5[1],ymm4[1],ymm5[3],ymm4[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,1,2,0,4,5,6,4] |
| ; AVX1-ONLY-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhps {{.*#+}} ymm1 = ymm7[2],ymm14[2],ymm7[3],ymm14[3],ymm7[6],ymm14[6],ymm7[7],ymm14[7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3] |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm11[1],ymm3[1],ymm11[3],ymm3[3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm3[1,1],ymm1[0,2],ymm3[5,5],ymm1[4,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps 208(%rax), %xmm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps 224(%rsi), %xmm2 |
| ; AVX1-ONLY-NEXT: vmovaps 224(%rdi), %xmm11 |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm11[1],xmm2[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm2[1,1],xmm0[0,2] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm3 |
| ; AVX1-ONLY-NEXT: vmovaps 224(%rcx), %xmm4 |
| ; AVX1-ONLY-NEXT: vmovaps 224(%rdx), %xmm1 |
| ; AVX1-ONLY-NEXT: vinsertps {{.*#+}} xmm0 = zero,xmm1[1],xmm4[1],zero |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm0[1,2],ymm3[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 228(%r8), %ymm13 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm13[3],ymm3[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 228(%r9), %ymm13 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm13[4,5],ymm3[6,7] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, 224(%rax), %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2,3,4],ymm0[5],ymm3[6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm1[1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm3 = xmm11[3,3],xmm2[3,3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5,6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 232(%r9), %xmm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm3[3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 232(%rax), %ymm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4],ymm0[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm14[3,3],ymm7[3,3],ymm14[7,7],ymm7[7,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm3 = ymm6[3,3],ymm5[3,3],ymm6[7,7],ymm5[7,7] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm3, %xmm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 220(%r8), %ymm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 220(%r9), %ymm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 216(%rax), %ymm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0],ymm0[1,2,3,4,5,6],ymm3[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm4[0],xmm1[1],xmm4[1] |
| ; AVX1-ONLY-NEXT: vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm11[0] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm1 = xmm1[2,0],xmm11[2,1] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,1,0,1] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, 224(%r8), %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 224(%r9), %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 224(%rax), %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6],ymm0[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm4[3,3],ymm3[3,3],ymm4[7,7],ymm3[7,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm6[3,3],ymm2[3,3],ymm6[7,7],ymm2[7,7] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 252(%r8), %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 252(%r9), %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 248(%rax), %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm3[0],ymm4[0],ymm3[2],ymm4[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm4[3,1],ymm0[0,2],ymm4[7,5],ymm0[4,6] |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} ymm1 = ymm6[0],ymm2[0],ymm6[1],ymm2[1],ymm6[4],ymm2[4],ymm6[5],ymm2[5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 236(%r8), %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 236(%r9), %xmm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm3[1,1],ymm4[1,1],ymm3[5,5],ymm4[5,5] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm2[1,1],ymm6[1,1],ymm2[5,5],ymm6[5,5] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 240(%r8), %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 240(%r9), %xmm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vbroadcastss 240(%rax), %ymm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = xmm2[2],mem[2],xmm2[3],mem[3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = mem[2,2,2,2] |
| ; AVX1-ONLY-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = mem[0,1,2],xmm1[3] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 8(%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm0 = ymm0[3,3],mem[3,3],ymm0[7,7],mem[7,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[3,3],ymm7[3,3],ymm1[7,7],ymm7[7,7] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,3],ymm1[1,2],ymm2[6,7],ymm1[5,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,2,3,1,4,6,7,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = xmm2[2],mem[2],xmm2[3],mem[3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vpermilps $170, (%rsp), %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = mem[2,2,2,2] |
| ; AVX1-ONLY-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = mem[0,1,2],xmm1[3] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 40(%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm0 = ymm0[3,3],mem[3,3],ymm0[7,7],mem[7,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,3],ymm1[1,2],ymm2[6,7],ymm1[5,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,2,3,1,4,6,7,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = xmm2[2],mem[2],xmm2[3],mem[3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = mem[2,2,2,2] |
| ; AVX1-ONLY-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = mem[0,1,2],xmm1[3] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 72(%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[3,3],ymm8[3,3],ymm0[7,7],ymm8[7,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[3,3],ymm6[3,3],ymm1[7,7],ymm6[7,7] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,3],ymm1[1,2],ymm2[6,7],ymm1[5,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,2,3,1,4,6,7,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = xmm2[2],mem[2],xmm2[3],mem[3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = mem[2,2,2,2] |
| ; AVX1-ONLY-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = mem[0,1,2],xmm1[3] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 104(%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm0[3,3],ymm12[3,3],ymm0[7,7],ymm12[7,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovaps %ymm15, %ymm4 |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[3,3],ymm15[3,3],ymm1[7,7],ymm15[7,7] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm1 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,3],ymm1[1,2],ymm2[6,7],ymm1[5,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,2,3,1,4,6,7,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = xmm2[2],mem[2],xmm2[3],mem[3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX1-ONLY-NEXT: vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = mem[2,2,2,2] |
| ; AVX1-ONLY-NEXT: vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm1 = mem[0,1,2],xmm1[3] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 136(%rax), %ymm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm0 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm0 = ymm14[3,3],mem[3,3],ymm14[7,7],mem[7,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm2 = ymm12[3,3],ymm1[3,3],ymm12[7,7],ymm1[7,7] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm2, %xmm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm2 = ymm2[3,3],mem[3,3],ymm2[7,7],mem[7,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm2 = ymm3[2,3],ymm2[1,2],ymm3[6,7],ymm2[5,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2,3,1,4,6,7,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0],ymm0[1,2,3,4],ymm2[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm2 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm2 = xmm3[2],mem[2],xmm3[3],mem[3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm0[5,6],ymm2[7] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm2 = xmm15[2,2,2,2] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} xmm2 = xmm13[0,1,2],xmm2[3] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 168(%rax), %ymm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3,4],ymm0[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm0 = ymm10[3,3],ymm11[3,3],ymm10[7,7],ymm11[7,7] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm2 = ymm9[3,3],ymm10[3,3],ymm9[7,7],ymm10[7,7] |
| ; AVX1-ONLY-NEXT: vextractf128 $1, %ymm2, %xmm2 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm2 = ymm2[3,3],mem[3,3],ymm2[7,7],mem[7,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm2 = ymm3[2,3],ymm2[1,2],ymm3[6,7],ymm2[5,6] |
| ; AVX1-ONLY-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,2,3] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2,3,1,4,6,7,5] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0],ymm0[1,2,3,4],ymm2[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm0 = xmm0[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm2 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm2 = xmm3[2],mem[2],xmm3[3],mem[3] |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2 |
| ; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm0[5,6],ymm2[7] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm2 = xmm8[2,2,2,2] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} xmm2 = xmm5[0,1,2],xmm2[3] |
| ; AVX1-ONLY-NEXT: vbroadcastsd 200(%rax), %ymm3 |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3,4],ymm0[5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} ymm0 = ymm0[0],ymm7[0],ymm0[1],ymm7[1],ymm0[4],ymm7[4],ymm0[5],ymm7[5] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm2[0],ymm7[0],ymm2[2],ymm7[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm3 = ymm7[3,1],ymm3[0,2],ymm7[7,5],ymm3[4,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5],ymm3[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm3 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm3 = xmm2[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm3 = xmm3[0,1,2],mem[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1,2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[4],mem[4],ymm0[5],mem[5] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm2[0],ymm7[0],ymm2[2],ymm7[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm3 = ymm7[3,1],ymm3[0,2],ymm7[7,5],ymm3[4,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5],ymm3[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, (%rsp), %xmm3, %xmm3 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm3 = xmm3[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm3 = xmm3[0,1,2],mem[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0],ymm3[1,2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[1],ymm6[1],ymm0[4],ymm6[4],ymm0[5],ymm6[5] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm3[0],ymm6[0],ymm3[2],ymm6[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm3 = ymm6[3,1],ymm3[0,2],ymm6[7,5],ymm3[4,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5],ymm3[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm3 = xmm3[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm3 = xmm3[0,1,2],mem[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0],ymm3[1,2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[1],ymm4[1],ymm0[4],ymm4[4],ymm0[5],ymm4[5] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[2],ymm4[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm3 = ymm4[3,1],ymm3[0,2],ymm4[7,5],ymm3[4,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5],ymm3[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm3 = xmm3[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm3 = xmm3[0,1,2],mem[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1,2,3],ymm0[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} ymm3 = ymm12[0],ymm1[0],ymm12[1],ymm1[1],ymm12[4],ymm1[4],ymm12[5],ymm1[5] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm12 = ymm4[0],ymm14[0],ymm4[2],ymm14[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm12 = ymm14[3,1],ymm12[0,2],ymm14[7,5],ymm12[4,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm12[0,1,2,3],ymm3[4,5],ymm12[6,7] |
| ; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX1-ONLY-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm12 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm12 = xmm1[3,3],mem[3,3] |
| ; AVX1-ONLY-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm12 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm12 = xmm12[0,1,2],mem[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm12[1,2,3],ymm3[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vunpcklps {{.*#+}} ymm12 = ymm9[0],ymm10[0],ymm9[1],ymm10[1],ymm9[4],ymm10[4],ymm9[5],ymm10[5] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm11[0],ymm1[0],ymm11[2],ymm1[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm10 = ymm1[3,1],ymm14[0,2],ymm1[7,5],ymm14[4,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm12[4,5],ymm10[6,7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm12 = xmm13[3,3],xmm15[3,3] |
| ; AVX1-ONLY-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm12 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm12 = xmm12[0,1,2],mem[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0],ymm12[1,2,3],ymm10[4,5,6,7] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm12 # 32-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # ymm12 = ymm1[0],mem[0],ymm1[1],mem[1],ymm1[4],mem[4],ymm1[5],mem[5] |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm4[0],ymm1[0],ymm4[2],ymm1[2] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} ymm9 = ymm1[3,1],ymm14[0,2],ymm1[7,5],ymm14[4,6] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm12[4,5],ymm9[6,7] |
| ; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm12 = xmm5[3,3],xmm8[3,3] |
| ; AVX1-ONLY-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm12 # 16-byte Folded Reload |
| ; AVX1-ONLY-NEXT: # xmm12 = xmm12[0,1,2],mem[3] |
| ; AVX1-ONLY-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0],ymm12[1,2,3],ymm9[4,5,6,7] |
| ; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX1-ONLY-NEXT: vmovaps %ymm9, 1440(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm10, 1216(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm3, 992(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 768(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm7, 544(%rax) |
| ; AVX1-ONLY-NEXT: vmovaps %ymm2, 320(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 96(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1504(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1408(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1312(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1280(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1184(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1088(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1056(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 960(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 864(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 832(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 736(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 640(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 608(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 512(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 416(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 384(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 288(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 192(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 160(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 64(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1696(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1664(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1472(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1376(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1344(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1248(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1152(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1120(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1024(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 928(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 896(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 800(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 704(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 672(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 576(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 480(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 448(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 352(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 256(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 224(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 128(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1760(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1728(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1568(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1536(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1632(%rax) |
| ; AVX1-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX1-ONLY-NEXT: vmovaps %ymm0, 1600(%rax) |
| ; AVX1-ONLY-NEXT: addq $3432, %rsp # imm = 0xD68 |
| ; AVX1-ONLY-NEXT: vzeroupper |
| ; AVX1-ONLY-NEXT: retq |
| ; |
| ; AVX2-SLOW-LABEL: store_i32_stride7_vf64: |
| ; AVX2-SLOW: # %bb.0: |
| ; AVX2-SLOW-NEXT: subq $2968, %rsp # imm = 0xB98 |
| ; AVX2-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-SLOW-NEXT: vmovaps (%rax), %xmm0 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vmovaps (%r8), %xmm5 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%r8), %xmm3 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%r9), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%r9), %xmm4 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm5[1],xmm1[2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps (%rcx), %xmm9 |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rcx), %xmm10 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%rdx), %xmm8 |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm1 = zero,xmm8[1],xmm9[1],zero |
