Skip to content

Commit f8b1c25

Browse files
zmodemzmodem
committed
Merging r370720 and r370721:
------------------------------------------------------------------------ r370720 | bjope | 2019-09-03 11:33:40 +0200 (Tue, 03 Sep 2019) | 13 lines [LV] Precommit test case showing miscompile from PR43166. NFC Summary: Precommit test case showing miscompile from PR43166. Reviewers: fhahn, Ayal Reviewed By: fhahn Subscribers: rkruppe, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D67072 ------------------------------------------------------------------------ ------------------------------------------------------------------------ r370721 | bjope | 2019-09-03 11:33:55 +0200 (Tue, 03 Sep 2019) | 20 lines [LV] Fix miscompiles by adding non-header PHI nodes to AllowedExit Summary: Fold-tail currently supports reduction last-vector-value live-out's, but has yet to support last-scalar-value live-outs, including non-header phi's. As it relies on AllowedExit in order to detect them and bail out we need to add the non-header PHI nodes to AllowedExit, otherwise we end up with miscompiles. Solves https://bugs.llvm.org/show_bug.cgi?id=43166 Reviewers: fhahn, Ayal Reviewed By: fhahn, Ayal Subscribers: anna, hiraditya, rkruppe, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D67074 ------------------------------------------------------------------------ llvm-svn: 371044
1 parent 5a661d6 commit f8b1c25

File tree

3 files changed

+168
-2
lines changed

3 files changed

+168
-2
lines changed

‎llvm/include/llvm/Transforms/Vectorize/LoopVectorizationLegality.h

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -452,8 +452,8 @@ class LoopVectorizationLegality {
452452
/// Holds the widest induction type encountered.
453453
Type *WidestIndTy = nullptr;
454454

455-
/// Allowed outside users. This holds the induction and reduction
456-
/// vars which can be accessed from outside the loop.
455+
/// Allowed outside users. This holds the variables that can be accessed from
456+
/// outside the loop.
457457
SmallPtrSet<Value *, 4> AllowedExit;
458458

459459
/// Can we assume the absence of NaNs.

‎llvm/lib/Transforms/Vectorize/LoopVectorizationLegality.cpp

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -631,6 +631,7 @@ bool LoopVectorizationLegality::canVectorizeInstrs() {
631631
// Unsafe cyclic dependencies with header phis are identified during
632632
// legalization for reduction, induction and first order
633633
// recurrences.
634+
AllowedExit.insert(&I);
634635
continue;
635636
}
636637

