From 92235e3bf467e1f3aa820cae3c406aac4e867a4b Mon Sep 17 00:00:00 2001 From: Simon Pilgrim Date: Sat, 30 Apr 2022 11:00:28 +0100 Subject: [PATCH] [X86] lowerShuffleAsRepeatedMaskAndLanePermute - permit 32-bit sublane permute for unary v32i8 cases Increase the likelihood that we can lower to a permd(pshufb()) pattern, but only after we've attempted with 64-bit sublane permutes first Fixes #55066 --- llvm/lib/Target/X86/X86ISelLowering.cpp | 24 ++-- .../X86/vector-interleaved-load-i8-stride-4.ll | 149 ++++++++++----------- llvm/test/CodeGen/X86/vector-shuffle-256-v32.ll | 142 ++++++-------------- llvm/test/CodeGen/X86/vector-shuffle-512-v64.ll | 48 +++---- llvm/test/CodeGen/X86/x86-interleaved-access.ll | 149 ++++++++++----------- 5 files changed, 213 insertions(+), 299 deletions(-) diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp index ec14ee9..c3d78dda 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -17397,16 +17397,24 @@ static SDValue lowerShuffleAsRepeatedMaskAndLanePermute( }; // On AVX2 targets we can permute 256-bit vectors as 64-bit sub-lanes - // (with PERMQ/PERMPD). On AVX512BW targets, permuting 32-bit sub-lanes, even - // with a variable shuffle, is worth it for 64xi8 vectors. Otherwise we can - // only permute whole 128-bit lanes. - int SubLaneScale = 1; - if (Subtarget.hasAVX2() && VT.is256BitVector()) - SubLaneScale = 2; + // (with PERMQ/PERMPD). On AVX2/AVX512BW targets, permuting 32-bit sub-lanes, + // even with a variable shuffle, can be worth it for v32i8/v64i8 vectors. + // Otherwise we can only permute whole 128-bit lanes. + int MinSubLaneScale = 1, MaxSubLaneScale = 1; + if (Subtarget.hasAVX2() && VT.is256BitVector()) { + bool OnlyLowestElts = isUndefOrInRange(Mask, 0, NumLaneElts); + MinSubLaneScale = 2; + MaxSubLaneScale = + (!OnlyLowestElts && V2.isUndef() && VT == MVT::v32i8) ? 4 : 2; + } if (Subtarget.hasBWI() && VT == MVT::v64i8) - SubLaneScale = 4; + MinSubLaneScale = MaxSubLaneScale = 4; + + for (int Scale = MinSubLaneScale; Scale <= MaxSubLaneScale; Scale *= 2) + if (SDValue Shuffle = ShuffleSubLanes(Scale)) + return Shuffle; - return ShuffleSubLanes(SubLaneScale); + return SDValue(); } static bool matchShuffleWithSHUFPD(MVT VT, SDValue &V1, SDValue &V2, diff --git a/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-4.ll b/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-4.ll index f309248..7d5e329 100644 --- a/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-4.ll +++ b/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-4.ll @@ -883,92 +883,79 @@ define void @load_i8_stride4_vf32(<128 x i8>* %in.vec, <32 x i8>* %out.vec0, <32 ; ; AVX2-LABEL: load_i8_stride4_vf32: ; AVX2: # %bb.0: -; AVX2-NEXT: vmovdqa {{.*#+}} xmm0 = -; AVX2-NEXT: vmovdqa 112(%rdi), %xmm10 -; AVX2-NEXT: vpshufb %xmm0, %xmm10, %xmm1 -; AVX2-NEXT: vmovdqa 96(%rdi), %xmm11 -; AVX2-NEXT: vpshufb %xmm0, %xmm11, %xmm3 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1] -; AVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm1 -; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u> -; AVX2-NEXT: vmovdqa 80(%rdi), %xmm13 -; AVX2-NEXT: vpshufb %xmm2, %xmm13, %xmm4 -; AVX2-NEXT: vmovdqa 64(%rdi), %xmm5 -; AVX2-NEXT: vpshufb %xmm2, %xmm5, %xmm6 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm6[0],xmm4[0],xmm6[1],xmm4[1] -; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4 -; AVX2-NEXT: vpblendd {{.*#+}} ymm8 = ymm4[0,1,2,3,4,5],ymm1[6,7] -; AVX2-NEXT: vmovdqa (%rdi), %xmm12 -; AVX2-NEXT: vmovdqa 16(%rdi), %xmm14 -; AVX2-NEXT: vmovdqa 32(%rdi), %xmm6 -; AVX2-NEXT: vmovdqa 48(%rdi), %xmm7 -; AVX2-NEXT: vpshufb %xmm0, %xmm7, %xmm1 -; AVX2-NEXT: vpshufb %xmm0, %xmm6, %xmm0 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] -; AVX2-NEXT: vpshufb %xmm2, %xmm14, %xmm1 -; AVX2-NEXT: vpshufb %xmm2, %xmm12, %xmm2 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1] -; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3] -; AVX2-NEXT: vpblendd {{.*#+}} ymm8 = ymm0[0,1,2,3],ymm8[4,5,6,7] +; AVX2-NEXT: vmovdqa 64(%rdi), %ymm9 +; AVX2-NEXT: vmovdqa 96(%rdi), %ymm11 +; AVX2-NEXT: vmovdqa (%rdi), %xmm2 +; AVX2-NEXT: vmovdqa 16(%rdi), %xmm3 +; AVX2-NEXT: vmovdqa 32(%rdi), %xmm4 +; AVX2-NEXT: vmovdqa 48(%rdi), %xmm5 +; AVX2-NEXT: vmovdqa {{.*#+}} xmm6 = +; AVX2-NEXT: vpshufb %xmm6, %xmm5, %xmm7 +; AVX2-NEXT: vpshufb %xmm6, %xmm4, %xmm6 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1] +; AVX2-NEXT: vmovdqa {{.*#+}} xmm7 = <0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %xmm7, %xmm3, %xmm0 +; AVX2-NEXT: vpshufb %xmm7, %xmm2, %xmm7 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm7[0],xmm0[0],xmm7[1],xmm0[1] +; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm6[2,3] +; AVX2-NEXT: vmovdqa {{.*#+}} ymm7 = <0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u,0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %ymm7, %ymm11, %ymm8 +; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm6 = [17179869184,17179869184,17179869184,17179869184] +; AVX2-NEXT: vpermd %ymm8, %ymm6, %ymm8 +; AVX2-NEXT: vpshufb %ymm7, %ymm9, %ymm7 +; AVX2-NEXT: vpermd %ymm7, %ymm6, %ymm7 +; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],ymm8[6,7] +; AVX2-NEXT: vpblendd {{.