| ; AVX2-SLOW-NEXT: vmovaps (%rdi), %xmm6 |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rdi), %xmm7 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%rsi), %xmm5 |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rsi), %xmm11 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm2 = xmm5[1,1,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm6[2],xmm2[3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rax), %xmm0 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm4[1,1,1,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm3[1],xmm1[2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm11[1,1,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm7[2],xmm1[3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rdx), %xmm7 |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm7[1],xmm10[1],zero |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 64(%r8), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 64(%r9), %xmm0 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rax), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rdi), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rsi), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rcx), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rdx), %xmm11 |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm11[1],xmm2[1],zero |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 96(%r8), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 96(%r9), %xmm0 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rax), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rdi), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rsi), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rcx), %xmm3 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rdx), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 128(%r8), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 128(%r9), %xmm0 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vmovaps 128(%rax), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 128(%rdi), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 128(%rsi), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vmovaps 128(%rcx), %xmm3 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 128(%rdx), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 160(%r8), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 160(%r9), %xmm0 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vmovaps 160(%rax), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 160(%rdi), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 160(%rsi), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vmovaps 160(%rcx), %xmm3 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 160(%rdx), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 192(%r9), %xmm0 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 192(%r8), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-SLOW-NEXT: vmovaps 192(%rax), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 192(%rdi), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 192(%rsi), %xmm1 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vmovaps 192(%rcx), %xmm3 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 192(%rdx), %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%rdi), %ymm0 |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%rsi), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps (%rdx), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%rcx), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[6],ymm1[6],ymm2[7],ymm1[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps (%r8), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps (%r9), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vmovaps 16(%rax), %xmm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rdi), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rsi), %ymm0 |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rdx), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%rcx), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 32(%r8), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 32(%r9), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vmovaps 48(%rax), %xmm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rdi), %ymm0 |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rsi), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rdx), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 64(%rcx), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 64(%r8), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 64(%r9), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vmovaps 80(%rax), %xmm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rdi), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rsi), %ymm0 |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rdx), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 96(%rcx), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 96(%r8), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 96(%r9), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vmovaps 112(%rax), %xmm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 128(%rdi), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 128(%rsi), %ymm0 |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps 128(%rdx), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 128(%rcx), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 128(%r8), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 128(%r9), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vmovaps 144(%rax), %xmm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 160(%rdi), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 160(%rsi), %ymm0 |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps 160(%rdx), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 160(%rcx), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 160(%r8), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 160(%r9), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vmovaps 176(%rax), %xmm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 192(%rdi), %ymm0 |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 192(%rsi), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps 192(%rdx), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 192(%rcx), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 192(%r8), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 192(%r9), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vmovaps 208(%rax), %xmm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 224(%rcx), %xmm0 |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm0, %xmm2 |
| ; AVX2-SLOW-NEXT: vmovaps 224(%rdx), %xmm1 |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm1, %xmm3 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1] |
| ; AVX2-SLOW-NEXT: vmovaps 224(%rsi), %xmm4 |
| ; AVX2-SLOW-NEXT: vmovaps 224(%rdi), %xmm2 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm15 = xmm2[0],xmm4[0],xmm2[1],xmm4[1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm15 = xmm15[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm15[0,1],ymm3[2,3],ymm15[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 224(%r8), %ymm15 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm15 = ymm3[0,1,2,3],ymm15[4,5],ymm3[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 224(%r9), %xmm3 |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm3, %ymm14 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps 224(%rax), %xmm15 |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm15, %ymm13 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm10 = ymm14[0,1,2,3,4,5],ymm13[6],ymm14[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm13 = xmm4[1,1,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm13 = xmm13[0,1],xmm2[2],xmm13[3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vinsertps {{.*#+}} xmm14 = zero,xmm1[1],xmm0[1],zero |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0],ymm14[1,2],ymm13[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 228(%r8), %ymm12 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2],ymm12[3],ymm13[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm13 = xmm3[1,1,1,1] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm13, %ymm0, %ymm13 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm13[4,5],ymm12[6,7] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm15, %ymm14, %ymm13 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm10 = ymm12[0],ymm13[1],ymm12[2,3,4],ymm13[5],ymm12[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm2 = xmm2[3,3],xmm4[3,3] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm1 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-SLOW-NEXT: vmovaps 224(%r8), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm3[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3],ymm0[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 232(%rax), %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4],ymm0[5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps 224(%rdi), %ymm4 |
| ; AVX2-SLOW-NEXT: vmovaps 224(%rsi), %ymm2 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4],ymm4[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm12 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovaps 224(%rdx), %ymm10 |
| ; AVX2-SLOW-NEXT: vmovaps 224(%rcx), %ymm1 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm13 = ymm10[1,1],ymm1[1,1],ymm10[5,5],ymm1[5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4],ymm13[5,6],ymm12[7] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 240(%r8), %ymm13 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0],ymm12[1,2,3,4,5,6],ymm13[7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 240(%r9), %xmm13 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0],ymm13[1],ymm12[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 240(%rax), %ymm13 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1],ymm13[2],ymm12[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm9, %xmm12 |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm8, %xmm13 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm13 = xmm6[0],xmm5[0],xmm6[1],xmm5[1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm13 = xmm13[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1],ymm12[2,3],ymm13[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm13 = xmm15[0],xmm0[0],xmm15[1],xmm0[1] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm13, %ymm0, %ymm13 |
| ; AVX2-SLOW-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm13[0],ymm14[0],ymm13[2],ymm14[2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm13[4,5,6],ymm12[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm5 = xmm6[3,3],xmm5[3,3] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} xmm6 = xmm8[2],xmm9[2],xmm8[3],xmm9[3] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6],ymm6[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm6 = xmm15[0,1,2],xmm6[3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 8(%rax), %ymm8 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm8[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm6[2,3,4],ymm5[5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm12, %xmm5 |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm7, %xmm6 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm6 = xmm13[0],xmm14[0],xmm13[1],xmm14[1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm6 = xmm9[0],xmm15[0],xmm9[1],xmm15[1] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm6 |
| ; AVX2-SLOW-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm8[0],ymm6[2],ymm8[2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm6[4,5,6],ymm5[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm5 = xmm13[3,3],xmm14[3,3] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} xmm6 = xmm7[2],xmm12[2],xmm7[3],xmm12[3] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6],ymm6[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm15[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm6 = xmm9[0,1,2],xmm6[3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 40(%rax), %ymm7 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm6[2,3,4],ymm5[5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm13, %xmm5 |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm11, %xmm6 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm6 = xmm12[0],xmm9[0],xmm12[1],xmm9[1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm6 = xmm8[0],xmm0[0],xmm8[1],xmm0[1] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm6 |
| ; AVX2-SLOW-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[2],ymm7[2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6],ymm5[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm5 = xmm12[3,3],xmm9[3,3] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} xmm6 = xmm11[2],xmm13[2],xmm11[3],xmm13[3] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6],ymm6[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm6 = xmm8[0,1,2],xmm6[3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 72(%rax), %ymm7 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm6[2,3,4],ymm5[5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm13, %xmm5 |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm14, %xmm6 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm6 = xmm12[0],xmm9[0],xmm12[1],xmm9[1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm6 = xmm8[0],xmm11[0],xmm8[1],xmm11[1] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm6 |
| ; AVX2-SLOW-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[2],ymm7[2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm6[4,5,6],ymm5[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm5 = xmm12[3,3],xmm9[3,3] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} xmm6 = xmm14[2],xmm13[2],xmm14[3],xmm13[3] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6],ymm6[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm11[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm6 = xmm8[0,1,2],xmm6[3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 104(%rax), %ymm7 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm6[2,3,4],ymm5[5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm12, %xmm5 |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm13, %xmm6 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm6 = xmm9[0],xmm8[0],xmm9[1],xmm8[1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps (%rsp), %xmm0 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm6 = xmm0[0],xmm11[0],xmm0[1],xmm11[1] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm6 |
| ; AVX2-SLOW-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[2],ymm7[2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6],ymm5[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm5 = xmm9[3,3],xmm8[3,3] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} xmm6 = xmm13[2],xmm12[2],xmm13[3],xmm12[3] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6],ymm6[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm11[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm6 = xmm0[0,1,2],xmm6[3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 136(%rax), %ymm7 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm6[2,3,4],ymm5[5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm12, %xmm5 |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm13, %xmm6 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm6 = xmm9[0],xmm8[0],xmm9[1],xmm8[1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm6 = xmm0[0],xmm11[0],xmm0[1],xmm11[1] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm6 |
| ; AVX2-SLOW-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[2],ymm7[2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6],ymm5[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm5 = xmm9[3,3],xmm8[3,3] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} xmm6 = xmm13[2],xmm12[2],xmm13[3],xmm12[3] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6],ymm6[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm11[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm6 = xmm0[0,1,2],xmm6[3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 168(%rax), %ymm7 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm6[2,3,4],ymm5[5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm11, %xmm5 |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vbroadcastss %xmm13, %xmm6 |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm6 = xmm9[0],xmm8[0],xmm9[1],xmm8[1] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} xmm6 = xmm12[0],xmm0[0],xmm12[1],xmm0[1] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm6 |
| ; AVX2-SLOW-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[2],ymm7[2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6],ymm5[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm5 = xmm9[3,3],xmm8[3,3] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} xmm6 = xmm13[2],xmm11[2],xmm13[3],xmm11[3] |
| ; AVX2-SLOW-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6],ymm6[7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm6 = xmm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} xmm6 = xmm12[0,1,2],xmm6[3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 200(%rax), %ymm7 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm6[2,3,4],ymm5[5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm5 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm5 = ymm12[2],mem[2],ymm12[3],mem[3],ymm12[6],mem[6],ymm12[7],mem[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm6 = ymm13[2],ymm11[2],ymm13[3],ymm11[3],ymm13[6],ymm11[6],ymm13[7],ymm11[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2],ymm5[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 220(%r8), %ymm6 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4],ymm6[5],ymm5[6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 220(%r9), %ymm6 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm6[6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 216(%rax), %ymm6 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0],ymm5[1,2,3,4,5,6],ymm6[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vbroadcastss 240(%rdx), %ymm5 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm6 = ymm1[3,1,2,0,7,5,6,4] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6],ymm6[7] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} ymm6 = ymm4[0],ymm2[0],ymm4[1],ymm2[1],ymm4[4],ymm2[4],ymm4[5],ymm2[5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5],ymm5[6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 236(%r8), %ymm6 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm3 = xmm3[2,2,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm5[0,1],ymm3[2,3],ymm5[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm5 = ymm4[2],ymm2[2],ymm4[3],ymm2[3],ymm4[6],ymm2[6],ymm4[7],ymm2[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm6 = ymm10[2],ymm1[2],ymm10[3],ymm1[3],ymm10[6],ymm1[6],ymm10[7],ymm1[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm5 = ymm5[0,1,2,3,4,5],mem[6,7] |
| ; AVX2-SLOW-NEXT: vpermilps {{.*#+}} ymm7 = mem[1,2,2,3,5,6,6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0],ymm5[1,2,3,4,5,6],ymm7[7] |
| ; AVX2-SLOW-NEXT: vmovaps 224(%rax), %ymm7 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm7[3],ymm3[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm7[2,3],ymm6[2,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0],ymm3[1],ymm5[2,3,4],ymm3[5],ymm5[6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm10[2],ymm1[3],ymm10[3],ymm1[6],ymm10[6],ymm1[7],ymm10[7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[6],ymm4[6],ymm2[7],ymm4[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 252(%r8), %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 252(%r9), %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 248(%rax), %ymm1 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm0 = ymm5[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4],ymm4[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm8[1,1],ymm3[1,1],ymm8[5,5],ymm3[5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm14[0,0,0,0,4,4,4,4] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm2 = ymm9[0,1,0,1,4,5,4,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 16(%rax), %ymm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm0 = ymm15[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm10[1],ymm0[2,3,4],ymm10[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1],ymm7[1,1],ymm1[5,5],ymm7[5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-SLOW-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-SLOW-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 48(%rax), %ymm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm0 = ymm6[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm0 = ymm0[0],mem[1],ymm0[2,3,4],mem[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm1 = ymm1[1,1],mem[1,1],ymm1[5,5],mem[5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-SLOW-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-SLOW-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 80(%rax), %ymm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm0 = mem[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm0 = ymm0[0],mem[1],ymm0[2,3,4],mem[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm1 = ymm1[1,1],mem[1,1],ymm1[5,5],mem[5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-SLOW-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-SLOW-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 112(%rax), %ymm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm0 = mem[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm0 = ymm0[0],mem[1],ymm0[2,3,4],mem[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm1 = ymm1[1,1],mem[1,1],ymm1[5,5],mem[5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-SLOW-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-SLOW-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 144(%rax), %ymm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm0 = mem[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm0 = ymm0[0],mem[1],ymm0[2,3,4],mem[5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm1 = ymm1[1,1],mem[1,1],ymm1[5,5],mem[5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-SLOW-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-SLOW-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 176(%rax), %ymm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,1],ymm12[1,1],ymm0[5,5],ymm12[5,5] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,1,1,1,5,5,5,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm11[1],ymm1[2,3,4],ymm11[5],ymm1[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-SLOW-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vbroadcastsd 208(%rax), %ymm2 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vbroadcastss 16(%rdx), %ymm0 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm3[3,1,2,0,7,5,6,4] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} ymm1 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; AVX2-SLOW-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm1 = xmm1[0,1,2],mem[3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm3[2],ymm8[2],ymm3[3],ymm8[3],ymm3[6],ymm8[6],ymm3[7],ymm8[7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[6],ymm4[6],ymm5[7],ymm4[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm9[3,3],ymm14[3,3],ymm9[7,7],ymm14[7,7] |
