Skip to content

Commit 1bb95a3

Browse files
RKSimonRKSimon
committed
[X86] combinePredicateReduction - attempt to fold subvector all_of(icmp_eq()) / any_of(icmp_ne()) to integers
Noticed while working on Issue #59867 and Issue #53419 - there's still more to do here, but for "all vector" comparisons, we should try to cast to a scalar integer for sub-128bit types
1 parent 8bb98b5 commit 1bb95a3

File tree

3 files changed

+43
-85
lines changed

3 files changed

+43
-85
lines changed

‎llvm/lib/Target/X86/X86ISelLowering.cpp

Lines changed: 12 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -44461,6 +44461,18 @@ static SDValue combinePredicateReduction(SDNode *Extract, SelectionDAG &DAG,
4446144461
if ((BinOp == ISD::AND && CC == ISD::CondCode::SETEQ) ||
4446244462
(BinOp == ISD::OR && CC == ISD::CondCode::SETNE)) {
4446344463
EVT VecVT = Match.getOperand(0).getValueType();
44464+
44465+
// If representable as a scalar integer:
44466+
// For all_of(setcc(x,y,eq)) - use (iX)x == (iX)y.
44467+
// For any_of(setcc(x,y,ne)) - use (iX)x != (iX)y.
44468+
EVT IntVT = EVT::getIntegerVT(Ctx, VecVT.getSizeInBits());
44469+
if (TLI.isTypeLegal(IntVT)) {
44470+
SDValue LHS = DAG.getFreeze(Match.getOperand(0));
44471+
SDValue RHS = DAG.getFreeze(Match.getOperand(1));
44472+
return DAG.getSetCC(DL, ExtractVT, DAG.getBitcast(IntVT, LHS),
44473+
DAG.getBitcast(IntVT, RHS), CC);
44474+
}
44475+
4446444476
EVT VecSVT = VecVT.getScalarType();
4446544477
if (VecSVT != MVT::i8 && (VecSVT.getSizeInBits() % 8) == 0) {
4446644478
NumElts *= VecSVT.getSizeInBits() / 8;

‎llvm/test/CodeGen/X86/pr53419.ll

Lines changed: 24 additions & 69 deletions
Original file line numberDiff line numberDiff line change
@@ -13,36 +13,17 @@ declare i1 @llvm.vector.reduce.and.v8i1(<8 x i1>)
1313
; FIXME: All four versions are semantically equivalent and should produce same asm as scalar version.
1414

1515
define i1 @intrinsic_v2i8(ptr align 1 %arg, ptr align 1 %arg1) {
16-
; SSE2-LABEL: intrinsic_v2i8:
17-
; SSE2: # %bb.0: # %bb
18-
; SSE2-NEXT: movzwl (%rsi), %eax
19-
; SSE2-NEXT: movd %eax, %xmm0
20-
; SSE2-NEXT: movzwl (%rdi), %eax
21-
; SSE2-NEXT: movd %eax, %xmm1
22-
; SSE2-NEXT: pcmpeqb %xmm0, %xmm1
23-
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
24-
; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,2,1,4,5,6,7]
25-
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
26-
; SSE2-NEXT: movmskpd %xmm0, %eax
27-
; SSE2-NEXT: cmpb $3, %al
28-
; SSE2-NEXT: sete %al
29-
; SSE2-NEXT: retq
30-
;
31-
; SSE42-LABEL: intrinsic_v2i8:
32-
; SSE42: # %bb.0: # %bb
33-
; SSE42-NEXT: pmovzxbq {{.*#+}} xmm0 = mem[0],zero,zero,zero,zero,zero,zero,zero,mem[1],zero,zero,zero,zero,zero,zero,zero
34-
; SSE42-NEXT: pmovzxbq {{.*#+}} xmm1 = mem[0],zero,zero,zero,zero,zero,zero,zero,mem[1],zero,zero,zero,zero,zero,zero,zero
35-
; SSE42-NEXT: psubq %xmm1, %xmm0
36-
; SSE42-NEXT: ptest %xmm0, %xmm0
37-
; SSE42-NEXT: sete %al
38-
; SSE42-NEXT: retq
16+
; SSE-LABEL: intrinsic_v2i8:
17+
; SSE: # %bb.0: # %bb
18+
; SSE-NEXT: movzwl (%rdi), %eax
19+
; SSE-NEXT: cmpw %ax, (%rsi)
20+
; SSE-NEXT: sete %al
21+
; SSE-NEXT: retq
3922
;
4023
; AVX-LABEL: intrinsic_v2i8:
4124
; AVX: # %bb.0: # %bb
42-
; AVX-NEXT: vpmovzxbq {{.*#+}} xmm0 = mem[0],zero,zero,zero,zero,zero,zero,zero,mem[1],zero,zero,zero,zero,zero,zero,zero
43-
; AVX-NEXT: vpmovzxbq {{.*#+}} xmm1 = mem[0],zero,zero,zero,zero,zero,zero,zero,mem[1],zero,zero,zero,zero,zero,zero,zero
44-
; AVX-NEXT: vpsubq %xmm1, %xmm0, %xmm0
45-
; AVX-NEXT: vptest %xmm0, %xmm0
25+
; AVX-NEXT: movzwl (%rdi), %eax
26+
; AVX-NEXT: cmpw %ax, (%rsi)
4627
; AVX-NEXT: sete %al
4728
; AVX-NEXT: retq
4829
;
@@ -63,10 +44,8 @@ define i1 @intrinsic_v2i8(ptr align 1 %arg, ptr align 1 %arg1) {
6344
; X86: # %bb.0: # %bb
6445
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
6546
; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
66-
; X86-NEXT: vpmovzxbq {{.*#+}} xmm0 = mem[0],zero,zero,zero,zero,zero,zero,zero,mem[1],zero,zero,zero,zero,zero,zero,zero
67-
; X86-NEXT: vpmovzxbq {{.*#+}} xmm1 = mem[0],zero,zero,zero,zero,zero,zero,zero,mem[1],zero,zero,zero,zero,zero,zero,zero
68-
; X86-NEXT: vpsubq %xmm1, %xmm0, %xmm0
69-
; X86-NEXT: vptest %xmm0, %xmm0
47+
; X86-NEXT: movzwl (%eax), %eax
48+
; X86-NEXT: cmpw %ax, (%ecx)
7049
; X86-NEXT: sete %al
7150
; X86-NEXT: retl
7251
bb:
@@ -78,33 +57,17 @@ bb:
7857
}
7958

