================
@@ -44,23 +44,19 @@ MDNode *LoopInfo::createPipeliningMetadata(const
LoopAttributes &Attrs,
else if (Attrs.PipelineInitiationInterval != 0)
Enabled = true;
+ SmallVector<Metadata *, 4> Args;
+ Args.append(LoopProperties.begin(), LoopProperties.end());
+
if (Enabled != true) {
- SmallVector<Metadata *, 4> NewLoopProperties;
if (Enabled == false) {
- NewLoopProperties.append(LoopProperties.begin(), LoopProperties.end());
----------------
kasuga-fj wrote:
> In principle, LoopVectorize should know that if the original loop had a
> progress guarantee, then the vectorized loop will as well, so it should add
> `llvm.loop.must_progress` no matter what.
I think this is completely correct. What I didn't understand is, why the
followup metadata of `LOOP_6` (`FOLLOW_VECTOR_6`) didn't have
`llvm.mustprogress` before this patch, but now it (`FOLLOWUP_VECTOR_3`) does.
I investigated a little deeper and found the cause; `FOLLOWUP_VECTOR_6`
actually had `mustprogress` (?!). That is, the test passed for both of the
following directives.
```
// Original.
// CHECK: ![[AFTER_VECTOR_6]] = distinct !{![[AFTER_VECTOR_6]],
![[ISVECTORIZED:.*]], ![[UNROLL_8:.*]]}
// This was also fine.
// CHECK: ![[AFTER_VECTOR_6]] = distinct !{![[AFTER_VECTOR_6]], [[MP]],
![[ISVECTORIZED:.*]], ![[UNROLL_8:.*]]}
```
Maybe FileCheck has a problem?
https://github.com/llvm/llvm-project/pull/131985
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