| ; AVX2-SLOW-NEXT: vpermilps {{.*#+}} ymm2 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3,4],ymm1[5,6],ymm2[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm9 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 48(%rdx), %ymm0 |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm7[3,1,2,0,7,5,6,4] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} ymm1 = ymm10[0],ymm15[0],ymm10[1],ymm15[1],ymm10[4],ymm15[4],ymm10[5],ymm15[5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; AVX2-SLOW-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm1 = xmm1[0,1,2],mem[3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm14 = ymm0[0],ymm1[1,2,3],ymm0[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm0 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm0 = ymm7[2],mem[2],ymm7[3],mem[3],ymm7[6],mem[6],ymm7[7],mem[7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm15[2],ymm10[2],ymm15[3],ymm10[3],ymm15[6],ymm10[6],ymm15[7],ymm10[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-SLOW-NEXT: vpermilps {{.*#+}} ymm2 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3,4],ymm1[5,6],ymm2[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm15 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 80(%rdx), %ymm0 |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm12[3,1,2,0,7,5,6,4] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm6[0],ymm2[1],ymm6[1],ymm2[4],ymm6[4],ymm2[5],ymm6[5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; AVX2-SLOW-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm1 = xmm1[0,1,2],mem[3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm13 = ymm0[0],ymm1[1,2,3],ymm0[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm0 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm0 = ymm12[2],mem[2],ymm12[3],mem[3],ymm12[6],mem[6],ymm12[7],mem[7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm6[2],ymm2[2],ymm6[3],ymm2[3],ymm6[6],ymm2[6],ymm6[7],ymm2[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-SLOW-NEXT: vpermilps {{.*#+}} ymm2 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3,4],ymm1[5,6],ymm2[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm12 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 112(%rdx), %ymm2 |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm1 = ymm0[3,1,2,0,7,5,6,4] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm1[0,1,2,3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm1[4,5],ymm6[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; AVX2-SLOW-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm1 = xmm1[0,1,2],mem[3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm1[1,2,3],ymm6[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-SLOW-NEXT: vpermilps {{.*#+}} ymm8 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm1 = ymm8[0],ymm1[1,2],ymm8[3,4],ymm1[5,6],ymm8[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm4 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 144(%rdx), %ymm1 |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm8 = ymm0[3,1,2,0,7,5,6,4] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5],ymm1[6],ymm8[7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} ymm8 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[4],ymm2[4],ymm1[5],ymm2[5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm8[4,5],ymm3[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps (%rsp), %xmm5 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm8 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm8 = xmm5[3,3],mem[3,3] |
| ; AVX2-SLOW-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm8 = xmm8[0,1,2],mem[3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm8[1,2,3],ymm3[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm8 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm8 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[6],ymm1[6],ymm2[7],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm8[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm8 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm8 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-SLOW-NEXT: vpermilps {{.*#+}} ymm7 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm8[1,2],ymm7[3,4],ymm8[5,6],ymm7[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm2 = ymm7[0],ymm0[1,2,3,4],ymm7[5,6,7] |
| ; AVX2-SLOW-NEXT: vbroadcastss 176(%rdx), %ymm7 |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps {{.*#+}} ymm8 = ymm0[3,1,2,0,7,5,6,4] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0,1,2,3,4,5],ymm7[6],ymm8[7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vunpcklps {{.*#+}} ymm8 = ymm1[0],ymm10[0],ymm1[1],ymm10[1],ymm1[4],ymm10[4],ymm1[5],ymm10[5] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5],ymm7[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm8 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm8 = xmm5[3,3],mem[3,3] |
| ; AVX2-SLOW-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm8 = xmm8[0,1,2],mem[3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm8[1,2,3],ymm7[4,5,6,7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm8 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm8 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7] |
| ; AVX2-SLOW-NEXT: vunpckhps {{.*#+}} ymm0 = ymm10[2],ymm1[2],ymm10[3],ymm1[3],ymm10[6],ymm1[6],ymm10[7],ymm1[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm8[3,4,5,6,7] |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm8 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm8 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-SLOW-NEXT: vpermilps {{.*#+}} ymm10 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm8 = ymm10[0],ymm8[1,2],ymm10[3,4],ymm8[5,6],ymm10[7] |
| ; AVX2-SLOW-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[2,1,2,3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm0 = ymm8[0],ymm0[1,2,3,4],ymm8[5,6,7] |
| ; AVX2-SLOW-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm8 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm8 = ymm11[0],mem[0],ymm11[1],mem[1],ymm11[4],mem[4],ymm11[5],mem[5] |
| ; AVX2-SLOW-NEXT: vpermilps $39, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # ymm10 = mem[3,1,2,0,7,5,6,4] |
| ; AVX2-SLOW-NEXT: vbroadcastss 208(%rdx), %ymm11 |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],ymm11[6],ymm10[7] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm8 = ymm10[0,1,2,3],ymm8[4,5],ymm10[6,7] |
| ; AVX2-SLOW-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-SLOW-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm10 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm10 = xmm1[3,3],mem[3,3] |
| ; AVX2-SLOW-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm10 # 16-byte Folded Reload |
| ; AVX2-SLOW-NEXT: # xmm10 = xmm10[0,1,2],mem[3] |
| ; AVX2-SLOW-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0],ymm10[1,2,3],ymm8[4,5,6,7] |
| ; AVX2-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-SLOW-NEXT: vmovaps %ymm8, 1440(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1312(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm7, 1216(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm2, 1088(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm3, 992(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm4, 864(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm6, 768(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm12, 640(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm13, 544(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm15, 416(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm14, 320(%rax) |
| ; AVX2-SLOW-NEXT: vmovaps %ymm9, 192(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 96(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1504(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1472(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1280(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1248(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1056(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1024(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 832(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 800(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 608(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 576(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 384(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 352(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 160(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 128(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1760(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1728(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1664(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1536(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1408(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1376(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1344(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1184(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1152(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1120(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 960(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 928(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 896(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 736(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 704(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 672(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 512(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 480(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 448(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 288(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 256(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 224(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 64(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1696(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1632(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1600(%rax) |
| ; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-SLOW-NEXT: vmovaps %ymm0, 1568(%rax) |
| ; AVX2-SLOW-NEXT: addq $2968, %rsp # imm = 0xB98 |
| ; AVX2-SLOW-NEXT: vzeroupper |
| ; AVX2-SLOW-NEXT: retq |
| ; |
| ; AVX2-FAST-LABEL: store_i32_stride7_vf64: |
| ; AVX2-FAST: # %bb.0: |
| ; AVX2-FAST-NEXT: subq $3096, %rsp # imm = 0xC18 |
| ; AVX2-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-NEXT: vmovaps (%rax), %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vmovaps (%r8), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%r8), %xmm3 |
| ; AVX2-FAST-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%r9), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%r9), %xmm4 |
| ; AVX2-FAST-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovaps (%rcx), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rcx), %xmm5 |
| ; AVX2-FAST-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%rdx), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertps {{.*#+}} xmm1 = zero,xmm1[1],xmm2[1],zero |
| ; AVX2-FAST-NEXT: vmovaps (%rdi), %xmm8 |
| ; AVX2-FAST-NEXT: vmovaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rdi), %xmm6 |
| ; AVX2-FAST-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%rsi), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rsi), %xmm7 |
| ; AVX2-FAST-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm2 = xmm2[1,1,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm8[2],xmm2[3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rax), %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm4[1,1,1,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm3[1],xmm1[2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm7[1,1,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm6[2],xmm1[3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-NEXT: vmovaps 32(%rdx), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm5[1],zero |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%r8), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%r9), %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-FAST-NEXT: vmovaps 64(%rax), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vmovaps 64(%rdi), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%rsi), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-NEXT: vmovaps 64(%rcx), %xmm3 |
| ; AVX2-FAST-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%rdx), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 96(%r8), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 96(%r9), %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-FAST-NEXT: vmovaps 96(%rax), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vmovaps 96(%rdi), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 96(%rsi), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-NEXT: vmovaps 96(%rcx), %xmm3 |
| ; AVX2-FAST-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 96(%rdx), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 128(%r8), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 128(%r9), %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-FAST-NEXT: vmovaps 128(%rax), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vmovaps 128(%rdi), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 128(%rsi), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-NEXT: vmovaps 128(%rcx), %xmm3 |
| ; AVX2-FAST-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 128(%rdx), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 160(%r8), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 160(%r9), %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-FAST-NEXT: vmovaps 160(%rax), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vmovaps 160(%rdi), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 160(%rsi), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-NEXT: vmovaps 160(%rcx), %xmm3 |
| ; AVX2-FAST-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 160(%rdx), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 192(%r9), %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 192(%r8), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-FAST-NEXT: vmovaps 192(%rax), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vmovaps 192(%rdi), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 192(%rsi), %xmm1 |
| ; AVX2-FAST-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-NEXT: vmovaps 192(%rcx), %xmm3 |
| ; AVX2-FAST-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 192(%rdx), %xmm2 |
| ; AVX2-FAST-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%rdi), %ymm0 |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%rsi), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps (%rdx), %ymm14 |
| ; AVX2-FAST-NEXT: vmovaps (%rcx), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm14[2],ymm1[2],ymm14[3],ymm1[3],ymm14[6],ymm1[6],ymm14[7],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps (%r8), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps (%r9), %ymm12 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm12[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vmovaps 16(%rax), %xmm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rdi), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rsi), %ymm0 |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps 32(%rdx), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%rcx), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps 32(%r8), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 32(%r9), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vmovaps 48(%rax), %xmm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%rdi), %ymm0 |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%rsi), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps 64(%rdx), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%rcx), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps 64(%r8), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 64(%r9), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vmovaps 80(%rax), %xmm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 96(%rdi), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 96(%rsi), %ymm0 |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps 96(%rdx), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 96(%rcx), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps 96(%r8), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 96(%r9), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vmovaps 112(%rax), %xmm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 128(%rdi), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 128(%rsi), %ymm0 |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps 128(%rdx), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 128(%rcx), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps 128(%r8), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 128(%r9), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vmovaps 144(%rax), %xmm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 160(%rdi), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 160(%rsi), %ymm0 |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps 160(%rdx), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 160(%rcx), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps 160(%r8), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 160(%r9), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vmovaps 176(%rax), %xmm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 192(%rdi), %ymm1 |
| ; AVX2-FAST-NEXT: vmovaps 192(%rsi), %ymm11 |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm11[2],ymm1[3],ymm11[3],ymm1[6],ymm11[6],ymm1[7],ymm11[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps %ymm1, %ymm13 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps 192(%rdx), %ymm8 |
| ; AVX2-FAST-NEXT: vmovaps 192(%rcx), %ymm9 |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[6],ymm9[6],ymm8[7],ymm9[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps 192(%r8), %ymm2 |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 192(%r9), %ymm1 |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vmovaps 208(%rax), %xmm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 224(%rdi), %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps 224(%rsi), %xmm1 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm2 = xmm1[1,1,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm0[2],xmm2[3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1] |
| ; AVX2-FAST-NEXT: vmovaps 224(%rcx), %xmm3 |
| ; AVX2-FAST-NEXT: vmovaps 224(%rdx), %xmm6 |
| ; AVX2-FAST-NEXT: vinsertps {{.*#+}} xmm5 = zero,xmm6[1],xmm3[1],zero |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm5[1,2],ymm2[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 228(%r8), %ymm4 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2],ymm4[3],ymm2[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps 224(%r9), %xmm7 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm15 = xmm7[1,1,1,1] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm15[4,5],ymm2[6,7] |
| ; AVX2-FAST-NEXT: vmovaps 224(%rax), %xmm15 |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm15, %ymm5, %ymm5 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm5[1],ymm2[2,3,4],ymm5[5],ymm2[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm3, %xmm2 |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm6, %xmm5 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm4 = xmm5[0],xmm2[0],xmm5[1],xmm2[1] |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm2 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| ; AVX2-FAST-NEXT: vbroadcastf128 {{.*#+}} ymm10 = [0,1,2,2,0,1,2,2] |
| ; AVX2-FAST-NEXT: # ymm10 = mem[0,1,0,1] |
| ; AVX2-FAST-NEXT: vpermps %ymm2, %ymm10, %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm4[2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastsd 224(%r8), %ymm4 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5],ymm2[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm7, %ymm4 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm4[5],ymm2[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm15, %ymm4 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6],ymm2[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} xmm1 = xmm6[2],xmm3[2],xmm6[3],xmm3[3] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vpermps %ymm1, %ymm10, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovaps 224(%r8), %ymm6 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm6[2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm7[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3],ymm0[4,5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 232(%rax), %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4],ymm0[5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps 224(%rdi), %ymm2 |
| ; AVX2-FAST-NEXT: vmovaps 224(%rsi), %ymm1 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm0 = ymm1[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3,4],ymm2[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm4 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovaps 224(%rdx), %ymm3 |
| ; AVX2-FAST-NEXT: vmovaps 224(%rcx), %ymm0 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm15 = ymm3[1,1],ymm0[1,1],ymm3[5,5],ymm0[5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4],ymm15[5,6],ymm4[7] |
| ; AVX2-FAST-NEXT: vbroadcastsd 240(%r8), %ymm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm15[0],ymm4[1,2,3,4,5,6],ymm15[7] |
| ; AVX2-FAST-NEXT: vbroadcastss 240(%r9), %xmm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm15[1],ymm4[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 240(%rax), %ymm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm15[2],ymm4[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm4 = ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[6],ymm8[6],ymm9[7],ymm8[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[3,3,3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm15 = ymm11[2],ymm13[2],ymm11[3],ymm13[3],ymm11[6],ymm13[6],ymm11[7],ymm13[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm15[0,1,2],ymm4[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 220(%r8), %ymm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4],ymm15[5],ymm4[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 220(%r9), %ymm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm15[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastsd 216(%rax), %ymm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm15[0],ymm4[1,2,3,4,5,6],ymm15[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vbroadcastss 240(%rdx), %ymm4 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm15 = ymm0[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm15[0,1,2,3,4,5],ymm4[6],ymm15[7] |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} ymm15 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[4],ymm1[4],ymm2[5],ymm1[5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm15[4,5],ymm4[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 236(%r8), %ymm15 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm15[1],ymm4[2,3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm7 = xmm7[2,2,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm7[2,3],ymm4[4,5,6,7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm7 = ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[6],ymm1[6],ymm2[7],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[2,2,2,2] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm15 = ymm3[2],ymm0[2],ymm3[3],ymm0[3],ymm3[6],ymm0[6],ymm3[7],ymm0[7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm15[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastsd {{.*#+}} ymm7 = [5,6,5,6,5,6,5,6] |
