[X86] combineINSERT_SUBVECTOR - peek through bitcasts to find a concatenation of subvector shuffles #131331
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…tenation of subvector shuffles Extend existing concat(shuffle(),shuffle(),..) shuffle combining to handle mixed bitcasts such as concat(shuffle(),bitcast(shuffle()),...)
|
@llvm/pr-subscribers-backend-x86 Author: Simon Pilgrim (RKSimon) ChangesExtend existing concat(shuffle(),shuffle(),..) shuffle combining to handle mixed bitcasts such as concat(shuffle(),bitcast(shuffle()),...) Patch is 1.30 MiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/131331.diff 8 Files Affected:
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index ac39c8fc0cba3..f02609a62425b 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -58818,7 +58818,7 @@ static SDValue combineINSERT_SUBVECTOR(SDNode *N, SelectionDAG &DAG,
// Attempt to recursively combine to a shuffle.
if (all_of(SubVectorOps, [](SDValue SubOp) {
- return isTargetShuffle(SubOp.getOpcode());
+ return isTargetShuffle(peekThroughBitcasts(SubOp).getOpcode());
})) {
SDValue Op(N, 0);
if (SDValue Res = combineX86ShufflesRecursively(Op, DAG, Subtarget))
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll
index fe567b90b3bcd..0df63422b5d84 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll
@@ -3342,629 +3342,615 @@ define void @store_i16_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512-LABEL: store_i16_stride5_vf32:
; AVX512: # %bb.0:
-; AVX512-NEXT: vmovdqa 32(%rdx), %xmm9
-; AVX512-NEXT: vmovdqa (%rcx), %xmm12
-; AVX512-NEXT: vmovdqa 32(%rcx), %xmm10
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm13 = [4,5,2,3,2,3,0,1,10,11,8,9,4,5,6,7]
-; AVX512-NEXT: vpshufb %xmm13, %xmm0, %xmm0
-; AVX512-NEXT: vmovdqa64 (%rdx), %ymm17
+; AVX512-NEXT: vmovdqa 32(%rdx), %xmm10
+; AVX512-NEXT: vmovdqa (%rcx), %xmm9
+; AVX512-NEXT: vmovdqa 32(%rcx), %xmm11
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [4,5,2,3,2,3,0,1,10,11,8,9,4,5,6,7]
+; AVX512-NEXT: vpshufb %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vmovdqa64 %xmm1, %xmm23
+; AVX512-NEXT: vmovdqa64 (%rdx), %ymm18
; AVX512-NEXT: vmovdqa 32(%rdx), %ymm5
-; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm17[3,2,3,3,7,6,7,7]
-; AVX512-NEXT: vmovdqa (%rcx), %ymm1
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm3 = ymm1[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3,4],ymm3[5,6,7,8],ymm2[9],ymm3[10],ymm2[11,12],ymm3[13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,2]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm2[0,1,2,3],zmm0[0,1,0,1]
-; AVX512-NEXT: vmovdqa (%rsi), %xmm14
-; AVX512-NEXT: vmovdqa 32(%rsi), %xmm8
-; AVX512-NEXT: vmovdqa (%rdi), %xmm15
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm18[3,2,3,3,7,6,7,7]
+; AVX512-NEXT: vmovdqa (%rcx), %ymm7
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm2 = ymm7[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2],ymm1[3,4],ymm2[5,6,7,8],ymm1[9],ymm2[10],ymm1[11,12],ymm2[13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,2]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[0,1,0,1]
+; AVX512-NEXT: vmovdqa (%rsi), %xmm3
+; AVX512-NEXT: vmovdqa 32(%rsi), %xmm1
+; AVX512-NEXT: vmovdqa (%rdi), %xmm8
; AVX512-NEXT: vmovdqa 32(%rdi), %xmm2
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm8[0],xmm2[1],xmm8[1],xmm2[2],xmm8[2],xmm2[3],xmm8[3]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[0,2,1,3]
; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,5,6]
-; AVX512-NEXT: vmovdqa64 (%rdi), %ymm20
-; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm20[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vmovdqa (%rdi), %ymm4
