[PATCH] D114779: [LV] Remove `LoopVectorizationCostModel::useEmulatedMaskMemRefHack()`

[Roman Lebedev via Phabricator via llvm-commits]

lebedev.ri added a comment.

Could you point me to the bugreport that you have opened before temporarily reverting this? :)

In D114779#3308717 <https://reviews.llvm.org/D114779#3308717>, @dmgreen wrote:

> OK Cool.
>
>> This is a correctness fix. It will not be at all surprising to learn that
>> some other architecture has started to unintentionally rely on this
>> erroneous behavior instead of implementing a correct and precise cost model.
>
> Sure. I just think that one of the architectures relying on this, at least in places, is X86.

Could you please be more specific?

> I was honestly surprised this didn't hurt performance in more places, I thought it was still pretty important. Hence me asking what kind of performance you had run.



> But it looks from these very tests that it's still needed for some things

Could you please be more specific? :)

> and I got reports of a few more. The X86 cases did seem to be hidden more by other costs, like the cost of a vector mul being 6 :)




Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D114779/new/

https://reviews.llvm.org/D114779