Skip to content

Commit e485d5e

Browse files
authored
[SLPVectorizer] Clear TreeEntryToStridedPtrInfoMap. (#160544)
We need to clear `TreeEntryToStridedPtrInfoMap` in `deleteTree`.
1 parent 6f1f00c commit e485d5e

File tree

2 files changed

+88
-0
lines changed

2 files changed

+88
-0
lines changed

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2105,6 +2105,7 @@ class BoUpSLP {
21052105
UserIgnoreList = nullptr;
21062106
PostponedGathers.clear();
21072107
ValueToGatherNodes.clear();
2108+
TreeEntryToStridedPtrInfoMap.clear();
21082109
}
21092110

21102111
unsigned getTreeSize() const { return VectorizableTree.size(); }
@@ -8948,6 +8949,8 @@ BoUpSLP::findExternalStoreUsersReorderIndices(TreeEntry *TE) const {
89488949
void BoUpSLP::buildTree(ArrayRef<Value *> Roots,
89498950
const SmallDenseSet<Value *> &UserIgnoreLst) {
89508951
deleteTree();
8952+
assert(TreeEntryToStridedPtrInfoMap.empty() &&
8953+
"TreeEntryToStridedPtrInfoMap is not cleared");
89518954
UserIgnoreList = &UserIgnoreLst;
89528955
if (!allSameType(Roots))
89538956
return;
@@ -8956,6 +8959,8 @@ void BoUpSLP::buildTree(ArrayRef<Value *> Roots,
89568959