+; AVX512-NEXT: vpshufd {{.*#+}} ymm13 = ymm4[2,3,2,3,6,7,6,7]
; AVX512-NEXT: vmovdqa (%rsi), %ymm6
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm7 = ymm6[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[2,3,2,2,6,7,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm7[0],ymm3[1],ymm7[2],ymm3[3],ymm7[4,5],ymm3[6],ymm7[7,8],ymm3[9],ymm7[10],ymm3[11],ymm7[12,13],ymm3[14],ymm7[15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,2]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm2[0,1,0,1]
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535]
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm3 = zmm0 ^ (zmm16 & (zmm3 ^ zmm0))
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm14 = ymm6[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[2,3,2,2,6,7,6,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm13 = ymm14[0],ymm13[1],ymm14[2],ymm13[3],ymm14[4,5],ymm13[6],ymm14[7,8],ymm13[9],ymm14[10],ymm13[11],ymm14[12,13],ymm13[14],ymm14[15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,2]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm13 = zmm13[0,1,2,3],zmm2[0,1,0,1]
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm17 = [65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535]
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm13 = zmm0 ^ (zmm17 & (zmm13 ^ zmm0))
; AVX512-NEXT: vpbroadcastq 24(%r8), %ymm0
; AVX512-NEXT: vpbroadcastq 32(%r8), %ymm2
; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm19
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm19 = zmm19 ^ (mem & (zmm19 ^ zmm3))
-; AVX512-NEXT: vmovdqa 32(%rsi), %ymm3
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm19 = zmm19 ^ (mem & (zmm19 ^ zmm13))
+; AVX512-NEXT: vmovdqa 32(%rsi), %ymm14
; AVX512-NEXT: vmovdqa {{.*#+}} ymm2 = [u,u,0,1,u,u,u,u,14,15,u,u,2,3,u,u,u,u,16,17,u,u,u,u,30,31,u,u,18,19,u,u]
-; AVX512-NEXT: vpshufb %ymm2, %ymm3, %ymm0
+; AVX512-NEXT: vpshufb %ymm2, %ymm14, %ymm0
+; AVX512-NEXT: vmovdqa64 %ymm2, %ymm21
+; AVX512-NEXT: vmovdqa64 32(%rdi), %ymm16
+; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm16[1,1,2,2]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0],ymm0[1],ymm13[2,3],ymm0[4],ymm13[5],ymm0[6],ymm13[7,8],ymm0[9],ymm13[10,11],ymm0[12],ymm13[13],ymm0[14],ymm13[15]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [6,7,u,u,10,11,6,7,u,u,8,9,u,u,12,13]
+; AVX512-NEXT: vpshufb %xmm0, %xmm1, %xmm1
+; AVX512-NEXT: vpbroadcastq 40(%rdi), %xmm15
+; AVX512-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm15[1],xmm1[2,3],xmm15[4],xmm1[5],xmm15[6],xmm1[7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
+; AVX512-NEXT: vinserti64x4 $1, %ymm13, %zmm1, %zmm13
+; AVX512-NEXT: vmovdqa 32(%rcx), %ymm15
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm2 = [u,u,12,13,u,u,0,1,u,u,u,u,14,15,u,u,u,u,28,29,u,u,16,17,u,u,u,u,30,31,u,u]
+; AVX512-NEXT: vpshufb %ymm2, %ymm15, %ymm1
; AVX512-NEXT: vmovdqa64 %ymm2, %ymm22
-; AVX512-NEXT: vmovdqa64 32(%rdi), %ymm18
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm18[1,1,2,2]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm7[0],ymm0[1],ymm7[2,3],ymm0[4],ymm7[5],ymm0[6],ymm7[7,8],ymm0[9],ymm7[10,11],ymm0[12],ymm7[13],ymm0[14],ymm7[15]
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm7 = [6,7,u,u,10,11,6,7,u,u,8,9,u,u,12,13]
-; AVX512-NEXT: vpshufb %xmm7, %xmm8, %xmm8
-; AVX512-NEXT: vpbroadcastq 40(%rdi), %xmm11
-; AVX512-NEXT: vpblendw {{.*#+}} xmm8 = xmm8[0],xmm11[1],xmm8[2,3],xmm11[4],xmm8[5],xmm11[6],xmm8[7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm11
-; AVX512-NEXT: vmovdqa 32(%rcx), %ymm0
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm4 = [u,u,12,13,u,u,0,1,u,u,u,u,14,15,u,u,u,u,28,29,u,u,16,17,u,u,u,u,30,31,u,u]