| ; AVX2-FAST-NEXT: vpermps 224(%r9), %ymm7, %ymm7 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0],ymm6[1,2,3,4,5,6],ymm7[7] |
| ; AVX2-FAST-NEXT: vmovaps 224(%rax), %ymm7 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2],ymm7[3],ymm4[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm7[2,3],ymm15[2,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm6[0],ymm4[1],ymm6[2,3,4],ymm4[5],ymm6[6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 252(%r8), %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 252(%r9), %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vbroadcastsd 248(%rax), %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm5, %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm4, %xmm1 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm8[0],xmm7[0],xmm8[1],xmm7[1] |
| ; AVX2-FAST-NEXT: vpermps %ymm1, %ymm10, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm3[0],xmm6[0],xmm3[1],xmm6[1] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm8[3,3],xmm7[3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} xmm1 = xmm4[2],xmm5[2],xmm4[3],xmm5[3] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vpermps %ymm1, %ymm10, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm6[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 8(%rax), %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm0 = mem[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm0 = ymm0[0],mem[1],ymm0[2,3,4],mem[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vshufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm1 = ymm14[1,1],mem[1,1],ymm14[5,5],mem[5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm12[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 16(%rax), %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm4, %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm8, %xmm1 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm5[0],xmm6[0],xmm5[1],xmm6[1] |
| ; AVX2-FAST-NEXT: vpermps %ymm1, %ymm10, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm3[0],xmm7[0],xmm3[1],xmm7[1] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm5[3,3],xmm6[3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} xmm1 = xmm8[2],xmm4[2],xmm8[3],xmm4[3] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vpermps %ymm1, %ymm10, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm7[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 40(%rax), %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm0 = mem[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm0 = ymm0[0],mem[1],ymm0[2,3,4],mem[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm1 = ymm1[1,1],mem[1,1],ymm1[5,5],mem[5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-FAST-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 48(%rax), %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm6, %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm7, %xmm1 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm5[0],xmm4[0],xmm5[1],xmm4[1] |
| ; AVX2-FAST-NEXT: vpermps %ymm1, %ymm10, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm8[0],xmm3[0],xmm8[1],xmm3[1] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm5[3,3],xmm4[3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} xmm1 = xmm7[2],xmm6[2],xmm7[3],xmm6[3] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vpermps %ymm1, %ymm10, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm3[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm8[0,1,2],xmm1[3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 72(%rax), %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm0 = ymm15[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm0 = ymm0[0],mem[1],ymm0[2,3,4],mem[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm1 = ymm1[1,1],mem[1,1],ymm1[5,5],mem[5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-FAST-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 80(%rax), %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm5, %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm6, %xmm1 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1] |
| ; AVX2-FAST-NEXT: vpermps %ymm1, %ymm10, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm8[0],xmm7[0],xmm8[1],xmm7[1] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm4[3,3],xmm3[3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} xmm1 = xmm6[2],xmm5[2],xmm6[3],xmm5[3] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vpermps %ymm1, %ymm10, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm7[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm8[0,1,2],xmm1[3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 104(%rax), %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm0 = mem[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm12[1],ymm0[2,3,4],ymm12[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1],ymm11[1,1],ymm1[5,5],ymm11[5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-FAST-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 112(%rax), %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm5, %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm6, %xmm1 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1] |
| ; AVX2-FAST-NEXT: vpermps %ymm1, %ymm10, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps (%rsp), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm8[0],xmm7[0],xmm8[1],xmm7[1] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm4[3,3],xmm3[3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} xmm1 = xmm6[2],xmm5[2],xmm6[3],xmm5[3] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vpermps %ymm1, %ymm10, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm7[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm8[0,1,2],xmm1[3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 136(%rax), %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm0 = mem[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm9[1],ymm0[2,3,4],ymm9[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm1 = ymm1[1,1],mem[1,1],ymm1[5,5],mem[5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-FAST-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 144(%rax), %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm5, %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm6, %xmm1 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1] |
| ; AVX2-FAST-NEXT: vpermps %ymm1, %ymm10, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm8[0],xmm7[0],xmm8[1],xmm7[1] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6],ymm0[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm4[3,3],xmm3[3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} xmm1 = xmm6[2],xmm5[2],xmm6[3],xmm5[3] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vpermps %ymm1, %ymm10, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm7[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm8[0,1,2],xmm1[3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 168(%rax), %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm0 = mem[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm0 = ymm0[0],mem[1],ymm0[2,3,4],mem[5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1],ymm14[1,1],ymm1[5,5],ymm14[5,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-FAST-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 176(%rax), %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm5, %xmm0 |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vbroadcastss %xmm4, %xmm1 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm0[0],xmm7[0],xmm0[1],xmm7[1] |
| ; AVX2-FAST-NEXT: vpermps %ymm1, %ymm10, %ymm1 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm8 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} xmm1 = xmm3[0],xmm6[0],xmm3[1],xmm6[1] |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[2],ymm2[2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm8[0,1,2,3],ymm1[4,5,6],ymm8[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm0 = xmm0[3,3],xmm7[3,3] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} xmm1 = xmm4[2],xmm5[2],xmm4[3],xmm5[3] |
| ; AVX2-FAST-NEXT: vpermps %ymm1, %ymm10, %ymm1 |
| ; AVX2-FAST-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} xmm1 = xmm6[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 200(%rax), %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3,4],ymm0[5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,1],ymm13[1,1],ymm0[5,5],ymm13[5,5] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm7[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm6[1],ymm1[2,3,4],ymm6[5],ymm1[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vbroadcastsd 208(%rax), %ymm2 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vbroadcastss 16(%rdx), %ymm0 |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm5[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; AVX2-FAST-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm1 = xmm1[0,1,2],mem[3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm0 = ymm5[2],mem[2],ymm5[3],mem[3],ymm5[6],mem[6],ymm5[7],mem[7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-FAST-NEXT: vpermilps {{.*#+}} ymm5 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm5[0],ymm1[1,2],ymm5[3,4],ymm1[5,6],ymm5[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vbroadcastss 48(%rdx), %ymm0 |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm5[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; AVX2-FAST-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm1 = xmm1[0,1,2],mem[3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm0 = ymm5[2],mem[2],ymm5[3],mem[3],ymm5[6],mem[6],ymm5[7],mem[7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-FAST-NEXT: vpermilps {{.*#+}} ymm5 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm5[0],ymm1[1,2],ymm5[3,4],ymm1[5,6],ymm5[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX2-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-NEXT: vbroadcastss 80(%rdx), %ymm0 |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm1 = ymm5[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6],ymm1[7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm15, %ymm3 |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm15[0],ymm2[1],ymm15[1],ymm2[4],ymm15[4],ymm2[5],ymm15[5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; AVX2-FAST-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm1 = xmm1[0,1,2],mem[3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm15 = ymm0[0],ymm1[1,2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm0 = ymm5[2],mem[2],ymm5[3],mem[3],ymm5[6],mem[6],ymm5[7],mem[7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-FAST-NEXT: vpermilps {{.*#+}} ymm5 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm1 = ymm5[0],ymm1[1,2],ymm5[3,4],ymm1[5,6],ymm5[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm10 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 112(%rdx), %ymm0 |
| ; AVX2-FAST-NEXT: vmovaps %ymm11, %ymm3 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm5 = ymm11[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5],ymm0[6],ymm5[7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} ymm5 = ymm12[0],ymm2[0],ymm12[1],ymm2[1],ymm12[4],ymm2[4],ymm12[5],ymm2[5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5],ymm0[6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm5 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm5 = xmm4[3,3],mem[3,3] |
| ; AVX2-FAST-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm5 = xmm5[0,1,2],mem[3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm11 = ymm0[0],ymm5[1,2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm0 = ymm3[2],mem[2],ymm3[3],mem[3],ymm3[6],mem[6],ymm3[7],mem[7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm5 = ymm2[2],ymm12[2],ymm2[3],ymm12[3],ymm2[6],ymm12[6],ymm2[7],ymm12[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm5 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-FAST-NEXT: vpermilps {{.*#+}} ymm12 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm5 = ymm12[0],ymm5[1,2],ymm12[3,4],ymm5[5,6],ymm12[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[2,1,2,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0],ymm0[1,2,3,4],ymm5[5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 144(%rdx), %ymm5 |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm12 = ymm2[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm5 = ymm12[0,1,2,3,4,5],ymm5[6],ymm12[7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} ymm12 = ymm9[0],ymm1[0],ymm9[1],ymm1[1],ymm9[4],ymm1[4],ymm9[5],ymm1[5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm12[4,5],ymm5[6,7] |
| ; AVX2-FAST-NEXT: vmovaps (%rsp), %xmm3 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm12 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm12 = xmm3[3,3],mem[3,3] |
| ; AVX2-FAST-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm12 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm12 = xmm12[0,1,2],mem[3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm12 = ymm5[0],ymm12[1,2,3],ymm5[4,5,6,7] |
| ; AVX2-FAST-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm5 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm5 = ymm2[2],mem[2],ymm2[3],mem[3],ymm2[6],mem[6],ymm2[7],mem[7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm9[2],ymm1[3],ymm9[3],ymm1[6],ymm9[6],ymm1[7],ymm9[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[3,3,3,3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm5[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm5 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-FAST-NEXT: vpermilps {{.*#+}} ymm9 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm5 = ymm9[0],ymm5[1,2],ymm9[3,4],ymm5[5,6],ymm9[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[2,1,2,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm3 = ymm5[0],ymm0[1,2,3,4],ymm5[5,6,7] |
| ; AVX2-FAST-NEXT: vbroadcastss 176(%rdx), %ymm5 |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm9 = ymm14[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm5 = ymm9[0,1,2,3,4,5],ymm5[6],ymm9[7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} ymm9 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[4],ymm0[4],ymm2[5],ymm0[5] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm9[4,5],ymm5[6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm9 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm9 = xmm8[3,3],mem[3,3] |
| ; AVX2-FAST-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm9 = xmm9[0,1,2],mem[3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm9[1,2,3],ymm5[4,5,6,7] |
| ; AVX2-FAST-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm9 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm9 = ymm14[2],mem[2],ymm14[3],mem[3],ymm14[6],mem[6],ymm14[7],mem[7] |
| ; AVX2-FAST-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[6],ymm2[6],ymm0[7],ymm2[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[3,3,3,3] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm9[3,4,5,6,7] |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm9 # 32-byte Folded Reload |
| ; AVX2-FAST-NEXT: # ymm9 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-FAST-NEXT: vpermilps {{.*#+}} ymm8 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0],ymm9[1,2],ymm8[3,4],ymm9[5,6],ymm8[7] |
| ; AVX2-FAST-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[2,1,2,3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm0 = ymm8[0],ymm0[1,2,3,4],ymm8[5,6,7] |
| ; AVX2-FAST-NEXT: vunpcklps {{.*#+}} ymm8 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5] |
| ; AVX2-FAST-NEXT: vshufps {{.*#+}} ymm9 = ymm13[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-NEXT: vbroadcastss 208(%rdx), %ymm7 |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm7 = ymm9[0,1,2,3,4,5],ymm7[6],ymm9[7] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5],ymm7[6,7] |
| ; AVX2-FAST-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-FAST-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm8 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm8 = xmm1[3,3],mem[3,3] |
| ; AVX2-FAST-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload |
| ; AVX2-FAST-NEXT: # xmm8 = xmm8[0,1,2],mem[3] |
| ; AVX2-FAST-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm8[1,2,3],ymm7[4,5,6,7] |
| ; AVX2-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-NEXT: vmovaps %ymm7, 1440(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1312(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm5, 1216(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm3, 1088(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm12, 992(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm4, 864(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm11, 768(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm10, 640(%rax) |
| ; AVX2-FAST-NEXT: vmovaps %ymm15, 544(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 416(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 320(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 192(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 96(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1504(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1472(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1408(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1344(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1280(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1248(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1184(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1120(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1056(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1024(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 960(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 896(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 832(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 800(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 736(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 672(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 608(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 576(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 512(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 448(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 384(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 352(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 288(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 224(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 160(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 128(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 64(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1760(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1728(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1664(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1536(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1376(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1152(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 928(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 704(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 480(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 256(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1696(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1632(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1568(%rax) |
| ; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-NEXT: vmovaps %ymm0, 1600(%rax) |
| ; AVX2-FAST-NEXT: addq $3096, %rsp # imm = 0xC18 |
| ; AVX2-FAST-NEXT: vzeroupper |
| ; AVX2-FAST-NEXT: retq |
| ; |
| ; AVX2-FAST-PERLANE-LABEL: store_i32_stride7_vf64: |
| ; AVX2-FAST-PERLANE: # %bb.0: |
| ; AVX2-FAST-PERLANE-NEXT: subq $2968, %rsp # imm = 0xB98 |