8059
define i1 @intrinsic_v4i8(ptr align 1 %arg, ptr align 1 %arg1) {
81-
; SSE2-LABEL: intrinsic_v4i8:
82-
; SSE2: # %bb.0: # %bb
83-
; SSE2-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
84-
; SSE2-NEXT: movd {{.*#+}} xmm1 = mem[0],zero,zero,zero
85-
; SSE2-NEXT: pcmpeqb %xmm0, %xmm1
86-
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
87-
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
88-
; SSE2-NEXT: movmskps %xmm0, %eax
89-
; SSE2-NEXT: cmpb $15, %al
90-
; SSE2-NEXT: sete %al
91-
; SSE2-NEXT: retq
92-
;
93-
; SSE42-LABEL: intrinsic_v4i8:
94-
; SSE42: # %bb.0: # %bb
95-
; SSE42-NEXT: pmovzxbd {{.*#+}} xmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
96-
; SSE42-NEXT: pmovzxbd {{.*#+}} xmm1 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
97-
; SSE42-NEXT: psubd %xmm1, %xmm0
98-
; SSE42-NEXT: ptest %xmm0, %xmm0
99-
; SSE42-NEXT: sete %al
100-
; SSE42-NEXT: retq
60+
; SSE-LABEL: intrinsic_v4i8:
61+
; SSE: # %bb.0: # %bb
62+
; SSE-NEXT: movl (%rdi), %eax
63+
; SSE-NEXT: cmpl %eax, (%rsi)
64+
; SSE-NEXT: sete %al
65+
; SSE-NEXT: retq
10166
;
10267
; AVX-LABEL: intrinsic_v4i8:
10368
; AVX: # %bb.0: # %bb
104-
; AVX-NEXT: vpmovzxbd {{.*#+}} xmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
105-
; AVX-NEXT: vpmovzxbd {{.*#+}} xmm1 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
106-
; AVX-NEXT: vpsubd %xmm1, %xmm0, %xmm0
107-
; AVX-NEXT: vptest %xmm0, %xmm0
69+
; AVX-NEXT: movl (%rdi), %eax
70+
; AVX-NEXT: cmpl %eax, (%rsi)
10871
; AVX-NEXT: sete %al
10972
; AVX-NEXT: retq
11073
;
@@ -123,10 +86,8 @@ define i1 @intrinsic_v4i8(ptr align 1 %arg, ptr align 1 %arg1) {
12386
; X86: # %bb.0: # %bb
12487
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
12588
; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
126-
; X86-NEXT: vpmovzxbd {{.*#+}} xmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
127-
; X86-NEXT: vpmovzxbd {{.*#+}} xmm1 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
128-
; X86-NEXT: vpsubd %xmm1, %xmm0, %xmm0
129-
; X86-NEXT: vptest %xmm0, %xmm0
89+
; X86-NEXT: movl (%eax), %eax
90+
; X86-NEXT: cmpl %eax, (%ecx)
13091
; X86-NEXT: sete %al
13192
; X86-NEXT: retl
13293
bb:
@@ -140,21 +101,15 @@ bb:
140101
define i1 @intrinsic_v8i8(ptr align 1 %arg, ptr align 1 %arg1) {
141102
; SSE-LABEL: intrinsic_v8i8:
142103
; SSE: # %bb.0: # %bb
143-
; SSE-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
144-
; SSE-NEXT: movq {{.*#+}} xmm1 = mem[0],zero
145-
; SSE-NEXT: pcmpeqb %xmm0, %xmm1
146-
; SSE-NEXT: pmovmskb %xmm1, %eax
147-
; SSE-NEXT: cmpb $-1, %al
104+
; SSE-NEXT: movq (%rdi), %rax
105+
; SSE-NEXT: cmpq %rax, (%rsi)
148106
; SSE-NEXT: sete %al
149107
; SSE-NEXT: retq
150108
;
151109
; AVX-LABEL: intrinsic_v8i8:
152110
; AVX: # %bb.0: # %bb
153-
; AVX-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
154-
; AVX-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
155-
; AVX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
156-
; AVX-NEXT: vpmovmskb %xmm0, %eax
157-
; AVX-NEXT: cmpb $-1, %al
111+
; AVX-NEXT: movq (%rdi), %rax
112+
; AVX-NEXT: cmpq %rax, (%rsi)
158113
; AVX-NEXT: sete %al
159114
; AVX-NEXT: retq
160115
;