-; AVX512-NEXT: vpshufb %ymm4, %ymm0, %ymm2
-; AVX512-NEXT: vmovdqa64 %ymm4, %ymm23
-; AVX512-NEXT: vpshufd {{.*#+}} ymm8 = ymm5[3,0,3,0,7,4,7,4]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm8[0],ymm2[1],ymm8[2],ymm2[3],ymm8[4,5],ymm2[6],ymm8[7,8],ymm2[9],ymm8[10],ymm2[11],ymm8[12,13],ymm2[14],ymm8[15]
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm8 = [10,11,u,u,6,7,u,u,8,9,8,9,u,u,8,9]
-; AVX512-NEXT: vpshufb %xmm8, %xmm10, %xmm10
-; AVX512-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[1,2,2,2]
-; AVX512-NEXT: vpblendw {{.*#+}} xmm9 = xmm10[0],xmm9[1],xmm10[2],xmm9[3],xmm10[4,5],xmm9[6],xmm10[7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,0]
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm9, %zmm2
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm11))
-; AVX512-NEXT: vmovdqa (%r8), %ymm10
-; AVX512-NEXT: vmovdqa 32(%r8), %ymm9
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm11 = [128,128,128,128,12,13,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128]
-; AVX512-NEXT: vpshufb %ymm11, %ymm9, %ymm4
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm21 = [65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,1,1]
-; AVX512-NEXT: vpandnq %ymm9, %ymm21, %ymm9
-; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm9, %zmm9
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 | (zmm2 & zmm21)
+; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm5[3,0,3,0,7,4,7,4]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0],ymm1[1],ymm12[2],ymm1[3],ymm12[4,5],ymm1[6],ymm12[7,8],ymm1[9],ymm12[10],ymm1[11],ymm12[12,13],ymm1[14],ymm12[15]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [10,11,u,u,6,7,u,u,8,9,8,9,u,u,8,9]
+; AVX512-NEXT: vpshufb %xmm1, %xmm11, %xmm11
+; AVX512-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[1,2,2,2]
+; AVX512-NEXT: vpblendw {{.*#+}} xmm10 = xmm11[0],xmm10[1],xmm11[2],xmm10[3],xmm11[4,5],xmm10[6],xmm11[7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,1,0,0]
+; AVX512-NEXT: vinserti64x4 $1, %ymm12, %zmm10, %zmm12
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm12 = zmm12 ^ (mem & (zmm12 ^ zmm13))
+; AVX512-NEXT: vmovdqa (%r8), %ymm11
+; AVX512-NEXT: vmovdqa 32(%r8), %ymm10
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm13 = [128,128,128,128,12,13,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128]
+; AVX512-NEXT: vpshufb %ymm13, %ymm10, %ymm2
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm20 = [65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,1,1,1]
+; AVX512-NEXT: vpandnq %ymm10, %ymm20, %ymm10
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm10, %zmm10
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm10 = zmm10 | (zmm12 & zmm20)
+; AVX512-NEXT: vpshufb %xmm0, %xmm3, %xmm0
+; AVX512-NEXT: vpbroadcastq 8(%rdi), %xmm2
+; AVX512-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2,3],xmm2[4],xmm0[5],xmm2[6],xmm0[7]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[0,2,1,3]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,5,6]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm2[0,1,0,1],zmm0[0,1,0,1]
; AVX512-NEXT: vmovdqa (%rdx), %xmm2
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm12[0],xmm2[0],xmm12[1],xmm2[1],xmm12[2],xmm2[2],xmm12[3],xmm2[3]
-; AVX512-NEXT: vpshufb %xmm13, %xmm4, %xmm4
-; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
-; AVX512-NEXT: vpshufb %xmm8, %xmm12, %xmm8
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm9[0],xmm2[0],xmm9[1],xmm2[1],xmm9[2],xmm2[2],xmm9[3],xmm2[3]
+; AVX512-NEXT: vmovdqa64 %xmm23, %xmm8
+; AVX512-NEXT: vpshufb %xmm8, %xmm3, %xmm3
+; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
+; AVX512-NEXT: vpshufb %xmm1, %xmm9, %xmm1