| ; AVX2-FAST-PERLANE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rax), %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r8), %xmm5 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%r8), %xmm3 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r9), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%r9), %xmm4 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,1,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm5[1],xmm1[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rcx), %xmm9 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rcx), %xmm10 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdx), %xmm8 |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm1 = zero,xmm8[1],xmm9[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdi), %xmm6 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rdi), %xmm7 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rsi), %xmm5 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rsi), %xmm11 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm2 = xmm5[1,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm6[2],xmm2[3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rax), %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm4[1,1,1,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm3[1],xmm1[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm1, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm11[1,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm7[2],xmm1[3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rdx), %xmm7 |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm7[1],xmm10[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%r8), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%r9), %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rax), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rdi), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rsi), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rcx), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rdx), %xmm11 |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm11[1],xmm2[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%r8), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%r9), %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rax), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rdi), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rsi), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rcx), %xmm3 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rdx), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 128(%r8), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 128(%r9), %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 128(%rax), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 128(%rdi), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 128(%rsi), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 128(%rcx), %xmm3 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 128(%rdx), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 160(%r8), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 160(%r9), %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 160(%rax), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 160(%rdi), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 160(%rsi), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 160(%rcx), %xmm3 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 160(%rdx), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 192(%r9), %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 192(%r8), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd %xmm0, %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 192(%rax), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 192(%rdi), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 192(%rsi), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm1[1,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 192(%rcx), %xmm3 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 192(%rdx), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm2[1],xmm3[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdi), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rsi), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdx), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rcx), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[6],ymm1[6],ymm2[7],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r8), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%r9), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 16(%rax), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rdi), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rsi), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rdx), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%rcx), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%r8), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 32(%r9), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 48(%rax), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rdi), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rsi), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rdx), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%rcx), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%r8), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 64(%r9), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 80(%rax), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rdi), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rsi), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rdx), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%rcx), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%r8), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 96(%r9), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 112(%rax), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 128(%rdi), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 128(%rsi), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 128(%rdx), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 128(%rcx), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 128(%r8), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 128(%r9), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 144(%rax), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 160(%rdi), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 160(%rsi), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 160(%rdx), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 160(%rcx), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 160(%r8), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 160(%r9), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 176(%rax), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 192(%rdi), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 192(%rsi), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 192(%rdx), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 192(%rcx), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 192(%r8), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 192(%r9), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1,2,2,5,5,6,6] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 208(%rax), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 224(%rcx), %xmm0 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm0, %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 224(%rdx), %xmm1 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm1, %xmm3 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 224(%rsi), %xmm4 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 224(%rdi), %xmm2 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm15 = xmm2[0],xmm4[0],xmm2[1],xmm4[1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm15 = xmm15[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm15[0,1],ymm3[2,3],ymm15[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 224(%r8), %ymm15 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm15 = ymm3[0,1,2,3],ymm15[4,5],ymm3[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 224(%r9), %xmm3 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm3, %ymm14 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 224(%rax), %xmm15 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm15, %ymm13 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm10 = ymm14[0,1,2,3,4,5],ymm13[6],ymm14[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm13 = xmm4[1,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm13 = xmm13[0,1],xmm2[2],xmm13[3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertps {{.*#+}} xmm14 = zero,xmm1[1],xmm0[1],zero |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0],ymm14[1,2],ymm13[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 228(%r8), %ymm12 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2],ymm12[3],ymm13[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm13 = xmm3[1,1,1,1] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm13, %ymm0, %ymm13 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm13[4,5],ymm12[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm15, %ymm14, %ymm13 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm10 = ymm12[0],ymm13[1],ymm12[2,3,4],ymm13[5],ymm12[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm2 = xmm2[3,3],xmm4[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 224(%r8), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm1 = xmm3[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3],ymm0[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 232(%rax), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4],ymm0[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 224(%rdi), %ymm4 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 224(%rsi), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4],ymm4[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm12 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 224(%rdx), %ymm10 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 224(%rcx), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm13 = ymm10[1,1],ymm1[1,1],ymm10[5,5],ymm1[5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4],ymm13[5,6],ymm12[7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 240(%r8), %ymm13 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0],ymm12[1,2,3,4,5,6],ymm13[7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 240(%r9), %xmm13 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0],ymm13[1],ymm12[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 240(%rax), %ymm13 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1],ymm13[2],ymm12[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm9, %xmm12 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm8, %xmm13 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm13 = xmm6[0],xmm5[0],xmm6[1],xmm5[1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm13 = xmm13[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1],ymm12[2,3],ymm13[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm13 = xmm15[0],xmm0[0],xmm15[1],xmm0[1] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm13, %ymm0, %ymm13 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm13[0],ymm14[0],ymm13[2],ymm14[2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm13[4,5,6],ymm12[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm5 = xmm6[3,3],xmm5[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} xmm6 = xmm8[2],xmm9[2],xmm8[3],xmm9[3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6],ymm6[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm6 = xmm15[0,1,2],xmm6[3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 8(%rax), %ymm8 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm8[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm6[2,3,4],ymm5[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm12, %xmm5 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm7, %xmm6 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm6 = xmm13[0],xmm14[0],xmm13[1],xmm14[1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm6 = xmm9[0],xmm15[0],xmm9[1],xmm15[1] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm8[0],ymm6[2],ymm8[2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm6[4,5,6],ymm5[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm5 = xmm13[3,3],xmm14[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} xmm6 = xmm7[2],xmm12[2],xmm7[3],xmm12[3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6],ymm6[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm15[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm6 = xmm9[0,1,2],xmm6[3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 40(%rax), %ymm7 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm6[2,3,4],ymm5[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm13, %xmm5 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm11, %xmm6 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm6 = xmm12[0],xmm9[0],xmm12[1],xmm9[1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm6 = xmm8[0],xmm0[0],xmm8[1],xmm0[1] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[2],ymm7[2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6],ymm5[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm5 = xmm12[3,3],xmm9[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} xmm6 = xmm11[2],xmm13[2],xmm11[3],xmm13[3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6],ymm6[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm6 = xmm8[0,1,2],xmm6[3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 72(%rax), %ymm7 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm6[2,3,4],ymm5[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm13, %xmm5 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm14, %xmm6 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm6 = xmm12[0],xmm9[0],xmm12[1],xmm9[1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm6 = xmm8[0],xmm11[0],xmm8[1],xmm11[1] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[2],ymm7[2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm6[4,5,6],ymm5[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm5 = xmm12[3,3],xmm9[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} xmm6 = xmm14[2],xmm13[2],xmm14[3],xmm13[3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6],ymm6[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm11[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm6 = xmm8[0,1,2],xmm6[3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 104(%rax), %ymm7 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm6[2,3,4],ymm5[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm12, %xmm5 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm13, %xmm6 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm6 = xmm9[0],xmm8[0],xmm9[1],xmm8[1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rsp), %xmm0 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm6 = xmm0[0],xmm11[0],xmm0[1],xmm11[1] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[2],ymm7[2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6],ymm5[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm5 = xmm9[3,3],xmm8[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} xmm6 = xmm13[2],xmm12[2],xmm13[3],xmm12[3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6],ymm6[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm11[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm6 = xmm0[0,1,2],xmm6[3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 136(%rax), %ymm7 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm6[2,3,4],ymm5[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm12, %xmm5 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm13, %xmm6 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm6 = xmm9[0],xmm8[0],xmm9[1],xmm8[1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm6 = xmm0[0],xmm11[0],xmm0[1],xmm11[1] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[2],ymm7[2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6],ymm5[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm5 = xmm9[3,3],xmm8[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} xmm6 = xmm13[2],xmm12[2],xmm13[3],xmm12[3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6],ymm6[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm11[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm6 = xmm0[0,1,2],xmm6[3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 168(%rax), %ymm7 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm6[2,3,4],ymm5[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm11, %xmm5 |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss %xmm13, %xmm6 |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm6 = xmm9[0],xmm8[0],xmm9[1],xmm8[1] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} xmm6 = xmm12[0],xmm0[0],xmm12[1],xmm0[1] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[2],ymm7[2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6],ymm5[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm5 = xmm9[3,3],xmm8[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} xmm6 = xmm13[2],xmm11[2],xmm13[3],xmm11[3] |
| ; AVX2-FAST-PERLANE-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,2,1] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5,6],ymm6[7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm6 = xmm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} xmm6 = xmm12[0,1,2],xmm6[3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 200(%rax), %ymm7 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm6[2,3,4],ymm5[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm5 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm5 = ymm12[2],mem[2],ymm12[3],mem[3],ymm12[6],mem[6],ymm12[7],mem[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm6 = ymm13[2],ymm11[2],ymm13[3],ymm11[3],ymm13[6],ymm11[6],ymm13[7],ymm11[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2],ymm5[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 220(%r8), %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4],ymm6[5],ymm5[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 220(%r9), %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm6[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 216(%rax), %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0],ymm5[1,2,3,4,5,6],ymm6[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 240(%rdx), %ymm5 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm6 = ymm1[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6],ymm6[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} ymm6 = ymm4[0],ymm2[0],ymm4[1],ymm2[1],ymm4[4],ymm2[4],ymm4[5],ymm2[5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5],ymm5[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 236(%r8), %ymm6 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm3 = xmm3[2,2,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm5[0,1],ymm3[2,3],ymm5[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm5 = ymm4[2],ymm2[2],ymm4[3],ymm2[3],ymm4[6],ymm2[6],ymm4[7],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm6 = ymm10[2],ymm1[2],ymm10[3],ymm1[3],ymm10[6],ymm1[6],ymm10[7],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm5 = ymm5[0,1,2,3,4,5],mem[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps {{.*#+}} ymm7 = mem[1,2,2,3,5,6,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0],ymm5[1,2,3,4,5,6],ymm7[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps 224(%rax), %ymm7 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1,2],ymm7[3],ymm3[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm7[2,3],ymm6[2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0],ymm3[1],ymm5[2,3,4],ymm3[5],ymm5[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm1[2],ymm10[2],ymm1[3],ymm10[3],ymm1[6],ymm10[6],ymm1[7],ymm10[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[6],ymm4[6],ymm2[7],ymm4[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 252(%r8), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 252(%r9), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 248(%rax), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm0 = ymm5[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4],ymm4[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm8[1,1],ymm3[1,1],ymm8[5,5],ymm3[5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm14[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm2 = ymm9[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 16(%rax), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm0 = ymm15[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm10[1],ymm0[2,3,4],ymm10[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1],ymm7[1,1],ymm1[5,5],ymm7[5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 48(%rax), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm0 = ymm6[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm0 = ymm0[0],mem[1],ymm0[2,3,4],mem[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm1 = ymm1[1,1],mem[1,1],ymm1[5,5],mem[5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 80(%rax), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm0 = mem[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm0 = ymm0[0],mem[1],ymm0[2,3,4],mem[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm1 = ymm1[1,1],mem[1,1],ymm1[5,5],mem[5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 112(%rax), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm0 = mem[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm0 = ymm0[0],mem[1],ymm0[2,3,4],mem[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm1 = ymm1[1,1],mem[1,1],ymm1[5,5],mem[5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 144(%rax), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm0 = mem[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $34, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm0 = ymm0[0],mem[1],ymm0[2,3,4],mem[5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm1 = ymm1[1,1],mem[1,1],ymm1[5,5],mem[5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5,6],ymm0[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 176(%rax), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,1],ymm12[1,1],ymm0[5,5],ymm12[5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,1,1,1,5,5,5,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm11[1],ymm1[2,3,4],ymm11[5],ymm1[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,2] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5,6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $0, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm1 = mem[0,0,0,0,4,4,4,4] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm2 = mem[0,1,0,1,4,5,4,5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastsd 208(%rax), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2,3],ymm1[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 16(%rdx), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm3[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} ymm1 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm1 = xmm1[0,1,2],mem[3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm3[2],ymm8[2],ymm3[3],ymm8[3],ymm3[6],ymm8[6],ymm3[7],ymm8[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[6],ymm4[6],ymm5[7],ymm4[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm9[3,3],ymm14[3,3],ymm9[7,7],ymm14[7,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps {{.