Skip to content

Commit d9ffd32

Browse files
committed
[X86] combineCMP - attempt to simplify KSHIFTR mask element extractions when just comparing against zero (REAPPLIED)
We can just bitcast the pre-shifted mask as an integer and use TEST/BT directly. Reapplied with fix for 239ab16 which didn't set the comparison type correctly
1 parent 600b463 commit d9ffd32

File tree

5 files changed

+54
-44
lines changed

5 files changed

+54
-44
lines changed

llvm/lib/Target/X86/X86ISelLowering.cpp

Lines changed: 32 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -53473,6 +53473,7 @@ static SDValue combineCMP(SDNode *N, SelectionDAG &DAG,
5347353473
SDLoc dl(N);
5347453474
SDValue Op = N->getOperand(0);
5347553475
EVT VT = Op.getValueType();
53476+
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5347653477

5347753478
// If we have a constant logical shift that's only used in a comparison
5347853479
// against zero turn it into an equivalent AND. This allows turning it into
@@ -53496,12 +53497,41 @@ static SDValue combineCMP(SDNode *N, SelectionDAG &DAG,
5349653497
}
5349753498
}
5349853499

53500+
// If we're extracting from a avx512 bool vector and comparing against zero,
53501+
// then try to just bitcast the vector to an integer to use TEST/BT directly.
53502+
// (and (extract_elt (kshiftr vXi1, C), 0), 1) -> (and (bc vXi1), 1<<C)
53503+
if (Op.getOpcode() == ISD::AND && isOneConstant(Op.getOperand(1)) &&
53504+
Op.hasOneUse() && onlyZeroFlagUsed(SDValue(N, 0))) {
53505+
SDValue Src = Op.getOperand(0);
53506+
if (Src.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
53507+
isNullConstant(Src.getOperand(1)) &&
53508+
Src.getOperand(0).getValueType().getScalarType() == MVT::i1) {
53509+
SDValue BoolVec = Src.getOperand(0);
53510+
unsigned ShAmt = 0;
53511+
if (BoolVec.getOpcode() == X86ISD::KSHIFTR) {
53512+
ShAmt = BoolVec.getConstantOperandVal(1);
53513+
BoolVec = BoolVec.getOperand(0);
53514+
}
53515+
BoolVec = widenMaskVector(BoolVec, false, Subtarget, DAG, dl);
53516+
EVT VecVT = BoolVec.getValueType();
53517+
unsigned BitWidth = VecVT.getVectorNumElements();
53518+
EVT BCVT = EVT::getIntegerVT(*DAG.getContext(), BitWidth);
53519+
if (TLI.isTypeLegal(VecVT) && TLI.isTypeLegal(BCVT)) {
53520+
APInt Mask = APInt::getOneBitSet(BitWidth, ShAmt);
53521+
Op = DAG.getBitcast(BCVT, BoolVec);
53522+
Op = DAG.getNode(ISD::AND, dl, BCVT, Op,
53523+
DAG.getConstant(Mask, dl, BCVT));
53524+
return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
53525+
DAG.getConstant(0, dl, BCVT));
53526+
}
53527+
}
53528+
}
53529+
5349953530
// Peek through any zero-extend if we're only testing for a zero result.
5350053531
if (Op.getOpcode() == ISD::ZERO_EXTEND && onlyZeroFlagUsed(SDValue(N, 0))) {
5350153532
SDValue Src = Op.getOperand(0);
5350253533
EVT SrcVT = Src.getValueType();
53503-
if (SrcVT.getScalarSizeInBits() >= 8 &&
53504-
DAG.getTargetLoweringInfo().isTypeLegal(SrcVT))
53534+
if (SrcVT.getScalarSizeInBits() >= 8 && TLI.isTypeLegal(SrcVT))
5350553535
return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Src,
5350653536
DAG.getConstant(0, dl, SrcVT));
5350753537
}