; AVX512-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,2,2,2]
-; AVX512-NEXT: vpblendw {{.*#+}} xmm2 = xmm8[0],xmm2[1],xmm8[2],xmm2[3],xmm8[4,5],xmm2[6],xmm8[7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,0]
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm2
-; AVX512-NEXT: vpshufb %xmm7, %xmm14, %xmm4
-; AVX512-NEXT: vpbroadcastq 8(%rdi), %xmm7
-; AVX512-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0],xmm7[1],xmm4[2,3],xmm7[4],xmm4[5],xmm7[6],xmm4[7]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[0,2,1,3]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,5,6]
-; AVX512-NEXT: vinserti32x4 $2, %xmm4, %zmm7, %zmm4
-; AVX512-NEXT: vpermq {{.*#+}} zmm4 = zmm4[0,1,0,1,4,5,4,5]
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm7 = [65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535]
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm4 = zmm2 ^ (zmm7 & (zmm4 ^ zmm2))
-; AVX512-NEXT: vpbroadcastq (%r8), %ymm2
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm10[0,1,1,1]
-; AVX512-NEXT: vinserti64x4 $1, %ymm8, %zmm2, %zmm2
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm4))
-; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm18[0,1,2,1,4,5,6,5]
-; AVX512-NEXT: vprolq $16, %ymm3, %ymm8
-; AVX512-NEXT: vpblendw {{.*#+}} ymm4 = ymm8[0,1],ymm4[2],ymm8[3],ymm4[4],ymm8[5,6],ymm4[7],ymm8[8,9],ymm4[10],ymm8[11],ymm4[12],ymm8[13,14],ymm4[15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm8 = ymm18[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm3 = ymm3[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,3,2,2,6,7,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm8[1],ymm3[2],ymm8[3],ymm3[4,5],ymm8[6],ymm3[7,8],ymm8[9],ymm3[10],ymm8[11],ymm3[12,13],ymm8[14],ymm3[15]
+; AVX512-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2],xmm2[3],xmm1[4,5],xmm2[6],xmm1[7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,0]
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535]
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (zmm2 & (zmm1 ^ zmm0))
+; AVX512-NEXT: vpbroadcastq (%r8), %ymm0
+; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm11[0,1,1,1]
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm1))
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm16[0,1,2,1,4,5,6,5]
+; AVX512-NEXT: vprolq $16, %ymm14, %ymm3
+; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0,1],ymm1[2],ymm3[3],ymm1[4],ymm3[5,6],ymm1[7],ymm3[8,9],ymm1[10],ymm3[11],ymm1[12],ymm3[13,14],ymm1[15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm16[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm8 = ymm14[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[2,3,2,2,6,7,6,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm8[0],ymm3[1],ymm8[2],ymm3[3],ymm8[4,5],ymm3[6],ymm8[7,8],ymm3[9],ymm8[10],ymm3[11],ymm8[12,13],ymm3[14],ymm8[15]
; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,2]
-; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm4
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1
; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25,18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25]
; AVX512-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm3, %ymm0, %ymm8
-; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm5[1,1,1,2,5,5,5,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0],ymm12[1],ymm8[2,3],ymm12[4],ymm8[5],ymm12[6],ymm8[7,8],ymm12[9],ymm8[10,11],ymm12[12],ymm8[13],ymm12[14],ymm8[15]
+; AVX512-NEXT: vpshufb %ymm3, %ymm15, %ymm8
+; AVX512-NEXT: vpshufd {{.*#+}} ymm9 = ymm5[1,1,1,2,5,5,5,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0],ymm9[1],ymm8[2,3],ymm9[4],ymm8[5],ymm9[6],ymm8[7,8],ymm9[9],ymm8[10,11],ymm9[12],ymm8[13],ymm9[14],ymm8[15]
; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[3,2,3,3,7,6,7,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2],ymm5[3,4],ymm0[5,6,7,8],ymm5[9],ymm0[10],ymm5[11,12],ymm0[13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,2]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm0
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm4 ^ (zmm7 & (zmm0 ^ zmm4))
-; AVX512-NEXT: vpbroadcastq 48(%r8), %ymm4
-; AVX512-NEXT: vpbroadcastq 56(%r8), %ymm5
-; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm0))
-; AVX512-NEXT: vmovdqa64 %ymm22, %ymm0
-; AVX512-NEXT: vpshufb %ymm0, %ymm6, %ymm0
-; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm20[1,1,2,2]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2,3],ymm0[4],ymm5[5],ymm0[6],ymm5[7,8],ymm0[9],ymm5[10,11],ymm0[12],ymm5[13],ymm0[14],ymm5[15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm20[0,1,2,1,4,5,6,5]
-; AVX512-NEXT: vprolq $16, %ymm6, %ymm6
-; AVX512-NEXT: vpblendw {{.*#+}} ymm5 = ymm6[0,1],ymm5[2],ymm6[3],ymm5[4],ymm6[5,6],ymm5[7],ymm6[8,9],ymm5[10],ymm6[11],ymm5[12],ymm6[13,14],ymm5[15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm0
-; AVX512-NEXT: vmovdqa64 %ymm23, %ymm5
-; AVX512-NEXT: vpshufb %ymm5, %ymm1, %ymm5
-; AVX512-NEXT: vpshufd {{.*#+}} ymm6 = ymm17[3,0,3,0,7,4,7,4]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2],ymm5[3],ymm6[4,5],ymm5[6],ymm6[7,8],ymm5[9],ymm6[10],ymm5[11],ymm6[12,13],ymm5[14],ymm6[15]
-; AVX512-NEXT: vpshufb %ymm3, %ymm1, %ymm1
-; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm17[1,1,1,2,5,5,5,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2,3],ymm3[4],ymm1[5],ymm3[6],ymm1[7,8],ymm3[9],ymm1[10,11],ymm3[12],ymm1[13],ymm3[14],ymm1[15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm5, %zmm1
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm1 = zmm0 ^ (zmm16 & (zmm1 ^ zmm0))
-; AVX512-NEXT: vpshufb %ymm11, %ymm10, %ymm0
-; AVX512-NEXT: vpbroadcastq 16(%r8), %ymm3
-; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm3, %ymm3
-; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 | (zmm1 & mem)
-; AVX512-NEXT: vmovdqa64 %zmm0, 64(%r9)
-; AVX512-NEXT: vmovdqa64 %zmm4, 256(%r9)
-; AVX512-NEXT: vmovdqa64 %zmm2, (%r9)
-; AVX512-NEXT: vmovdqa64 %zmm9, 192(%r9)
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm9 = ymm15[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm5 = ymm9[0],ymm5[1],ymm9[2],ymm5[3,4],ymm9[5,6,7,8],ymm5[9],ymm9[10],ymm5[11,12],ymm9[13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,3,2]
+; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm8, %zmm5
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm5 = zmm1 ^ (zmm2 & (zmm5 ^ zmm1))
+; AVX512-NEXT: vpbroadcastq 48(%r8), %ymm1
+; AVX512-NEXT: vpbroadcastq 56(%r8), %ymm2
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm5))
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm4[0,1,2,1,4,5,6,5]
+; AVX512-NEXT: vprolq $16, %ymm6, %ymm5
+; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2],ymm5[3],ymm2[4],ymm5[5,6],ymm2[7],ymm5[8,9],ymm2[10],ymm5[11],ymm2[12],ymm5[13,14],ymm2[15]
+; AVX512-NEXT: vmovdqa64 %ymm21, %ymm5
+; AVX512-NEXT: vpshufb %ymm5, %ymm6, %ymm5
+; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,1,2,2]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[2,3],ymm5[4],ymm4[5],ymm5[6],ymm4[7,8],ymm5[9],ymm4[10,11],ymm5[12],ymm4[13],ymm5[14],ymm4[15]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm4[0,1,2,3],zmm2[2,3,2,3]
+; AVX512-NEXT: vpshufb %ymm3, %ymm7,...
[truncated]
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Extend existing concat(shuffle(),shuffle(),..) shuffle combining to handle mixed bitcasts such as concat(shuffle(),bitcast(shuffle()),...)