*#+}} ymm2 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3,4],ymm1[5,6],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm9 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 48(%rdx), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm7[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} ymm1 = ymm10[0],ymm15[0],ymm10[1],ymm15[1],ymm10[4],ymm15[4],ymm10[5],ymm15[5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm1 = xmm1[0,1,2],mem[3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm14 = ymm0[0],ymm1[1,2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm0 = ymm7[2],mem[2],ymm7[3],mem[3],ymm7[6],mem[6],ymm7[7],mem[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm15[2],ymm10[2],ymm15[3],ymm10[3],ymm15[6],ymm10[6],ymm15[7],ymm10[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps {{.*#+}} ymm2 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3,4],ymm1[5,6],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm15 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 80(%rdx), %ymm0 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm12[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm6[0],ymm2[1],ymm6[1],ymm2[4],ymm6[4],ymm2[5],ymm6[5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5],ymm0[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm1 = xmm1[0,1,2],mem[3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm13 = ymm0[0],ymm1[1,2,3],ymm0[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm0 = ymm12[2],mem[2],ymm12[3],mem[3],ymm12[6],mem[6],ymm12[7],mem[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm6[2],ymm2[2],ymm6[3],ymm2[3],ymm6[6],ymm2[6],ymm6[7],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps {{.*#+}} ymm2 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3,4],ymm1[5,6],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm12 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 112(%rdx), %ymm2 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm1 = ymm0[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm1[0,1,2,3,4,5],ymm2[6],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} ymm1 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm1[4,5],ymm6[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm1 = xmm1[3,3],mem[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm1 = xmm1[0,1,2],mem[3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm1[1,2,3],ymm6[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm1 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps {{.*#+}} ymm8 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm1 = ymm8[0],ymm1[1,2],ymm8[3,4],ymm1[5,6],ymm8[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm4 = ymm1[0],ymm0[1,2,3,4],ymm1[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 144(%rdx), %ymm1 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm8 = ymm0[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5],ymm1[6],ymm8[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} ymm8 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[4],ymm2[4],ymm1[5],ymm2[5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm8[4,5],ymm3[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps (%rsp), %xmm5 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm8 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm8 = xmm5[3,3],mem[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm8 = xmm8[0,1,2],mem[3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm8[1,2,3],ymm3[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm8 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm8 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[6],ymm1[6],ymm2[7],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm8[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm8 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm8 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps {{.*#+}} ymm7 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm8[1,2],ymm7[3,4],ymm8[5,6],ymm7[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm2 = ymm7[0],ymm0[1,2,3,4],ymm7[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 176(%rdx), %ymm7 |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} ymm8 = ymm0[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0,1,2,3,4,5],ymm7[6],ymm8[7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{.*#+}} ymm8 = ymm1[0],ymm10[0],ymm1[1],ymm10[1],ymm1[4],ymm10[4],ymm1[5],ymm10[5] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5],ymm7[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm8 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm8 = xmm5[3,3],mem[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm8 = xmm8[0,1,2],mem[3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm8[1,2,3],ymm7[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm8 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm8 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpckhps {{.*#+}} ymm0 = ymm10[2],ymm1[2],ymm10[3],ymm1[3],ymm10[6],ymm1[6],ymm10[7],ymm1[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[3,3,3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm8[3,4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm8 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm8 = ymm1[3,3],mem[3,3],ymm1[7,7],mem[7,7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps {{.*#+}} ymm10 = mem[2,3,2,3,6,7,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm8 = ymm10[0],ymm8[1,2],ymm10[3,4],ymm8[5,6],ymm10[7] |
| ; AVX2-FAST-PERLANE-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[2,1,2,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm0 = ymm8[0],ymm0[1,2,3,4],ymm8[5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm8 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm8 = ymm11[0],mem[0],ymm11[1],mem[1],ymm11[4],mem[4],ymm11[5],mem[5] |
| ; AVX2-FAST-PERLANE-NEXT: vpermilps $39, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # ymm10 = mem[3,1,2,0,7,5,6,4] |
| ; AVX2-FAST-PERLANE-NEXT: vbroadcastss 208(%rdx), %ymm11 |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],ymm11[6],ymm10[7] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm8 = ymm10[0,1,2,3],ymm8[4,5],ymm10[6,7] |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vshufps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm10 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm10 = xmm1[3,3],mem[3,3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm10 # 16-byte Folded Reload |
| ; AVX2-FAST-PERLANE-NEXT: # xmm10 = xmm10[0,1,2],mem[3] |
| ; AVX2-FAST-PERLANE-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0],ymm10[1,2,3],ymm8[4,5,6,7] |
| ; AVX2-FAST-PERLANE-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm8, 1440(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1312(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm7, 1216(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm2, 1088(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm3, 992(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm4, 864(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm6, 768(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm12, 640(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm13, 544(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm15, 416(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm14, 320(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm9, 192(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 96(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1504(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1472(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1280(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1248(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1056(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1024(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 832(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 800(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 608(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 576(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 384(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 352(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 160(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 128(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1760(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1728(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1664(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1536(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1408(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1376(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1344(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1184(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1152(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1120(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 960(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 928(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 896(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 736(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 704(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 672(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 512(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 480(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 448(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 288(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 256(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 224(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 64(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 32(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, (%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1696(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1632(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1600(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload |
| ; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 1568(%rax) |
| ; AVX2-FAST-PERLANE-NEXT: addq $2968, %rsp # imm = 0xB98 |
| ; AVX2-FAST-PERLANE-NEXT: vzeroupper |
| ; AVX2-FAST-PERLANE-NEXT: retq |
| ; |
| ; AVX512F-LABEL: store_i32_stride7_vf64: |
| ; AVX512F: # %bb.0: |
| ; AVX512F-NEXT: subq $3016, %rsp # imm = 0xBC8 |
| ; AVX512F-NEXT: vmovdqa64 (%rdx), %zmm2 |
| ; AVX512F-NEXT: vmovdqa64 (%rcx), %zmm0 |
| ; AVX512F-NEXT: vmovdqa64 (%r8), %zmm22 |
| ; AVX512F-NEXT: vmovdqa64 64(%r8), %zmm19 |
| ; AVX512F-NEXT: vmovdqa64 128(%r8), %zmm13 |
| ; AVX512F-NEXT: vmovdqa64 (%r9), %zmm4 |
| ; AVX512F-NEXT: vmovdqa64 64(%r9), %zmm1 |
| ; AVX512F-NEXT: vmovdqa64 128(%r9), %zmm3 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm5 = <u,u,0,16,u,u,u,u,u,1,17,u,u,u,u,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm2, %zmm6 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm5, %zmm6 |
| ; AVX512F-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm6 = <u,u,u,u,0,16,u,u,u,u,u,1,17,u,u,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm22, %zmm7 |
| ; AVX512F-NEXT: vpermt2d %zmm4, %zmm6, %zmm7 |
| ; AVX512F-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm7 = <2,18,u,u,u,u,u,3,19,u,u,u,u,u,4,20> |
| ; AVX512F-NEXT: vmovdqa64 %zmm2, %zmm8 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm7, %zmm8 |
| ; AVX512F-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm8 = <u,u,2,18,u,u,u,u,u,3,19,u,u,u,u,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm22, %zmm9 |
| ; AVX512F-NEXT: vpermt2d %zmm4, %zmm8, %zmm9 |
| ; AVX512F-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm4, %zmm10 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm4 = <4,20,u,u,u,u,u,5,21,u,u,u,u,u,6,22> |
| ; AVX512F-NEXT: vmovdqa64 %zmm22, %zmm9 |
| ; AVX512F-NEXT: vpermt2d %zmm10, %zmm4, %zmm9 |
| ; AVX512F-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm9 = <u,u,u,u,u,7,23,u,u,u,u,u,8,24,u,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm22, %zmm11 |
| ; AVX512F-NEXT: vpermt2d %zmm10, %zmm9, %zmm11 |
| ; AVX512F-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm9, %zmm14 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm9 = <u,u,u,9,25,u,u,u,u,u,10,26,u,u,u,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm22, %zmm11 |
| ; AVX512F-NEXT: vpermt2d %zmm10, %zmm9, %zmm11 |
| ; AVX512F-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,11,27,u,u,u,u,u,12,28,u,u,u,u,u,13> |
| ; AVX512F-NEXT: vmovdqa64 %zmm22, %zmm12 |
| ; AVX512F-NEXT: vpermt2d %zmm10, %zmm11, %zmm12 |
| ; AVX512F-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm31 = <13,u,u,u,u,u,30,14,u,u,u,u,u,31,15,u> |
| ; AVX512F-NEXT: vpermt2d %zmm22, %zmm31, %zmm10 |
| ; AVX512F-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm19, %zmm10 |
| ; AVX512F-NEXT: vpermt2d %zmm1, %zmm6, %zmm10 |
| ; AVX512F-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm19, %zmm10 |
| ; AVX512F-NEXT: vpermt2d %zmm1, %zmm8, %zmm10 |
| ; AVX512F-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm10 |
| ; AVX512F-NEXT: vmovdqa64 %zmm19, %zmm1 |
| ; AVX512F-NEXT: vpermt2d %zmm10, %zmm4, %zmm1 |
| ; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm19, %zmm1 |
| ; AVX512F-NEXT: vpermt2d %zmm10, %zmm14, %zmm1 |
| ; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm19, %zmm1 |
| ; AVX512F-NEXT: vpermt2d %zmm10, %zmm9, %zmm1 |
| ; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm19, %zmm1 |
| ; AVX512F-NEXT: vpermt2d %zmm10, %zmm11, %zmm1 |
| ; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vpermt2d %zmm19, %zmm31, %zmm10 |
| ; AVX512F-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm13, %zmm1 |
| ; AVX512F-NEXT: vpermt2d %zmm3, %zmm6, %zmm1 |
| ; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm13, %zmm1 |
| ; AVX512F-NEXT: vpermt2d %zmm3, %zmm8, %zmm1 |
| ; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm13, %zmm1 |
| ; AVX512F-NEXT: vpermt2d %zmm3, %zmm4, %zmm1 |
| ; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vpermi2d %zmm3, %zmm13, %zmm14 |
| ; AVX512F-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vpermi2d %zmm3, %zmm13, %zmm9 |
| ; AVX512F-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vpermi2d %zmm3, %zmm13, %zmm11 |
| ; AVX512F-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vpermt2d %zmm13, %zmm31, %zmm3 |
| ; AVX512F-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm1 = <u,u,u,u,u,5,21,u,u,u,u,u,6,22,u,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm2, %zmm3 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm1, %zmm2 |
| ; AVX512F-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm10 = <u,u,u,7,23,u,u,u,u,u,8,24,u,u,u,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm3, %zmm2 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm10, %zmm2 |
| ; AVX512F-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm21 = <u,9,25,u,u,u,u,u,10,26,u,u,u,u,u,11> |
| ; AVX512F-NEXT: vmovdqa64 %zmm3, %zmm2 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm21, %zmm2 |
| ; AVX512F-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm31 = <11,u,u,u,u,u,28,12,u,u,u,u,u,29,13,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm2 |
| ; AVX512F-NEXT: vpermt2d %zmm3, %zmm31, %zmm2 |
| ; AVX512F-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,u,u,u,14,30,u,u,u,u,u,15,31,u,u,u> |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm2, %zmm3 |
| ; AVX512F-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 64(%rdx), %zmm17 |
| ; AVX512F-NEXT: vmovdqa64 64(%rcx), %zmm0 |
| ; AVX512F-NEXT: vmovdqa64 %zmm17, %zmm3 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm5, %zmm3 |
| ; AVX512F-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm17, %zmm3 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm7, %zmm3 |
| ; AVX512F-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm17, %zmm3 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm1, %zmm3 |
| ; AVX512F-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm17, %zmm3 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm10, %zmm3 |
| ; AVX512F-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm17, %zmm3 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm21, %zmm3 |
| ; AVX512F-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm3 |
| ; AVX512F-NEXT: vpermt2d %zmm17, %zmm31, %zmm3 |
| ; AVX512F-NEXT: vmovdqu64 %zmm3, (%rsp) # 64-byte Spill |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm2, %zmm17 |
| ; AVX512F-NEXT: vmovdqa64 192(%rdx), %zmm25 |
| ; AVX512F-NEXT: vmovdqa64 192(%rcx), %zmm0 |
| ; AVX512F-NEXT: vmovdqa64 %zmm25, %zmm3 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm10, %zmm3 |
| ; AVX512F-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm25, %zmm3 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm21, %zmm3 |
| ; AVX512F-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm3 |
| ; AVX512F-NEXT: vpermt2d %zmm25, %zmm31, %zmm3 |
| ; AVX512F-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm25, %zmm3 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm2, %zmm3 |
| ; AVX512F-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 128(%rdx), %zmm23 |
| ; AVX512F-NEXT: vmovdqa64 128(%rcx), %zmm3 |
| ; AVX512F-NEXT: vpermi2d %zmm3, %zmm23, %zmm10 |
| ; AVX512F-NEXT: vpermi2d %zmm3, %zmm23, %zmm21 |
| ; AVX512F-NEXT: vpermi2d %zmm23, %zmm3, %zmm31 |
| ; AVX512F-NEXT: vmovdqa64 %zmm23, %zmm8 |
| ; AVX512F-NEXT: vmovdqa64 %zmm23, %zmm6 |
| ; AVX512F-NEXT: vmovdqa64 %zmm23, %zmm4 |
| ; AVX512F-NEXT: vpermt2d %zmm3, %zmm2, %zmm23 |
| ; AVX512F-NEXT: vpermt2d %zmm3, %zmm5, %zmm8 |
| ; AVX512F-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vpermt2d %zmm3, %zmm7, %zmm6 |
| ; AVX512F-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vpermt2d %zmm3, %zmm1, %zmm4 |
| ; AVX512F-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vpermi2d %zmm0, %zmm25, %zmm5 |
| ; AVX512F-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vpermi2d %zmm0, %zmm25, %zmm7 |
| ; AVX512F-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm1, %zmm25 |
| ; AVX512F-NEXT: vmovdqa64 (%rdi), %zmm24 |
| ; AVX512F-NEXT: vmovdqa64 (%rsi), %zmm0 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm9 = <0,16,u,u,u,u,u,1,17,u,u,u,u,u,2,18> |
| ; AVX512F-NEXT: vmovdqa64 %zmm24, %zmm1 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm9, %zmm1 |
| ; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm15 = <u,u,u,u,u,3,19,u,u,u,u,u,4,20,u,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm24, %zmm1 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm15, %zmm1 |
| ; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm29 = <u,u,u,5,21,u,u,u,u,u,6,22,u,u,u,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm24, %zmm1 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm29, %zmm1 |
| ; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm8 = <u,7,23,u,u,u,u,u,8,24,u,u,u,u,u,9> |
| ; AVX512F-NEXT: vmovdqa64 %zmm24, %zmm1 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm8, %zmm1 |
| ; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm6 = <9,u,u,u,u,u,26,10,u,u,u,u,u,27,11,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm28 |
| ; AVX512F-NEXT: vpermt2d %zmm24, %zmm6, %zmm28 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,u,u,u,12,28,u,u,u,u,u,13,29,u,u,u> |
| ; AVX512F-NEXT: vmovdqa64 %zmm24, %zmm3 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm2, %zmm3 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm30 = <u,u,14,30,u,u,u,u,u,15,31,u,u,u,u,u> |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm30, %zmm24 |
| ; AVX512F-NEXT: vmovdqa64 64(%rdi), %zmm5 |
| ; AVX512F-NEXT: vmovdqa64 64(%rsi), %zmm0 |
| ; AVX512F-NEXT: vmovdqa64 %zmm5, %zmm27 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm9, %zmm27 |
| ; AVX512F-NEXT: vmovdqa64 %zmm5, %zmm1 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm15, %zmm1 |
| ; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm5, %zmm1 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm29, %zmm1 |
| ; AVX512F-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 %zmm5, %zmm26 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm8, %zmm26 |
| ; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm19 |
| ; AVX512F-NEXT: vpermt2d %zmm5, %zmm6, %zmm19 |
| ; AVX512F-NEXT: vmovdqa64 %zmm5, %zmm20 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm2, %zmm20 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm30, %zmm5 |
| ; AVX512F-NEXT: vmovdqa64 192(%rdi), %zmm12 |
| ; AVX512F-NEXT: vmovdqa64 192(%rsi), %zmm11 |
| ; AVX512F-NEXT: vmovdqa64 %zmm12, %zmm1 |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm8, %zmm1 |
| ; AVX512F-NEXT: vmovdqa64 %zmm11, %zmm14 |
| ; AVX512F-NEXT: vpermt2d %zmm12, %zmm6, %zmm14 |
| ; AVX512F-NEXT: vmovdqa64 %zmm12, %zmm7 |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm2, %zmm7 |
| ; AVX512F-NEXT: vmovdqa64 %zmm12, %zmm13 |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm30, %zmm13 |
| ; AVX512F-NEXT: vmovdqa64 128(%rdi), %zmm16 |
| ; AVX512F-NEXT: vmovdqa64 128(%rsi), %zmm0 |
| ; AVX512F-NEXT: vpermi2d %zmm0, %zmm16, %zmm8 |
| ; AVX512F-NEXT: vpermi2d %zmm16, %zmm0, %zmm6 |
| ; AVX512F-NEXT: vpermi2d %zmm0, %zmm16, %zmm2 |