*#+}} ymm10 = ymm0[0,1,2,3],ymm7[4,5,6,7] ; AVX2-NEXT: vmovdqa {{.*#+}} xmm0 = -; AVX2-NEXT: vpshufb %xmm0, %xmm10, %xmm1 -; AVX2-NEXT: vpshufb %xmm0, %xmm11, %xmm2 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1] -; AVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm1 -; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u> -; AVX2-NEXT: vpshufb %xmm2, %xmm13, %xmm3 -; AVX2-NEXT: vpshufb %xmm2, %xmm5, %xmm4 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1] -; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3 -; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7] -; AVX2-NEXT: vpshufb %xmm0, %xmm7, %xmm3 -; AVX2-NEXT: vpshufb %xmm0, %xmm6, %xmm0 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1] -; AVX2-NEXT: vpshufb %xmm2, %xmm14, %xmm3 -; AVX2-NEXT: vpshufb %xmm2, %xmm12, %xmm2 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1] -; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3] -; AVX2-NEXT: vpblendd {{.*#+}} ymm9 = ymm0[0,1,2,3],ymm1[4,5,6,7] +; AVX2-NEXT: vpshufb %xmm0, %xmm5, %xmm7 +; AVX2-NEXT: vpshufb %xmm0, %xmm4, %xmm0 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1] +; AVX2-NEXT: vmovdqa {{.*#+}} xmm7 = <1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %xmm7, %xmm3, %xmm1 +; AVX2-NEXT: vpshufb %xmm7, %xmm2, %xmm7 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm7[0],xmm1[0],xmm7[1],xmm1[1] +; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3] +; AVX2-NEXT: vmovdqa {{.*#+}} ymm1 = <1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %ymm1, %ymm11, %ymm7 +; AVX2-NEXT: vpermd %ymm7, %ymm6, %ymm7 +; AVX2-NEXT: vpshufb %ymm1, %ymm9, %ymm1 +; AVX2-NEXT: vpermd %ymm1, %ymm6, %ymm1 +; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm7[6,7] +; AVX2-NEXT: vpblendd {{.*#+}} ymm8 = ymm0[0,1,2,3],ymm1[4,5,6,7] ; AVX2-NEXT: vmovdqa {{.*#+}} xmm0 = -; AVX2-NEXT: vpshufb %xmm0, %xmm10, %xmm1 -; AVX2-NEXT: vpshufb %xmm0, %xmm11, %xmm2 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1] -; AVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm1 -; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <2,6,10,14,u,u,u,u,u,u,u,u,u,u,u,u> -; AVX2-NEXT: vpshufb %xmm2, %xmm13, %xmm3 -; AVX2-NEXT: vpshufb %xmm2, %xmm5, %xmm4 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1] -; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3 -; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7] -; AVX2-NEXT: vpshufb %xmm0, %xmm7, %xmm3 -; AVX2-NEXT: vpshufb %xmm0, %xmm6, %xmm0 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1] -; AVX2-NEXT: vpshufb %xmm2, %xmm14, %xmm3 -; AVX2-NEXT: vpshufb %xmm2, %xmm12, %xmm2 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1] -; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3] +; AVX2-NEXT: vpshufb %xmm0, %xmm5, %xmm1 +; AVX2-NEXT: vpshufb %xmm0, %xmm4, %xmm0 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] +; AVX2-NEXT: vmovdqa {{.*#+}} xmm1 = <2,6,10,14,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %xmm1, %xmm3, %xmm7 +; AVX2-NEXT: vpshufb %xmm1, %xmm2, %xmm1 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm7[0],xmm1[1],xmm7[1] +; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3] +; AVX2-NEXT: vmovdqa {{.*#+}} ymm1 = <2,6,10,14,u,u,u,u,u,u,u,u,u,u,u,u,2,6,10,14,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %ymm1, %ymm11, %ymm7 +; AVX2-NEXT: vpermd %ymm7, %ymm6, %ymm7 +; AVX2-NEXT: vpshufb %ymm1, %ymm9, %ymm1 +; AVX2-NEXT: vpermd %ymm1, %ymm6, %ymm1 +; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm7[6,7] ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] ; AVX2-NEXT: vmovdqa {{.*#+}} xmm1 = -; AVX2-NEXT: vpshufb %xmm1, %xmm10, %xmm2 -; AVX2-NEXT: vpshufb %xmm1, %xmm11, %xmm3 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1] -; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2 -; AVX2-NEXT: vmovdqa {{.*#+}} xmm3 = <3,7,11,15,u,u,u,u,u,u,u,u,u,u,u,u> -; AVX2-NEXT: vpshufb %xmm3, %xmm13, %xmm4 -; AVX2-NEXT: vpshufb %xmm3, %xmm5, %xmm5 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1] -; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4 -; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5],ymm2[6,7] -; AVX2-NEXT: vpshufb %xmm1, %xmm7, %xmm4 -; AVX2-NEXT: vpshufb %xmm1, %xmm6, %xmm1 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1] -; AVX2-NEXT: vpshufb %xmm3, %xmm14, %xmm4 -; AVX2-NEXT: vpshufb %xmm3, %xmm12, %xmm3 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1] -; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3] +; AVX2-NEXT: vpshufb %xmm1, %xmm5, %xmm5 +; AVX2-NEXT: vpshufb %xmm1, %xmm4, %xmm1 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1] +; AVX2-NEXT: vmovdqa {{.*#+}} xmm4 = <3,7,11,15,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %xmm4, %xmm3, %xmm3 +; AVX2-NEXT: vpshufb %xmm4, %xmm2, %xmm2 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1] +; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3] +; AVX2-NEXT: vmovdqa {{.