Skip to content

Commit c9f3f38

Browse files
HanKuanChenpuja2196
HanKuanChen
authored and
puja2196
committed
[SLP][REVEC] getTypeSizeInBits should apply to scalar type instead of FixedVectorType. (#110610)
reference: llvm/llvm-project#109835
1 parent 6396a65 commit c9f3f38

File tree

2 files changed

+175
-1
lines changed

2 files changed

+175
-1
lines changed

‎llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

Lines changed: 5 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -11786,14 +11786,18 @@ InstructionCost BoUpSLP::getTreeCost(ArrayRef<Value *> VectorizedVals) {
1178611786
auto BWIt = MinBWs.find(&Root);
1178711787
if (BWIt != MinBWs.end()) {
1178811788
Type *DstTy = Root.Scalars.front()->getType();
11789-
unsigned OriginalSz = DL->getTypeSizeInBits(DstTy);
11789+
unsigned OriginalSz = DL->getTypeSizeInBits(DstTy->getScalarType());
1179011790
unsigned SrcSz =
1179111791
ReductionBitWidth == 0 ? BWIt->second.first : ReductionBitWidth;
1179211792
if (OriginalSz != SrcSz) {
1179311793
unsigned Opcode = Instruction::Trunc;
1179411794
if (OriginalSz > SrcSz)
1179511795
Opcode = BWIt->second.second ? Instruction::SExt : Instruction::ZExt;
1179611796
Type *SrcTy = IntegerType::get(DstTy->getContext(), SrcSz);
11797+
if (auto *VecTy = dyn_cast<FixedVectorType>(DstTy)) {
11798+
assert(SLPReVec && "Only supported by REVEC.");
11799+
SrcTy = getWidenedType(SrcTy, VecTy->getNumElements());
11800+
}
1179711801
Cost += TTI->getCastInstrCost(Opcode, DstTy, SrcTy,
1179811802
TTI::CastContextHint::None,
1179911803
TTI::TCK_RecipThroughput);

