| Matt Arsenault | aa6bdf9 | 2019-06-12 14:05:58 +0000 | [diff] [blame] | 1 | ; RUN: opt -basicaa -loop-versioning -S < %s | FileCheck %s |
| 2 | |
| 3 | target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128" |
| 4 | |
| 5 | ; Do not version this loop because of a convergent operation |
| 6 | |
| 7 | ; CHECK-LABEL: @f( |
| 8 | ; CHECK: call i32 @llvm.convergent( |
| 9 | ; CHECK-NOT: call i32 @llvm.convergent( |
| 10 | define void @f(i32* %a, i32* %b, i32* %c) #0 { |
| 11 | entry: |
| 12 | br label %for.body |
| 13 | |
| 14 | for.body: ; preds = %for.body, %entry |
| 15 | %ind = phi i64 [ 0, %entry ], [ %add, %for.body ] |
| 16 | |
| 17 | %arrayidxA = getelementptr inbounds i32, i32* %a, i64 %ind |
| 18 | %loadA = load i32, i32* %arrayidxA, align 4 |
| 19 | |
| 20 | %arrayidxB = getelementptr inbounds i32, i32* %b, i64 %ind |
| 21 | %loadB = load i32, i32* %arrayidxB, align 4 |
| 22 | %convergentB = call i32 @llvm.convergent(i32 %loadB) |
| 23 | |
| 24 | %mulC = mul i32 %loadA, %convergentB |
| 25 | |
| 26 | %arrayidxC = getelementptr inbounds i32, i32* %c, i64 %ind |
| 27 | store i32 %mulC, i32* %arrayidxC, align 4 |
| 28 | |
| 29 | %add = add nuw nsw i64 %ind, 1 |
| 30 | %exitcond = icmp eq i64 %add, 20 |
| 31 | br i1 %exitcond, label %for.end, label %for.body |
| 32 | |
| 33 | for.end: ; preds = %for.body |
| 34 | ret void |
| 35 | } |
| 36 | |
| 37 | declare i32 @llvm.convergent(i32) #1 |
| 38 | |
| 39 | attributes #0 = { nounwind convergent } |
| 40 | attributes #1 = { nounwind readnone convergent } |