caio/clang-p2996
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[X86] combineCMP - attempt to simplify KSHIFTR mask element extractions when just comparing against zero

Browse Source 
We can just bitcast the pre-shifted mask as an integer and use TEST/BT directly.

This can be extended further to better handle sub-i8 mask cases, but just getting rid of KSHIFTR nodes makes a notable difference.
This commit is contained in:
Simon Pilgrim
2023-08-31 13:13:52 +01:00
parent e8af24736a
commit 239ab16ec1
5 changed files with 67 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
llvm/lib/Target/X86/X86ISelLowering.cpp
View File

@@ -53487,6 +53487,7 @@ static SDValue combineCMP(SDNode *N, SelectionDAG &DAG) {
SDLoc dl(N);
SDValue Op = N->getOperand(0);
EVT VT = Op.getValueType();
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
// If we have a constant logical shift that's only used in a comparison
// against zero turn it into an equivalent AND. This allows turning it into
@@ -53510,12 +53511,42 @@ static SDValue combineCMP(SDNode *N, SelectionDAG &DAG) {
}
}
// If we're extracting from a avx512 bool vector and comparing against zero,
// then try to just bitcast the vector to an integer to use TEST/BT directly.
// TODO: Handle v2i1/v4i1 bool vector cases.
// (and (extract_elt (kshiftr vXi1, C), 0), 1) -> (and (bc vXi1), 1<<C)
if (Op.getOpcode() == ISD::AND && isOneConstant(Op.getOperand(1)) &&
Op.hasOneUse() && onlyZeroFlagUsed(SDValue(N, 0))) {
SDValue Src = Op.getOperand(0);
if (Src.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
isNullConstant(Src.getOperand(1)) &&
Src.getOperand(0).getValueType().getScalarType() == MVT::i1) {
SDValue BoolVec = Src.getOperand(0);
EVT VecVT = BoolVec.getValueType();
unsigned BitWidth = VecVT.getVectorNumElements();
EVT BCVT = EVT::getIntegerVT(*DAG.getContext(), BitWidth);
if (TLI.isTypeLegal(VecVT) && TLI.isTypeLegal(BCVT)) {
unsigned ShAmt = 0;
if (BoolVec.getOpcode() == X86ISD::KSHIFTR &&
BoolVec.getConstantOperandAPInt(1).ult(BitWidth)) {
ShAmt = BoolVec.getConstantOperandVal(1);
BoolVec = BoolVec.getOperand(0);
}
APInt Mask = APInt::getOneBitSet(BitWidth, ShAmt);
Op = DAG.getBitcast(BCVT, BoolVec);
Op = DAG.getNode(ISD::AND, dl, BCVT, Op,
DAG.getConstant(Mask, dl, BCVT));
return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
DAG.getConstant(0, dl, VT));
}
}
}
// Peek through any zero-extend if we're only testing for a zero result.
if (Op.getOpcode() == ISD::ZERO_EXTEND && onlyZeroFlagUsed(SDValue(N, 0))) {
SDValue Src = Op.getOperand(0);
EVT SrcVT = Src.getValueType();
if (SrcVT.getScalarSizeInBits() >= 8 &&
DAG.getTargetLoweringInfo().isTypeLegal(SrcVT))
if (SrcVT.getScalarSizeInBits() >= 8 && TLI.isTypeLegal(SrcVT))
return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Src,
DAG.getConstant(0, dl, SrcVT));
}

12
llvm/test/CodeGen/X86/avx512-insert-extract.ll
View File

@@ -175,9 +175,8 @@ define <16 x i32> @test11(<16 x i32>%a, <16 x i32>%b) nounwind {
; KNL-LABEL: test11:
; KNL: ## %bb.0:
; KNL-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; KNL-NEXT: kshiftrw $4, %k0, %k0
; KNL-NEXT: kmovw %k0, %eax
; KNL-NEXT: testb $1, %al
; KNL-NEXT: testb $16, %al
; KNL-NEXT: je LBB10_2
; KNL-NEXT: ## %bb.1: ## %A
; KNL-NEXT: vmovdqa64 %zmm1, %zmm0
@@ -189,9 +188,8 @@ define <16 x i32> @test11(<16 x i32>%a, <16 x i32>%b) nounwind {
; SKX-LABEL: test11:
; SKX: ## %bb.0:
; SKX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; SKX-NEXT: kshiftrw $4, %k0, %k0
; SKX-NEXT: kmovd %k0, %eax
; SKX-NEXT: testb $1, %al
; SKX-NEXT: testb $16, %al
; SKX-NEXT: je LBB10_2
; SKX-NEXT: ## %bb.1: ## %A
; SKX-NEXT: vmovdqa64 %zmm1, %zmm0
@@ -276,9 +274,8 @@ define i64 @test14(<8 x i64>%a, <8 x i64>%b, i64 %a1, i64 %b1) nounwind {
; KNL: ## %bb.0:
; KNL-NEXT: movq %rdi, %rax
; KNL-NEXT: vpcmpgtq %zmm0, %zmm1, %k0
; KNL-NEXT: kshiftrw $4, %k0, %k0
; KNL-NEXT: kmovw %k0, %ecx
; KNL-NEXT: testb $1, %cl
; KNL-NEXT: testb $16, %cl
; KNL-NEXT: cmoveq %rsi, %rax
; KNL-NEXT: vzeroupper
; KNL-NEXT: retq
@@ -287,9 +284,8 @@ define i64 @test14(<8 x i64>%a, <8 x i64>%b, i64 %a1, i64 %b1) nounwind {
; SKX: ## %bb.0:
; SKX-NEXT: movq %rdi, %rax
; SKX-NEXT: vpcmpgtq %zmm0, %zmm1, %k0
; SKX-NEXT: kshiftrb $4, %k0, %k0
; SKX-NEXT: kmovd %k0, %ecx
; SKX-NEXT: testb $1, %cl
; SKX-NEXT: testb $16, %cl
; SKX-NEXT: cmoveq %rsi, %rax
; SKX-NEXT: vzeroupper
; SKX-NEXT: retq

