• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: opt -basicaa -loop-accesses -analyze < %s | FileCheck %s
2; RUN: opt -passes='require<scalar-evolution>,require<aa>,loop(print-access-info)' -disable-output  < %s 2>&1 | FileCheck %s
3
4; This loop:
5;
6;   int **A;
7;   for (i)
8;     for (j) {
9;        A[i][j] = A[i-1][j] * B[j]
10;        B[j+1] = 2       // backward dep between this and the previous
11;     }
12;
13; is transformed by Load-PRE to stash away A[i] for the next iteration of the
14; outer loop:
15;
16;   Curr = A[0];          // Prev_0
17;   for (i: 1..N) {
18;     Prev = Curr;        // Prev = PHI (Prev_0, Curr)
19;     Curr = A[i];
20;     for (j: 0..N) {
21;        Curr[j] = Prev[j] * B[j]
22;        B[j+1] = 2       // backward dep between this and the previous
23;     }
24;   }
25;
26; Since A[i] and A[i-1] are likely to be independent, getUnderlyingObjects
27; should not assume that Curr and Prev share the same underlying object.
28;
29; If it did we would try to dependence-analyze Curr and Prev and the analysis
30; would fail with non-constant distance.
31;
32; To illustrate one of the negative consequences of this, if the loop has a
33; backward dependence we won't detect this but instead fully fall back on
34; memchecks (that is what LAA does after encountering a case of non-constant
35; distance).
36
37target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
38target triple = "x86_64-apple-macosx10.10.0"
39
40; CHECK: for_j.body:
41; CHECK-NEXT: Report: unsafe dependent memory operations in loop
42; CHECK-NEXT: Dependences:
43; CHECK-NEXT: Backward:
44; CHECK-NEXT: %loadB = load i8, i8* %gepB, align 1 ->
45; CHECK-NEXT: store i8 2, i8* %gepB_plus_one, align 1
46
47define void @f(i8** noalias %A, i8* noalias %B, i64 %N) {
48for_i.preheader:
49  %prev_0 = load i8*, i8** %A, align 8
50  br label %for_i.body
51
52for_i.body:
53  %i = phi i64 [1, %for_i.preheader], [%i.1, %for_j.end]
54  %prev = phi i8* [%prev_0, %for_i.preheader], [%curr, %for_j.end]
55  %gep = getelementptr inbounds i8*, i8** %A, i64 %i
56  %curr = load i8*, i8** %gep, align 8
57  br label %for_j.preheader
58
59for_j.preheader:
60  br label %for_j.body
61
62for_j.body:
63  %j = phi i64 [0, %for_j.preheader], [%j.1, %for_j.body]
64
65  %gepPrev = getelementptr inbounds i8, i8* %prev, i64 %j
66  %gepCurr = getelementptr inbounds i8, i8* %curr, i64 %j
67  %gepB = getelementptr inbounds i8, i8* %B, i64 %j
68
69  %loadPrev = load i8, i8* %gepPrev, align 1
70  %loadB = load i8, i8* %gepB, align 1
71
72  %mul = mul i8 %loadPrev, %loadB
73
74  store i8 %mul, i8* %gepCurr, align 1
75
76  %gepB_plus_one = getelementptr inbounds i8, i8* %gepB, i64 1
77  store i8 2, i8* %gepB_plus_one, align 1
78
79  %j.1 = add nuw i64 %j, 1
80  %exitcondj = icmp eq i64 %j.1, %N
81  br i1 %exitcondj, label %for_j.end, label %for_j.body
82
83for_j.end:
84
85  %i.1 = add nuw i64 %i, 1
86  %exitcond = icmp eq i64 %i.1, %N
87  br i1 %exitcond, label %for_i.end, label %for_i.body
88
89for_i.end:
90  ret void
91}
92