llvm/test/CodeGen/X86/avx512-insert-extract.ll

Lines changed: 4 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -175,9 +175,8 @@ define <16 x i32> @test11(<16 x i32>%a, <16 x i32>%b) nounwind {
175175
; KNL-LABEL: test11:
176176
; KNL: ## %bb.0:
177177
; KNL-NEXT: vpcmpltud %zmm1, %zmm0, %k0
178-
; KNL-NEXT: kshiftrw $4, %k0, %k0
179178
; KNL-NEXT: kmovw %k0, %eax
180-
; KNL-NEXT: testb $1, %al
179+
; KNL-NEXT: testb $16, %al
181180
; KNL-NEXT: je LBB10_2
182181
; KNL-NEXT: ## %bb.1: ## %A
183182
; KNL-NEXT: vmovdqa64 %zmm1, %zmm0
@@ -189,9 +188,8 @@ define <16 x i32> @test11(<16 x i32>%a, <16 x i32>%b) nounwind {
189188
; SKX-LABEL: test11:
190189
; SKX: ## %bb.0:
191190
; SKX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
192-
; SKX-NEXT: kshiftrw $4, %k0, %k0
193191
; SKX-NEXT: kmovd %k0, %eax
194-
; SKX-NEXT: testb $1, %al
192+
; SKX-NEXT: testb $16, %al
195193
; SKX-NEXT: je LBB10_2
196194
; SKX-NEXT: ## %bb.1: ## %A
197195
; 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 {
276274
; KNL: ## %bb.0:
277275
; KNL-NEXT: movq %rdi, %rax
278276
; KNL-NEXT: vpcmpgtq %zmm0, %zmm1, %k0
279-
; KNL-NEXT: kshiftrw $4, %k0, %k0
280277
; KNL-NEXT: kmovw %k0, %ecx
281-
; KNL-NEXT: testb $1, %cl
278+
; KNL-NEXT: testb $16, %cl
282279
; KNL-NEXT: cmoveq %rsi, %rax
283280
; KNL-NEXT: vzeroupper
284281
; KNL-NEXT: retq
@@ -287,9 +284,8 @@ define i64 @test14(<8 x i64>%a, <8 x i64>%b, i64 %a1, i64 %b1) nounwind {
287284
; SKX: ## %bb.0:
288285
; SKX-NEXT: movq %rdi, %rax
289286
; SKX-NEXT: vpcmpgtq %zmm0, %zmm1, %k0
290-
; SKX-NEXT: kshiftrb $4, %k0, %k0
291287
; SKX-NEXT: kmovd %k0, %ecx
292-
; SKX-NEXT: testb $1, %cl
288+
; SKX-NEXT: testb $16, %cl
293289
; SKX-NEXT: cmoveq %rsi, %rax
294290
; SKX-NEXT: vzeroupper
295291
; SKX-NEXT: retq

llvm/test/CodeGen/X86/movmsk-cmp.ll

Lines changed: 2 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -4350,10 +4350,8 @@ define i32 @PR39665_c_ray(<2 x double> %x, <2 x double> %y) {
43504350
; KNL-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
43514351
; KNL-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
43524352
; KNL-NEXT: vcmpltpd %zmm0, %zmm1, %k0
4353-
; KNL-NEXT: kshiftrw $1, %k0, %k1
4354-
; KNL-NEXT: kmovw %k1, %eax
43554353
; KNL-NEXT: kmovw %k0, %ecx
4356-
; KNL-NEXT: testb $1, %al
4354+
; KNL-NEXT: testb $2, %cl
43574355
; KNL-NEXT: movl $42, %eax
43584356
; KNL-NEXT: movl $99, %edx
43594357
; KNL-NEXT: cmovel %edx, %eax
@@ -4365,10 +4363,8 @@ define i32 @PR39665_c_ray(<2 x double> %x, <2 x double> %y) {
43654363
; SKX-LABEL: PR39665_c_ray:
43664364
; SKX: # %bb.0:
43674365
; SKX-NEXT: vcmpltpd %xmm0, %xmm1, %k0
4368-
; SKX-NEXT: kshiftrb $1, %k0, %k1
4369-
; SKX-NEXT: kmovd %k1, %eax
43704366
; SKX-NEXT: kmovd %k0, %ecx
4371-
; SKX-NEXT: testb $1, %al
4367+
; SKX-NEXT: testb $2, %cl
43724368
; SKX-NEXT: movl $42, %eax
43734369
; SKX-NEXT: movl $99, %edx
43744370
; SKX-NEXT: cmovel %edx, %eax