| ; AVX512F-NEXT: vmovdqa64 %zmm16, %zmm18 |
| ; AVX512F-NEXT: vmovdqa64 %zmm16, %zmm4 |
| ; AVX512F-NEXT: vmovdqa64 %zmm16, %zmm22 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm30, %zmm16 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm9, %zmm18 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm15, %zmm4 |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm29, %zmm22 |
| ; AVX512F-NEXT: vpermi2d %zmm11, %zmm12, %zmm9 |
| ; AVX512F-NEXT: vpermi2d %zmm11, %zmm12, %zmm15 |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm29, %zmm12 |
| ; AVX512F-NEXT: movw $3096, %ax # imm = 0xC18 |
| ; AVX512F-NEXT: kmovw %eax, %k1 |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1} |
| ; AVX512F-NEXT: movw $-31994, %ax # imm = 0x8306 |
| ; AVX512F-NEXT: kmovw %eax, %k2 |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm0, %zmm14 {%k2} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm0, %zmm29 {%k1} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm0, %zmm26 {%k1} |
| ; AVX512F-NEXT: vmovdqa32 %zmm10, %zmm8 {%k1} |
| ; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512F-NEXT: vmovdqa64 (%rax), %zmm0 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm10 = <22,u,u,u,u,5,6,23,u,u,u,u,12,13,24,u> |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm10, %zmm11 |
| ; AVX512F-NEXT: movw $28897, %cx # imm = 0x70E1 |
| ; AVX512F-NEXT: kmovw %ecx, %k1 |
| ; AVX512F-NEXT: vmovdqa32 %zmm11, %zmm29 {%k1} |
| ; AVX512F-NEXT: vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa64 64(%rax), %zmm11 |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm10, %zmm29 |
| ; AVX512F-NEXT: vmovdqa32 %zmm29, %zmm26 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 128(%rax), %zmm30 |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm30, %zmm10, %zmm29 |
| ; AVX512F-NEXT: vmovdqa32 %zmm29, %zmm8 {%k1} |
| ; AVX512F-NEXT: movw $6192, %cx # imm = 0x1830 |
| ; AVX512F-NEXT: kmovw %ecx, %k1 |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm7, %zmm29 {%k1} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm7, %zmm13 {%k1} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm7, %zmm28 {%k2} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm7, %zmm19 {%k2} |
| ; AVX512F-NEXT: vmovdqa32 %zmm21, %zmm6 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm7 = <u,u,u,3,4,25,u,u,u,u,10,11,26,u,u,u> |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm7, %zmm10 |
| ; AVX512F-NEXT: movw $7224, %cx # imm = 0x1C38 |
| ; AVX512F-NEXT: kmovw %ecx, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm10, %zmm28 {%k2} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm3, %zmm21 {%k1} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm7, %zmm10 |
| ; AVX512F-NEXT: vmovdqa32 %zmm10, %zmm19 {%k2} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm30, %zmm7, %zmm10 |
| ; AVX512F-NEXT: vmovdqa32 %zmm10, %zmm6 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm7 = <u,1,2,27,u,u,u,u,8,9,28,u,u,u,u,15> |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm7, %zmm10 |
| ; AVX512F-NEXT: movw $-30962, %cx # imm = 0x870E |
| ; AVX512F-NEXT: kmovw %ecx, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm10, %zmm21 {%k2} |
| ; AVX512F-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqu64 (%rsp), %zmm3 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm20, %zmm3 {%k1} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm7, %zmm10 |
| ; AVX512F-NEXT: vmovdqa32 %zmm10, %zmm3 {%k2} |
| ; AVX512F-NEXT: vmovdqu64 %zmm3, (%rsp) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqa32 %zmm2, %zmm31 {%k1} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm30, %zmm7, %zmm2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm2, %zmm31 {%k2} |
| ; AVX512F-NEXT: movw $1548, %cx # imm = 0x60C |
| ; AVX512F-NEXT: kmovw %ecx, %k2 |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm2, %zmm3 {%k2} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm2, %zmm24 {%k1} |
| ; AVX512F-NEXT: vmovdqa32 %zmm17, %zmm5 {%k1} |
| ; AVX512F-NEXT: vmovdqa32 %zmm23, %zmm16 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm2 = <0,29,u,u,u,u,6,7,30,u,u,u,u,13,14,31> |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm2, %zmm7 |
| ; AVX512F-NEXT: movw $-7741, %cx # imm = 0xE1C3 |
| ; AVX512F-NEXT: kmovw %ecx, %k1 |
| ; AVX512F-NEXT: vmovdqa32 %zmm7, %zmm24 {%k1} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm2, %zmm7 |
| ; AVX512F-NEXT: vmovdqa32 %zmm7, %zmm5 {%k1} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm30, %zmm2, %zmm7 |
| ; AVX512F-NEXT: vmovdqa32 %zmm7, %zmm16 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,u,u,u,4,5,16,u,u,u,u,11,12,17,u,u> |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm2, %zmm7 |
| ; AVX512F-NEXT: movw $14448, %cx # imm = 0x3870 |
| ; AVX512F-NEXT: kmovw %ecx, %k3 |
| ; AVX512F-NEXT: vmovdqa32 %zmm7, %zmm3 {%k3} |
| ; AVX512F-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm7, %zmm27 {%k2} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm7, %zmm18 {%k2} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm7, %zmm9 {%k2} |
| ; AVX512F-NEXT: movw $12384, %cx # imm = 0x3060 |
| ; AVX512F-NEXT: kmovw %ecx, %k1 |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm3, %zmm20 {%k1} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm2, %zmm7 |
| ; AVX512F-NEXT: vmovdqa32 %zmm7, %zmm27 {%k3} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm30, %zmm2, %zmm7 |
| ; AVX512F-NEXT: vmovdqa32 %zmm7, %zmm18 {%k3} |
| ; AVX512F-NEXT: vmovdqa64 192(%r8), %zmm7 |
| ; AVX512F-NEXT: vmovdqa64 192(%r9), %zmm10 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm21 = <u,u,u,u,0,16,u,u,u,u,u,1,17,u,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm10, %zmm7, %zmm21 |
| ; AVX512F-NEXT: vmovdqa64 192(%rax), %zmm17 |
| ; AVX512F-NEXT: vpermt2d %zmm17, %zmm2, %zmm21 |
| ; AVX512F-NEXT: vmovdqa32 %zmm21, %zmm9 {%k3} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,u,2,3,18,u,u,u,u,9,10,19,u,u,u,u> |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm2, %zmm21 |
| ; AVX512F-NEXT: movw $3612, %ax # imm = 0xE1C |
| ; AVX512F-NEXT: kmovw %eax, %k2 |
| ; AVX512F-NEXT: vmovdqa32 %zmm21, %zmm20 {%k2} |
| ; AVX512F-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm3, %zmm20 {%k1} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm2, %zmm21 |
| ; AVX512F-NEXT: vmovdqa32 %zmm21, %zmm20 {%k2} |
| ; AVX512F-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm4, %zmm20 {%k1} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm30, %zmm2, %zmm21 |
| ; AVX512F-NEXT: vmovdqa32 %zmm21, %zmm20 {%k2} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm15, %zmm21 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm15 = <u,u,2,18,u,u,u,u,u,3,19,u,u,u,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm10, %zmm7, %zmm15 |
| ; AVX512F-NEXT: vpermt2d %zmm17, %zmm2, %zmm15 |
| ; AVX512F-NEXT: vmovdqa32 %zmm15, %zmm21 {%k2} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,1,2,3,4,23,u,u,8,9,10,11,24,u,u,15> |
| ; AVX512F-NEXT: vpermi2d %zmm7, %zmm1, %zmm2 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm3 = <0,1,2,25,u,u,6,7,8,9,26,u,u,13,14,15> |
| ; AVX512F-NEXT: vpermi2d %zmm7, %zmm14, %zmm3 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm14 = <0,27,u,u,4,5,6,7,28,u,u,11,12,13,14,29> |
| ; AVX512F-NEXT: vpermi2d %zmm7, %zmm29, %zmm14 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm15 = <u,u,2,3,4,5,30,u,u,9,10,11,12,31,u,u> |
| ; AVX512F-NEXT: vpermi2d %zmm7, %zmm13, %zmm15 |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm13, %zmm29 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm13 = <0,1,20,u,u,u,u,7,8,21,u,u,u,u,14,15> |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm0, %zmm13, %zmm23 |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512F-NEXT: vmovdqa32 %zmm0, %zmm22 {%k1} |
| ; AVX512F-NEXT: vmovdqa32 %zmm25, %zmm12 {%k1} |
| ; AVX512F-NEXT: movw $15480, %ax # imm = 0x3C78 |
| ; AVX512F-NEXT: kmovw %eax, %k1 |
| ; AVX512F-NEXT: vmovdqa32 %zmm29, %zmm23 {%k1} |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm11, %zmm13, %zmm0 |
| ; AVX512F-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm25 |
| ; AVX512F-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512F-NEXT: vpermt2d %zmm30, %zmm13, %zmm0 |
| ; AVX512F-NEXT: vmovdqa32 %zmm22, %zmm0 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 %zmm0, %zmm4 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm0 = <4,20,u,u,u,u,u,5,21,u,u,u,u,u,6,22> |
| ; AVX512F-NEXT: vpermt2d %zmm10, %zmm0, %zmm7 |
| ; AVX512F-NEXT: vpermt2d %zmm17, %zmm13, %zmm7 |
| ; AVX512F-NEXT: vmovdqa32 %zmm12, %zmm7 {%k1} |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm0 = <u,1,2,3,4,5,23,u,8,9,10,11,12,24,u,15> |
| ; AVX512F-NEXT: vpermi2d %zmm10, %zmm2, %zmm0 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm1 = <0,1,2,3,25,u,6,7,8,9,10,26,u,13,14,15> |
| ; AVX512F-NEXT: vpermi2d %zmm10, %zmm3, %zmm1 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm2 = <0,1,27,u,4,5,6,7,8,28,u,11,12,13,14,15> |
| ; AVX512F-NEXT: vpermi2d %zmm10, %zmm14, %zmm2 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm11 = <29,u,2,3,4,5,6,30,u,9,10,11,12,13,31,u> |
| ; AVX512F-NEXT: vpermi2d %zmm10, %zmm15, %zmm11 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm10 = [22,1,2,3,4,5,6,23,8,9,10,11,12,13,24,15] |
| ; AVX512F-NEXT: vpermi2d %zmm17, %zmm0, %zmm10 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,1,2,3,4,25,6,7,8,9,10,11,26,13,14,15] |
| ; AVX512F-NEXT: vpermi2d %zmm17, %zmm1, %zmm0 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,1,2,27,4,5,6,7,8,9,28,11,12,13,14,15] |
| ; AVX512F-NEXT: vpermi2d %zmm17, %zmm2, %zmm1 |
| ; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,29,2,3,4,5,6,7,30,9,10,11,12,13,14,31] |
| ; AVX512F-NEXT: vpermi2d %zmm17, %zmm11, %zmm2 |
| ; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512F-NEXT: vmovdqa64 %zmm7, 1472(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm21, 1408(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm9, 1344(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm16, 1280(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm31, 1216(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm6, 1152(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm8, 1088(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm4, 1024(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm20, 960(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm18, 896(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm5, 832(%rax) |
| ; AVX512F-NEXT: vmovups (%rsp), %zmm3 # 64-byte Reload |
| ; AVX512F-NEXT: vmovaps %zmm3, 768(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm19, 704(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm26, 640(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm25, 576(%rax) |
| ; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512F-NEXT: vmovaps %zmm3, 512(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm27, 448(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm24, 384(%rax) |
| ; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512F-NEXT: vmovaps %zmm3, 320(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm28, 256(%rax) |
| ; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512F-NEXT: vmovaps %zmm3, 192(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm23, 128(%rax) |
| ; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512F-NEXT: vmovaps %zmm3, 64(%rax) |
| ; AVX512F-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512F-NEXT: vmovaps %zmm3, (%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm2, 1728(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm1, 1664(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm0, 1600(%rax) |
| ; AVX512F-NEXT: vmovdqa64 %zmm10, 1536(%rax) |
| ; AVX512F-NEXT: addq $3016, %rsp # imm = 0xBC8 |
| ; AVX512F-NEXT: vzeroupper |
| ; AVX512F-NEXT: retq |
| ; |
| ; AVX512BW-LABEL: store_i32_stride7_vf64: |
| ; AVX512BW: # %bb.0: |
| ; AVX512BW-NEXT: subq $3016, %rsp # imm = 0xBC8 |
| ; AVX512BW-NEXT: vmovdqa64 (%rdx), %zmm2 |
| ; AVX512BW-NEXT: vmovdqa64 (%rcx), %zmm0 |
| ; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm22 |
| ; AVX512BW-NEXT: vmovdqa64 64(%r8), %zmm19 |
| ; AVX512BW-NEXT: vmovdqa64 128(%r8), %zmm13 |
| ; AVX512BW-NEXT: vmovdqa64 (%r9), %zmm4 |
| ; AVX512BW-NEXT: vmovdqa64 64(%r9), %zmm1 |
| ; AVX512BW-NEXT: vmovdqa64 128(%r9), %zmm3 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm5 = <u,u,0,16,u,u,u,u,u,1,17,u,u,u,u,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm6 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm5, %zmm6 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm6 = <u,u,u,u,0,16,u,u,u,u,u,1,17,u,u,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm22, %zmm7 |
| ; AVX512BW-NEXT: vpermt2d %zmm4, %zmm6, %zmm7 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm7 = <2,18,u,u,u,u,u,3,19,u,u,u,u,u,4,20> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm8 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm7, %zmm8 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm8 = <u,u,2,18,u,u,u,u,u,3,19,u,u,u,u,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm22, %zmm9 |
| ; AVX512BW-NEXT: vpermt2d %zmm4, %zmm8, %zmm9 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm10 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm4 = <4,20,u,u,u,u,u,5,21,u,u,u,u,u,6,22> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm22, %zmm9 |
| ; AVX512BW-NEXT: vpermt2d %zmm10, %zmm4, %zmm9 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm9 = <u,u,u,u,u,7,23,u,u,u,u,u,8,24,u,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm22, %zmm11 |
| ; AVX512BW-NEXT: vpermt2d %zmm10, %zmm9, %zmm11 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm14 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm9 = <u,u,u,9,25,u,u,u,u,u,10,26,u,u,u,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm22, %zmm11 |
| ; AVX512BW-NEXT: vpermt2d %zmm10, %zmm9, %zmm11 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm11 = <u,11,27,u,u,u,u,u,12,28,u,u,u,u,u,13> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm22, %zmm12 |
| ; AVX512BW-NEXT: vpermt2d %zmm10, %zmm11, %zmm12 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm31 = <13,u,u,u,u,u,30,14,u,u,u,u,u,31,15,u> |
| ; AVX512BW-NEXT: vpermt2d %zmm22, %zmm31, %zmm10 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm10 |
| ; AVX512BW-NEXT: vpermt2d %zmm1, %zmm6, %zmm10 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm10 |
| ; AVX512BW-NEXT: vpermt2d %zmm1, %zmm8, %zmm10 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm10 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm1 |
| ; AVX512BW-NEXT: vpermt2d %zmm10, %zmm4, %zmm1 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm1 |
| ; AVX512BW-NEXT: vpermt2d %zmm10, %zmm14, %zmm1 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm1 |
| ; AVX512BW-NEXT: vpermt2d %zmm10, %zmm9, %zmm1 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm1 |
| ; AVX512BW-NEXT: vpermt2d %zmm10, %zmm11, %zmm1 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vpermt2d %zmm19, %zmm31, %zmm10 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm1 |
| ; AVX512BW-NEXT: vpermt2d %zmm3, %zmm6, %zmm1 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm1 |
| ; AVX512BW-NEXT: vpermt2d %zmm3, %zmm8, %zmm1 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm1 |
| ; AVX512BW-NEXT: vpermt2d %zmm3, %zmm4, %zmm1 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vpermi2d %zmm3, %zmm13, %zmm14 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vpermi2d %zmm3, %zmm13, %zmm9 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vpermi2d %zmm3, %zmm13, %zmm11 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vpermt2d %zmm13, %zmm31, %zmm3 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm1 = <u,u,u,u,u,5,21,u,u,u,u,u,6,22,u,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm3 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm1, %zmm2 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm10 = <u,u,u,7,23,u,u,u,u,u,8,24,u,u,u,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm2 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm10, %zmm2 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm21 = <u,9,25,u,u,u,u,u,10,26,u,u,u,u,u,11> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm2 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm21, %zmm2 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm31 = <11,u,u,u,u,u,28,12,u,u,u,u,u,29,13,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm2 |
| ; AVX512BW-NEXT: vpermt2d %zmm3, %zmm31, %zmm2 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,u,u,u,14,30,u,u,u,u,u,15,31,u,u,u> |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm2, %zmm3 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 64(%rdx), %zmm17 |
| ; AVX512BW-NEXT: vmovdqa64 64(%rcx), %zmm0 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm17, %zmm3 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm5, %zmm3 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm17, %zmm3 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm7, %zmm3 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm17, %zmm3 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm1, %zmm3 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm17, %zmm3 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm10, %zmm3 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm17, %zmm3 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm21, %zmm3 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm3 |
| ; AVX512BW-NEXT: vpermt2d %zmm17, %zmm31, %zmm3 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm3, (%rsp) # 64-byte Spill |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm2, %zmm17 |
| ; AVX512BW-NEXT: vmovdqa64 192(%rdx), %zmm25 |
| ; AVX512BW-NEXT: vmovdqa64 192(%rcx), %zmm0 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm3 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm10, %zmm3 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm3 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm21, %zmm3 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm3 |
| ; AVX512BW-NEXT: vpermt2d %zmm25, %zmm31, %zmm3 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm3 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm2, %zmm3 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 128(%rdx), %zmm23 |
| ; AVX512BW-NEXT: vmovdqa64 128(%rcx), %zmm3 |
| ; AVX512BW-NEXT: vpermi2d %zmm3, %zmm23, %zmm10 |
| ; AVX512BW-NEXT: vpermi2d %zmm3, %zmm23, %zmm21 |
| ; AVX512BW-NEXT: vpermi2d %zmm23, %zmm3, %zmm31 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm23, %zmm8 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm23, %zmm6 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm23, %zmm4 |
| ; AVX512BW-NEXT: vpermt2d %zmm3, %zmm2, %zmm23 |
| ; AVX512BW-NEXT: vpermt2d %zmm3, %zmm5, %zmm8 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vpermt2d %zmm3, %zmm7, %zmm6 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vpermt2d %zmm3, %zmm1, %zmm4 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vpermi2d %zmm0, %zmm25, %zmm5 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vpermi2d %zmm0, %zmm25, %zmm7 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm1, %zmm25 |
| ; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm24 |
| ; AVX512BW-NEXT: vmovdqa64 (%rsi), %zmm0 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm9 = <0,16,u,u,u,u,u,1,17,u,u,u,u,u,2,18> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm24, %zmm1 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm9, %zmm1 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm15 = <u,u,u,u,u,3,19,u,u,u,u,u,4,20,u,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm24, %zmm1 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm15, %zmm1 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm29 = <u,u,u,5,21,u,u,u,u,u,6,22,u,u,u,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm24, %zmm1 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm29, %zmm1 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm8 = <u,7,23,u,u,u,u,u,8,24,u,u,u,u,u,9> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm24, %zmm1 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm8, %zmm1 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm6 = <9,u,u,u,u,u,26,10,u,u,u,u,u,27,11,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm28 |
| ; AVX512BW-NEXT: vpermt2d %zmm24, %zmm6, %zmm28 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,u,u,u,12,28,u,u,u,u,u,13,29,u,u,u> |
| ; AVX512BW-NEXT: vmovdqa64 %zmm24, %zmm3 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm2, %zmm3 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm30 = <u,u,14,30,u,u,u,u,u,15,31,u,u,u,u,u> |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm30, %zmm24 |
| ; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm5 |
| ; AVX512BW-NEXT: vmovdqa64 64(%rsi), %zmm0 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm27 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm9, %zmm27 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm1 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm15, %zmm1 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm1 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm29, %zmm1 |
| ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm26 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm8, %zmm26 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm19 |
