From 03c4e2663ce4c3431afd075ae72e2357a7d085c4 Mon Sep 17 00:00:00 2001 From: Simon Pilgrim Date: Sat, 27 Apr 2019 20:44:08 +0000 Subject: [PATCH] Revert rL359389: [X86][SSE] Add support for <64 x i1> bool reduction Minor generalization of the existing <32 x i1> pre-AVX2 split code. ........ Causing irregular buildbot failures. llvm-svn: 359391 --- llvm/lib/Target/X86/X86ISelLowering.cpp | 22 ++--- llvm/test/CodeGen/X86/vector-reduce-and-bool.ll | 117 +++++++++++++++++----- llvm/test/CodeGen/X86/vector-reduce-or-bool.ll | 123 ++++++++++++++++++------ 3 files changed, 195 insertions(+), 67 deletions(-) diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp index 2043ab0..420397f 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -34415,31 +34415,29 @@ static SDValue combineHorizontalPredicateResult(SDNode *Extract, SDLoc DL(Extract); EVT MatchVT = Match.getValueType(); unsigned NumElts = MatchVT.getVectorNumElements(); - const TargetLowering &TLI = DAG.getTargetLoweringInfo(); if (ExtractVT == MVT::i1) { // Special case for (pre-legalization) vXi1 reductions. - if (NumElts > 64 || !isPowerOf2_32(NumElts)) + if (NumElts > 32) return SDValue(); - if (TLI.isTypeLegal(MatchVT)) { + if (DAG.getTargetLoweringInfo().isTypeLegal(MatchVT)) { // If this is a legal AVX512 predicate type then we can just bitcast. EVT MovmskVT = EVT::getIntegerVT(*DAG.getContext(), NumElts); Movmsk = DAG.getBitcast(MovmskVT, Match); } else { // Use combineBitcastvxi1 to create the MOVMSK. - unsigned MaxElts = Subtarget.hasInt256() ? 32 : 16; - while (NumElts > MaxElts) { + if (NumElts == 32 && !Subtarget.hasInt256()) { SDValue Lo, Hi; std::tie(Lo, Hi) = DAG.SplitVector(Match, DL); Match = DAG.getNode(BinOp, DL, Lo.getValueType(), Lo, Hi); - NumElts /= 2; + NumElts = 16; } EVT MovmskVT = EVT::getIntegerVT(*DAG.getContext(), NumElts); Movmsk = combineBitcastvxi1(DAG, MovmskVT, Match, DL, Subtarget); } if (!Movmsk) return SDValue(); - Movmsk = DAG.getZExtOrTrunc(Movmsk, DL, NumElts > 32 ? MVT::i64 : MVT::i32); + Movmsk = DAG.getZExtOrTrunc(Movmsk, DL, MVT::i32); } else { // Bail with AVX512VL (which uses predicate registers). if (Subtarget.hasVLX()) @@ -34480,25 +34478,25 @@ static SDValue combineHorizontalPredicateResult(SDNode *Extract, Movmsk = getPMOVMSKB(DL, BitcastLogicOp, DAG, Subtarget); NumElts = MaskSrcVT.getVectorNumElements(); } - assert((NumElts <= 32 || NumElts == 64) && "Not expecting more than 64 elements"); + assert(NumElts <= 32 && "Not expecting more than 32 elements"); SDValue CmpC; ISD::CondCode CondCode; - MVT CmpVT = NumElts == 64 ? MVT::i64 : MVT::i32; if (BinOp == ISD::OR) { // any_of -> MOVMSK != 0 - CmpC = DAG.getConstant(0, DL, CmpVT); + CmpC = DAG.getConstant(0, DL, MVT::i32); CondCode = ISD::CondCode::SETNE; } else { // all_of -> MOVMSK == ((1 << NumElts) - 1) - CmpC = DAG.getConstant((1ULL << NumElts) - 1, DL, CmpVT); + CmpC = DAG.getConstant((1ULL << NumElts) - 1, DL, MVT::i32); CondCode = ISD::CondCode::SETEQ; } // The setcc produces an i8 of 0/1, so extend that to the result width and // negate to get the final 0/-1 mask value. + const TargetLowering &TLI = DAG.getTargetLoweringInfo(); EVT SetccVT = - TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), CmpVT); + TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i32); SDValue Setcc = DAG.getSetCC(DL, SetccVT, Movmsk, CmpC, CondCode); SDValue Zext = DAG.getZExtOrTrunc(Setcc, DL, ExtractVT); SDValue Zero = DAG.getConstant(0, DL, ExtractVT); diff --git a/llvm/test/CodeGen/X86/vector-reduce-and-bool.ll b/llvm/test/CodeGen/X86/vector-reduce-and-bool.ll index 6fe7e31..5f8eb82 100644 --- a/llvm/test/CodeGen/X86/vector-reduce-and-bool.ll +++ b/llvm/test/CodeGen/X86/vector-reduce-and-bool.ll @@ -650,36 +650,77 @@ define i1 @trunc_v32i16_v32i1(<32 x i16>) { } define i1 @trunc_v64i8_v64i1(<64 x i8>) { -; SSE-LABEL: trunc_v64i8_v64i1: -; SSE: # %bb.0: -; SSE-NEXT: pand %xmm3, %xmm1 -; SSE-NEXT: pand %xmm2, %xmm1 -; SSE-NEXT: pand %xmm0, %xmm1 -; SSE-NEXT: psllw $7, %xmm1 -; SSE-NEXT: pmovmskb %xmm1, %eax -; SSE-NEXT: cmpw $-1, %ax -; SSE-NEXT: sete %al -; SSE-NEXT: retq +; SSE2-LABEL: trunc_v64i8_v64i1: +; SSE2: # %bb.0: +; SSE2-NEXT: pand %xmm3, %xmm1 +; SSE2-NEXT: pand %xmm2, %xmm1 +; SSE2-NEXT: pand %xmm0, %xmm1 +; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1] +; SSE2-NEXT: pand %xmm1, %xmm0 +; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3] +; SSE2-NEXT: pand %xmm0, %xmm1 +; SSE2-NEXT: movdqa %xmm1, %xmm0 +; SSE2-NEXT: psrld $16, %xmm0 +; SSE2-NEXT: pand %xmm1, %xmm0 +; SSE2-NEXT: movdqa %xmm0, %xmm1 +; SSE2-NEXT: psrlw $8, %xmm1 +; SSE2-NEXT: pand %xmm0, %xmm1 +; SSE2-NEXT: movd %xmm1, %eax +; SSE2-NEXT: # kill: def $al killed $al killed $eax +; SSE2-NEXT: retq +; +; SSE41-LABEL: trunc_v64i8_v64i1: +; SSE41: # %bb.0: +; SSE41-NEXT: pand %xmm3, %xmm1 +; SSE41-NEXT: pand %xmm2, %xmm1 +; SSE41-NEXT: pand %xmm0, %xmm1 +; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1] +; SSE41-NEXT: pand %xmm1, %xmm0 +; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3] +; SSE41-NEXT: pand %xmm0, %xmm1 +; SSE41-NEXT: movdqa %xmm1, %xmm0 +; SSE41-NEXT: psrld $16, %xmm0 +; SSE41-NEXT: pand %xmm1, %xmm0 +; SSE41-NEXT: movdqa %xmm0, %xmm1 +; SSE41-NEXT: psrlw $8, %xmm1 +; SSE41-NEXT: pand %xmm0, %xmm1 +; SSE41-NEXT: pextrb $0, %xmm1, %eax +; SSE41-NEXT: # kill: def $al killed $al killed $eax +; SSE41-NEXT: retq ; ; AVX1-LABEL: trunc_v64i8_v64i1: ; AVX1: # %bb.0: ; AVX1-NEXT: vandps %ymm1, %ymm0, %ymm0 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 ; AVX1-NEXT: vandps %xmm1, %xmm0, %xmm0 -; AVX1-NEXT: vpsllw $7, %xmm0, %xmm0 -; AVX1-NEXT: vpmovmskb %xmm0, %eax -; AVX1-NEXT: cmpw $-1, %ax -; AVX1-NEXT: sete %al +; AVX1-NEXT: vpermilps {{.