From 90e121fbe6b892a0fd4b6d1f971677691547d2d8 Mon Sep 17 00:00:00 2001 From: Simon Pilgrim Date: Thu, 27 Jun 2019 11:16:03 +0000 Subject: [PATCH] [X86][AVX] SimplifyDemandedVectorElts - combine PERMPD(x) -> EXTRACTF128(X) If we only use the bottom lane, see if we can simplify this to extract_subvector - which is always at least as quick as PERMPD/PERMQ. llvm-svn: 364518 --- llvm/lib/Target/X86/X86ISelLowering.cpp | 16 ++++++++++++++++ llvm/test/CodeGen/X86/var-permute-128.ll | 20 ++++++++++---------- llvm/test/CodeGen/X86/vector-shuffle-256-v16.ll | 2 +- llvm/test/CodeGen/X86/vector-shuffle-512-v32.ll | 2 +- llvm/test/CodeGen/X86/vector-shuffle-v48.ll | 5 +++-- 5 files changed, 31 insertions(+), 14 deletions(-) diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp index 4bd5a41..b15edb2 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -33979,6 +33979,22 @@ bool X86TargetLowering::SimplifyDemandedVectorEltsForTargetNode( insertSubVector(UndefVec, ExtOp, 0, TLO.DAG, DL, ExtSizeInBits); return TLO.CombineTo(Op, Insert); } + case X86ISD::VPERMI: { + // Simplify PERMPD/PERMQ to extract_subvector. + // TODO: This should be done in shuffle combining. + if (VT == MVT::v4f64 || VT == MVT::v4i64) { + SmallVector Mask; + DecodeVPERMMask(NumElts, Op.getConstantOperandVal(1), Mask); + if (isUndefOrEqual(Mask[0], 2) && isUndefOrEqual(Mask[1], 3)) { + SDLoc DL(Op); + SDValue Ext = extractSubVector(Op.getOperand(0), 2, TLO.DAG, DL, 128); + SDValue UndefVec = TLO.DAG.getUNDEF(VT); + SDValue Insert = insertSubVector(UndefVec, Ext, 0, TLO.DAG, DL, 128); + return TLO.CombineTo(Op, Insert); + } + } + break; + } // Target Shuffles. case X86ISD::PSHUFB: case X86ISD::UNPCKL: diff --git a/llvm/test/CodeGen/X86/var-permute-128.ll b/llvm/test/CodeGen/X86/var-permute-128.ll index 5e25346..3a1214e 100644 --- a/llvm/test/CodeGen/X86/var-permute-128.ll +++ b/llvm/test/CodeGen/X86/var-permute-128.ll @@ -1026,33 +1026,33 @@ define <16 x i8> @var_shuffle_v16i8_from_v32i8_v16i8(<32 x i8> %v, <16 x i8> %in ; ; AVX2-LABEL: var_shuffle_v16i8_from_v32i8_v16i8: ; AVX2: # %bb.0: -; AVX2-NEXT: vpshufb %xmm1, %xmm0, %xmm2 -; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3] +; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm2 +; AVX2-NEXT: vpshufb %xmm1, %xmm2, %xmm2 ; AVX2-NEXT: vpshufb %xmm1, %xmm0, %xmm0 ; AVX2-NEXT: vpcmpgtb {{.*}}(%rip), %xmm1, %xmm1 -; AVX2-NEXT: vpblendvb %xmm1, %xmm0, %xmm2, %xmm0 +; AVX2-NEXT: vpblendvb %xmm1, %xmm2, %xmm0, %xmm0 ; AVX2-NEXT: vzeroupper ; AVX2-NEXT: retq ; ; AVX512-LABEL: var_shuffle_v16i8_from_v32i8_v16i8: ; AVX512: # %bb.0: -; AVX512-NEXT: vpshufb %xmm1, %xmm0, %xmm2 -; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3] +; AVX512-NEXT: vextracti128 $1, %ymm0, %xmm2 +; AVX512-NEXT: vpshufb %xmm1, %xmm2, %xmm2 ; AVX512-NEXT: vpshufb %xmm1, %xmm0, %xmm0 ; AVX512-NEXT: vpcmpgtb {{.*}}(%rip), %xmm1, %xmm1 -; AVX512-NEXT: vpblendvb %xmm1, %xmm0, %xmm2, %xmm0 +; AVX512-NEXT: vpblendvb %xmm1, %xmm2, %xmm0, %xmm0 ; AVX512-NEXT: vzeroupper ; AVX512-NEXT: retq ; ; AVX512VLBW-LABEL: var_shuffle_v16i8_from_v32i8_v16i8: ; AVX512VLBW: # %bb.0: ; AVX512VLBW-NEXT: # kill: def $xmm1 killed $xmm1 def $ymm1 -; AVX512VLBW-NEXT: vpshufb %xmm1, %xmm0, %xmm2 -; AVX512VLBW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3] +; AVX512VLBW-NEXT: vextracti128 $1, %ymm0, %xmm2 +; AVX512VLBW-NEXT: vpshufb %xmm1, %xmm2, %xmm2 ; AVX512VLBW-NEXT: vpshufb %xmm1, %xmm0, %xmm0 ; AVX512VLBW-NEXT: vpcmpgtb {{.