89578960
void BoUpSLP::buildTree(ArrayRef<Value *> Roots) {
89588961
deleteTree();
8962+
assert(TreeEntryToStridedPtrInfoMap.empty() &&
8963+
"TreeEntryToStridedPtrInfoMap is not cleared");
89598964
if (!allSameType(Roots))
89608965
return;
89618966
buildTreeRec(Roots, 0, EdgeInfo());
Lines changed: 83 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,83 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
2+
; RUN: opt -mtriple=riscv64 -mattr=+m,+v -passes=slp-vectorizer -S < %s | FileCheck %s
3+
4+
; CHECK-NOT: TreeEntryToStridedPtrInfoMap is not cleared
5+
define void @const_stride_1_no_reordering(ptr %pl, ptr %ps) {
6+
; CHECK-LABEL: define void @const_stride_1_no_reordering(
7+
; CHECK-SAME: ptr [[PL:%.*]], ptr [[PS:%.*]]) #[[ATTR0:[0-9]+]] {
8+
; CHECK-NEXT: [[GEP_L0:%.*]] = getelementptr inbounds i8, ptr [[PL]], i64 0
9+
; CHECK-NEXT: [[GEP_S0:%.*]] = getelementptr inbounds i8, ptr [[PS]], i64 0
10+
; CHECK-NEXT: [[TMP1:%.*]] = load <16 x i8>, ptr [[GEP_L0]], align 1
11+
; CHECK-NEXT: store <16 x i8> [[TMP1]], ptr [[GEP_S0]], align 1
12+
; CHECK-NEXT: ret void
13+
;
14+
%gep_l0 = getelementptr inbounds i8, ptr %pl, i64 0
15+
%gep_l1 = getelementptr inbounds i8, ptr %pl, i64 1
16+
%gep_l2 = getelementptr inbounds i8, ptr %pl, i64 2
17+
%gep_l3 = getelementptr inbounds i8, ptr %pl, i64 3
18+
%gep_l4 = getelementptr inbounds i8, ptr %pl, i64 4
19+
%gep_l5 = getelementptr inbounds i8, ptr %pl, i64 5
20+
%gep_l6 = getelementptr inbounds i8, ptr %pl, i64 6
21+
%gep_l7 = getelementptr inbounds i8, ptr %pl, i64 7
22+
%gep_l8 = getelementptr inbounds i8, ptr %pl, i64 8
23+
%gep_l9 = getelementptr inbounds i8, ptr %pl, i64 9
24+
%gep_l10 = getelementptr inbounds i8, ptr %pl, i64 10
25+
%gep_l11 = getelementptr inbounds i8, ptr %pl, i64 11
26+
%gep_l12 = getelementptr inbounds i8, ptr %pl, i64 12
27+
%gep_l13 = getelementptr inbounds i8, ptr %pl, i64 13
28+
%gep_l14 = getelementptr inbounds i8, ptr %pl, i64 14
29+
%gep_l15 = getelementptr inbounds i8, ptr %pl, i64 15
30+
31+
%load0 = load i8, ptr %gep_l0
32+
%load1 = load i8, ptr %gep_l1
33+
%load2 = load i8, ptr %gep_l2
34+
%load3 = load i8, ptr %gep_l3
35+
%load4 = load i8, ptr %gep_l4
36+
%load5 = load i8, ptr %gep_l5
37+
%load6 = load i8, ptr %gep_l6
38+
%load7 = load i8, ptr %gep_l7
39+
%load8 = load i8, ptr %gep_l8
40+
%load9 = load i8, ptr %gep_l9
41+
%load10 = load i8, ptr %gep_l10
42+
%load11 = load i8, ptr %gep_l11
43+
%load12 = load i8, ptr %gep_l12
44+
%load13 = load i8, ptr %gep_l13
45+
%load14 = load i8, ptr %gep_l14
46+
%load15 = load i8, ptr %gep_l15
47+
48+
%gep_s0 = getelementptr inbounds i8, ptr %ps, i64 0
49+
%gep_s1 = getelementptr inbounds i8, ptr %ps, i64 1
50+
%gep_s2 = getelementptr inbounds i8, ptr %ps, i64 2
51+
%gep_s3 = getelementptr inbounds i8, ptr %ps, i64 3
52+
%gep_s4 = getelementptr inbounds i8, ptr %ps, i64 4
53+
%gep_s5 = getelementptr inbounds i8, ptr %ps, i64 5
54+
%gep_s6 = getelementptr inbounds i8, ptr %ps, i64 6
55+
%gep_s7 = getelementptr inbounds i8, ptr %ps, i64 7
56+
%gep_s8 = getelementptr inbounds i8, ptr %ps, i64 8
57+
%gep_s9 = getelementptr inbounds i8, ptr %ps, i64 9
58+
%gep_s10 = getelementptr inbounds i8, ptr %ps, i64 10
59+
%gep_s11 = getelementptr inbounds i8, ptr %ps, i64 11
60+
%gep_s12 = getelementptr inbounds i8, ptr %ps, i64 12
61+
%gep_s13 = getelementptr inbounds i8, ptr %ps, i64 13
62+
%gep_s14 = getelementptr inbounds i8, ptr %ps, i64 14
63+
%gep_s15 = getelementptr inbounds i8, ptr %ps, i64 15
64+
65+
store i8 %load0, ptr %gep_s0
66+
store i8 %load1, ptr %gep_s1
67+
store i8 %load2, ptr %gep_s2
68+
store i8 %load3, ptr %gep_s3
69+
store i8 %load4, ptr %gep_s4
70+
store i8 %load5, ptr %gep_s5
71+
store i8 %load6, ptr %gep_s6
72+
store i8 %load7, ptr %gep_s7
73+
store i8 %load8, ptr %gep_s8
74+
store i8 %load9, ptr %gep_s9
75+
store i8 %load10, ptr %gep_s10
76+
store i8 %load11, ptr %gep_s11
77+
store i8 %load12, ptr %gep_s12
78+
store i8 %load13, ptr %gep_s13
79+
store i8 %load14, ptr %gep_s14
80+
store i8 %load15, ptr %gep_s15
81+
82+
ret void
83+
}

0 commit comments

Comments
 (0)