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

Wed Feb 9 11:44:10 PST 2022

lebedev.ri added a comment.

Hi,

In D114779#3306221 <https://reviews.llvm.org/D114779#3306221>, @bmahjour wrote:

> Even though the initial intention of the "hack" might have been to prevent predication on some X86 subtargets, it was added in a way that affected all targets, so it could expose cost model issues on various targets if suddenly removed. I wasn't aware of this and just noticed it....has any performance runs been done on other targets (eg. Power)?

In D114779#3308611 <https://reviews.llvm.org/D114779#3308611>, @dmgreen wrote:

> I'm interested - what kind of performance measuring did you do for this patch, and what were the numbers it gave?

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.

If you have identified one of such places, please feel free to file a bug,
and optionally ask for a temporary revert until you've had a chance to fix said bug.
Though, it would be best to just fix the uncovered issues, it's not like the branch is soon.

Repository:
  rG LLVM Github Monorepo

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

https://reviews.llvm.org/D114779