Skip to content

Commit 8cda2f3

Browse files
artagnonartagnon
committed
LoopVectorize: fix phi cost when it is scalar after vectorization
In getInstructionCost, when a phi node is identified as a fixed-order recurrence, and the vectorization factor indicates that we're vectorizing, we return the cost of a shuffle. However, if it also satisfies the condition isScalarAfterVectorization, attempting to get the shuffle cost will result in a crash from a cast. Fix this. Fixes #72969.
1 parent 1afcfdf commit 8cda2f3

File tree

2 files changed

+114
-3
lines changed

2 files changed

+114
-3
lines changed

‎llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -6885,7 +6885,10 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF,
68856885
auto *Phi = cast<PHINode>(I);
68866886

68876887
// First-order recurrences are replaced by vector shuffles inside the loop.
6888-
if (VF.isVector() && Legal->isFixedOrderRecurrence(Phi)) {
6888+
// However, if the Phi is a scalar after vectorization, don't get shuffle
6889+
// cost.
6890+
if (VF.isVector() && Legal->isFixedOrderRecurrence(Phi) &&
6891+
!isScalarAfterVectorization(Phi, VF)) {
68896892
SmallVector<int> Mask(VF.getKnownMinValue());
68906893
std::iota(Mask.begin(), Mask.end(), VF.getKnownMinValue() - 1);
68916894
return TTI.getShuffleCost(TargetTransformInfo::SK_Splice,

‎llvm/test/Transforms/LoopVectorize/X86/pr72969.ll

Lines changed: 110 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1,9 +1,111 @@
1-
; RUN: not --crash opt -mtriple=x86_64 -mattr=-avx,-avx2,-avx512f,+sse,-sse2,-sse3,-sse4.2 -passes=loop-vectorize -S < %s
2-
; RUN: not --crash opt -mtriple=x86_64 -mattr=-avx,-avx2,-avx512f,+sse,-sse2,-sse3,-sse4.2 -passes=loop-vectorize -force-vector-width=4 -S < %s
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
2+
; RUN: opt -mtriple=x86_64 -mattr=-avx,-avx2,-avx512f,+sse,-sse2,-sse3,-sse4.2 -passes=loop-vectorize -S < %s | FileCheck --check-prefix=NOVEC %s
3+
; RUN: opt -mtriple=x86_64 -mattr=-avx,-avx2,-avx512f,+sse,-sse2,-sse3,-sse4.2 -passes=loop-vectorize -force-vector-width=4 -S < %s | FileCheck --check-prefix=VEC %s
34