*#+}} ymm2 = <3,7,11,15,u,u,u,u,u,u,u,u,u,u,u,u,3,7,11,15,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %ymm2, %ymm11, %ymm3 +; AVX2-NEXT: vpermd %ymm3, %ymm6, %ymm3 +; AVX2-NEXT: vpshufb %ymm2, %ymm9, %ymm2 +; AVX2-NEXT: vpermd %ymm2, %ymm6, %ymm2 +; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7] ; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] -; AVX2-NEXT: vmovdqa %ymm8, (%rsi) -; AVX2-NEXT: vmovdqa %ymm9, (%rdx) +; AVX2-NEXT: vmovdqa %ymm10, (%rsi) +; AVX2-NEXT: vmovdqa %ymm8, (%rdx) ; AVX2-NEXT: vmovdqa %ymm0, (%rcx) ; AVX2-NEXT: vmovdqa %ymm1, (%r8) ; AVX2-NEXT: vzeroupper diff --git a/llvm/test/CodeGen/X86/vector-shuffle-256-v32.ll b/llvm/test/CodeGen/X86/vector-shuffle-256-v32.ll index 893dfc2..b1ea67e 100644 --- a/llvm/test/CodeGen/X86/vector-shuffle-256-v32.ll +++ b/llvm/test/CodeGen/X86/vector-shuffle-256-v32.ll @@ -4,8 +4,8 @@ ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX2OR512VL,AVX2,AVX2-FAST,AVX2-FAST-ALL ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX2OR512VL,AVX2,AVX2-FAST,AVX2-FAST-PERLANE ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=ALL,AVX2OR512VL,AVX512VL,AVX512VLBW,AVX512VLBW-SLOW -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX2OR512VL,AVX512VL,AVX512VLBW,AVX512VLBW-FAST,AVX512VLBW-FAST-ALL -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512bw,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX2OR512VL,AVX512VL,AVX512VLBW,AVX512VLBW-FAST,AVX512VLBW-FAST-PERLANE +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX2OR512VL,AVX512VL,AVX512VLBW,AVX512VLBW-FAST +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512bw,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX2OR512VL,AVX512VL,AVX512VLBW,AVX512VLBW-FAST ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512vbmi | FileCheck %s --check-prefixes=ALL,AVX2OR512VL,AVX512VL,AVX512VLVBMI,AVX512VLVBMI-SLOW ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512vbmi,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX2OR512VL,AVX512VL,AVX512VLVBMI,AVX512VLVBMI-FAST-ALL ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512vbmi,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX2OR512VL,AVX512VL,AVX512VLVBMI,AVX512VLVBMI-FAST-PERLANE @@ -3391,89 +3391,32 @@ define <32 x i8> @shuffle_v32i8_42_45_12_13_35_35_60_40_17_22_29_44_33_12_48_51_ ; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0 ; AVX1-NEXT: retq ; -; AVX2-SLOW-LABEL: shuffle_v32i8_42_45_12_13_35_35_60_40_17_22_29_44_33_12_48_51_20_19_52_19_49_54_37_32_48_42_59_07_36_34_36_39: -; AVX2-SLOW: # %bb.0: -; AVX2-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm1[10,13],zero,zero,ymm1[3,3],zero,ymm1[8],zero,zero,zero,ymm1[12,1],zero,zero,zero,zero,zero,ymm1[20],zero,ymm1[17,22],zero,zero,ymm1[16],zero,ymm1[27],zero,zero,zero,zero,zero -; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,0,1] -; AVX2-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,zero,zero,ymm1[12],zero,zero,zero,zero,zero,zero,zero,ymm1[0,3],zero,zero,zero,zero,zero,zero,ymm1[21,16],zero,ymm1[26],zero,zero,ymm1[20,18,20,23] -; AVX2-SLOW-NEXT: vpor %ymm1, %ymm2, %ymm1 -; AVX2-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[u,u,12,13,u,u,u,u,u,u,u,u,u,12,u,u,20,19,u,19,u,u,u,u,u,u,u,u,u,u,u,u] -; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] -; AVX2-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,1,6,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,23,u,u,u,u] -; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1],ymm0[2],ymm2[3,4,5],ymm0[6],ymm2[7] -; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm2 = [255,255,0,0,255,255,255,255,0,0,0,255,255,0,255,255,0,0,255,0,255,255,255,255,255,255,255,0,255,255,255,255] -; AVX2-SLOW-NEXT: vpblendvb %ymm2, %ymm1, %ymm0, %ymm0 -; AVX2-SLOW-NEXT: retq -; -; AVX2-FAST-ALL-LABEL: shuffle_v32i8_42_45_12_13_35_35_60_40_17_22_29_44_33_12_48_51_20_19_52_19_49_54_37_32_48_42_59_07_36_34_36_39: -; AVX2-FAST-ALL: # %bb.0: -; AVX2-FAST-ALL-NEXT: vpshufb {{.*#+}} ymm2 = ymm1[10,13],zero,zero,ymm1[3,3],zero,ymm1[8],zero,zero,zero,ymm1[12,1],zero,zero,zero,zero,zero,ymm1[20],zero,ymm1[17,22],zero,zero,ymm1[16],zero,ymm1[27],zero,zero,zero,zero,zero -; AVX2-FAST-ALL-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,0,1] -; AVX2-FAST-ALL-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,zero,zero,ymm1[12],zero,zero,zero,zero,zero,zero,zero,ymm1[0,3],zero,zero,zero,zero,zero,zero,ymm1[21,16],zero,ymm1[26],zero,zero,ymm1[20,18,20,23] -; AVX2-FAST-ALL-NEXT: vpor %ymm1, %ymm2, %ymm1 -; AVX2-FAST-ALL-NEXT: vmovdqa {{.*#+}} ymm2 = <3,4,5,7,5,4,1,u> -; AVX2-FAST-ALL-NEXT: vpermd %ymm0, %ymm2, %ymm0 -; AVX2-FAST-ALL-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,0,1,u,u,u,u,5,10,13,u,u,0,u,u,16,23,u,23,u,u,u,u,u,u,u,27,u,u,u,u] -; AVX2-FAST-ALL-NEXT: vmovdqa {{.