Skip to content

Commit a378d4f

Browse files
RKSimonRKSimon
committed
[CostModel] improveShuffleKindFromMask - recognise a SK_PermuteSingleSrc incorrectly tagged as SK_PermuteTwoSrc
If a SK_PermuteTwoSrc shuffle kind's mask only references the first operand, then treat this as SK_PermuteSingleSrc Part of #145335
1 parent 576085c commit a378d4f

File tree

3 files changed

+23
-33
lines changed

3 files changed

+23
-33
lines changed

‎llvm/include/llvm/CodeGen/BasicTTIImpl.h

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1122,6 +1122,9 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
11221122
break;
11231123
}
11241124
case TTI::SK_PermuteTwoSrc: {
1125+
if (all_of(Mask, [NumSrcElts](int M) { return M < NumSrcElts; }))
1126+
return improveShuffleKindFromMask(TTI::SK_PermuteSingleSrc, Mask, SrcTy,
1127+
Index, SubTy);
11251128
int NumSubElts;
11261129
if (Mask.size() > 2 && ShuffleVectorInst::isInsertSubvectorMask(
11271130
Mask, NumSrcElts, NumSubElts, Index)) {

‎llvm/test/Transforms/SLPVectorizer/AArch64/InstructionsState-is-invalid-0.ll

Lines changed: 10 additions & 13 deletions
Original file line numberDiff line numberDiff line change
@@ -7,20 +7,17 @@ target triple = "aarch64-unknown-linux-gnu"
77
define void @foo(ptr %0) {
88
; CHECK-LABEL: @foo(
99
; CHECK-NEXT: vector.scevcheck:
10-
; CHECK-NEXT: [[SCEVGEP:%.*]] = getelementptr i8, ptr [[TMP0:%.*]], i64 4
11-
; CHECK-NEXT: [[SCEVGEP3:%.*]] = getelementptr i8, ptr null, i64 4
12-
; CHECK-NEXT: [[TMP1:%.*]] = insertelement <4 x ptr> poison, ptr [[TMP0]], i32 1
13-
; CHECK-NEXT: [[TMP2:%.*]] = insertelement <4 x ptr> [[TMP1]], ptr [[SCEVGEP]], i32 0
14-
; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <4 x ptr> [[TMP2]], <4 x ptr> poison, <4 x i32> <i32 0, i32 0, i32 0, i32 1>
15-
; CHECK-NEXT: [[TMP4:%.*]] = icmp ult <4 x ptr> [[TMP3]], zeroinitializer
16-
; CHECK-NEXT: [[TMP5:%.*]] = and <4 x i1> [[TMP4]], zeroinitializer
10+
; CHECK-NEXT: [[TMP1:%.*]] = insertelement <2 x ptr> <ptr poison, ptr null>, ptr [[TMP0:%.*]], i32 0
11+
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i8, <2 x ptr> [[TMP1]], <2 x i64> splat (i64 4)
12+
; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <2 x ptr> [[TMP2]], <2 x ptr> poison, <4 x i32> <i32 0, i32 0, i32 0, i32 1>
1713
; CHECK-NEXT: [[TMP6:%.*]] = insertelement <4 x ptr> poison, ptr [[TMP0]], i32 0
18-
; CHECK-NEXT: [[TMP7:%.*]] = insertelement <4 x ptr> [[TMP6]], ptr [[SCEVGEP3]], i32 1
19-
; CHECK-NEXT: [[TMP8:%.*]] = shufflevector <4 x ptr> [[TMP7]], <4 x ptr> poison, <4 x i32> <i32 0, i32 0, i32 0, i32 1>
20-
; CHECK-NEXT: [[TMP9:%.*]] = icmp ult <4 x ptr> [[TMP8]], zeroinitializer
21-
; CHECK-NEXT: [[TMP10:%.*]] = and <4 x i1> [[TMP9]], zeroinitializer
22-
; CHECK-NEXT: [[RDX_OP:%.*]] = or <4 x i1> [[TMP5]], [[TMP10]]
23-
; CHECK-NEXT: [[OP_RDX:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[RDX_OP]])
14+
; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <4 x ptr> [[TMP6]], <4 x ptr> poison, <4 x i32> zeroinitializer
15+
; CHECK-NEXT: [[TMP11:%.*]] = call <8 x ptr> @llvm.vector.insert.v8p0.v4p0(<8 x ptr> poison, <4 x ptr> [[TMP3]], i64 0)
16+
; CHECK-NEXT: [[TMP7:%.*]] = call <8 x ptr> @llvm.vector.insert.v8p0.v4p0(<8 x ptr> [[TMP11]], <4 x ptr> [[TMP5]], i64 4)
17+
; CHECK-NEXT: [[TMP8:%.*]] = shufflevector <8 x ptr> [[TMP7]], <8 x ptr> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 4, i32 5, i32 6, i32 7, i32 3>
18+
; CHECK-NEXT: [[TMP9:%.*]] = icmp ult <8 x ptr> [[TMP8]], zeroinitializer
19+
; CHECK-NEXT: [[TMP10:%.*]] = and <8 x i1> [[TMP9]], zeroinitializer
20+
; CHECK-NEXT: [[OP_RDX:%.*]] = call i1 @llvm.vector.reduce.or.v8i1(<8 x i1> [[TMP10]])
2421
; CHECK-NEXT: br i1 [[OP_RDX]], label [[DOTLR_PH:%.*]], label [[VECTOR_PH:%.*]]
2522
; CHECK: vector.ph:
2623
; CHECK-NEXT: ret void

