[PATCH] D153537: [AMDGPU] Do not wait for vscnt on function entry and return

[Austin Kerbow via Phabricator via llvm-commits]

kerbowa added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp:1231
       !ST->hasAutoWaitcntBeforeBarrier() && !ST->supportsBackOffBarrier()) {
-    Wait = Wait.combined(allZeroWaitcnt());
+    Wait = Wait.combined(AMDGPU::Waitcnt::allZero(ST->hasVscnt()));
   }
----------------
foad wrote:
> @kerbowa unlike the rest of SIInsertWaitcnts, I assume this part does want to wait for vscnt==0 since it is handling memory dependencies?
Technically it should not matter on HW with VScnt since they all can back off barriers, so this 'if' should never be true on Navi. There is an exception currently with gfx11 because of the memory model description bug with cumode, but that //should// be temporary.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D153537/new/

https://reviews.llvm.org/D153537