[PATCH] D121167: [NewPM] Actually recompute GlobalsAA before module optimization pipeline

Tue Mar 8 10:58:49 PST 2022

aeubanks added a comment.

In D121167#3366362 <https://reviews.llvm.org/D121167#3366362>, @nikic wrote:

> We have some places doing a Require<GlobalsAA>, Invalidate<AAManager> dance, can that be replaced with this as well? Or is the situation there that we do need to invalidate the AAManager because otherwise the new GlobalsAA just isn't going to be used?

Yes that's right, just recomputing GlobalsAA won't cause AAManager to pick it up. Here we're still using the same GlobalsAA instance that the AAManager already has a reference to.
I'll look into adding GlobalsAA into AAManager without invalidating anything.

================
Comment at: llvm/lib/Passes/PassBuilderPipelines.cpp:1169
   // memory operations.
-  MPM.addPass(RequireAnalysisPass<GlobalsAA, Module>());
+  MPM.addPass(RecomputeGlobalsAAPass());

----------------
fhahn wrote:
> 
> Is there a drawback to just doing `MPM.addPass(InvalidateAnalysisPass<GlobalsAA>());` before re-requiring `GlobalsAA`? At least for the test case, it looks like this is sufficient.
Yeah I briefly mentioned it in the summary, we end up having to invalidate any analysis that depends on any alias analysis. This way we don't have to do that.
(I got annoyed updating new-pm-defaults.ll and co with all the newly invalidated analyses, which might not even be deterministic in the order they're invalidated, which is why I went down this route)

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D121167/new/

https://reviews.llvm.org/D121167