45
@h = global i64 0
56

67
define void @test(ptr %p) {
8+
; NOVEC-LABEL: define void @test(
9+
; NOVEC-SAME: ptr [[P:%.*]]) #[[ATTR0:[0-9]+]] {
10+
; NOVEC-NEXT: entry:
11+
; NOVEC-NEXT: br label [[FOR_BODY:%.*]]
12+
; NOVEC: for.body:
13+
; NOVEC-NEXT: [[IDX_EXT_MERGE:%.*]] = phi i64 [ 1, [[ENTRY:%.*]] ], [ [[IDX:%.*]], [[FOR_BODY]] ]
14+
; NOVEC-NEXT: [[INC_MERGE:%.*]] = phi i16 [ 1, [[ENTRY]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
15+
; NOVEC-NEXT: [[IDX_MERGE:%.*]] = phi i64 [ 0, [[ENTRY]] ], [ [[IDX_EXT_MERGE]], [[FOR_BODY]] ]
16+
; NOVEC-NEXT: [[ADD:%.*]] = shl i64 [[IDX_MERGE]], 1
17+
; NOVEC-NEXT: [[ARRAYIDX:%.*]] = getelementptr i64, ptr [[P]], i64 [[ADD]]
18+
; NOVEC-NEXT: store i64 0, ptr [[ARRAYIDX]], align 8
19+
; NOVEC-NEXT: [[INC]] = add i16 [[INC_MERGE]], 1
20+
; NOVEC-NEXT: [[IDX]] = zext i16 [[INC]] to i64
21+
; NOVEC-NEXT: [[GEP:%.*]] = getelementptr i64, ptr [[P]], i64 [[IDX]]
22+
; NOVEC-NEXT: [[CMP:%.*]] = icmp ugt ptr [[GEP]], @h
23+
; NOVEC-NEXT: br i1 [[CMP]], label [[EXIT:%.*]], label [[FOR_BODY]]
24+
; NOVEC: exit:
25+
; NOVEC-NEXT: ret void
26+
;
27+
; VEC-LABEL: define void @test(
28+
; VEC-SAME: ptr [[P:%.*]]) #[[ATTR0:[0-9]+]] {
29+
; VEC-NEXT: entry:
30+
; VEC-NEXT: [[P1:%.*]] = ptrtoint ptr [[P]] to i64
31+
; VEC-NEXT: [[TMP0:%.*]] = add i64 [[P1]], 16
32+
; VEC-NEXT: [[UMAX2:%.*]] = call i64 @llvm.umax.i64(i64 [[TMP0]], i64 add (i64 ptrtoint (ptr @h to i64), i64 1))
33+
; VEC-NEXT: [[TMP1:%.*]] = add i64 [[UMAX2]], -9
34+
; VEC-NEXT: [[TMP2:%.*]] = sub i64 [[TMP1]], [[P1]]
35+
; VEC-NEXT: [[TMP3:%.*]] = lshr i64 [[TMP2]], 3
36+
; VEC-NEXT: [[TMP4:%.*]] = add nuw nsw i64 [[TMP3]], 1
37+
; VEC-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP4]], 4
38+
; VEC-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_SCEVCHECK:%.*]]
39+
; VEC: vector.scevcheck:
40+
; VEC-NEXT: [[TMP5:%.*]] = add i64 [[P1]], 16
41+
; VEC-NEXT: [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 [[TMP5]], i64 add (i64 ptrtoint (ptr @h to i64), i64 1))
42+
; VEC-NEXT: [[TMP6:%.*]] = add i64 [[UMAX]], -9
43+
; VEC-NEXT: [[TMP7:%.*]] = sub i64 [[TMP6]], [[P1]]
44+
; VEC-NEXT: [[TMP8:%.*]] = lshr i64 [[TMP7]], 3
45+
; VEC-NEXT: [[TMP9:%.*]] = icmp ugt i64 [[TMP8]], 65535
46+
; VEC-NEXT: [[TMP10:%.*]] = trunc i64 [[TMP8]] to i16
47+
; VEC-NEXT: [[TMP11:%.*]] = add i16 2, [[TMP10]]
48+
; VEC-NEXT: [[TMP12:%.*]] = icmp ult i16 [[TMP11]], 2
49+
; VEC-NEXT: [[TMP13:%.*]] = icmp ugt i64 [[TMP8]], 65535
50+
; VEC-NEXT: [[TMP14:%.*]] = or i1 [[TMP12]], [[TMP13]]
51+
; VEC-NEXT: br i1 [[TMP14]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]]
52+
; VEC: vector.ph:
53+
; VEC-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[TMP4]], 4
54+
; VEC-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP4]], [[N_MOD_VF]]
55+
; VEC-NEXT: [[DOTCAST:%.*]] = trunc i64 [[N_VEC]] to i16
56+
; VEC-NEXT: [[IND_END:%.*]] = add i16 1, [[DOTCAST]]
57+
; VEC-NEXT: br label [[VECTOR_BODY:%.*]]
58+
; VEC: vector.body:
59+
; VEC-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
60+
; VEC-NEXT: [[VECTOR_RECUR:%.*]] = phi <4 x i64> [ <i64 poison, i64 poison, i64 poison, i64 1>, [[VECTOR_PH]] ], [ [[TMP28:%.*]], [[VECTOR_BODY]] ]
61+
; VEC-NEXT: [[VEC_IND:%.*]] = phi <4 x i16> [ <i16 1, i16 2, i16 3, i16 4>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
62+
; VEC-NEXT: [[TMP15:%.*]] = add i64 [[INDEX]], 0
63+
; VEC-NEXT: [[TMP16:%.*]] = add i64 [[INDEX]], 1
64+
; VEC-NEXT: [[TMP17:%.*]] = add i64 [[INDEX]], 2
65+
