[PATCH] D96639: [AMDGPU] Add two TSFlags: IsAtomicNoRtn and IsAtomicRtn

[Stanislav Mekhanoshin via Phabricator via llvm-commits]

rampitec added a comment.

In D96639#2561212 <https://reviews.llvm.org/D96639#2561212>, @rampitec wrote:

> @t-tye memory legalizer changes are due to the fact that SIMemoryLegalizer::isAtomicRet() was returning false for the ds_wrxchg_rtn_b32.
>
> Please verify if the tests now look correctly or not. I am checking memory model for GFX10 and I do not see vmcnt required for these situations, but I do not see vscnt which was generated before there as well. Looks like this has uncovered a bug in either legalizer (likely) or memory model description. On top of that I do not see memory model describing atomicrmw acquire agent local combination.

I.e. I understand why vscnt is replaced with vmcnt, it was treated as store and now as load. This is fine. I don't understand what does it do with vmem at all.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D96639/new/

https://reviews.llvm.org/D96639