*#+}} ymm2 = [255,255,0,0,255,255,255,255,0,0,0,255,255,0,255,255,0,0,255,0,255,255,255,255,255,255,255,0,255,255,255,255] -; AVX2-FAST-ALL-NEXT: vpblendvb %ymm2, %ymm1, %ymm0, %ymm0 -; AVX2-FAST-ALL-NEXT: retq -; -; AVX2-FAST-PERLANE-LABEL: shuffle_v32i8_42_45_12_13_35_35_60_40_17_22_29_44_33_12_48_51_20_19_52_19_49_54_37_32_48_42_59_07_36_34_36_39: -; AVX2-FAST-PERLANE: # %bb.0: -; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm2 = ymm1[10,13],zero,zero,ymm1[3,3],zero,ymm1[8],zero,zero,zero,ymm1[12,1],zero,zero,zero,zero,zero,ymm1[20],zero,ymm1[17,22],zero,zero,ymm1[16],zero,ymm1[27],zero,zero,zero,zero,zero -; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,0,1] -; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,zero,zero,ymm1[12],zero,zero,zero,zero,zero,zero,zero,ymm1[0,3],zero,zero,zero,zero,zero,zero,ymm1[21,16],zero,ymm1[26],zero,zero,ymm1[20,18,20,23] -; AVX2-FAST-PERLANE-NEXT: vpor %ymm1, %ymm2, %ymm1 -; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[u,u,12,13,u,u,u,u,u,u,u,u,u,12,u,u,20,19,u,19,u,u,u,u,u,u,u,u,u,u,u,u] -; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] -; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,1,6,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,23,u,u,u,u] -; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1],ymm0[2],ymm2[3,4,5],ymm0[6],ymm2[7] -; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm2 = [255,255,0,0,255,255,255,255,0,0,0,255,255,0,255,255,0,0,255,0,255,255,255,255,255,255,255,0,255,255,255,255] -; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm2, %ymm1, %ymm0, %ymm0 -; AVX2-FAST-PERLANE-NEXT: retq -; -; AVX512VLBW-SLOW-LABEL: shuffle_v32i8_42_45_12_13_35_35_60_40_17_22_29_44_33_12_48_51_20_19_52_19_49_54_37_32_48_42_59_07_36_34_36_39: -; AVX512VLBW-SLOW: # %bb.0: -; AVX512VLBW-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[u,u,12,13,u,u,u,u,u,u,u,u,u,12,u,u,20,19,u,19,u,u,u,u,u,u,u,u,u,u,u,u] -; AVX512VLBW-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] -; AVX512VLBW-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,1,6,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,23,u,u,u,u] -; AVX512VLBW-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm0[2],ymm2[3,4,5],ymm0[6],ymm2[7] -; AVX512VLBW-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[10,13],zero,zero,ymm1[3,3],zero,ymm1[8],zero,zero,zero,ymm1[12,1],zero,zero,zero,zero,zero,ymm1[20],zero,ymm1[17,22],zero,zero,ymm1[16],zero,ymm1[27],zero,zero,zero,zero,zero -; AVX512VLBW-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,0,1] -; AVX512VLBW-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,zero,zero,ymm1[12],zero,zero,zero,zero,zero,zero,zero,ymm1[0,3],zero,zero,zero,zero,zero,zero,ymm1[21,16],zero,ymm1[26],zero,zero,ymm1[20,18,20,23] -; AVX512VLBW-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm0 -; AVX512VLBW-SLOW-NEXT: movl $134948620, %eax # imm = 0x80B270C -; AVX512VLBW-SLOW-NEXT: kmovd %eax, %k1 -; AVX512VLBW-SLOW-NEXT: vmovdqu8 %ymm2, %ymm0 {%k1} -; AVX512VLBW-SLOW-NEXT: retq +; AVX2-LABEL: shuffle_v32i8_42_45_12_13_35_35_60_40_17_22_29_44_33_12_48_51_20_19_52_19_49_54_37_32_48_42_59_07_36_34_36_39: +; AVX2: # %bb.0: +; AVX2-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,12,12,13,1,6,13,7,u,u,u,u,u,u,u,u,u,u,u,u,17,22,29,23,20,19,u,19,u,u,u,u] +; AVX2-NEXT: vmovdqa {{.*#+}} ymm2 = <0,u,5,0,6,u,1,u> +; AVX2-NEXT: vpermd %ymm0, %ymm2, %ymm0 +; AVX2-NEXT: vpshufb {{.*#+}} ymm2 = ymm1[10,13],zero,zero,ymm1[3,3],zero,ymm1[8],zero,zero,zero,ymm1[12,1],zero,zero,zero,zero,zero,ymm1[20],zero,ymm1[17,22],zero,zero,ymm1[16],zero,ymm1[27],zero,zero,zero,zero,zero +; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,0,1] +; AVX2-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,zero,zero,ymm1[12],zero,zero,zero,zero,zero,zero,zero,ymm1[0,3],zero,zero,zero,zero,zero,zero,ymm1[21,16],zero,ymm1[26],zero,zero,ymm1[20,18,20,23] +; AVX2-NEXT: vpor %ymm1, %ymm2, %ymm1 +; AVX2-NEXT: vmovdqa {{.*#+}} ymm2 = [255,255,0,0,255,255,255,255,0,0,0,255,255,0,255,255,0,0,255,0,255,255,255,255,255,255,255,0,255,255,255,255] +; AVX2-NEXT: vpblendvb %ymm2, %ymm1, %ymm0, %ymm0 +; AVX2-NEXT: retq ; -; AVX512VLBW-FAST-ALL-LABEL: shuffle_v32i8_42_45_12_13_35_35_60_40_17_22_29_44_33_12_48_51_20_19_52_19_49_54_37_32_48_42_59_07_36_34_36_39: -; AVX512VLBW-FAST-ALL: # %bb.0: -; AVX512VLBW-FAST-ALL-NEXT: vmovdqa {{.*#+}} ymm2 = <3,4,5,7,5,4,1,u> -; AVX512VLBW-FAST-ALL-NEXT: vpermd %ymm0, %ymm2, %ymm2 -; AVX512VLBW-FAST-ALL-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[10,13],zero,zero,ymm1[3,3],zero,ymm1[8],zero,zero,zero,ymm1[12,1],zero,zero,zero,zero,zero,ymm1[20],zero,ymm1[17,22],zero,zero,ymm1[16],zero,ymm1[27],zero,zero,zero,zero,zero -; AVX512VLBW-FAST-ALL-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,0,1] -; AVX512VLBW-FAST-ALL-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,zero,zero,ymm1[12],zero,zero,zero,zero,zero,zero,zero,ymm1[0,3],zero,zero,zero,zero,zero,zero,ymm1[21,16],zero,ymm1[26],zero,zero,ymm1[20,18,20,23] -; AVX512VLBW-FAST-ALL-NEXT: vpor %ymm0, %ymm1, %ymm0 -; AVX512VLBW-FAST-ALL-NEXT: movl $134948620, %eax # imm = 0x80B270C -; AVX512VLBW-FAST-ALL-NEXT: kmovd %eax, %k1 -; AVX512VLBW-FAST-ALL-NEXT: vpshufb {{.