10
llvm/test/CodeGen/X86/movmsk-cmp.ll
View File

@@ -4350,10 +4350,9 @@ define i32 @PR39665_c_ray(<2 x double> %x, <2 x double> %y) {
; KNL-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
; KNL-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
; KNL-NEXT: vcmpltpd %zmm0, %zmm1, %k0
; KNL-NEXT: kshiftrw $1, %k0, %k1
; KNL-NEXT: kmovw %k1, %eax
; KNL-NEXT: kmovw %k0, %ecx
; KNL-NEXT: testb $1, %al
; KNL-NEXT: kmovw %k0, %eax
; KNL-NEXT: testb $2, %al
; KNL-NEXT: movl $42, %eax
; KNL-NEXT: movl $99, %edx
; KNL-NEXT: cmovel %edx, %eax
@@ -4365,10 +4364,9 @@ define i32 @PR39665_c_ray(<2 x double> %x, <2 x double> %y) {
; SKX-LABEL: PR39665_c_ray:
; SKX: # %bb.0:
; SKX-NEXT: vcmpltpd %xmm0, %xmm1, %k0
; SKX-NEXT: kshiftrb $1, %k0, %k1
; SKX-NEXT: kmovd %k1, %eax
; SKX-NEXT: kmovd %k0, %ecx
; SKX-NEXT: testb $1, %al
; SKX-NEXT: kmovd %k0, %eax
; SKX-NEXT: testb $2, %al
; SKX-NEXT: movl $42, %eax
; SKX-NEXT: movl $99, %edx
; SKX-NEXT: cmovel %edx, %eax

40
llvm/test/CodeGen/X86/pr33349.ll
View File

@@ -11,32 +11,30 @@ target triple = "x86_64-unknown-linux-gnu"
; KNL-NEXT: vpslld $31, %xmm0, %xmm0
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
; KNL-NEXT: kshiftrw $2, %k0, %k1
; KNL-NEXT: kshiftrw $1, %k1, %k2
; KNL-NEXT: kmovw %k1, %eax
; KNL-NEXT: testb $1, %al
; KNL-NEXT: kmovw %k1, %ecx
; KNL-NEXT: testb $1, %cl
; KNL-NEXT: fld1
; KNL-NEXT: fldz
; KNL-NEXT: fld %st(0)
; KNL-NEXT: fcmovne %st(2), %st
; KNL-NEXT: kmovw %k2, %eax
; KNL-NEXT: testb $1, %al
; KNL-NEXT: testb $2, %al
; KNL-NEXT: fld %st(1)
; KNL-NEXT: fcmovne %st(3), %st
; KNL-NEXT: kshiftrw $1, %k0, %k1
; KNL-NEXT: kmovw %k1, %eax
; KNL-NEXT: testb $1, %al
; KNL-NEXT: kmovw %k0, %eax
; KNL-NEXT: kmovw %k0, %ecx
; KNL-NEXT: testb $1, %cl
; KNL-NEXT: fld %st(2)
; KNL-NEXT: fcmovne %st(4), %st
; KNL-NEXT: kmovw %k0, %eax
; KNL-NEXT: testb $1, %al
; KNL-NEXT: testb $2, %al
; KNL-NEXT: fxch %st(3)
; KNL-NEXT: fcmovne %st(4), %st
; KNL-NEXT: fstp %st(4)
; KNL-NEXT: fxch %st(3)
; KNL-NEXT: fstpt (%rdi)
; KNL-NEXT: fxch %st(1)
; KNL-NEXT: fstpt 10(%rdi)
; KNL-NEXT: fxch %st(1)
; KNL-NEXT: fstpt (%rdi)
; KNL-NEXT: fxch %st(1)
; KNL-NEXT: fstpt 30(%rdi)
; KNL-NEXT: fstpt 20(%rdi)
; KNL-NEXT: vzeroupper
@@ -47,32 +45,30 @@ target triple = "x86_64-unknown-linux-gnu"
; SKX-NEXT: vpslld $31, %xmm0, %xmm0
; SKX-NEXT: vpmovd2m %xmm0, %k0
; SKX-NEXT: kshiftrb $2, %k0, %k1
; SKX-NEXT: kshiftrb $1, %k1, %k2
; SKX-NEXT: kmovd %k1, %eax
; SKX-NEXT: testb $1, %al
; SKX-NEXT: kmovd %k1, %ecx
; SKX-NEXT: testb $1, %cl
; SKX-NEXT: fld1
; SKX-NEXT: fldz
; SKX-NEXT: fld %st(0)
; SKX-NEXT: fcmovne %st(2), %st
; SKX-NEXT: kmovd %k2, %eax
; SKX-NEXT: testb $1, %al
; SKX-NEXT: testb $2, %al
; SKX-NEXT: fld %st(1)
; SKX-NEXT: fcmovne %st(3), %st
; SKX-NEXT: kshiftrb $1, %k0, %k1
; SKX-NEXT: kmovd %k1, %eax
; SKX-NEXT: testb $1, %al
; SKX-NEXT: kmovd %k0, %eax
; SKX-NEXT: kmovd %k0, %ecx
; SKX-NEXT: testb $1, %cl
; SKX-NEXT: fld %st(2)
; SKX-NEXT: fcmovne %st(4), %st
; SKX-NEXT: kmovd %k0, %eax
; SKX-NEXT: testb $1, %al
; SKX-NEXT: testb $2, %al
; SKX-NEXT: fxch %st(3)
; SKX-NEXT: fcmovne %st(4), %st
; SKX-NEXT: fstp %st(4)
; SKX-NEXT: fxch %st(3)
; SKX-NEXT: fstpt (%rdi)
; SKX-NEXT: fxch %st(1)
; SKX-NEXT: fstpt 10(%rdi)
; SKX-NEXT: fxch %st(1)
; SKX-NEXT: fstpt (%rdi)
; SKX-NEXT: fxch %st(1)
; SKX-NEXT: fstpt 30(%rdi)
; SKX-NEXT: fstpt 20(%rdi)
; SKX-NEXT: retq