*}}(%rip), %ymm1, %k1 -; AVX512VLBW-NEXT: vmovdqu8 %ymm0, %ymm2 {%k1} -; AVX512VLBW-NEXT: vmovdqa %xmm2, %xmm0 +; AVX512VLBW-NEXT: vmovdqu8 %ymm2, %ymm0 {%k1} +; AVX512VLBW-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0 ; AVX512VLBW-NEXT: vzeroupper ; AVX512VLBW-NEXT: retq ; diff --git a/llvm/test/CodeGen/X86/vector-shuffle-256-v16.ll b/llvm/test/CodeGen/X86/vector-shuffle-256-v16.ll index 7ef4fa4..843e06e 100644 --- a/llvm/test/CodeGen/X86/vector-shuffle-256-v16.ll +++ b/llvm/test/CodeGen/X86/vector-shuffle-256-v16.ll @@ -4646,7 +4646,7 @@ define <16 x i16> @PR34369(<16 x i16> %vec, <16 x i16> %mask) { ; ; AVX2-LABEL: PR34369: ; AVX2: # %bb.0: -; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm0[2,3,0,1] +; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm2 ; AVX2-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[8,9,10,11,4,5,10,11,8,9,10,11,4,5,4,5] ; AVX2-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[6,7,0,1,0,1,u,u,10,11,4,5,4,5,u,u,30,31,16,17,28,29,16,17,18,19,20,21,24,25,24,25] ; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm0[0,1,2],ymm2[3],ymm0[4,5,6],ymm2[7],ymm0[8,9,10],ymm2[11],ymm0[12,13,14],ymm2[15] diff --git a/llvm/test/CodeGen/X86/vector-shuffle-512-v32.ll b/llvm/test/CodeGen/X86/vector-shuffle-512-v32.ll index b0f1573..a77099a 100644 --- a/llvm/test/CodeGen/X86/vector-shuffle-512-v32.ll +++ b/llvm/test/CodeGen/X86/vector-shuffle-512-v32.ll @@ -66,7 +66,7 @@ define <32 x i16> @shuffle_v32i16_0f_1f_0e_16_0d_1d_04_1e_0b_1b_0a_1a_09_19_08_1 ; KNL-NEXT: vpermq {{.*#+}} ymm2 = ymm1[2,3,0,1] ; KNL-NEXT: vpblendw {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm2[4,5],ymm1[6],ymm2[7],ymm1[8,9,10,11],ymm2[12,13],ymm1[14],ymm2[15] ; KNL-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,14,15,u,u,12,13,u,u,10,11,u,u,8,9,u,u,22,23,u,u,20,21,u,u,18,19,u,u,u,u] -; KNL-NEXT: vpermq {{.*#+}} ymm4 = ymm0[2,3,0,1] +; KNL-NEXT: vextracti128 $1, %ymm0, %xmm4 ; KNL-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm4[5,6,7],ymm0[8,9,10,11,12],ymm4[13,14,15] ; KNL-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[14,15,u,u,12,13,u,u,10,11,u,u,8,9,u,u,22,23,u,u,20,21,u,u,18,19,u,u,16,17,u,u] ; KNL-NEXT: vpblendw {{.*#+}} ymm1 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7],ymm0[8],ymm1[9],ymm0[10],ymm1[11],ymm0[12],ymm1[13],ymm0[14],ymm1[15] diff --git a/llvm/test/CodeGen/X86/vector-shuffle-v48.ll b/llvm/test/CodeGen/X86/vector-shuffle-v48.ll index aab0690..37a55e7 100644 --- a/llvm/test/CodeGen/X86/vector-shuffle-v48.ll +++ b/llvm/test/CodeGen/X86/vector-shuffle-v48.ll @@ -1,11 +1,12 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py -; RUN: llc -mtriple=x86_64-pc-linux -mattr=+avx2 < %s | FileCheck %s +; RUN: llc < %s -mtriple=x86_64-pc-linux -mattr=+avx2 | FileCheck %s + define <32 x i8> @foo(<48 x i8>* %x0, <16 x i32> %x1, <16 x i32> %x2) { ; CHECK-LABEL: foo: ; CHECK: # %bb.0: ; CHECK-NEXT: vmovdqu 32(%rdi), %xmm0 ; CHECK-NEXT: vmovdqu (%rdi), %ymm1 -; CHECK-NEXT: vpermq {{.*#+}} ymm2 = ymm1[2,3,0,1] +; CHECK-NEXT: vmovdqu 16(%rdi), %xmm2 ; CHECK-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,u,u,u,u,u,u,u,0,2,3,5,6] ; CHECK-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[0,1,3,4,6,7,9,10,12,13,15,u,u,u,u,u,24,25,27,28,30,31,u,u,u,u,u,u,u,u,u,u] ; CHECK-NEXT: vmovdqa {{.*#+}} ymm3 = <255,255,255,255,255,255,255,255,255,255,255,0,0,0,0,0,255,255,255,255,255,255,u,u,u,u,u,u,u,u,u,u> -- 2.7.4