[PATCH] D133203: [TypePromotionPass] Don't treat phi's as ToPromote

[Dave Green via Phabricator via llvm-commits]

dmgreen added a comment.

In D133203#3769617 <https://reviews.llvm.org/D133203#3769617>, @samparker wrote:

>> The atomic loops are only expected to be executed once most of the time.
>
> A load compare loop is going to occur more often that just atomic operations though, such as a basic search loop. I was surprised that this didn't come up in the TypePromotion tests, so I have added some <https://github.com/llvm/llvm-project/commit/e893345589c69efd8ec2b6048b6e70b3dee808be>.

OK. Hmm. Those tests don't seem to be effected by this patch. I presume they were meant to be? I've added some other tests that were.

I've been trying to think what to do about it. Best I have so far is to check if there are sinks inside the loops with sources outside. That way we are (potentially) pushing more extends out of the loop, in a way that ISel will do better with.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D133203/new/

https://reviews.llvm.org/D133203