‎llvm/test/Transforms/SLPVectorizer/SystemZ/revec-fix-109835.ll

Lines changed: 170 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,170 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2+
; RUN: opt -mtriple=s390x-unknown-linux -mcpu=z16 -passes=slp-vectorizer -S -slp-revec %s | FileCheck %s
3+
4+
@g_155 = external dso_local global i64, align 8
5+
@g_855 = external dso_local global i8, align 2
6+
@g_3_1_0 = external dso_local global i32, align 4
7+
@g_7 = external dso_local global [5 x i32], align 4
8+
9+
; Function Attrs: nofree norecurse noreturn nounwind memory(readwrite, argmem: none)
10+
define void @func_1() {
11+
; CHECK-LABEL: @func_1(
12+
; CHECK-NEXT: [[DOTPRE:%.*]] = load i32, ptr @g_7, align 4
13+
; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i32>, ptr getelementptr inbounds nuw (i8, ptr @g_7, i64 4), align 4
14+
; CHECK-NEXT: br label [[DOTLOOPEXIT:%.*]]
15+
; CHECK: .loopexit:
16+
; CHECK-NEXT: [[TMP2:%.*]] = phi i32 [ [[OP_RDX15:%.*]], [[DOTLOOPEXIT]] ], [ [[DOTPRE]], [[TMP0:%.*]] ]
17+
; CHECK-NEXT: [[TMP3:%.*]] = phi <4 x i32> [ [[TMP71:%.*]], [[DOTLOOPEXIT]] ], [ [[TMP1]], [[TMP0]] ]
18+
; CHECK-NEXT: [[TMP4:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
19+
; CHECK-NEXT: [[TMP5:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
20+
; CHECK-NEXT: [[TMP6:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
21+
; CHECK-NEXT: [[TMP7:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
22+
; CHECK-NEXT: [[TMP8:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
23+
; CHECK-NEXT: [[TMP9:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
24+
; CHECK-NEXT: [[TMP10:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
25+
; CHECK-NEXT: [[TMP11:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
26+
; CHECK-NEXT: [[TMP12:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
27+
; CHECK-NEXT: [[TMP13:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
28+
; CHECK-NEXT: [[TMP14:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
29+
; CHECK-NEXT: [[TMP15:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
30+
; CHECK-NEXT: [[TMP16:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
31+
; CHECK-NEXT: [[TMP17:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
32+
; CHECK-NEXT: [[TMP18:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
33+
; CHECK-NEXT: [[TMP19:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
34+
; CHECK-NEXT: [[TMP20:%.*]] = insertelement <4 x i32> poison, i32 [[TMP4]], i32 0
35+
; CHECK-NEXT: [[TMP21:%.*]] = insertelement <4 x i32> [[TMP20]], i32 [[TMP9]], i32 1
36+
; CHECK-NEXT: [[TMP22:%.*]] = insertelement <4 x i32> [[TMP21]], i32 [[TMP14]], i32 2
37+
; CHECK-NEXT: [[TMP23:%.*]] = insertelement <4 x i32> [[TMP22]], i32 [[TMP19]], i32 3
38+
; CHECK-NEXT: [[TMP24:%.*]] = icmp eq <4 x i32> [[TMP23]], zeroinitializer
39+
; CHECK-NEXT: [[TMP25:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
40+
; CHECK-NEXT: [[TMP26:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
41+
; CHECK-NEXT: [[TMP27:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
42+
; CHECK-NEXT: [[TMP28:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
43+
; CHECK-NEXT: [[TMP29:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
44+
; CHECK-NEXT: [[DOTNOT2_410:%.*]] = icmp eq i32 [[TMP29]], 0
45+
; CHECK-NEXT: [[TMP30:%.*]] = zext i1 [[DOTNOT2_410]] to i32
46+
; CHECK-NEXT: [[TMP31:%.*]] = zext <4 x i1> [[TMP24]] to <4 x i32>
47+
; CHECK-NEXT: [[TMP32:%.*]] = call i32 @llvm.vector.reduce.xor.v4i32(<4 x i32> [[TMP31]])
48+
; CHECK-NEXT: [[OP_RDX:%.*]] = xor i32 [[TMP32]], [[TMP30]]
49+
