| Adam Nemet | 32e6a34 | 2016-12-21 04:07:40 +0000 | [diff] [blame] | 1 | ; RUN: opt -loop-distribute -enable-loop-distribute -S -verify-loop-info -verify-dom-info < %s \ |
| Adam Nemet | 938d3d6 | 2015-05-14 12:05:18 +0000 | [diff] [blame] | 2 | ; RUN: | FileCheck %s |
| 3 | |
| 4 | ; Distributing this loop to avoid the dependence cycle would require to |
| 5 | ; reorder S1 and S2 to form the two partitions: {S2} | {S1, S3}. The analysis |
| 6 | ; provided by LoopAccessAnalysis does not allow us to reorder memory |
| 7 | ; operations so make sure we bail on this loop. |
| 8 | ; |
| 9 | ; for (i = 0; i < n; i++) { |
| 10 | ; S1: d = D[i]; |
| 11 | ; S2: A[i + 1] = A[i] * B[i]; |
| 12 | ; S3: C[i] = d * E[i]; |
| 13 | ; } |
| 14 | |
| 15 | target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128" |
| 16 | target triple = "x86_64-apple-macosx10.10.0" |
| 17 | |
| 18 | define void @f(i32* noalias %a, |
| 19 | i32* noalias %b, |
| 20 | i32* noalias %c, |
| 21 | i32* noalias %d, |
| 22 | i32* noalias %e) { |
| 23 | entry: |
| 24 | br label %for.body |
| 25 | |
| 26 | ; CHECK: entry: |
| 27 | ; CHECK: br label %for.body |
| 28 | ; CHECK: for.body: |
| 29 | ; CHECK: br i1 %exitcond, label %for.end, label %for.body |
| 30 | ; CHECK: for.end: |
| 31 | ; CHECK: ret void |
| 32 | |
| 33 | for.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 | %arrayidxD = getelementptr inbounds i32, i32* %d, i64 %ind |
| 45 | %loadD = load i32, i32* %arrayidxD, align 4 |
| 46 | |
| 47 | %add = add nuw nsw i64 %ind, 1 |
| 48 | %arrayidxA_plus_4 = getelementptr inbounds i32, i32* %a, i64 %add |
| 49 | store i32 %mulA, i32* %arrayidxA_plus_4, align 4 |
| 50 | |
| 51 | %arrayidxC = getelementptr inbounds i32, i32* %c, i64 %ind |
| 52 | |
| 53 | %arrayidxE = getelementptr inbounds i32, i32* %e, i64 %ind |
| 54 | %loadE = load i32, i32* %arrayidxE, align 4 |
| 55 | |
| 56 | %mulC = mul i32 %loadD, %loadE |
| 57 | |
| 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 | |
| 63 | for.end: ; preds = %for.body |
| 64 | ret void |
| 65 | } |