‎llvm/test/Transforms/LoopVectorize/pr43166-fold-tail-by-masking.ll

Lines changed: 165 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,165 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2+
; RUN: opt < %s -loop-vectorize -force-vector-width=4 -S | FileCheck %s
3+
4+
5+
; Test cases below are reduced (and slightly modified) reproducers based on a
6+
; problem seen when compiling a C program like this:
7+
;
8+
; #include <stdint.h>
9+
; #include <stdio.h>
10+
;
11+
; int y = 0;
12+
; int b = 1;
13+
; int d = 1;
14+
;
15+
; int main() {
16+
; #pragma clang loop vectorize_width(4)
17+
; for (int i = 0; i < 3; ++i) {
18+
; b = (y == 0) ? d : (d / y);
19+
; }
20+
;
21+
; if (b == 1)
22+
; printf("GOOD!\n");
23+
; else
24+
; printf("BAD!\n");
25+
; }
26+
;
27+
; When compiled+executed using
28+
; build-all/bin/clang -O1 lv-bug.c && ./a.out
29+
; the result is "GOOD!"
30+
;
31+
; When compiled+executed using
32+
; build-all/bin/clang -O1 lv-bug.c -fvectorize && ./a.out
33+
; the result is "BAD!"
34+
35+
36+
; This test case miscompiled with clang 8.0.0 (see PR43166), now we get
37+
; loop not vectorized: Cannot fold tail by masking in the presence of live outs.
38+
; instead.
39+
define i64 @test1(i64 %y) {
40+
; CHECK-LABEL: @test1(
41+
; CHECK-NEXT: entry:
42+
; CHECK-NEXT: br label [[FOR_BODY:%.*]]
43+
; CHECK: for.body:
44+
; CHECK-NEXT: [[I:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[COND_END:%.*]] ]
45+
; CHECK-NEXT: [[CMP:%.*]] = icmp eq i64 [[Y:%.*]], 0
46+
; CHECK-NEXT: br i1 [[CMP]], label [[COND_END]], label [[COND_FALSE:%.*]]
47+
; CHECK: cond.false:
48+
; CHECK-NEXT: [[DIV:%.*]] = xor i64 3, [[Y]]
49+
; CHECK-NEXT: br label [[COND_END]]
50+
; CHECK: cond.end:
51+
; CHECK-NEXT: [[COND:%.*]] = phi i64 [ [[DIV]], [[COND_FALSE]] ], [ 77, [[FOR_BODY]] ]
52+
; CHECK-NEXT: [[INC]] = add nuw nsw i32 [[I]], 1
53+
; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i32 [[INC]], 3
54+
; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY]]
55+
; CHECK: for.cond.cleanup:
56+
; CHECK-NEXT: [[COND_LCSSA:%.*]] = phi i64 [ [[COND]], [[COND_END]] ]
57+
; CHECK-NEXT: ret i64 [[COND_LCSSA]]
58+
;
59+
entry:
60+
br label %for.body
61+
62+
for.body:
63+
%i = phi i32 [ 0, %entry ], [ %inc, %cond.end ]
64+
%cmp = icmp eq i64 %y, 0
65+
br i1 %cmp, label %cond.end, label %cond.false
66+
67+
cond.false:
68+
%div = xor i64 3, %y
69+
br label %cond.end
70+
71+
cond.end:
72+
%cond = phi i64 [ %div, %cond.false ], [ 77, %for.body ]
73+
%inc = add nuw nsw i32 %i, 1
74+
%exitcond = icmp eq i32 %inc, 3
75+
br i1 %exitcond, label %for.cond.cleanup, label %for.body
76+
77+
for.cond.cleanup:
78+
ret i64 %cond
79+
}
80+
81+
; This test case miscompiled with clang 8.0.0 (see PR43166), now we get
82+
; loop not vectorized: Cannot fold tail by masking in the presence of live outs.
83+
; instead.
84+
define i64 @test2(i64 %y) {
85+
; CHECK-LABEL: @test2(
86+
; CHECK-NEXT: entry:
87+
; CHECK-NEXT: br label [[FOR_BODY:%.*]]
88+
; CHECK: for.body:
89+
; CHECK-NEXT: [[I:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[COND_END:%.*]] ]
90+
; CHECK-NEXT: [[CMP:%.*]] = icmp eq i64 [[Y:%.*]], 0
91+
; CHECK-NEXT: br i1 [[CMP]], label [[COND_END]], label [[COND_FALSE:%.*]]
92+
; CHECK: cond.false:
93+
; CHECK-NEXT: br label [[COND_END]]
94+
; CHECK: cond.end:
95+
; CHECK-NEXT: [[COND:%.*]] = phi i64 [ 55, [[COND_FALSE]] ], [ 77, [[FOR_BODY]] ]
96+
; CHECK-NEXT: [[INC]] = add nuw nsw i32 [[I]], 1
97+
; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i32 [[INC]], 3
98+
; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY]]
99+
; CHECK: for.cond.cleanup:
100+
; CHECK-NEXT: [[COND_LCSSA:%.*]] = phi i64 [ [[COND]], [[COND_END]] ]
101+
; CHECK-NEXT: ret i64 [[COND_LCSSA]]
102+
;
103+
entry:
104+
br label %for.body
105+
106+
for.body:
107+
%i = phi i32 [ 0, %entry ], [ %inc, %cond.end ]
108+
%cmp = icmp eq i64 %y, 0
109+
br i1 %cmp, label %cond.end, label %cond.false
110+
111+
cond.false:
112+
br label %cond.end
113+
114+
cond.end:
115+
%cond = phi i64 [ 55, %cond.false ], [ 77, %for.body ]
116+
%inc = add nuw nsw i32 %i, 1
117+
%exitcond = icmp eq i32 %inc, 3
118+
br i1 %exitcond, label %for.cond.cleanup, label %for.body
119+
120+
for.cond.cleanup:
121+
ret i64 %cond
122+
}
123+
124+
; This test case miscompiled with clang 8.0.0 (see PR43166), now we get
125+
; loop not vectorized: Cannot fold tail by masking in the presence of live outs.
126+
; instead.
127+
define i32 @test3(i64 %y) {
128+
; CHECK-LABEL: @test3(
129+
; CHECK-NEXT: entry:
130+
; CHECK-NEXT: br label [[FOR_BODY:%.*]]
131+
; CHECK: for.body:
132+
; CHECK-NEXT: [[I:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[COND_END:%.*]] ]
133+
; CHECK-NEXT: [[CMP:%.*]] = icmp eq i64 [[Y:%.*]], 0
134+
; CHECK-NEXT: br i1 [[CMP]], label [[COND_END]], label [[COND_FALSE:%.*]]
135+
; CHECK: cond.false:
136+
; CHECK-NEXT: br label [[COND_END]]
137+
; CHECK: cond.end:
138+
; CHECK-NEXT: [[COND:%.*]] = phi i32 [ 55, [[COND_FALSE]] ], [ [[I]], [[FOR_BODY]] ]
139+
; CHECK-NEXT: [[INC]] = add nuw nsw i32 [[I]], 1
140+
; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i32 [[INC]], 3
141+
; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY]]
142+
; CHECK: for.cond.cleanup:
143+
; CHECK-NEXT: [[COND_LCSSA:%.*]] = phi i32 [ [[COND]], [[COND_END]] ]
144+
; CHECK-NEXT: ret i32 [[COND_LCSSA]]
145+
;
146+
entry:
147+
br label %for.body
148+
149+
for.body:
150+
%i = phi i32 [ 0, %entry ], [ %inc, %cond.end ]
151+
%cmp = icmp eq i64 %y, 0
152+
br i1 %cmp, label %cond.end, label %cond.false
153+
154+
cond.false:
155+
br label %cond.end
156+
157+
cond.end:
158+
%cond = phi i32 [ 55, %cond.false ], [ %i, %for.body ]
159+
%inc = add nuw nsw i32 %i, 1
160+
%exitcond = icmp eq i32 %inc, 3
161+
br i1 %exitcond, label %for.cond.cleanup, label %for.body
162+
163+
for.cond.cleanup:
164+
ret i32 %cond
165+
}

0 commit comments

Comments
 (0)