*#+}} xmm1 = xmm0[2,3,0,1] +; AVX1-NEXT: vandps %xmm1, %xmm0, %xmm0 +; AVX1-NEXT: vpermilps {{.*#+}} xmm1 = xmm0[1,1,2,3] +; AVX1-NEXT: vandps %xmm1, %xmm0, %xmm0 +; AVX1-NEXT: vpsrld $16, %xmm0, %xmm1 +; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm0 +; AVX1-NEXT: vpsrlw $8, %xmm0, %xmm1 +; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm0 +; AVX1-NEXT: vpextrb $0, %xmm0, %eax +; AVX1-NEXT: # kill: def $al killed $al killed $eax ; AVX1-NEXT: vzeroupper ; AVX1-NEXT: retq ; ; AVX2-LABEL: trunc_v64i8_v64i1: ; AVX2: # %bb.0: ; AVX2-NEXT: vpand %ymm1, %ymm0, %ymm0 -; AVX2-NEXT: vpsllw $7, %ymm0, %ymm0 -; AVX2-NEXT: vpmovmskb %ymm0, %eax -; AVX2-NEXT: cmpl $-1, %eax -; AVX2-NEXT: sete %al +; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 +; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm0 +; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1] +; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm0 +; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3] +; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm0 +; AVX2-NEXT: vpsrld $16, %xmm0, %xmm1 +; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm0 +; AVX2-NEXT: vpsrlw $8, %xmm0, %xmm1 +; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm0 +; AVX2-NEXT: vpextrb $0, %xmm0, %eax +; AVX2-NEXT: # kill: def $al killed $al killed $eax ; AVX2-NEXT: vzeroupper ; AVX2-NEXT: retq ; @@ -687,8 +728,20 @@ define i1 @trunc_v64i8_v64i1(<64 x i8>) { ; AVX512: # %bb.0: ; AVX512-NEXT: vpsllw $7, %zmm0, %zmm0 ; AVX512-NEXT: vpmovb2m %zmm0, %k0 -; AVX512-NEXT: kortestq %k0, %k0 -; AVX512-NEXT: sete %al +; AVX512-NEXT: kshiftrq $32, %k0, %k1 +; AVX512-NEXT: kandq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $16, %k0, %k1 +; AVX512-NEXT: kandq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $8, %k0, %k1 +; AVX512-NEXT: kandq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $4, %k0, %k1 +; AVX512-NEXT: kandq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $2, %k0, %k1 +; AVX512-NEXT: kandq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $1, %k0, %k1 +; AVX512-NEXT: kandq %k1, %k0, %k0 +; AVX512-NEXT: kmovd %k0, %eax +; AVX512-NEXT: # kill: def $al killed $al killed $eax ; AVX512-NEXT: vzeroupper ; AVX512-NEXT: retq %a = trunc <64 x i8> %0 to <64 x i1> @@ -1344,8 +1397,9 @@ define i1 @icmp_v64i8_v64i1(<64 x i8>) { ; SSE-NEXT: por %xmm0, %xmm1 ; SSE-NEXT: pcmpeqb %xmm3, %xmm1 ; SSE-NEXT: pmovmskb %xmm1, %eax -; SSE-NEXT: cmpw $-1, %ax +; SSE-NEXT: cmpl $65535, %eax # imm = 0xFFFF ; SSE-NEXT: sete %al +; SSE-NEXT: negb %al ; SSE-NEXT: retq ; ; AVX1-LABEL: icmp_v64i8_v64i1: @@ -1356,10 +1410,10 @@ define i1 @icmp_v64i8_v64i1(<64 x i8>) { ; AVX1-NEXT: vpcmpeqb %xmm2, %xmm1, %xmm1 ; AVX1-NEXT: vpcmpeqb %xmm2, %xmm0, %xmm0 ; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm0 -; AVX1-NEXT: vpsllw $7, %xmm0, %xmm0 ; AVX1-NEXT: vpmovmskb %xmm0, %eax -; AVX1-NEXT: cmpw $-1, %ax +; AVX1-NEXT: cmpl $65535, %eax # imm = 0xFFFF ; AVX1-NEXT: sete %al +; AVX1-NEXT: negb %al ; AVX1-NEXT: vzeroupper ; AVX1-NEXT: retq ; @@ -1371,14 +1425,27 @@ define i1 @icmp_v64i8_v64i1(<64 x i8>) { ; AVX2-NEXT: vpmovmskb %ymm0, %eax ; AVX2-NEXT: cmpl $-1, %eax ; AVX2-NEXT: sete %al +; AVX2-NEXT: negb %al ; AVX2-NEXT: vzeroupper ; AVX2-NEXT: retq ; ; AVX512-LABEL: icmp_v64i8_v64i1: ; AVX512: # %bb.0: ; AVX512-NEXT: vptestnmb %zmm0, %zmm0, %k0 -; AVX512-NEXT: kortestq %k0, %k0 -; AVX512-NEXT: sete %al +; AVX512-NEXT: kshiftrq $32, %k0, %k1 +; AVX512-NEXT: kandq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $16, %k0, %k1 +; AVX512-NEXT: kandq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $8, %k0, %k1 +; AVX512-NEXT: kandq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $4, %k0, %k1 +; AVX512-NEXT: kandq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $2, %k0, %k1 +; AVX512-NEXT: kandq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $1, %k0, %k1 +; AVX512-NEXT: kandq %k1, %k0, %k0 +; AVX512-NEXT: kmovd %k0, %eax +; AVX512-NEXT: # kill: def $al killed $al killed $eax ; AVX512-NEXT: vzeroupper ; AVX512-NEXT: retq %a = icmp eq <64 x i8> %0, zeroinitializer diff --git a/llvm/test/CodeGen/X86/vector-reduce-or-bool.ll b/llvm/test/CodeGen/X86/vector-reduce-or-bool.ll index 8926d53..f501a8b 100644 --- a/llvm/test/CodeGen/X86/vector-reduce-or-bool.ll +++ b/llvm/test/CodeGen/X86/vector-reduce-or-bool.ll @@ -644,36 +644,77 @@ define i1 @trunc_v32i16_v32i1(<32 x i16>) { } define i1 @trunc_v64i8_v64i1(<64 x i8>) { -; SSE-LABEL: trunc_v64i8_v64i1: -; SSE: # %bb.0: -; SSE-NEXT: por %xmm3, %xmm1 -; SSE-NEXT: por %xmm2, %xmm1 -; SSE-NEXT: por %xmm0, %xmm1 -; SSE-NEXT: psllw $7, %xmm1 -; SSE-NEXT: pmovmskb %xmm1, %eax -; SSE-NEXT: testw %ax, %ax -; SSE-NEXT: setne %al -; SSE-NEXT: retq +; SSE2-LABEL: trunc_v64i8_v64i1: +; SSE2: # %bb.0: +; SSE2-NEXT: por %xmm3, %xmm1 +; SSE2-NEXT: por %xmm2, %xmm1 +; SSE2-NEXT: por %xmm0, %xmm1 +; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1] +; SSE2-NEXT: por %xmm1, %xmm0 +; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3] +; SSE2-NEXT: por %xmm0, %xmm1 +; SSE2-NEXT: movdqa %xmm1, %xmm0 +; SSE2-NEXT: psrld $16, %xmm0 +; SSE2-NEXT: por %xmm1, %xmm0 +; SSE2-NEXT: movdqa %xmm0, %xmm1 +; SSE2-NEXT: psrlw $8, %xmm1 +; SSE2-NEXT: por %xmm0, %xmm1 +; SSE2-NEXT: movd %xmm1, %eax +; SSE2-NEXT: # kill: def $al killed $al killed $eax +; SSE2-NEXT: retq +; +; SSE41-LABEL: trunc_v64i8_v64i1: +; SSE41: # %bb.0: +; SSE41-NEXT: por %xmm3, %xmm1 +; SSE41-NEXT: por %xmm2, %xmm1 +; SSE41-NEXT: por %xmm0, %xmm1 +; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1] +; SSE41-NEXT: por %xmm1, %xmm0 +; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3] +; SSE41-NEXT: por %xmm0, %xmm1 +; SSE41-NEXT: movdqa %xmm1, %xmm0 +; SSE41-NEXT: psrld $16, %xmm0 +; SSE41-NEXT: por %xmm1, %xmm0 +; SSE41-NEXT: movdqa %xmm0, %xmm1 +; SSE41-NEXT: psrlw $8, %xmm1 +; SSE41-NEXT: por %xmm0, %xmm1 +; SSE41-NEXT: pextrb $0, %xmm1, %eax +; SSE41-NEXT: # kill: def $al killed $al killed $eax +; SSE41-NEXT: retq ; ; AVX1-LABEL: trunc_v64i8_v64i1: ; AVX1: # %bb.0: ; AVX1-NEXT: vorps %ymm1, %ymm0, %ymm0 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1 ; AVX1-NEXT: vorps %xmm1, %xmm0, %xmm0 -; AVX1-NEXT: vpsllw $7, %xmm0, %xmm0 -; AVX1-NEXT: vpmovmskb %xmm0, %eax -; AVX1-NEXT: testw %ax, %ax -; AVX1-NEXT: setne %al +; AVX1-NEXT: vpermilps {{.*#+}} xmm1 = xmm0[2,3,0,1] +; AVX1-NEXT: vorps %xmm1, %xmm0, %xmm0 +; AVX1-NEXT: vpermilps {{.*#+}} xmm1 = xmm0[1,1,2,3] +; AVX1-NEXT: vorps %xmm1, %xmm0, %xmm0 +; AVX1-NEXT: vpsrld $16, %xmm0, %xmm1 +; AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0 +; AVX1-NEXT: vpsrlw $8, %xmm0, %xmm1 +; AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0 +; AVX1-NEXT: vpextrb $0, %xmm0, %eax +; AVX1-NEXT: # kill: def $al killed $al killed $eax ; AVX1-NEXT: vzeroupper ; AVX1-NEXT: retq ; ; AVX2-LABEL: trunc_v64i8_v64i1: ; AVX2: # %bb.0: ; AVX2-NEXT: vpor %ymm1, %ymm0, %ymm0 -; AVX2-NEXT: vpsllw $7, %ymm0, %ymm0 -; AVX2-NEXT: vpmovmskb %ymm0, %eax -; AVX2-NEXT: testl %eax, %eax -; AVX2-NEXT: setne %al +; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 +; AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0 +; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1] +; AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0 +; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3] +; AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0 +; AVX2-NEXT: vpsrld $16, %xmm0, %xmm1 +; AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0 +; AVX2-NEXT: vpsrlw $8, %xmm0, %xmm1 +; AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0 +; AVX2-NEXT: vpextrb $0, %xmm0, %eax +; AVX2-NEXT: # kill: def $al killed $al killed $eax ; AVX2-NEXT: vzeroupper ; AVX2-NEXT: retq ; @@ -681,8 +722,20 @@ define i1 @trunc_v64i8_v64i1(<64 x i8>) { ; AVX512: # %bb.0: ; AVX512-NEXT: vpsllw $7, %zmm0, %zmm0 ; AVX512-NEXT: vpmovb2m %zmm0, %k0 -; AVX512-NEXT: kortestq %k0, %k0 -; AVX512-NEXT: setne %al +; AVX512-NEXT: kshiftrq $32, %k0, %k1 +; AVX512-NEXT: korq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $16, %k0, %k1 +; AVX512-NEXT: korq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $8, %k0, %k1 +; AVX512-NEXT: korq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $4, %k0, %k1 +; AVX512-NEXT: korq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $2, %k0, %k1 +; AVX512-NEXT: korq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $1, %k0, %k1 +; AVX512-NEXT: korq %k1, %k0, %k0 +; AVX512-NEXT: kmovd %k0, %eax +; AVX512-NEXT: # kill: def $al killed $al killed $eax ; AVX512-NEXT: vzeroupper ; AVX512-NEXT: retq %a = trunc <64 x i8> %0 to <64 x i1> @@ -1335,10 +1388,9 @@ define i1 @icmp_v64i8_v64i1(<64 x i8>) { ; SSE-NEXT: por %xmm3, %xmm1 ; SSE-NEXT: por %xmm2, %xmm1 ; SSE-NEXT: por %xmm0, %xmm1 -; SSE-NEXT: psllw $7, %xmm1 ; SSE-NEXT: pmovmskb %xmm1, %eax -; SSE-NEXT: testw %ax, %ax -; SSE-NEXT: setne %al +; SSE-NEXT: negl %eax +; SSE-NEXT: sbbb %al, %al ; SSE-NEXT: retq ; ; AVX1-LABEL: icmp_v64i8_v64i1: @@ -1353,10 +1405,9 @@ define i1 @icmp_v64i8_v64i1(<64 x i8>) { ; AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX1-NEXT: vpor %xmm0, %xmm3, %xmm0 ; AVX1-NEXT: vpor %xmm0, %xmm4, %xmm0 -; AVX1-NEXT: vpsllw $7, %xmm0, %xmm0 ; AVX1-NEXT: vpmovmskb %xmm0, %eax -; AVX1-NEXT: testw %ax, %ax -; AVX1-NEXT: setne %al +; AVX1-NEXT: negl %eax +; AVX1-NEXT: sbbb %al, %al ; AVX1-NEXT: vzeroupper ; AVX1-NEXT: retq ; @@ -1367,16 +1418,28 @@ define i1 @icmp_v64i8_v64i1(<64 x i8>) { ; AVX2-NEXT: vpcmpeqb %ymm2, %ymm0, %ymm0 ; AVX2-NEXT: vpor %ymm1, %ymm0, %ymm0 ; AVX2-NEXT: vpmovmskb %ymm0, %eax -; AVX2-NEXT: testl %eax, %eax -; AVX2-NEXT: setne %al +; AVX2-NEXT: negl %eax +; AVX2-NEXT: sbbb %al, %al ; AVX2-NEXT: vzeroupper ; AVX2-NEXT: retq ; ; AVX512-LABEL: icmp_v64i8_v64i1: ; AVX512: # %bb.0: ; AVX512-NEXT: vptestnmb %zmm0, %zmm0, %k0 -; AVX512-NEXT: kortestq %k0, %k0 -; AVX512-NEXT: setne %al +; AVX512-NEXT: kshiftrq $32, %k0, %k1 +; AVX512-NEXT: korq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $16, %k0, %k1 +; AVX512-NEXT: korq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $8, %k0, %k1 +; AVX512-NEXT: korq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $4, %k0, %k1 +; AVX512-NEXT: korq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $2, %k0, %k1 +; AVX512-NEXT: korq %k1, %k0, %k0 +; AVX512-NEXT: kshiftrq $1, %k0, %k1 +; AVX512-NEXT: korq %k1, %k0, %k0 +; AVX512-NEXT: kmovd %k0, %eax +; AVX512-NEXT: # kill: def $al killed $al killed $eax ; AVX512-NEXT: vzeroupper ; AVX512-NEXT: retq %a = icmp eq <64 x i8> %0, zeroinitializer -- 2.7.4