*#+}} ymm0 {%k1} = ymm2[u,u,0,1,u,u,u,u,5,10,13,u,u,0,u,u,16,23,u,23,u,u,u,u,u,u,u,27,u,u,u,u] -; AVX512VLBW-FAST-ALL-NEXT: retq -; -; AVX512VLBW-FAST-PERLANE-LABEL: shuffle_v32i8_42_45_12_13_35_35_60_40_17_22_29_44_33_12_48_51_20_19_52_19_49_54_37_32_48_42_59_07_36_34_36_39: -; AVX512VLBW-FAST-PERLANE: # %bb.0: -; AVX512VLBW-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[u,u,12,13,u,u,u,u,u,u,u,u,u,12,u,u,20,19,u,19,u,u,u,u,u,u,u,u,u,u,u,u] -; AVX512VLBW-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] -; AVX512VLBW-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,1,6,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,23,u,u,u,u] -; AVX512VLBW-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm0[2],ymm2[3,4,5],ymm0[6],ymm2[7] -; AVX512VLBW-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[10,13],zero,zero,ymm1[3,3],zero,ymm1[8],zero,zero,zero,ymm1[12,1],zero,zero,zero,zero,zero,ymm1[20],zero,ymm1[17,22],zero,zero,ymm1[16],zero,ymm1[27],zero,zero,zero,zero,zero -; AVX512VLBW-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,0,1] -; AVX512VLBW-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,zero,zero,ymm1[12],zero,zero,zero,zero,zero,zero,zero,ymm1[0,3],zero,zero,zero,zero,zero,zero,ymm1[21,16],zero,ymm1[26],zero,zero,ymm1[20,18,20,23] -; AVX512VLBW-FAST-PERLANE-NEXT: vpor %ymm0, %ymm1, %ymm0 -; AVX512VLBW-FAST-PERLANE-NEXT: movl $134948620, %eax # imm = 0x80B270C -; AVX512VLBW-FAST-PERLANE-NEXT: kmovd %eax, %k1 -; AVX512VLBW-FAST-PERLANE-NEXT: vmovdqu8 %ymm2, %ymm0 {%k1} -; AVX512VLBW-FAST-PERLANE-NEXT: retq +; AVX512VLBW-LABEL: shuffle_v32i8_42_45_12_13_35_35_60_40_17_22_29_44_33_12_48_51_20_19_52_19_49_54_37_32_48_42_59_07_36_34_36_39: +; AVX512VLBW: # %bb.0: +; AVX512VLBW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,12,12,13,1,6,13,7,u,u,u,u,u,u,u,u,u,u,u,u,17,22,29,23,20,19,u,19,u,u,u,u] +; AVX512VLBW-NEXT: vmovdqa {{.*#+}} ymm2 = <0,u,5,0,6,u,1,u> +; AVX512VLBW-NEXT: vpermd %ymm0, %ymm2, %ymm2 +; AVX512VLBW-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[10,13],zero,zero,ymm1[3,3],zero,ymm1[8],zero,zero,zero,ymm1[12,1],zero,zero,zero,zero,zero,ymm1[20],zero,ymm1[17,22],zero,zero,ymm1[16],zero,ymm1[27],zero,zero,zero,zero,zero +; AVX512VLBW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,0,1] +; AVX512VLBW-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,zero,zero,ymm1[12],zero,zero,zero,zero,zero,zero,zero,ymm1[0,3],zero,zero,zero,zero,zero,zero,ymm1[21,16],zero,ymm1[26],zero,zero,ymm1[20,18,20,23] +; AVX512VLBW-NEXT: vpor %ymm0, %ymm1, %ymm0 +; AVX512VLBW-NEXT: movl $134948620, %eax # imm = 0x80B270C +; AVX512VLBW-NEXT: kmovd %eax, %k1 +; AVX512VLBW-NEXT: vmovdqu8 %ymm2, %ymm0 {%k1} +; AVX512VLBW-NEXT: retq ; ; AVX512VLVBMI-LABEL: shuffle_v32i8_42_45_12_13_35_35_60_40_17_22_29_44_33_12_48_51_20_19_52_19_49_54_37_32_48_42_59_07_36_34_36_39: ; AVX512VLVBMI: # %bb.0: @@ -3499,14 +3442,13 @@ define <32 x i8> @shuffle_v32i8_42_45_12_13_35_35_60_40_17_22_29_44_33_12_48_51_ ; ; XOPAVX2-LABEL: shuffle_v32i8_42_45_12_13_35_35_60_40_17_22_29_44_33_12_48_51_20_19_52_19_49_54_37_32_48_42_59_07_36_34_36_39: ; XOPAVX2: # %bb.0: +; XOPAVX2-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,12,12,13,1,6,13,7,u,u,u,u,u,u,u,u,u,u,u,u,17,22,29,23,20,19,u,19,u,u,u,u] +; XOPAVX2-NEXT: vmovdqa {{.*#+}} ymm2 = <0,u,5,0,6,u,1,u> +; XOPAVX2-NEXT: vpermd %ymm0, %ymm2, %ymm0 ; XOPAVX2-NEXT: vpshufb {{.*#+}} ymm2 = ymm1[10,13],zero,zero,ymm1[3,3],zero,ymm1[8],zero,zero,zero,ymm1[12,1],zero,zero,zero,zero,zero,ymm1[20],zero,ymm1[17,22],zero,zero,ymm1[16],zero,ymm1[27],zero,zero,zero,zero,zero ; XOPAVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,0,1] ; XOPAVX2-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,zero,zero,ymm1[12],zero,zero,zero,zero,zero,zero,zero,ymm1[0,3],zero,zero,zero,zero,zero,zero,ymm1[21,16],zero,ymm1[26],zero,zero,ymm1[20,18,20,23] ; XOPAVX2-NEXT: vpor %ymm1, %ymm2, %ymm1 -; XOPAVX2-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[u,u,12,13,u,u,u,u,u,u,u,u,u,12,u,u,20,19,u,19,u,u,u,u,u,u,u,u,u,u,u,u] -; XOPAVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] -; XOPAVX2-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,1,6,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,23,u,u,u,u] -; XOPAVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1],ymm0[2],ymm2[3,4,5],ymm0[6],ymm2[7] ; XOPAVX2-NEXT: vmovdqa {{.*#+}} ymm2 = [255,255,0,0,255,255,255,255,0,0,0,255,255,0,255,255,0,0,255,0,255,255,255,255,255,255,255,0,255,255,255,255] ; XOPAVX2-NEXT: vpblendvb %ymm2, %ymm1, %ymm0, %ymm0 ; XOPAVX2-NEXT: retq @@ -5130,20 +5072,16 @@ define <32 x i8> @PR55066(<32 x i8> %a0) { ; ; AVX2-LABEL: PR55066: ; AVX2: # %bb.0: -; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm0[2,3,2,3] -; AVX2-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,0,4,8,12,u,u,u,u,0,4,8,12,u,u,u,u,16,20,24,28,u,u,u,u,16,20,24,28] -; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1] -; AVX2-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,u,u,u,u,0,4,8,12,u,u,u,u,16,20,24,28,u,u,u,u,16,20,24,28,u,u,u,u] -; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7] +; AVX2-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u,u,u,u,u] +; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm1 = [17179869184,17179869184,17179869184,17179869184] +; AVX2-NEXT: vpermd %ymm0, %ymm1, %ymm0 ; AVX2-NEXT: retq ; ; AVX512VLBW-LABEL: PR55066: ; AVX512VLBW: # %bb.0: -; AVX512VLBW-NEXT: vpermq {{.