[PATCH] D157599: [AMDGPU] Do not release VGPRs at -O0

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp:1039
     if (ST->getGeneration() >= AMDGPUSubtarget::GFX11 &&
+        MI.getMF()->getTarget().getOptLevel() != CodeGenOpt::None &&
         ScoreBrackets.getScoreRange(VS_CNT) != 0 &&
----------------
arsenm wrote:
> foad wrote:
> > arsenm wrote:
> > > should also account for optnone on the function
> > Is there a helper I can use that takes both of these into account? In the whole of the backend I only see one explicit check for F.hasOptNone(), in AMDGPURegBankSelect.
> I'm not sure, we don't have many of these skip optimizations in mandatory passes types of checks. Usually optnone is considered by skipFunction at the start of the pass, and the pass isn't added at -O0 to begin with
I'd just add an optnone member and set it at the start of the function from optnone/opt level


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D157599/new/

https://reviews.llvm.org/D157599