; CHECK-NEXT: [[OP_RDX15]] = xor i32 [[OP_RDX]], [[TMP2]]
50+
; CHECK-NEXT: [[TMP33:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
51+
; CHECK-NEXT: [[TMP34:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
52+
; CHECK-NEXT: [[TMP35:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
53+
; CHECK-NEXT: [[TMP36:%.*]] = load volatile i32, ptr @g_3_1_0, align 4
54+
; CHECK-NEXT: [[TMP37:%.*]] = insertelement <4 x i32> poison, i32 [[TMP5]], i32 0
55+
; CHECK-NEXT: [[TMP38:%.*]] = insertelement <4 x i32> [[TMP37]], i32 [[TMP6]], i32 1
56+
; CHECK-NEXT: [[TMP39:%.*]] = insertelement <4 x i32> [[TMP38]], i32 [[TMP7]], i32 2
57+
; CHECK-NEXT: [[TMP40:%.*]] = insertelement <4 x i32> [[TMP39]], i32 [[TMP8]], i32 3
58+
; CHECK-NEXT: [[TMP41:%.*]] = icmp eq <4 x i32> [[TMP40]], zeroinitializer
59+
; CHECK-NEXT: [[TMP42:%.*]] = zext <4 x i1> [[TMP41]] to <4 x i32>
60+
; CHECK-NEXT: [[TMP43:%.*]] = xor <4 x i32> [[TMP3]], [[TMP42]]
61+
; CHECK-NEXT: [[TMP44:%.*]] = insertelement <4 x i32> poison, i32 [[TMP10]], i32 0
62+
; CHECK-NEXT: [[TMP45:%.*]] = insertelement <4 x i32> [[TMP44]], i32 [[TMP11]], i32 1
63+
; CHECK-NEXT: [[TMP46:%.*]] = insertelement <4 x i32> [[TMP45]], i32 [[TMP12]], i32 2
64+
; CHECK-NEXT: [[TMP47:%.*]] = insertelement <4 x i32> [[TMP46]], i32 [[TMP13]], i32 3
65+
; CHECK-NEXT: [[TMP48:%.*]] = icmp eq <4 x i32> [[TMP47]], zeroinitializer
66+
; CHECK-NEXT: [[TMP49:%.*]] = zext <4 x i1> [[TMP48]] to <4 x i32>
67+
; CHECK-NEXT: [[TMP50:%.*]] = xor <4 x i32> [[TMP43]], [[TMP49]]
68+
; CHECK-NEXT: [[TMP51:%.*]] = insertelement <4 x i32> poison, i32 [[TMP15]], i32 0
69+
; CHECK-NEXT: [[TMP52:%.*]] = insertelement <4 x i32> [[TMP51]], i32 [[TMP16]], i32 1
70+
; CHECK-NEXT: [[TMP53:%.*]] = insertelement <4 x i32> [[TMP52]], i32 [[TMP17]], i32 2
71+
; CHECK-NEXT: [[TMP54:%.*]] = insertelement <4 x i32> [[TMP53]], i32 [[TMP18]], i32 3
72+
; CHECK-NEXT: [[TMP55:%.*]] = icmp eq <4 x i32> [[TMP54]], zeroinitializer
73+
; CHECK-NEXT: [[TMP56:%.*]] = zext <4 x i1> [[TMP55]] to <4 x i32>
74+
; CHECK-NEXT: [[TMP57:%.*]] = xor <4 x i32> [[TMP50]], [[TMP56]]
75+
; CHECK-NEXT: [[TMP58:%.*]] = insertelement <4 x i32> poison, i32 [[TMP25]], i32 0
76+
; CHECK-NEXT: [[TMP59:%.*]] = insertelement <4 x i32> [[TMP58]], i32 [[TMP26]], i32 1
77+
; CHECK-NEXT: [[TMP60:%.*]] = insertelement <4 x i32> [[TMP59]], i32 [[TMP27]], i32 2
78+
; CHECK-NEXT: [[TMP61:%.*]] = insertelement <4 x i32> [[TMP60]], i32 [[TMP28]], i32 3
79+
; CHECK-NEXT: [[TMP62:%.*]] = icmp eq <4 x i32> [[TMP61]], zeroinitializer
80+
; CHECK-NEXT: [[TMP63:%.*]] = zext <4 x i1> [[TMP62]] to <4 x i32>
81+
; CHECK-NEXT: [[TMP64:%.*]] = xor <4 x i32> [[TMP57]], [[TMP63]]
82+
; CHECK-NEXT: [[TMP65:%.*]] = insertelement <4 x i32> poison, i32 [[TMP33]], i32 0
83+
; CHECK-NEXT: [[TMP66:%.*]] = insertelement <4 x i32> [[TMP65]], i32 [[TMP34]], i32 1
84+
; CHECK-NEXT: [[TMP67:%.*]] = insertelement <4 x i32> [[TMP66]], i32 [[TMP35]], i32 2
85+
; CHECK-NEXT: [[TMP68:%.*]] = insertelement <4 x i32> [[TMP67]], i32 [[TMP36]], i32 3
86+
; CHECK-NEXT: [[TMP69:%.*]] = icmp eq <4 x i32> [[TMP68]], zeroinitializer
87+
; CHECK-NEXT: [[TMP70:%.*]] = zext <4 x i1> [[TMP69]] to <4 x i32>
88+
; CHECK-NEXT: [[TMP71]] = xor <4 x i32> [[TMP64]], [[TMP70]]
89+
; CHECK-NEXT: br label [[DOTLOOPEXIT]]
90+
;
91+
%.pre = load i32, ptr @g_7, align 4
92+
%1 = load <4 x i32>, ptr getelementptr inbounds nuw (i8, ptr @g_7, i64 4), align 4
93+
br label %.loopexit
94+
95+
.loopexit: ; preds = %.loopexit, %0
96+
%2 = phi i32 [ %op.rdx15, %.loopexit ], [ %.pre, %0 ]
97+
%3 = phi <4 x i32> [ %71, %.loopexit ], [ %1, %0 ]
98+