*#+}} ymm1 = ymm0[2,3,2,3] -; AVX512VLBW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,0,4,8,12,u,u,u,u,0,4,8,12,u,u,u,u,16,20,24,28,u,u,u,u,16,20,24,28] -; AVX512VLBW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1] -; AVX512VLBW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,u,u,u,u,0,4,8,12,u,u,u,u,16,20,24,28,u,u,u,u,16,20,24,28,u,u,u,u] -; AVX512VLBW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7] +; AVX512VLBW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u,u,u,u,u] +; AVX512VLBW-NEXT: vpbroadcastq {{.*#+}} ymm1 = [17179869184,17179869184,17179869184,17179869184] +; AVX512VLBW-NEXT: vpermd %ymm0, %ymm1, %ymm0 ; AVX512VLBW-NEXT: retq ; ; AVX512VLVBMI-LABEL: PR55066: @@ -5161,11 +5099,9 @@ define <32 x i8> @PR55066(<32 x i8> %a0) { ; ; XOPAVX2-LABEL: PR55066: ; XOPAVX2: # %bb.0: -; XOPAVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm0[2,3,2,3] -; XOPAVX2-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,0,4,8,12,u,u,u,u,0,4,8,12,u,u,u,u,16,20,24,28,u,u,u,u,16,20,24,28] -; XOPAVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1] -; XOPAVX2-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,u,u,u,u,0,4,8,12,u,u,u,u,16,20,24,28,u,u,u,u,16,20,24,28,u,u,u,u] -; XOPAVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7] +; XOPAVX2-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u,u,u,u,u] +; XOPAVX2-NEXT: vpbroadcastq {{.*#+}} ymm1 = [17179869184,17179869184,17179869184,17179869184] +; XOPAVX2-NEXT: vpermd %ymm0, %ymm1, %ymm0 ; XOPAVX2-NEXT: retq %shuffle = shufflevector <32 x i8> %a0, <32 x i8> poison, <32 x i32> ret <32 x i8> %shuffle diff --git a/llvm/test/CodeGen/X86/vector-shuffle-512-v64.ll b/llvm/test/CodeGen/X86/vector-shuffle-512-v64.ll index 2b71022..1abc8aa 100644 --- a/llvm/test/CodeGen/X86/vector-shuffle-512-v64.ll +++ b/llvm/test/CodeGen/X86/vector-shuffle-512-v64.ll @@ -210,19 +210,17 @@ define <64 x i8> @shuffle_v64i8_02_03_04_05_06_07_00_01_10_11_12_13_14_15_08_09_ define <64 x i8> @shuffle_v64i8_01_03_02_05_07_06_09_11_10_13_15_14_17_19_18_21_23_22_25_27_26_29_31_30_33_35_34_37_39_38_41_43_42_45_47_46_49_51_50_53_55_54_57_59_58_61_63_62_01_03_02_05_01_03_02_05_01_03_02_05_01_03_02_05(<64 x i8> %a) { ; AVX512F-LABEL: shuffle_v64i8_01_03_02_05_07_06_09_11_10_13_15_14_17_19_18_21_23_22_25_27_26_29_31_30_33_35_34_37_39_38_41_43_42_45_47_46_49_51_50_53_55_54_57_59_58_61_63_62_01_03_02_05_01_03_02_05_01_03_02_05_01_03_02_05: ; AVX512F: # %bb.0: -; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm1 -; AVX512F-NEXT: vpshufb {{.*#+}} xmm2 = xmm1[10,13,15,14,u,u,u,u,u,u,u,u,u,u,u,u] -; AVX512F-NEXT: vperm2i128 {{.*#+}} ymm3 = ymm1[2,3],ymm0[0,1] -; AVX512F-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,1,3,2,5,7,6,9,11,10,13,15,14,17,19,18,21,17,19,18,21,17,19,18,21,17,19,18,21] -; AVX512F-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1,2,3,4,5,6,7] -; AVX512F-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,u,u,u,u,u,u,1,3,2,5,7,6,9,11] -; AVX512F-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm1 -; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm3 -; AVX512F-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u,u,u,u,u,u,u,u,u,u,u,1,3,2,5] -; AVX512F-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[1,3,2,5,7,6,9,11,10,13,15,14,u,u,u,u,23,22,25,27,26,29,31,30,u,u,u,u,u,u,u,u] -; AVX512F-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2],ymm3[3],ymm0[4,5,6,7] -; AVX512F-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7] -; AVX512F-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0 +; AVX512F-NEXT: vmovdqa {{.*#+}} ymm1 = [0,1,2,4,5,6,22,23] +; AVX512F-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[1,3,2,5,7,6,9,11,10,13,15,14,u,u,u,u,17,19,18,21,23,22,25,27,26,29,31,30,u,u,u,u] +; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm3 +; AVX512F-NEXT: vpshufb {{.*#+}} xmm4 = xmm3[u,u,u,u,u,u,u,u,1,3,2,5,7,6,9,11] +; AVX512F-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4 +; AVX512F-NEXT: vpermt2d %zmm4, %zmm1, %zmm2 +; AVX512F-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm3[2,3,4,5,6,7] +; AVX512F-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[10,13,15,14,1,3,2,5,u,u,u,u,u,u,u,u,26,29,31,30,17,19,18,21,23,22,25,27,u,u,u,u] +; AVX512F-NEXT: vmovdqa {{.*#+}} ymm1 = [0,5,6,4,1,1,1,1] +; AVX512F-NEXT: vpermd %ymm0, %ymm1, %ymm0 +; AVX512F-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0 ; AVX512F-NEXT: retq ; ; AVX512BW-LABEL: shuffle_v64i8_01_03_02_05_07_06_09_11_10_13_15_14_17_19_18_21_23_22_25_27_26_29_31_30_33_35_34_37_39_38_41_43_42_45_47_46_49_51_50_53_55_54_57_59_58_61_63_62_01_03_02_05_01_03_02_05_01_03_02_05_01_03_02_05: @@ -234,19 +232,17 @@ define <64 x i8> @shuffle_v64i8_01_03_02_05_07_06_09_11_10_13_15_14_17_19_18_21_ ; ; AVX512DQ-LABEL: shuffle_v64i8_01_03_02_05_07_06_09_11_10_13_15_14_17_19_18_21_23_22_25_27_26_29_31_30_33_35_34_37_39_38_41_43_42_45_47_46_49_51_50_53_55_54_57_59_58_61_63_62_01_03_02_05_01_03_02_05_01_03_02_05_01_03_02_05: ; AVX512DQ: # %bb.0: -; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm1 -; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm2 = xmm1[10,13,15,14,u,u,u,u,u,u,u,u,u,u,u,u] -; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm3 = ymm1[2,3],ymm0[0,1] -; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,1,3,2,5,7,6,9,11,10,13,15,14,17,19,18,21,17,19,18,21,17,19,18,21,17,19,18,21] -; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1,2,3,4,5,6,7] -; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,u,u,u,u,u,u,1,3,2,5,7,6,9,11] -; AVX512DQ-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm1 -; AVX512DQ-NEXT: vextracti128 $1, %ymm0, %xmm3 -; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u,u,u,u,u,u,u,u,u,u,u,u,1,3,2,5] -; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[1,3,2,5,7,6,9,11,10,13,15,14,u,u,u,u,23,22,25,27,26,29,31,30,u,u,u,u,u,u,u,u] -; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2],ymm3[3],ymm0[4,5,6,7] -; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7] -; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0 +; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm1 = [0,1,2,4,5,6,22,23] +; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[1,3,2,5,7,6,9,11,10,13,15,14,u,u,u,u,17,19,18,21,23,22,25,27,26,29,31,30,u,u,u,u] +; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm3 +; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm4 = xmm3[u,u,u,u,u,u,u,u,1,3,2,5,7,6,9,11] +; AVX512DQ-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4 +; AVX512DQ-NEXT: vpermt2d %zmm4, %zmm1, %zmm2 +; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm3[2,3,4,5,6,7] +; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[10,13,15,14,1,3,2,5,u,u,u,u,u,u,u,u,26,29,31,30,17,19,18,21,23,22,25,27,u,u,u,u] +; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm1 = [0,5,6,4,1,1,1,1] +; AVX512DQ-NEXT: vpermd %ymm0, %ymm1, %ymm0 +; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0 ; AVX512DQ-NEXT: retq ; ; AVX512VBMI-LABEL: shuffle_v64i8_01_03_02_05_07_06_09_11_10_13_15_14_17_19_18_21_23_22_25_27_26_29_31_30_33_35_34_37_39_38_41_43_42_45_47_46_49_51_50_53_55_54_57_59_58_61_63_62_01_03_02_05_01_03_02_05_01_03_02_05_01_03_02_05: diff --git a/llvm/test/CodeGen/X86/x86-interleaved-access.ll b/llvm/test/CodeGen/X86/x86-interleaved-access.ll index bae9ab6..db35cdc 100644 --- a/llvm/test/CodeGen/X86/x86-interleaved-access.ll +++ b/llvm/test/CodeGen/X86/x86-interleaved-access.ll @@ -670,93 +670,80 @@ define <32 x i1> @interleaved_load_vf32_i8_stride4(<128 x i8>* %ptr) { ; ; AVX2-LABEL: interleaved_load_vf32_i8_stride4: ; AVX2: # %bb.0: -; AVX2-NEXT: vmovdqa {{.*#+}} xmm0 = -; AVX2-NEXT: vmovdqa 112(%rdi), %xmm9 -; AVX2-NEXT: vpshufb %xmm0, %xmm9, %xmm1 -; AVX2-NEXT: vmovdqa 96(%rdi), %xmm10 -; AVX2-NEXT: vpshufb %xmm0, %xmm10, %xmm3 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1] -; AVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm1 -; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u> -; AVX2-NEXT: vmovdqa 80(%rdi), %xmm12 -; AVX2-NEXT: vpshufb %xmm2, %xmm12, %xmm4 -; AVX2-NEXT: vmovdqa 64(%rdi), %xmm5 -; AVX2-NEXT: vpshufb %xmm2, %xmm5, %xmm6 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm6[0],xmm4[0],xmm6[1],xmm4[1] -; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4 -; AVX2-NEXT: vpblendd {{.*#+}} ymm8 = ymm4[0,1,2,3,4,5],ymm1[6,7] -; AVX2-NEXT: vmovdqa (%rdi), %xmm11 -; AVX2-NEXT: vmovdqa 16(%rdi), %xmm13 -; AVX2-NEXT: vmovdqa 32(%rdi), %xmm6 -; AVX2-NEXT: vmovdqa 48(%rdi), %xmm7 -; AVX2-NEXT: vpshufb %xmm0, %xmm7, %xmm1 -; AVX2-NEXT: vpshufb %xmm0, %xmm6, %xmm0 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] -; AVX2-NEXT: vpshufb %xmm2, %xmm13, %xmm1 -; AVX2-NEXT: vpshufb %xmm2, %xmm11, %xmm2 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1] +; AVX2-NEXT: vmovdqa 64(%rdi), %ymm8 +; AVX2-NEXT: vmovdqa 96(%rdi), %ymm10 +; AVX2-NEXT: vmovdqa (%rdi), %xmm2 +; AVX2-NEXT: vmovdqa 16(%rdi), %xmm3 +; AVX2-NEXT: vmovdqa 32(%rdi), %xmm4 +; AVX2-NEXT: vmovdqa 48(%rdi), %xmm5 +; AVX2-NEXT: vmovdqa {{.*#+}} xmm6 = +; AVX2-NEXT: vpshufb %xmm6, %xmm5, %xmm7 +; AVX2-NEXT: vpshufb %xmm6, %xmm4, %xmm6 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1] +; AVX2-NEXT: vmovdqa {{.*#+}} xmm7 = <0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %xmm7, %xmm3, %xmm0 +; AVX2-NEXT: vpshufb %xmm7, %xmm2, %xmm7 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm7[0],xmm0[0],xmm7[1],xmm0[1] +; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm6[2,3] +; AVX2-NEXT: vmovdqa {{.*#+}} ymm7 = <0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u,0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %ymm7, %ymm10, %ymm9 +; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm6 = [17179869184,17179869184,17179869184,17179869184] +; AVX2-NEXT: vpermd %ymm9, %ymm6, %ymm9 +; AVX2-NEXT: vpshufb %ymm7, %ymm8, %ymm7 +; AVX2-NEXT: vpermd %ymm7, %ymm6, %ymm7 +; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],ymm9[6,7] +; AVX2-NEXT: vpblendd {{.*#+}} ymm9 = ymm0[0,1,2,3],ymm7[4,5,6,7] +; AVX2-NEXT: vmovdqa {{.*#+}} xmm7 = +; AVX2-NEXT: vpshufb %xmm7, %xmm5, %xmm0 +; AVX2-NEXT: vpshufb %xmm7, %xmm4, %xmm7 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm7[0],xmm0[0],xmm7[1],xmm0[1] +; AVX2-NEXT: vmovdqa {{.*#+}} xmm7 = <1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %xmm7, %xmm3, %xmm1 +; AVX2-NEXT: vpshufb %xmm7, %xmm2, %xmm7 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm7[0],xmm1[0],xmm7[1],xmm1[1] ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3] -; AVX2-NEXT: vpblendd {{.