1; RUN: opt -loop-accesses -analyze < %s | FileCheck %s
2
3; Check that loop-indepedent forward dependences are discovered properly.
4;
5; FIXME: This does not actually always work which is pretty confusing.  Right
6; now there is hack in LAA that tries to figure out loop-indepedent forward
7; dependeces *outside* of the MemoryDepChecker logic (i.e. proper dependence
8; analysis).
9;
10; Therefore if there is only loop-independent dependences for an array
11; (i.e. the same index is used), we don't discover the forward dependence.
12; So, at ***, we add another non-I-based access of A to trigger
13; MemoryDepChecker analysis for accesses of A.
14;
15;   for (unsigned i = 0; i < 100; i++) {
16;     A[i + 1] = B[i] + 1;   // ***
17;     A[i] = B[i] + 2;
18;     C[i] = A[i] * 2;
19;   }
20
21target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
22
23define void @f(i32* noalias %A, i32* noalias %B, i32* noalias %C, i64 %N) {
24
25; CHECK: Dependences:
26; CHECK-NEXT:   Forward:
27; CHECK-NEXT:       store i32 %b_p1, i32* %Aidx, align 4 ->
28; CHECK-NEXT:       %a = load i32, i32* %Aidx, align 4
29; CHECK:        ForwardButPreventsForwarding:
30; CHECK-NEXT:       store i32 %b_p2, i32* %Aidx_next, align 4 ->
31; CHECK-NEXT:       %a = load i32, i32* %Aidx, align 4
32; CHECK:        Forward:
33; CHECK-NEXT:       store i32 %b_p2, i32* %Aidx_next, align 4 ->
34; CHECK-NEXT:       store i32 %b_p1, i32* %Aidx, align 4
35
36entry:
37  br label %for.body
38
39for.body:                                         ; preds = %for.body, %entry
40  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
41  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
42
43  %Bidx = getelementptr inbounds i32, i32* %B, i64 %indvars.iv
44  %Cidx = getelementptr inbounds i32, i32* %C, i64 %indvars.iv
45  %Aidx_next = getelementptr inbounds i32, i32* %A, i64 %indvars.iv.next
46  %Aidx = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
47
48  %b = load i32, i32* %Bidx, align 4
49  %b_p2 = add i32 %b, 1
50  store i32 %b_p2, i32* %Aidx_next, align 4
51
52  %b_p1 = add i32 %b, 2
53  store i32 %b_p1, i32* %Aidx, align 4
54
55  %a = load i32, i32* %Aidx, align 4
56  %c = mul i32 %a, 2
57  store i32 %c, i32* %Cidx, align 4
58
59  %exitcond = icmp eq i64 %indvars.iv.next, %N
60  br i1 %exitcond, label %for.end, label %for.body
61
62for.end:                                          ; preds = %for.body
63  ret void
64}
65