18
llvm/test/CodeGen/X86/pr34177.ll
View File

@@ -49,24 +49,22 @@ define void @test(<4 x i64> %a, <4 x x86_fp80> %b, ptr %c) local_unnamed_addr {
; AVX512VL-LABEL: test:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpcmpeqq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %k0
; AVX512VL-NEXT: kshiftrb $1, %k0, %k1
; AVX512VL-NEXT: kshiftrb $2, %k0, %k2
; AVX512VL-NEXT: kmovd %k0, %eax
; AVX512VL-NEXT: testb $1, %al
; AVX512VL-NEXT: kshiftrb $2, %k0, %k1
; AVX512VL-NEXT: kmovd %k0, %ecx
; AVX512VL-NEXT: testb $2, %cl
; AVX512VL-NEXT: fld1
; AVX512VL-NEXT: fldz
; AVX512VL-NEXT: fld %st(0)
; AVX512VL-NEXT: fcmovne %st(2), %st
; AVX512VL-NEXT: kmovd %k1, %eax
; AVX512VL-NEXT: testb $1, %al
; AVX512VL-NEXT: fld %st(1)
; AVX512VL-NEXT: fcmovne %st(3), %st
; AVX512VL-NEXT: kshiftrb $1, %k2, %k0
; AVX512VL-NEXT: kmovd %k0, %eax
; AVX512VL-NEXT: testb $1, %al
; AVX512VL-NEXT: kmovd %k1, %eax
; AVX512VL-NEXT: kmovd %k1, %ecx
; AVX512VL-NEXT: testb $2, %cl
; AVX512VL-NEXT: fld %st(2)
; AVX512VL-NEXT: fcmovne %st(4), %st
; AVX512VL-NEXT: kmovd %k2, %eax
; AVX512VL-NEXT: testb $1, %al
; AVX512VL-NEXT: fxch %st(3)
; AVX512VL-NEXT: fcmovne %st(4), %st
@@ -81,10 +79,10 @@ define void @test(<4 x i64> %a, <4 x x86_fp80> %b, ptr %c) local_unnamed_addr {
; AVX512VL-NEXT: fstpt 10(%rdi)
; AVX512VL-NEXT: fxch %st(1)
; AVX512VL-NEXT: fadd %st, %st(0)
; AVX512VL-NEXT: fstpt (%rdi)
; AVX512VL-NEXT: fadd %st, %st(0)
; AVX512VL-NEXT: fstpt 20(%rdi)
; AVX512VL-NEXT: fadd %st, %st(0)
; AVX512VL-NEXT: fstpt (%rdi)
; AVX512VL-NEXT: fadd %st, %st(0)
; AVX512VL-NEXT: fstpt 60(%rdi)
; AVX512VL-NEXT: fadd %st, %st(0)
; AVX512VL-NEXT: fstpt 40(%rdi)