*#+}} ymm8 = ymm0[0,1,2,3],ymm8[4,5,6,7] -; AVX2-NEXT: vmovdqa {{.*#+}} xmm1 = -; AVX2-NEXT: vpshufb %xmm1, %xmm9, %xmm2 -; AVX2-NEXT: vpshufb %xmm1, %xmm10, %xmm0 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1] -; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 -; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u> -; AVX2-NEXT: vpshufb %xmm2, %xmm12, %xmm3 -; AVX2-NEXT: vpshufb %xmm2, %xmm5, %xmm4 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1] -; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3 -; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm0[6,7] -; AVX2-NEXT: vpshufb %xmm1, %xmm7, %xmm3 -; AVX2-NEXT: vpshufb %xmm1, %xmm6, %xmm1 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1] -; AVX2-NEXT: vpshufb %xmm2, %xmm13, %xmm3 -; AVX2-NEXT: vpshufb %xmm2, %xmm11, %xmm2 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1] -; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3] -; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7] -; AVX2-NEXT: vpcmpeqb %ymm0, %ymm8, %ymm8 +; AVX2-NEXT: vmovdqa {{.*#+}} ymm1 = <1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %ymm1, %ymm10, %ymm7 +; AVX2-NEXT: vpermd %ymm7, %ymm6, %ymm7 +; AVX2-NEXT: vpshufb %ymm1, %ymm8, %ymm1 +; AVX2-NEXT: vpermd %ymm1, %ymm6, %ymm1 +; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm7[6,7] +; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] +; AVX2-NEXT: vpcmpeqb %ymm0, %ymm9, %ymm9 ; AVX2-NEXT: vmovdqa {{.*#+}} xmm0 = -; AVX2-NEXT: vpshufb %xmm0, %xmm9, %xmm1 -; AVX2-NEXT: vpshufb %xmm0, %xmm10, %xmm2 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1] -; AVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm1 -; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = <2,6,10,14,u,u,u,u,u,u,u,u,u,u,u,u> -; AVX2-NEXT: vpshufb %xmm2, %xmm12, %xmm3 -; AVX2-NEXT: vpshufb %xmm2, %xmm5, %xmm4 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1] -; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm0, %ymm3 -; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7] -; AVX2-NEXT: vpshufb %xmm0, %xmm7, %xmm3 -; AVX2-NEXT: vpshufb %xmm0, %xmm6, %xmm0 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1] -; AVX2-NEXT: vpshufb %xmm2, %xmm13, %xmm3 -; AVX2-NEXT: vpshufb %xmm2, %xmm11, %xmm2 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1] -; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3] +; AVX2-NEXT: vpshufb %xmm0, %xmm5, %xmm1 +; AVX2-NEXT: vpshufb %xmm0, %xmm4, %xmm0 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] +; AVX2-NEXT: vmovdqa {{.*#+}} xmm1 = <2,6,10,14,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %xmm1, %xmm3, %xmm7 +; AVX2-NEXT: vpshufb %xmm1, %xmm2, %xmm1 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm7[0],xmm1[1],xmm7[1] +; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3] +; AVX2-NEXT: vmovdqa {{.*#+}} ymm1 = <2,6,10,14,u,u,u,u,u,u,u,u,u,u,u,u,2,6,10,14,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %ymm1, %ymm10, %ymm7 +; AVX2-NEXT: vpermd %ymm7, %ymm6, %ymm7 +; AVX2-NEXT: vpshufb %ymm1, %ymm8, %ymm1 +; AVX2-NEXT: vpermd %ymm1, %ymm6, %ymm1 +; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm7[6,7] ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7] ; AVX2-NEXT: vmovdqa {{.*#+}} xmm1 = -; AVX2-NEXT: vpshufb %xmm1, %xmm9, %xmm2 -; AVX2-NEXT: vpshufb %xmm1, %xmm10, %xmm3 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1] -; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2 -; AVX2-NEXT: vmovdqa {{.*#+}} xmm3 = <3,7,11,15,u,u,u,u,u,u,u,u,u,u,u,u> -; AVX2-NEXT: vpshufb %xmm3, %xmm12, %xmm4 -; AVX2-NEXT: vpshufb %xmm3, %xmm5, %xmm5 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1] -; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4 -; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5],ymm2[6,7] -; AVX2-NEXT: vpshufb %xmm1, %xmm7, %xmm4 -; AVX2-NEXT: vpshufb %xmm1, %xmm6, %xmm1 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1] -; AVX2-NEXT: vpshufb %xmm3, %xmm13, %xmm4 -; AVX2-NEXT: vpshufb %xmm3, %xmm11, %xmm3 -; AVX2-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1] -; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3] +; AVX2-NEXT: vpshufb %xmm1, %xmm5, %xmm5 +; AVX2-NEXT: vpshufb %xmm1, %xmm4, %xmm1 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1] +; AVX2-NEXT: vmovdqa {{.*#+}} xmm4 = <3,7,11,15,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %xmm4, %xmm3, %xmm3 +; AVX2-NEXT: vpshufb %xmm4, %xmm2, %xmm2 +; AVX2-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1] +; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3] +; AVX2-NEXT: vmovdqa {{.*#+}} ymm2 = <3,7,11,15,u,u,u,u,u,u,u,u,u,u,u,u,3,7,11,15,u,u,u,u,u,u,u,u,u,u,u,u> +; AVX2-NEXT: vpshufb %ymm2, %ymm10, %ymm3 +; AVX2-NEXT: vpermd %ymm3, %ymm6, %ymm3 +; AVX2-NEXT: vpshufb %ymm2, %ymm8, %ymm2 +; AVX2-NEXT: vpermd %ymm2, %ymm6, %ymm2 +; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7] ; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7] ; AVX2-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0 -; AVX2-NEXT: vpxor %ymm0, %ymm8, %ymm0 +; AVX2-NEXT: vpxor %ymm0, %ymm9, %ymm0 ; AVX2-NEXT: vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0 ; AVX2-NEXT: retq ; -- 2.7.4