Skip to content

Commit 86f655c

Browse files
fhahnfhahn
committed
[LAA] Add tests showing unnecessary RT check due to applying loop guards
Test courtesy to @bjope showing a regression due to ecae3ed.
1 parent 97e35e0 commit 86f655c

File tree

1 file changed

+140
-0
lines changed

1 file changed

+140
-0
lines changed

‎llvm/test/Analysis/LoopAccessAnalysis/is-safe-dep-distance-with-loop-guards.ll

Lines changed: 140 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,140 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --version 4
2+
; RUN: opt -passes='print<access-info>' -disable-output %s 2>&1 | FileCheck %s
3+
4+
target datalayout = "S16-p:16:16-i1:16-i8:8-i32:16-i64:16-i128:16"
5+
6+
define void @safe_deps_1_due_to_dependence_distance(i16 %n, ptr %p) {
7+
; CHECK-LABEL: 'safe_deps_1_due_to_dependence_distance'
8+
; CHECK-NEXT: loop:
9+
; CHECK-NEXT: Memory dependences are safe with run-time checks
10+
; CHECK-NEXT: Dependences:
11+
; CHECK-NEXT: Run-time memory checks:
12+
; CHECK-NEXT: Check 0:
13+
; CHECK-NEXT: Comparing group ([[GRP1:0x[0-9a-f]+]]):
14+
; CHECK-NEXT: %gep.iv = getelementptr inbounds i32, ptr %p, i16 %iv
15+
; CHECK-NEXT: Against group ([[GRP2:0x[0-9a-f]+]]):
16+
; CHECK-NEXT: %gep.off.iv = getelementptr i32, ptr %gep.off, i16 %iv
17+
; CHECK-NEXT: Grouped accesses:
18+
; CHECK-NEXT: Group [[GRP1]]:
19+
; CHECK-NEXT: (Low: %p High: ((4 * %n) + %p))
20+
; CHECK-NEXT: Member: {%p,+,4}<nuw><%loop>
21+
; CHECK-NEXT: Group [[GRP2]]:
22+
; CHECK-NEXT: (Low: ((4 * %n) + %p) High: ((8 * %n) + %p))
23+
; CHECK-NEXT: Member: {((4 * %n) + %p),+,4}<nw><%loop>
24+
; CHECK-EMPTY:
25+
; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
26+
; CHECK-NEXT: SCEV assumptions:
27+
; CHECK-EMPTY:
28+
; CHECK-NEXT: Expressions re-written:
29+
;
30+
entry:
31+
%cmp42 = icmp sgt i16 %n, 0
32+
br i1 %cmp42, label %ph, label %exit
33+
34+
ph:
35+
%gep.off = getelementptr i32, ptr %p, i16 %n
36+
br label %loop
37+
38+
loop:
39+
%iv = phi i16 [ 0, %ph ], [ %iv.next, %loop ]
40+
%gep.iv = getelementptr inbounds i32, ptr %p, i16 %iv
41+
store i32 0, ptr %gep.iv, align 1
42+
%gep.off.iv = getelementptr i32, ptr %gep.off, i16 %iv
43+
store i32 1, ptr %gep.off.iv, align 1
44+
%iv.next = add i16 %iv, 1
45+
%exitcond.not = icmp eq i16 %iv.next, %n
46+
br i1 %exitcond.not, label %exit, label %loop
47+
48+
exit:
49+
ret void
50+
}
51+
52+
define void @safe_deps_2_due_to_dependence_distance(i16 %n, ptr %p3, i16 noundef %q, ptr %p1, ptr %p2) {
53+
; CHECK-LABEL: 'safe_deps_2_due_to_dependence_distance'
54+
; CHECK-NEXT: loop:
55+
; CHECK-NEXT: Memory dependences are safe with run-time checks
56+
; CHECK-NEXT: Dependences:
57+
; CHECK-NEXT: Run-time memory checks:
58+
; CHECK-NEXT: Check 0:
59+
; CHECK-NEXT: Comparing group ([[GRP3:0x[0-9a-f]+]]):
60+
; CHECK-NEXT: %arrayidx22 = getelementptr inbounds [2 x i32], ptr %alloca, i16 %iv
61+
; CHECK-NEXT: Against group ([[GRP4:0x[0-9a-f]+]]):
62+
; CHECK-NEXT: %arrayidx33 = getelementptr inbounds i8, ptr %arrayidx22, i16 4
63+
; CHECK-NEXT: Check 1:
64+
; CHECK-NEXT: Comparing group ([[GRP3]]):
65+
; CHECK-NEXT: %arrayidx22 = getelementptr inbounds [2 x i32], ptr %alloca, i16 %iv
66+
; CHECK-NEXT: Against group ([[GRP5:0x[0-9a-f]+]]):
67+
; CHECK-NEXT: %arrayidx42 = getelementptr inbounds [2 x i32], ptr %arrayidx40, i16 %iv
68+
; CHECK-NEXT: Check 2:
69+
; CHECK-NEXT: Comparing group ([[GRP3]]):
70+
