• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: opt -basicaa -loop-distribute -enable-loop-distribute -S < %s | \
2; RUN:     FileCheck %s --check-prefix=ALL --check-prefix=STRIDE_SPEC
3
4; RUN: opt -basicaa -loop-distribute -enable-loop-distribute -S -enable-mem-access-versioning=0 < %s | \
5; RUN:     FileCheck %s --check-prefix=ALL --check-prefix=NO_STRIDE_SPEC
6
7; If we don't speculate stride for 1 we can't distribute along the line
8; because we could have a backward dependence:
9;
10;   for (i = 0; i < n; i++) {
11;     A[i + 1] = A[i] * B[i];
12;     =======================
13;     C[i] = D[i] * A[stride * i];
14;   }
15
16target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
17target triple = "x86_64-apple-macosx10.10.0"
18
19; ALL-LABEL: @f(
20define void @f(i32* noalias %a,
21               i32* noalias %b,
22               i32* noalias %c,
23               i32* noalias %d,
24               i64 %stride) {
25entry:
26  br label %for.body
27
28; STRIDE_SPEC: %ident.check = icmp ne i64 %stride, 1
29
30; STRIDE_SPEC: for.body.ldist1:
31; NO_STRIDE_SPEC-NOT: for.body.ldist1:
32
33for.body:                                         ; preds = %for.body, %entry
34  %ind = phi i64 [ 0, %entry ], [ %add, %for.body ]
35
36  %arrayidxA = getelementptr inbounds i32, i32* %a, i64 %ind
37  %loadA = load i32, i32* %arrayidxA, align 4
38
39  %arrayidxB = getelementptr inbounds i32, i32* %b, i64 %ind
40  %loadB = load i32, i32* %arrayidxB, align 4
41
42  %mulA = mul i32 %loadB, %loadA
43
44  %add = add nuw nsw i64 %ind, 1
45  %arrayidxA_plus_4 = getelementptr inbounds i32, i32* %a, i64 %add
46  store i32 %mulA, i32* %arrayidxA_plus_4, align 4
47
48  %arrayidxD = getelementptr inbounds i32, i32* %d, i64 %ind
49  %loadD = load i32, i32* %arrayidxD, align 4
50
51  %mul = mul i64 %ind, %stride
52  %arrayidxStridedA = getelementptr inbounds i32, i32* %a, i64 %mul
53  %loadStridedA = load i32, i32* %arrayidxStridedA, align 4
54
55  %mulC = mul i32 %loadD, %loadStridedA
56
57  %arrayidxC = getelementptr inbounds i32, i32* %c, i64 %ind
58  store i32 %mulC, i32* %arrayidxC, align 4
59
60  %exitcond = icmp eq i64 %add, 20
61  br i1 %exitcond, label %for.end, label %for.body
62
63for.end:                                          ; preds = %for.body
64  ret void
65}
66