From e53d4869a0d862934f8c74e4b74feb2b3430186a Mon Sep 17 00:00:00 2001 From: Simon Pilgrim Date: Tue, 5 May 2020 11:57:25 +0100 Subject: [PATCH] [X86][AVX] combineVectorSignBitsTruncation - avoid complex vXi64->vXi32 PACKSS truncations (PR45794) Unless we're truncating an 'all-bits' result, using PACKSS for vXi64->vXi32 truncation causes problems with later combines as ComputeNumSignBits struggles to see through BITCASTs to smaller types. If we don't use PACKSS in these cases then we fallback to shuffles which are usually just as good. --- llvm/lib/Target/X86/X86ISelLowering.cpp | 7 +++++++ llvm/test/CodeGen/X86/known-signbits-vector.ll | 21 +++------------------ llvm/test/CodeGen/X86/min-legal-vector-width.ll | 9 +++++---- 3 files changed, 15 insertions(+), 22 deletions(-) diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp index 86a1dad..206abcb 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -43516,6 +43516,13 @@ static SDValue combineVectorSignBitsTruncation(SDNode *N, const SDLoc &DL, // Use PACKSS if the input has sign-bits that extend all the way to the // packed/truncated value. e.g. Comparison result, sext_in_reg, etc. unsigned NumSignBits = DAG.ComputeNumSignBits(In); + + // Don't use PACKSS for vXi64 -> vXi32 truncations unless we're dealing with + // a sign splat. ComputeNumSignBits struggles to see through BITCASTs later + // on and combines/simplifications can't then use it. + if (SVT == MVT::i32 && NumSignBits != InSVT.getSizeInBits()) + return SDValue(); + if (NumSignBits > (InSVT.getSizeInBits() - NumPackedSignBits)) return truncateVectorWithPACK(X86ISD::PACKSS, VT, In, DL, DAG, Subtarget); diff --git a/llvm/test/CodeGen/X86/known-signbits-vector.ll b/llvm/test/CodeGen/X86/known-signbits-vector.ll index 675faab..8eb242d 100644 --- a/llvm/test/CodeGen/X86/known-signbits-vector.ll +++ b/llvm/test/CodeGen/X86/known-signbits-vector.ll @@ -114,15 +114,9 @@ define <4 x float> @signbits_ashr_sitofp_1(<4 x i64> %a0) nounwind { ; X86-LABEL: signbits_ashr_sitofp_1: ; X86: # %bb.0: ; X86-NEXT: vextractf128 $1, %ymm0, %xmm1 -; X86-NEXT: vpsrad $31, %xmm1, %xmm2 ; X86-NEXT: vpsrad $16, %xmm1, %xmm1 -; X86-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] -; X86-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3],xmm1[4,5],xmm2[6,7] -; X86-NEXT: vpsrad $31, %xmm0, %xmm2 ; X86-NEXT: vpsrad $16, %xmm0, %xmm0 -; X86-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] -; X86-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7] -; X86-NEXT: vpackssdw %xmm1, %xmm0, %xmm0 +; X86-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,3],xmm1[1,3] ; X86-NEXT: vcvtdq2ps %xmm0, %xmm0 ; X86-NEXT: vzeroupper ; X86-NEXT: retl @@ -130,27 +124,18 @@ define <4 x float> @signbits_ashr_sitofp_1(<4 x i64> %a0) nounwind { ; X64-AVX1-LABEL: signbits_ashr_sitofp_1: ; X64-AVX1: # %bb.0: ; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 -; X64-AVX1-NEXT: vpsrad $31, %xmm1, %xmm2 ; X64-AVX1-NEXT: vpsrad $16, %xmm1, %xmm1 -; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,3,3] -; X64-AVX1-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3],xmm1[4,5],xmm2[6,7] -; X64-AVX1-NEXT: vpsrad $31, %xmm0, %xmm2 ; X64-AVX1-NEXT: vpsrad $16, %xmm0, %xmm0 -; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,3,3] -; X64-AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7] -; X64-AVX1-NEXT: vpackssdw %xmm1, %xmm0, %xmm0 +; X64-AVX1-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,3],xmm1[1,3] ; X64-AVX1-NEXT: vcvtdq2ps %xmm0, %xmm0 ; X64-AVX1-NEXT: vzeroupper ; X64-AVX1-NEXT: retq ; ; X64-AVX2-LABEL: signbits_ashr_sitofp_1: ; X64-AVX2: # %bb.0: -; X64-AVX2-NEXT: vpsrad $31, %ymm0, %ymm1 ; X64-AVX2-NEXT: vpsrad $16, %ymm0, %ymm0 -; X64-AVX2-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[1,1,3,3,5,5,7,7] -; X64-AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7] ; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 -; X64-AVX2-NEXT: vpackssdw %xmm1, %xmm0, %xmm0 +; X64-AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,3],xmm1[1,3] ; X64-AVX2-NEXT: vcvtdq2ps %xmm0, %xmm0 ; X64-AVX2-NEXT: vzeroupper ; X64-AVX2-NEXT: retq diff --git a/llvm/test/CodeGen/X86/min-legal-vector-width.ll b/llvm/test/CodeGen/X86/min-legal-vector-width.ll index 443d57a..d548639 100644 --- a/llvm/test/CodeGen/X86/min-legal-vector-width.ll +++ b/llvm/test/CodeGen/X86/min-legal-vector-width.ll @@ -905,10 +905,11 @@ define <32 x i8> @trunc_v32i16_v32i8_zeroes(<32 x i16>* %x) nounwind "min-legal- define <8 x i32> @trunc_v8i64_v8i32_sign(<8 x i64>* %x) nounwind "min-legal-vector-width"="256" { ; CHECK-LABEL: trunc_v8i64_v8i32_sign: ; CHECK: # %bb.0: -; CHECK-NEXT: vpsraq $48, 32(%rdi), %ymm1 -; CHECK-NEXT: vpsraq $48, (%rdi), %ymm2 -; CHECK-NEXT: vmovdqa {{.*#+}} ymm0 = [0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30] -; CHECK-NEXT: vpermi2w %ymm1, %ymm2, %ymm0 +; CHECK-NEXT: vpsraq $48, 32(%rdi), %ymm0 +; CHECK-NEXT: vpsraq $48, (%rdi), %ymm1 +; CHECK-NEXT: vpmovqd %ymm1, %xmm1 +; CHECK-NEXT: vpmovqd %ymm0, %xmm0 +; CHECK-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0 ; CHECK-NEXT: retq %a = load <8 x i64>, <8 x i64>* %x %b = ashr <8 x i64> %a, -- 2.7.4