‎llvm/test/CodeGen/X86/vector-reduce-and-bool.ll

Lines changed: 7 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -1567,36 +1567,27 @@ define i8 @icmp0_v8i1(<8 x i8>) {
15671567
; SSE2-LABEL: icmp0_v8i1:
15681568
; SSE2: # %bb.0:
15691569
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1570-
; SSE2-NEXT: pcmpeqd %xmm1, %xmm1
1571-
; SSE2-NEXT: pxor %xmm0, %xmm1
1572-
; SSE2-NEXT: psllw $15, %xmm1
1573-
; SSE2-NEXT: packsswb %xmm1, %xmm1
1574-
; SSE2-NEXT: pmovmskb %xmm1, %eax
1575-
; SSE2-NEXT: cmpb $-1, %al
1570+
; SSE2-NEXT: psllw $15, %xmm0
1571+
; SSE2-NEXT: pmovmskb %xmm0, %eax
1572+
; SSE2-NEXT: testl $43690, %eax # imm = 0xAAAA
15761573
; SSE2-NEXT: sete %al
15771574
; SSE2-NEXT: retq
15781575
;
15791576
; SSE41-LABEL: icmp0_v8i1:
15801577
; SSE41: # %bb.0:
15811578
; SSE41-NEXT: pmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
1582-
; SSE41-NEXT: pcmpeqd %xmm1, %xmm1
1583-
; SSE41-NEXT: pxor %xmm0, %xmm1
1584-
; SSE41-NEXT: psllw $15, %xmm1
1585-
; SSE41-NEXT: packsswb %xmm1, %xmm1
1586-
; SSE41-NEXT: pmovmskb %xmm1, %eax
1587-
; SSE41-NEXT: cmpb $-1, %al
1579+
; SSE41-NEXT: psllw $15, %xmm0
1580+
; SSE41-NEXT: pmovmskb %xmm0, %eax
1581+
; SSE41-NEXT: testl $43690, %eax # imm = 0xAAAA
15881582
; SSE41-NEXT: sete %al
15891583
; SSE41-NEXT: retq
15901584
;
15911585
; AVX-LABEL: icmp0_v8i1:
15921586
; AVX: # %bb.0:
15931587
; AVX-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
1594-
; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
1595-
; AVX-NEXT: vpxor %xmm1, %xmm0, %xmm0
15961588
; AVX-NEXT: vpsllw $15, %xmm0, %xmm0
1597-
; AVX-NEXT: vpacksswb %xmm0, %xmm0, %xmm0
15981589
; AVX-NEXT: vpmovmskb %xmm0, %eax
1599-
; AVX-NEXT: cmpb $-1, %al
1590+
; AVX-NEXT: testl $43690, %eax # imm = 0xAAAA
16001591
; AVX-NEXT: sete %al
16011592
; AVX-NEXT: retq
16021593
;

0 commit comments

Comments
 (0)