llvm/test/CodeGen/X86/pr33349.ll

Lines changed: 10 additions & 18 deletions
Original file line numberDiff line numberDiff line change
@@ -11,32 +11,28 @@ target triple = "x86_64-unknown-linux-gnu"
1111
; KNL-NEXT: vpslld $31, %xmm0, %xmm0
1212
; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
1313
; KNL-NEXT: kshiftrw $2, %k0, %k1
14-
; KNL-NEXT: kshiftrw $1, %k1, %k2
1514
; KNL-NEXT: kmovw %k1, %eax
1615
; KNL-NEXT: testb $1, %al
1716
; KNL-NEXT: fld1
1817
; KNL-NEXT: fldz
1918
; KNL-NEXT: fld %st(0)
2019
; KNL-NEXT: fcmovne %st(2), %st
21-
; KNL-NEXT: kmovw %k2, %eax
22-
; KNL-NEXT: testb $1, %al
20+
; KNL-NEXT: testb $2, %al
2321
; KNL-NEXT: fld %st(1)
2422
; KNL-NEXT: fcmovne %st(3), %st
25-
; KNL-NEXT: kshiftrw $1, %k0, %k1
26-
; KNL-NEXT: kmovw %k1, %eax
23+
; KNL-NEXT: kmovw %k0, %eax
2724
; KNL-NEXT: testb $1, %al
2825
; KNL-NEXT: fld %st(2)
2926
; KNL-NEXT: fcmovne %st(4), %st
30-
; KNL-NEXT: kmovw %k0, %eax
31-
; KNL-NEXT: testb $1, %al
27+
; KNL-NEXT: testb $2, %al
3228
; KNL-NEXT: fxch %st(3)
3329
; KNL-NEXT: fcmovne %st(4), %st
3430
; KNL-NEXT: fstp %st(4)
3531
; KNL-NEXT: fxch %st(3)
36-
; KNL-NEXT: fstpt (%rdi)
37-
; KNL-NEXT: fxch %st(1)
3832
; KNL-NEXT: fstpt 10(%rdi)
3933
; KNL-NEXT: fxch %st(1)
34+
; KNL-NEXT: fstpt (%rdi)
35+
; KNL-NEXT: fxch %st(1)
4036
; KNL-NEXT: fstpt 30(%rdi)
4137
; KNL-NEXT: fstpt 20(%rdi)
4238
; KNL-NEXT: vzeroupper
@@ -47,32 +43,28 @@ target triple = "x86_64-unknown-linux-gnu"
4743
; SKX-NEXT: vpslld $31, %xmm0, %xmm0
4844
; SKX-NEXT: vpmovd2m %xmm0, %k0
4945
; SKX-NEXT: kshiftrb $2, %k0, %k1
50-
; SKX-NEXT: kshiftrb $1, %k1, %k2
5146
; SKX-NEXT: kmovd %k1, %eax
5247
; SKX-NEXT: testb $1, %al
5348
; SKX-NEXT: fld1
5449
; SKX-NEXT: fldz
5550
; SKX-NEXT: fld %st(0)
5651
; SKX-NEXT: fcmovne %st(2), %st
57-
; SKX-NEXT: kmovd %k2, %eax
58-
; SKX-NEXT: testb $1, %al
52+
; SKX-NEXT: testb $2, %al
5953
; SKX-NEXT: fld %st(1)
6054
; SKX-NEXT: fcmovne %st(3), %st
61-
; SKX-NEXT: kshiftrb $1, %k0, %k1
62-
; SKX-NEXT: kmovd %k1, %eax
55+
; SKX-NEXT: kmovd %k0, %eax
6356
; SKX-NEXT: testb $1, %al
6457
; SKX-NEXT: fld %st(2)
6558
; SKX-NEXT: fcmovne %st(4), %st
66-
; SKX-NEXT: kmovd %k0, %eax
67-
; SKX-NEXT: testb $1, %al
59+
; SKX-NEXT: testb $2, %al
6860
; SKX-NEXT: fxch %st(3)
6961
; SKX-NEXT: fcmovne %st(4), %st
7062
; SKX-NEXT: fstp %st(4)
7163
; SKX-NEXT: fxch %st(3)
72-
; SKX-NEXT: fstpt (%rdi)
73-
; SKX-NEXT: fxch %st(1)
7464
; SKX-NEXT: fstpt 10(%rdi)
7565
; SKX-NEXT: fxch %st(1)
66+
; SKX-NEXT: fstpt (%rdi)
67+
; SKX-NEXT: fxch %st(1)
7668
; SKX-NEXT: fstpt 30(%rdi)
7769
; SKX-NEXT: fstpt 20(%rdi)
7870
; SKX-NEXT: retq