‎llvm/test/Transforms/VectorCombine/X86/permute-of-binops.ll

Lines changed: 10 additions & 20 deletions
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,6 @@
11
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
2-
; RUN: opt < %s -passes=vector-combine -S -mtriple=x86_64-- -mattr=sse2 | FileCheck %s --check-prefixes=CHECK,SSE
3-
; RUN: opt < %s -passes=vector-combine -S -mtriple=x86_64-- -mattr=avx2 | FileCheck %s --check-prefixes=CHECK,AVX
2+
; RUN: opt < %s -passes=vector-combine -S -mtriple=x86_64-- -mattr=sse2 | FileCheck %s
3+
; RUN: opt < %s -passes=vector-combine -S -mtriple=x86_64-- -mattr=avx2 | FileCheck %s
44

55
; Fold "shuffle (binop (shuffle, shuffle)), undef" --> "binop (shuffle), (shuffle)"
66

@@ -83,25 +83,15 @@ define <4 x double> @fadd_v4f64_multiuse_op(<4 x double> %a, <4 x double> %b) {
8383
ret <4 x double> %post
8484
}
8585

86-
; Negative test - multiple use of inner shuffle (only fold if the moved shuffle is cheaper).
8786
define <4 x double> @fadd_v4f64_multiuse_shuffle(<4 x double> %a, <4 x double> %b) {
88-
; SSE-LABEL: define <4 x double> @fadd_v4f64_multiuse_shuffle(
89-
; SSE-SAME: <4 x double> [[A:%.*]], <4 x double> [[B:%.*]]) #[[ATTR0]] {
90-
; SSE-NEXT: [[A1:%.*]] = shufflevector <4 x double> [[A]], <4 x double> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
91-
; SSE-NEXT: [[TMP1:%.*]] = shufflevector <4 x double> [[A]], <4 x double> poison, <4 x i32> <i32 2, i32 3, i32 0, i32 1>
92-
; SSE-NEXT: [[TMP2:%.*]] = shufflevector <4 x double> [[B]], <4 x double> poison, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
93-
; SSE-NEXT: [[POST:%.*]] = fadd <4 x double> [[TMP1]], [[TMP2]]
94-
; SSE-NEXT: call void @use_v4f64(<4 x double> [[A1]])
95-
; SSE-NEXT: ret <4 x double> [[POST]]
96-
;
97-
; AVX-LABEL: define <4 x double> @fadd_v4f64_multiuse_shuffle(
98-
; AVX-SAME: <4 x double> [[A:%.*]], <4 x double> [[B:%.*]]) #[[ATTR0]] {
99-
; AVX-NEXT: [[A1:%.*]] = shufflevector <4 x double> [[A]], <4 x double> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
100-
; AVX-NEXT: [[B1:%.*]] = shufflevector <4 x double> [[B]], <4 x double> poison, <4 x i32> <i32 1, i32 0, i32 1, i32 0>
101-
; AVX-NEXT: [[OP:%.*]] = fadd <4 x double> [[A1]], [[B1]]
102-
; AVX-NEXT: [[POST:%.*]] = shufflevector <4 x double> [[OP]], <4 x double> poison, <4 x i32> <i32 1, i32 0, i32 3, i32 2>
103-
; AVX-NEXT: call void @use_v4f64(<4 x double> [[A1]])
104-
; AVX-NEXT: ret <4 x double> [[POST]]
87+
; CHECK-LABEL: define <4 x double> @fadd_v4f64_multiuse_shuffle(
88+
; CHECK-SAME: <4 x double> [[A:%.*]], <4 x double> [[B:%.*]]) #[[ATTR0]] {
89+
; CHECK-NEXT: [[A1:%.*]] = shufflevector <4 x double> [[A]], <4 x double> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
90+
; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <4 x double> [[A]], <4 x double> poison, <4 x i32> <i32 2, i32 3, i32 0, i32 1>
91+
; CHECK-NEXT: [[TMP2:%.*]] = shufflevector <4 x double> [[B]], <4 x double> poison, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
92+
; CHECK-NEXT: [[POST:%.*]] = fadd <4 x double> [[TMP1]], [[TMP2]]
93+
; CHECK-NEXT: call void @use_v4f64(<4 x double> [[A1]])
94+
; CHECK-NEXT: ret <4 x double> [[POST]]
10595
;
10696
%a1 = shufflevector <4 x double> %a, <4 x double> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
10797
%b1 = shufflevector <4 x double> %b, <4 x double> poison, <4 x i32> <i32 1, i32 0, i32 1, i32 0>

0 commit comments

Comments
 (0)