| ; AVX512BW-NEXT: vpermt2d %zmm5, %zmm6, %zmm19 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm20 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm2, %zmm20 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm30, %zmm5 |
| ; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm12 |
| ; AVX512BW-NEXT: vmovdqa64 192(%rsi), %zmm11 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm1 |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm8, %zmm1 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm14 |
| ; AVX512BW-NEXT: vpermt2d %zmm12, %zmm6, %zmm14 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm7 |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm2, %zmm7 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm13 |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm30, %zmm13 |
| ; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm16 |
| ; AVX512BW-NEXT: vmovdqa64 128(%rsi), %zmm0 |
| ; AVX512BW-NEXT: vpermi2d %zmm0, %zmm16, %zmm8 |
| ; AVX512BW-NEXT: vpermi2d %zmm16, %zmm0, %zmm6 |
| ; AVX512BW-NEXT: vpermi2d %zmm0, %zmm16, %zmm2 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm18 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm4 |
| ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm22 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm30, %zmm16 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm9, %zmm18 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm15, %zmm4 |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm29, %zmm22 |
| ; AVX512BW-NEXT: vpermi2d %zmm11, %zmm12, %zmm9 |
| ; AVX512BW-NEXT: vpermi2d %zmm11, %zmm12, %zmm15 |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm29, %zmm12 |
| ; AVX512BW-NEXT: movw $3096, %ax # imm = 0xC18 |
| ; AVX512BW-NEXT: kmovd %eax, %k1 |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1} |
| ; AVX512BW-NEXT: movw $-31994, %ax # imm = 0x8306 |
| ; AVX512BW-NEXT: kmovd %eax, %k2 |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm14 {%k2} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm29 {%k1} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm26 {%k1} |
| ; AVX512BW-NEXT: vmovdqa32 %zmm10, %zmm8 {%k1} |
| ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512BW-NEXT: vmovdqa64 (%rax), %zmm0 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm10 = <22,u,u,u,u,5,6,23,u,u,u,u,12,13,24,u> |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm10, %zmm11 |
| ; AVX512BW-NEXT: movw $28897, %cx # imm = 0x70E1 |
| ; AVX512BW-NEXT: kmovd %ecx, %k1 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm11, %zmm29 {%k1} |
| ; AVX512BW-NEXT: vmovdqu64 %zmm29, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa64 64(%rax), %zmm11 |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm10, %zmm29 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm29, %zmm26 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 128(%rax), %zmm30 |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm30, %zmm10, %zmm29 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm29, %zmm8 {%k1} |
| ; AVX512BW-NEXT: movw $6192, %cx # imm = 0x1830 |
| ; AVX512BW-NEXT: kmovd %ecx, %k1 |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm29 {%k1} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm13 {%k1} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm28 {%k2} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm19 {%k2} |
| ; AVX512BW-NEXT: vmovdqa32 %zmm21, %zmm6 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm7 = <u,u,u,3,4,25,u,u,u,u,10,11,26,u,u,u> |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm7, %zmm10 |
| ; AVX512BW-NEXT: movw $7224, %cx # imm = 0x1C38 |
| ; AVX512BW-NEXT: kmovd %ecx, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm10, %zmm28 {%k2} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm3, %zmm21 {%k1} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm7, %zmm10 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm10, %zmm19 {%k2} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm30, %zmm7, %zmm10 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm10, %zmm6 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm7 = <u,1,2,27,u,u,u,u,8,9,28,u,u,u,u,15> |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm7, %zmm10 |
| ; AVX512BW-NEXT: movw $-30962, %cx # imm = 0x870E |
| ; AVX512BW-NEXT: kmovd %ecx, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm10, %zmm21 {%k2} |
| ; AVX512BW-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqu64 (%rsp), %zmm3 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm20, %zmm3 {%k1} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm7, %zmm10 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm10, %zmm3 {%k2} |
| ; AVX512BW-NEXT: vmovdqu64 %zmm3, (%rsp) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm31 {%k1} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm30, %zmm7, %zmm2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm31 {%k2} |
| ; AVX512BW-NEXT: movw $1548, %cx # imm = 0x60C |
| ; AVX512BW-NEXT: kmovd %ecx, %k2 |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm3 {%k2} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm24 {%k1} |
| ; AVX512BW-NEXT: vmovdqa32 %zmm17, %zmm5 {%k1} |
| ; AVX512BW-NEXT: vmovdqa32 %zmm23, %zmm16 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = <0,29,u,u,u,u,6,7,30,u,u,u,u,13,14,31> |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm2, %zmm7 |
| ; AVX512BW-NEXT: movw $-7741, %cx # imm = 0xE1C3 |
| ; AVX512BW-NEXT: kmovd %ecx, %k1 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm24 {%k1} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm2, %zmm7 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm5 {%k1} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm30, %zmm2, %zmm7 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm16 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,u,u,u,4,5,16,u,u,u,u,11,12,17,u,u> |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm2, %zmm7 |
| ; AVX512BW-NEXT: movw $14448, %cx # imm = 0x3870 |
| ; AVX512BW-NEXT: kmovd %ecx, %k3 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm3 {%k3} |
| ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm27 {%k2} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm18 {%k2} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm9 {%k2} |
| ; AVX512BW-NEXT: movw $12384, %cx # imm = 0x3060 |
| ; AVX512BW-NEXT: kmovd %ecx, %k1 |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm3, %zmm20 {%k1} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm2, %zmm7 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm27 {%k3} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm30, %zmm2, %zmm7 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm18 {%k3} |
| ; AVX512BW-NEXT: vmovdqa64 192(%r8), %zmm7 |
| ; AVX512BW-NEXT: vmovdqa64 192(%r9), %zmm10 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm21 = <u,u,u,u,0,16,u,u,u,u,u,1,17,u,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm10, %zmm7, %zmm21 |
| ; AVX512BW-NEXT: vmovdqa64 192(%rax), %zmm17 |
| ; AVX512BW-NEXT: vpermt2d %zmm17, %zmm2, %zmm21 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm21, %zmm9 {%k3} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,u,2,3,18,u,u,u,u,9,10,19,u,u,u,u> |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm2, %zmm21 |
| ; AVX512BW-NEXT: movw $3612, %ax # imm = 0xE1C |
| ; AVX512BW-NEXT: kmovd %eax, %k2 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm21, %zmm20 {%k2} |
| ; AVX512BW-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm3, %zmm20 {%k1} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm2, %zmm21 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm21, %zmm20 {%k2} |
| ; AVX512BW-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm4, %zmm20 {%k1} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm30, %zmm2, %zmm21 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm21, %zmm20 {%k2} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm15, %zmm21 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm15 = <u,u,2,18,u,u,u,u,u,3,19,u,u,u,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm10, %zmm7, %zmm15 |
| ; AVX512BW-NEXT: vpermt2d %zmm17, %zmm2, %zmm15 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm15, %zmm21 {%k2} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,1,2,3,4,23,u,u,8,9,10,11,24,u,u,15> |
| ; AVX512BW-NEXT: vpermi2d %zmm7, %zmm1, %zmm2 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm3 = <0,1,2,25,u,u,6,7,8,9,26,u,u,13,14,15> |
| ; AVX512BW-NEXT: vpermi2d %zmm7, %zmm14, %zmm3 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm14 = <0,27,u,u,4,5,6,7,28,u,u,11,12,13,14,29> |
| ; AVX512BW-NEXT: vpermi2d %zmm7, %zmm29, %zmm14 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm15 = <u,u,2,3,4,5,30,u,u,9,10,11,12,31,u,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm7, %zmm13, %zmm15 |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm13, %zmm29 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm13 = <0,1,20,u,u,u,u,7,8,21,u,u,u,u,14,15> |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm0, %zmm13, %zmm23 |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm22 {%k1} |
| ; AVX512BW-NEXT: vmovdqa32 %zmm25, %zmm12 {%k1} |
| ; AVX512BW-NEXT: movw $15480, %ax # imm = 0x3C78 |
| ; AVX512BW-NEXT: kmovd %eax, %k1 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm29, %zmm23 {%k1} |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm11, %zmm13, %zmm0 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm25 |
| ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload |
| ; AVX512BW-NEXT: vpermt2d %zmm30, %zmm13, %zmm0 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm22, %zmm0 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm4 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm0 = <4,20,u,u,u,u,u,5,21,u,u,u,u,u,6,22> |
| ; AVX512BW-NEXT: vpermt2d %zmm10, %zmm0, %zmm7 |
| ; AVX512BW-NEXT: vpermt2d %zmm17, %zmm13, %zmm7 |
| ; AVX512BW-NEXT: vmovdqa32 %zmm12, %zmm7 {%k1} |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm0 = <u,1,2,3,4,5,23,u,8,9,10,11,12,24,u,15> |
| ; AVX512BW-NEXT: vpermi2d %zmm10, %zmm2, %zmm0 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm1 = <0,1,2,3,25,u,6,7,8,9,10,26,u,13,14,15> |
| ; AVX512BW-NEXT: vpermi2d %zmm10, %zmm3, %zmm1 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = <0,1,27,u,4,5,6,7,8,28,u,11,12,13,14,15> |
| ; AVX512BW-NEXT: vpermi2d %zmm10, %zmm14, %zmm2 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm11 = <29,u,2,3,4,5,6,30,u,9,10,11,12,13,31,u> |
| ; AVX512BW-NEXT: vpermi2d %zmm10, %zmm15, %zmm11 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm10 = [22,1,2,3,4,5,6,23,8,9,10,11,12,13,24,15] |
| ; AVX512BW-NEXT: vpermi2d %zmm17, %zmm0, %zmm10 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,1,2,3,4,25,6,7,8,9,10,11,26,13,14,15] |
| ; AVX512BW-NEXT: vpermi2d %zmm17, %zmm1, %zmm0 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,1,2,27,4,5,6,7,8,9,28,11,12,13,14,15] |
| ; AVX512BW-NEXT: vpermi2d %zmm17, %zmm2, %zmm1 |
| ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,29,2,3,4,5,6,7,30,9,10,11,12,13,14,31] |
| ; AVX512BW-NEXT: vpermi2d %zmm17, %zmm11, %zmm2 |
| ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax |
| ; AVX512BW-NEXT: vmovdqa64 %zmm7, 1472(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm21, 1408(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm9, 1344(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm16, 1280(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm31, 1216(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm6, 1152(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm8, 1088(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm4, 1024(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm20, 960(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm18, 896(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm5, 832(%rax) |
| ; AVX512BW-NEXT: vmovups (%rsp), %zmm3 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovaps %zmm3, 768(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm19, 704(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm26, 640(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm25, 576(%rax) |
| ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovaps %zmm3, 512(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm27, 448(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm24, 384(%rax) |
| ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovaps %zmm3, 320(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm28, 256(%rax) |
| ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovaps %zmm3, 192(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm23, 128(%rax) |
| ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovaps %zmm3, 64(%rax) |
| ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload |
| ; AVX512BW-NEXT: vmovaps %zmm3, (%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm2, 1728(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm1, 1664(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm0, 1600(%rax) |
| ; AVX512BW-NEXT: vmovdqa64 %zmm10, 1536(%rax) |
| ; AVX512BW-NEXT: addq $3016, %rsp # imm = 0xBC8 |
| ; AVX512BW-NEXT: vzeroupper |
| ; AVX512BW-NEXT: retq |
| %in.vec0 = load <64 x i32>, ptr %in.vecptr0, align 64 |
| %in.vec1 = load <64 x i32>, ptr %in.vecptr1, align 64 |
| %in.vec2 = load <64 x i32>, ptr %in.vecptr2, align 64 |
| %in.vec3 = load <64 x i32>, ptr %in.vecptr3, align 64 |
| %in.vec4 = load <64 x i32>, ptr %in.vecptr4, align 64 |
| %in.vec5 = load <64 x i32>, ptr %in.vecptr5, align 64 |
| %in.vec6 = load <64 x i32>, ptr %in.vecptr6, align 64 |
| %1 = shufflevector <64 x i32> %in.vec0, <64 x i32> %in.vec1, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127> |
| %2 = shufflevector <64 x i32> %in.vec2, <64 x i32> %in.vec3, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127> |
| %3 = shufflevector <64 x i32> %in.vec4, <64 x i32> %in.vec5, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127> |
| %4 = shufflevector <128 x i32> %1, <128 x i32> %2, <256 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223, i32 224, i32 225, i32 226, i32 227, i32 228, i32 229, i32 230, i32 231, i32 232, i32 233, i32 234, i32 235, i32 236, i32 237, i32 238, i32 239, i32 240, i32 241, i32 242, i32 243, i32 244, i32 245, i32 246, i32 247, i32 248, i32 249, i32 250, i32 251, i32 252, i32 253, i32 254, i32 255> |
| %5 = shufflevector <64 x i32> %in.vec6, <64 x i32> poison, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef> |
| %6 = shufflevector <128 x i32> %3, <128 x i32> %5, <192 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191> |
| %7 = shufflevector <192 x i32> %6, <192 x i32> poison, <256 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef> |
| %8 = shufflevector <256 x i32> %4, <256 x i32> %7, <448 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223, i32 224, i32 225, i32 226, i32 227, i32 228, i32 229, i32 230, i32 231, i32 232, i32 233, i32 234, i32 235, i32 236, i32 237, i32 238, i32 239, i32 240, i32 241, i32 242, i32 243, i32 244, i32 245, i32 246, i32 247, i32 248, i32 249, i32 250, i32 251, i32 252, i32 253, i32 254, i32 255, i32 256, i32 257, i32 258, i32 259, i32 260, i32 261, i32 262, i32 263, i32 264, i32 265, i32 266, i32 267, i32 268, i32 269, i32 270, i32 271, i32 272, i32 273, i32 274, i32 275, i32 276, i32 277, i32 278, i32 279, i32 280, i32 281, i32 282, i32 283, i32 284, i32 285, i32 286, i32 287, i32 288, i32 289, i32 290, i32 291, i32 292, i32 293, i32 294, i32 295, i32 296, i32 297, i32 298, i32 299, i32 300, i32 301, i32 302, i32 303, i32 304, i32 305, i32 306, i32 307, i32 308, i32 309, i32 310, i32 311, i32 312, i32 313, i32 314, i32 315, i32 316, i32 317, i32 318, i32 319, i32 320, i32 321, i32 322, i32 323, i32 324, i32 325, i32 326, i32 327, i32 328, i32 329, i32 330, i32 331, i32 332, i32 333, i32 334, i32 335, i32 336, i32 337, i32 338, i32 339, i32 340, i32 341, i32 342, i32 343, i32 344, i32 345, i32 346, i32 347, i32 348, i32 349, i32 350, i32 351, i32 352, i32 353, i32 354, i32 355, i32 356, i32 357, i32 358, i32 359, i32 360, i32 361, i32 362, i32 363, i32 364, i32 365, i32 366, i32 367, i32 368, i32 369, i32 370, i32 371, i32 372, i32 373, i32 374, i32 375, i32 376, i32 377, i32 378, i32 379, i32 380, i32 381, i32 382, i32 383, i32 384, i32 385, i32 386, i32 387, i32 388, i32 389, i32 390, i32 391, i32 392, i32 393, i32 394, i32 395, i32 396, i32 397, i32 398, i32 399, i32 400, i32 401, i32 402, i32 403, i32 404, i32 405, i32 406, i32 407, i32 408, i32 409, i32 410, i32 411, i32 412, i32 413, i32 414, i32 415, i32 416, i32 417, i32 418, i32 419, i32 420, i32 421, i32 422, i32 423, i32 424, i32 425, i32 426, i32 427, i32 428, i32 429, i32 430, i32 431, i32 432, i32 433, i32 434, i32 435, i32 436, i32 437, i32 438, i32 439, i32 440, i32 441, i32 442, i32 443, i32 444, i32 445, i32 446, i32 447> |
| %interleaved.vec = shufflevector <448 x i32> %8, <448 x i32> poison, <448 x i32> <i32 0, i32 64, i32 128, i32 192, i32 256, i32 320, i32 384, i32 1, i32 65, i32 129, i32 193, i32 257, i32 321, i32 385, i32 2, i32 66, i32 130, i32 194, i32 258, i32 322, i32 386, i32 3, i32 67, i32 131, i32 195, i32 259, i32 323, i32 387, i32 4, i32 68, i32 132, i32 196, i32 260, i32 324, i32 388, i32 5, i32 69, i32 133, i32 197, i32 261, i32 325, i32 389, i32 6, i32 70, i32 134, i32 198, i32 262, i32 326, i32 390, i32 7, i32 71, i32 135, i32 199, i32 263, i32 327, i32 391, i32 8, i32 72, i32 136, i32 200, i32 264, i32 328, i32 392, i32 9, i32 73, i32 137, i32 201, i32 265, i32 329, i32 393, i32 10, i32 74, i32 138, i32 202, i32 266, i32 330, i32 394, i32 11, i32 75, i32 139, i32 203, i32 267, i32 331, i32 395, i32 12, i32 76, i32 140, i32 204, i32 268, i32 332, i32 396, i32 13, i32 77, i32 141, i32 205, i32 269, i32 333, i32 397, i32 14, i32 78, i32 142, i32 206, i32 270, i32 334, i32 398, i32 15, i32 79, i32 143, i32 207, i32 271, i32 335, i32 399, i32 16, i32 80, i32 144, i32 208, i32 272, i32 336, i32 400, i32 17, i32 81, i32 145, i32 209, i32 273, i32 337, i32 401, i32 18, i32 82, i32 146, i32 210, i32 274, i32 338, i32 402, i32 19, i32 83, i32 147, i32 211, i32 275, i32 339, i32 403, i32 20, i32 84, i32 148, i32 212, i32 276, i32 340, i32 404, i32 21, i32 85, i32 149, i32 213, i32 277, i32 341, i32 405, i32 22, i32 86, i32 150, i32 214, i32 278, i32 342, i32 406, i32 23, i32 87, i32 151, i32 215, i32 279, i32 343, i32 407, i32 24, i32 88, i32 152, i32 216, i32 280, i32 344, i32 408, i32 25, i32 89, i32 153, i32 217, i32 281, i32 345, i32 409, i32 26, i32 90, i32 154, i32 218, i32 282, i32 346, i32 410, i32 27, i32 91, i32 155, i32 219, i32 283, i32 347, i32 411, i32 28, i32 92, i32 156, i32 220, i32 284, i32 348, i32 412, i32 29, i32 93, i32 157, i32 221, i32 285, i32 349, i32 413, i32 30, i32 94, i32 158, i32 222, i32 286, i32 350, i32 414, i32 31, i32 95, i32 159, i32 223, i32 287, i32 351, i32 415, i32 32, i32 96, i32 160, i32 224, i32 288, i32 352, i32 416, i32 33, i32 97, i32 161, i32 225, i32 289, i32 353, i32 417, i32 34, i32 98, i32 162, i32 226, i32 290, i32 354, i32 418, i32 35, i32 99, i32 163, i32 227, i32 291, i32 355, i32 419, i32 36, i32 100, i32 164, i32 228, i32 292, i32 356, i32 420, i32 37, i32 101, i32 165, i32 229, i32 293, i32 357, i32 421, i32 38, i32 102, i32 166, i32 230, i32 294, i32 358, i32 422, i32 39, i32 103, i32 167, i32 231, i32 295, i32 359, i32 423, i32 40, i32 104, i32 168, i32 232, i32 296, i32 360, i32 424, i32 41, i32 105, i32 169, i32 233, i32 297, i32 361, i32 425, i32 42, i32 106, i32 170, i32 234, i32 298, i32 362, i32 426, i32 43, i32 107, i32 171, i32 235, i32 299, i32 363, i32 427, i32 44, i32 108, i32 172, i32 236, i32 300, i32 364, i32 428, i32 45, i32 109, i32 173, i32 237, i32 301, i32 365, i32 429, i32 46, i32 110, i32 174, i32 238, i32 302, i32 366, i32 430, i32 47, i32 111, i32 175, i32 239, i32 303, i32 367, i32 431, i32 48, i32 112, i32 176, i32 240, i32 304, i32 368, i32 432, i32 49, i32 113, i32 177, i32 241, i32 305, i32 369, i32 433, i32 50, i32 114, i32 178, i32 242, i32 306, i32 370, i32 434, i32 51, i32 115, i32 179, i32 243, i32 307, i32 371, i32 435, i32 52, i32 116, i32 180, i32 244, i32 308, i32 372, i32 436, i32 53, i32 117, i32 181, i32 245, i32 309, i32 373, i32 437, i32 54, i32 118, i32 182, i32 246, i32 310, i32 374, i32 438, i32 55, i32 119, i32 183, i32 247, i32 311, i32 375, i32 439, i32 56, i32 120, i32 184, i32 248, i32 312, i32 376, i32 440, i32 57, i32 121, i32 185, i32 249, i32 313, i32 377, i32 441, i32 58, i32 122, i32 186, i32 250, i32 314, i32 378, i32 442, i32 59, i32 123, i32 187, i32 251, i32 315, i32 379, i32 443, i32 60, i32 124, i32 188, i32 252, i32 316, i32 380, i32 444, i32 61, i32 125, i32 189, i32 253, i32 317, i32 381, i32 445, i32 62, i32 126, i32 190, i32 254, i32 318, i32 382, i32 446, i32 63, i32 127, i32 191, i32 255, i32 319, i32 383, i32 447> |
| store <448 x i32> %interleaved.vec, ptr %out.vec, align 64 |
| ret void |
| } |
| ;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line: |
| ; AVX: {{.*}} |
| ; AVX1: {{.*}} |
| ; AVX2: {{.*}} |
| ; AVX2-ONLY: {{.*}} |
| ; AVX512BW-ONLY-FAST: {{.*}} |
| ; AVX512BW-ONLY-SLOW: {{.*}} |
| ; AVX512DQ-FAST: {{.*}} |
| ; AVX512DQ-SLOW: {{.*}} |
| ; AVX512DQBW-FAST: {{.*}} |
| ; AVX512DQBW-SLOW: {{.*}} |
| ; AVX512F-ONLY-FAST: {{.*}} |
| ; AVX512F-ONLY-SLOW: {{.*}} |
| ; FALLBACK0: {{.*}} |
| ; FALLBACK1: {{.*}} |
| ; FALLBACK10: {{.*}} |
| ; FALLBACK11: {{.*}} |
| ; FALLBACK12: {{.*}} |
| ; FALLBACK2: {{.*}} |
| ; FALLBACK3: {{.*}} |
| ; FALLBACK4: {{.*}} |
| ; FALLBACK5: {{.*}} |
| ; FALLBACK6: {{.*}} |
| ; FALLBACK7: {{.*}} |
| ; FALLBACK8: {{.*}} |
| ; FALLBACK9: {{.*}} |