; VEC-NEXT: [[TMP18:%.*]] = add i64 [[INDEX]], 3
66+
; VEC-NEXT: [[TMP19:%.*]] = shl i64 [[TMP15]], 1
67+
; VEC-NEXT: [[TMP20:%.*]] = shl i64 [[TMP16]], 1
68+
; VEC-NEXT: [[TMP21:%.*]] = shl i64 [[TMP17]], 1
69+
; VEC-NEXT: [[TMP22:%.*]] = shl i64 [[TMP18]], 1
70+
; VEC-NEXT: [[TMP23:%.*]] = getelementptr i64, ptr [[P]], i64 [[TMP19]]
71+
; VEC-NEXT: [[TMP24:%.*]] = getelementptr i64, ptr [[P]], i64 [[TMP20]]
72+
; VEC-NEXT: [[TMP25:%.*]] = getelementptr i64, ptr [[P]], i64 [[TMP21]]
73+
; VEC-NEXT: [[TMP26:%.*]] = getelementptr i64, ptr [[P]], i64 [[TMP22]]
74+
; VEC-NEXT: store i64 0, ptr [[TMP23]], align 8
75+
; VEC-NEXT: store i64 0, ptr [[TMP24]], align 8
76+
; VEC-NEXT: store i64 0, ptr [[TMP25]], align 8
77+
; VEC-NEXT: store i64 0, ptr [[TMP26]], align 8
78+
; VEC-NEXT: [[TMP27:%.*]] = add <4 x i16> [[VEC_IND]], <i16 1, i16 1, i16 1, i16 1>
79+
; VEC-NEXT: [[TMP28]] = zext <4 x i16> [[TMP27]] to <4 x i64>
80+
; VEC-NEXT: [[TMP29:%.*]] = shufflevector <4 x i64> [[VECTOR_RECUR]], <4 x i64> [[TMP28]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
81+
; VEC-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
82+
; VEC-NEXT: [[VEC_IND_NEXT]] = add <4 x i16> [[VEC_IND]], <i16 4, i16 4, i16 4, i16 4>
83+
; VEC-NEXT: [[TMP30:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
84+
; VEC-NEXT: br i1 [[TMP30]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
85+
; VEC: middle.block:
86+
; VEC-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP4]], [[N_VEC]]
87+
; VEC-NEXT: [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i64> [[TMP28]], i32 3
88+
; VEC-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
89+
; VEC: scalar.ph:
90+
; VEC-NEXT: [[SCALAR_RECUR_INIT:%.*]] = phi i64 [ 1, [[VECTOR_SCEVCHECK]] ], [ 1, [[ENTRY:%.*]] ], [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ]
91+
; VEC-NEXT: [[BC_RESUME_VAL:%.*]] = phi i16 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 1, [[ENTRY]] ], [ 1, [[VECTOR_SCEVCHECK]] ]
92+
; VEC-NEXT: [[BC_RESUME_VAL3:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ], [ 0, [[VECTOR_SCEVCHECK]] ]
93+
; VEC-NEXT: br label [[FOR_BODY:%.*]]
94+
; VEC: for.body:
95+
; VEC-NEXT: [[SCALAR_RECUR:%.*]] = phi i64 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[IDX:%.*]], [[FOR_BODY]] ]
96+
; VEC-NEXT: [[INC_MERGE:%.*]] = phi i16 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
97+
; VEC-NEXT: [[IDX_MERGE:%.*]] = phi i64 [ [[BC_RESUME_VAL3]], [[SCALAR_PH]] ], [ [[SCALAR_RECUR]], [[FOR_BODY]] ]
98+
; VEC-NEXT: [[ADD:%.*]] = shl i64 [[IDX_MERGE]], 1
99+
; VEC-NEXT: [[ARRAYIDX:%.*]] = getelementptr i64, ptr [[P]], i64 [[ADD]]
100+
; VEC-NEXT: store i64 0, ptr [[ARRAYIDX]], align 8
101+
; VEC-NEXT: [[INC]] = add i16 [[INC_MERGE]], 1
102+
; VEC-NEXT: [[IDX]] = zext i16 [[INC]] to i64
103+
; VEC-NEXT: [[GEP:%.*]] = getelementptr i64, ptr [[P]], i64 [[IDX]]
104+
; VEC-NEXT: [[CMP:%.*]] = icmp ugt ptr [[GEP]], @h
105+
; VEC-NEXT: br i1 [[CMP]], label [[EXIT]], label [[FOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
106+
; VEC: exit:
107+
; VEC-NEXT: ret void
108+
;
7109
entry:
8110
br label %for.body
9111

@@ -23,3 +125,9 @@ for.body:
23125
exit:
24126
ret void
25127
}
128+
;.
129+
; VEC: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
130+
; VEC: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
131+
; VEC: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
132+
; VEC: [[LOOP3]] = distinct !{[[LOOP3]], [[META1]]}
133+
;.

0 commit comments

Comments
 (0)