%4 = load volatile i32, ptr @g_3_1_0, align 4
99+
%5 = load volatile i32, ptr @g_3_1_0, align 4
100+
%6 = load volatile i32, ptr @g_3_1_0, align 4
101+
%7 = load volatile i32, ptr @g_3_1_0, align 4
102+
%8 = load volatile i32, ptr @g_3_1_0, align 4
103+
%9 = load volatile i32, ptr @g_3_1_0, align 4
104+
%10 = load volatile i32, ptr @g_3_1_0, align 4
105+
%11 = load volatile i32, ptr @g_3_1_0, align 4
106+
%12 = load volatile i32, ptr @g_3_1_0, align 4
107+
%13 = load volatile i32, ptr @g_3_1_0, align 4
108+
%14 = load volatile i32, ptr @g_3_1_0, align 4
109+
%15 = load volatile i32, ptr @g_3_1_0, align 4
110+
%16 = load volatile i32, ptr @g_3_1_0, align 4
111+
%17 = load volatile i32, ptr @g_3_1_0, align 4
112+
%18 = load volatile i32, ptr @g_3_1_0, align 4
113+
%19 = load volatile i32, ptr @g_3_1_0, align 4
114+
%20 = insertelement <4 x i32> poison, i32 %4, i32 0
115+
%21 = insertelement <4 x i32> %20, i32 %9, i32 1
116+
%22 = insertelement <4 x i32> %21, i32 %14, i32 2
117+
%23 = insertelement <4 x i32> %22, i32 %19, i32 3
118+
%24 = icmp eq <4 x i32> %23, zeroinitializer
119+
%25 = load volatile i32, ptr @g_3_1_0, align 4
120+
%26 = load volatile i32, ptr @g_3_1_0, align 4
121+
%27 = load volatile i32, ptr @g_3_1_0, align 4
122+
%28 = load volatile i32, ptr @g_3_1_0, align 4
123+
%29 = load volatile i32, ptr @g_3_1_0, align 4
124+
%.not2.410 = icmp eq i32 %29, 0
125+
%30 = zext i1 %.not2.410 to i32
126+
%31 = zext <4 x i1> %24 to <4 x i32>
127+
%32 = call i32 @llvm.vector.reduce.xor.v4i32(<4 x i32> %31)
128+
%op.rdx = xor i32 %32, %30
129+
%op.rdx15 = xor i32 %op.rdx, %2
130+
%33 = load volatile i32, ptr @g_3_1_0, align 4
131+
%34 = load volatile i32, ptr @g_3_1_0, align 4
132+
%35 = load volatile i32, ptr @g_3_1_0, align 4
133+
%36 = load volatile i32, ptr @g_3_1_0, align 4
134+
%37 = insertelement <4 x i32> poison, i32 %5, i32 0
135+
%38 = insertelement <4 x i32> %37, i32 %6, i32 1
136+
%39 = insertelement <4 x i32> %38, i32 %7, i32 2
137+
%40 = insertelement <4 x i32> %39, i32 %8, i32 3
138+
%41 = icmp eq <4 x i32> %40, zeroinitializer
139+
%42 = zext <4 x i1> %41 to <4 x i32>
140+
%43 = xor <4 x i32> %3, %42
141+
%44 = insertelement <4 x i32> poison, i32 %10, i32 0
142+
%45 = insertelement <4 x i32> %44, i32 %11, i32 1
143+
%46 = insertelement <4 x i32> %45, i32 %12, i32 2
144+
%47 = insertelement <4 x i32> %46, i32 %13, i32 3
145+
%48 = icmp eq <4 x i32> %47, zeroinitializer
146+
%49 = zext <4 x i1> %48 to <4 x i32>
147+
%50 = xor <4 x i32> %43, %49
148+
%51 = insertelement <4 x i32> poison, i32 %15, i32 0
149+
%52 = insertelement <4 x i32> %51, i32 %16, i32 1
150+
%53 = insertelement <4 x i32> %52, i32 %17, i32 2
151+
%54 = insertelement <4 x i32> %53, i32 %18, i32 3
152+
%55 = icmp eq <4 x i32> %54, zeroinitializer
153+
%56 = zext <4 x i1> %55 to <4 x i32>
154+
%57 = xor <4 x i32> %50, %56
155+
%58 = insertelement <4 x i32> poison, i32 %25, i32 0
156+
%59 = insertelement <4 x i32> %58, i32 %26, i32 1
157+
%60 = insertelement <4 x i32> %59, i32 %27, i32 2
158+
%61 = insertelement <4 x i32> %60, i32 %28, i32 3
159+
%62 = icmp eq <4 x i32> %61, zeroinitializer
160+
%63 = zext <4 x i1> %62 to <4 x i32>
161+
%64 = xor <4 x i32> %57, %63
162+
%65 = insertelement <4 x i32> poison, i32 %33, i32 0
163+
%66 = insertelement <4 x i32> %65, i32 %34, i32 1
164+
%67 = insertelement <4 x i32> %66, i32 %35, i32 2
165+
%68 = insertelement <4 x i32> %67, i32 %36, i32 3
166+
%69 = icmp eq <4 x i32> %68, zeroinitializer
167+
%70 = zext <4 x i1> %69 to <4 x i32>
168+
%71 = xor <4 x i32> %64, %70
169+
br label %.loopexit
170+
}

0 commit comments

Comments
 (0)