llvm/test/CodeGen/X86/pr34177.ll

Lines changed: 6 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -49,24 +49,20 @@ define void @test(<4 x i64> %a, <4 x x86_fp80> %b, ptr %c) local_unnamed_addr {
4949
; AVX512VL-LABEL: test:
5050
; AVX512VL: # %bb.0:
5151
; AVX512VL-NEXT: vpcmpeqq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %k0
52-
; AVX512VL-NEXT: kshiftrb $1, %k0, %k1
53-
; AVX512VL-NEXT: kshiftrb $2, %k0, %k2
52+
; AVX512VL-NEXT: kshiftrb $2, %k0, %k1
5453
; AVX512VL-NEXT: kmovd %k0, %eax
55-
; AVX512VL-NEXT: testb $1, %al
54+
; AVX512VL-NEXT: testb $2, %al
5655
; AVX512VL-NEXT: fld1
5756
; AVX512VL-NEXT: fldz
5857
; AVX512VL-NEXT: fld %st(0)
5958
; AVX512VL-NEXT: fcmovne %st(2), %st
60-
; AVX512VL-NEXT: kmovd %k1, %eax
6159
; AVX512VL-NEXT: testb $1, %al
6260
; AVX512VL-NEXT: fld %st(1)
6361
; AVX512VL-NEXT: fcmovne %st(3), %st
64-
; AVX512VL-NEXT: kshiftrb $1, %k2, %k0
65-
; AVX512VL-NEXT: kmovd %k0, %eax
66-
; AVX512VL-NEXT: testb $1, %al
62+
; AVX512VL-NEXT: kmovd %k1, %eax
63+
; AVX512VL-NEXT: testb $2, %al
6764
; AVX512VL-NEXT: fld %st(2)
6865
; AVX512VL-NEXT: fcmovne %st(4), %st
69-
; AVX512VL-NEXT: kmovd %k2, %eax
7066
; AVX512VL-NEXT: testb $1, %al
7167
; AVX512VL-NEXT: fxch %st(3)
7268
; AVX512VL-NEXT: fcmovne %st(4), %st
@@ -81,10 +77,10 @@ define void @test(<4 x i64> %a, <4 x x86_fp80> %b, ptr %c) local_unnamed_addr {
8177
; AVX512VL-NEXT: fstpt 10(%rdi)
8278
; AVX512VL-NEXT: fxch %st(1)
8379
; AVX512VL-NEXT: fadd %st, %st(0)
84-
; AVX512VL-NEXT: fstpt (%rdi)
85-
; AVX512VL-NEXT: fadd %st, %st(0)
8680
; AVX512VL-NEXT: fstpt 20(%rdi)
8781
; AVX512VL-NEXT: fadd %st, %st(0)
82+
; AVX512VL-NEXT: fstpt (%rdi)
83+
; AVX512VL-NEXT: fadd %st, %st(0)
8884
; AVX512VL-NEXT: fstpt 60(%rdi)
8985
; AVX512VL-NEXT: fadd %st, %st(0)
9086
; AVX512VL-NEXT: fstpt 40(%rdi)

0 commit comments

Comments
 (0)