| Adam Nemet | b0c4eae | 2016-03-17 20:32:37 +0000 | [diff] [blame] | 1 | ; RUN: opt -basicaa -loop-vectorize -force-vector-width=2 \ |
| 2 | ; RUN: -force-vector-interleave=1 -S < %s \ |
| 3 | ; RUN: | FileCheck %s -check-prefix=BOTH -check-prefix=LV |
| 4 | ; RUN: opt -basicaa -scoped-noalias -loop-vectorize -dse -force-vector-width=2 \ |
| 5 | ; RUN: -force-vector-interleave=1 -S < %s \ |
| 6 | ; RUN: | FileCheck %s -check-prefix=BOTH -check-prefix=DSE |
| 7 | |
| 8 | target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128" |
| 9 | |
| 10 | ; This loop needs to be versioned with memchecks between {A, B} x {C} before |
| 11 | ; it can be vectorized. |
| 12 | ; |
| 13 | ; for (i = 0; i < n; i++) { |
| 14 | ; C[i] = A[i] + 1; |
| 15 | ; C[i] += B[i]; |
| 16 | ; } |
| 17 | ; |
| 18 | ; Check that the corresponding noalias metadata is added to the vector loop |
| 19 | ; but not to the scalar loop. |
| 20 | ; |
| 21 | ; Since in the versioned vector loop C and B can no longer alias, the first |
| 22 | ; store to C[i] can be DSE'd. |
| 23 | |
| 24 | |
| 25 | define void @f(i32* %a, i32* %b, i32* %c) { |
| 26 | entry: |
| 27 | br label %for.body |
| 28 | |
| 29 | ; BOTH: vector.memcheck: |
| 30 | ; BOTH: vector.body: |
| 31 | for.body: ; preds = %for.body, %entry |
| 32 | %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ] |
| 33 | |
| 34 | %arrayidxA = getelementptr inbounds i32, i32* %a, i64 %ind |
| 35 | ; Scope 1 |
| 36 | ; LV: = load {{.*}} !alias.scope !0 |
| 37 | %loadA = load i32, i32* %arrayidxA, align 4 |
| 38 | |
| 39 | %add = add nuw i32 %loadA, 2 |
| 40 | |
| 41 | %arrayidxC = getelementptr inbounds i32, i32* %c, i64 %ind |
| 42 | ; Noalias with scope 1 and 6 |
| 43 | ; LV: store {{.*}} !alias.scope !3, !noalias !5 |
| 44 | ; DSE-NOT: store |
| 45 | store i32 %add, i32* %arrayidxC, align 4 |
| 46 | |
| 47 | %arrayidxB = getelementptr inbounds i32, i32* %b, i64 %ind |
| 48 | ; Scope 6 |
| 49 | ; LV: = load {{.*}} !alias.scope !7 |
| 50 | %loadB = load i32, i32* %arrayidxB, align 4 |
| 51 | |
| 52 | %add2 = add nuw i32 %add, %loadB |
| 53 | |
| 54 | ; Noalias with scope 1 and 6 |
| 55 | ; LV: store {{.*}} !alias.scope !3, !noalias !5 |
| 56 | ; DSE: store |
| 57 | store i32 %add2, i32* %arrayidxC, align 4 |
| 58 | |
| 59 | %inc = add nuw nsw i64 %ind, 1 |
| 60 | %exitcond = icmp eq i64 %inc, 20 |
| 61 | br i1 %exitcond, label %for.end, label %for.body |
| 62 | |
| 63 | ; BOTH: for.body: |
| 64 | ; BOTH-NOT: !alias.scope |
| 65 | ; BOTH-NOT: !noalias |
| 66 | |
| 67 | for.end: ; preds = %for.body |
| 68 | ret void |
| 69 | } |
| 70 | |
| 71 | ; LV: !0 = !{!1} |
| 72 | ; LV: !1 = distinct !{!1, !2} |
| 73 | ; LV: !2 = distinct !{!2, !"LVerDomain"} |
| 74 | ; LV: !3 = !{!4} |
| 75 | ; LV: !4 = distinct !{!4, !2} |
| 76 | ; LV: !5 = !{!1, !6} |
| 77 | ; LV: !6 = distinct !{!6, !2} |
| 78 | ; LV: !7 = !{!6} |