; CHECK-NEXT: %arrayidx22 = getelementptr inbounds [2 x i32], ptr %alloca, i16 %iv
71+
; CHECK-NEXT: Against group ([[GRP6:0x[0-9a-f]+]]):
72+
; CHECK-NEXT: %arrayidx53 = getelementptr inbounds i8, ptr %arrayidx42, i16 4
73+
; CHECK-NEXT: Check 3:
74+
; CHECK-NEXT: Comparing group ([[GRP4]]):
75+
; CHECK-NEXT: %arrayidx33 = getelementptr inbounds i8, ptr %arrayidx22, i16 4
76+
; CHECK-NEXT: Against group ([[GRP5]]):
77+
; CHECK-NEXT: %arrayidx42 = getelementptr inbounds [2 x i32], ptr %arrayidx40, i16 %iv
78+
; CHECK-NEXT: Check 4:
79+
; CHECK-NEXT: Comparing group ([[GRP4]]):
80+
; CHECK-NEXT: %arrayidx33 = getelementptr inbounds i8, ptr %arrayidx22, i16 4
81+
; CHECK-NEXT: Against group ([[GRP6]]):
82+
; CHECK-NEXT: %arrayidx53 = getelementptr inbounds i8, ptr %arrayidx42, i16 4
83+
; CHECK-NEXT: Check 5:
84+
; CHECK-NEXT: Comparing group ([[GRP5]]):
85+
; CHECK-NEXT: %arrayidx42 = getelementptr inbounds [2 x i32], ptr %arrayidx40, i16 %iv
86+
; CHECK-NEXT: Against group ([[GRP6]]):
87+
; CHECK-NEXT: %arrayidx53 = getelementptr inbounds i8, ptr %arrayidx42, i16 4
88+
; CHECK-NEXT: Grouped accesses:
89+
; CHECK-NEXT: Group [[GRP3]]:
90+
; CHECK-NEXT: (Low: %alloca High: (-4 + (8 * %n) + %alloca))
91+
; CHECK-NEXT: Member: {%alloca,+,8}<nuw><%loop>
92+
; CHECK-NEXT: Group [[GRP4]]:
93+
; CHECK-NEXT: (Low: (4 + %alloca) High: ((8 * %n) + %alloca))
94+
; CHECK-NEXT: Member: {(4 + %alloca),+,8}<nuw><%loop>
95+
; CHECK-NEXT: Group [[GRP5]]:
96+
; CHECK-NEXT: (Low: ((8 * %n) + %alloca) High: (-4 + (16 * %n) + %alloca))
97+
; CHECK-NEXT: Member: {((8 * %n) + %alloca),+,8}<nw><%loop>
98+
; CHECK-NEXT: Group [[GRP6]]:
99+
; CHECK-NEXT: (Low: (4 + (8 * %n) + %alloca) High: ((16 * %n) + %alloca))
100+
; CHECK-NEXT: Member: {(4 + (8 * %n) + %alloca),+,8}<nw><%loop>
101+
; CHECK-EMPTY:
102+
; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
103+
; CHECK-NEXT: SCEV assumptions:
104+
; CHECK-NEXT: {((8 * %n) + %alloca),+,8}<nw><%loop> Added Flags: <nusw>
105+
; CHECK-NEXT: {(4 + (8 * %n) + %alloca),+,8}<nw><%loop> Added Flags: <nusw>
106+
; CHECK-EMPTY:
107+
; CHECK-NEXT: Expressions re-written:
108+
;
109+
entry:
110+
%0 = shl i16 %n, 1
111+
%alloca = alloca [2 x i32], i16 %0
112+
%arrayidx1 = getelementptr inbounds i32, ptr %p1, i16 %q
113+
%arrayidx2 = getelementptr inbounds i8, ptr %p3, i16 2
114+
%arrayidx4 = getelementptr inbounds i32, ptr %p2, i16 %q
115+
%cmp42 = icmp sgt i16 %n, 0
116+
br i1 %cmp42, label %ph, label %exit
117+
118+
ph:
119+
%arrayidx40 = getelementptr inbounds [2 x i32], ptr %alloca, i16 %n
120+
br label %loop
121+
122+
loop:
123+
%iv = phi i16 [ 0, %ph ], [ %iv.next, %loop ]
124+
%arrayidx6 = getelementptr inbounds i32, ptr %arrayidx1, i16 %iv
125+
%arrayidx11 = getelementptr inbounds i32, ptr %arrayidx4, i16 %iv
126+
%arrayidx22 = getelementptr inbounds [2 x i32], ptr %alloca, i16 %iv
127+
store i32 10, ptr %arrayidx22
128+
%arrayidx33 = getelementptr inbounds i8, ptr %arrayidx22, i16 4
129+
store i32 16, ptr %arrayidx33
130+
%arrayidx42 = getelementptr inbounds [2 x i32], ptr %arrayidx40, i16 %iv
131+
store i32 19, ptr %arrayidx42
132+
%arrayidx53 = getelementptr inbounds i8, ptr %arrayidx42, i16 4
133+
store i32 23, ptr %arrayidx53
134+
%iv.next = add nuw nsw i16 %iv, 1
135+
%exitcond.not = icmp eq i16 %iv.next, %n
136+
br i1 %exitcond.not, label %exit, label %loop
137+
138+
exit:
139+
ret void
140+
}

0 commit comments

Comments
 (0)