[PATCH] D84959: [NewPM][LVI] Abandon LVI after CVP

Fri Jul 31 13:24:19 PDT 2020

tejohnson added a comment.

In D84959#2188097 <https://reviews.llvm.org/D84959#2188097>, @aeubanks wrote:

> In D84959#2188060 <https://reviews.llvm.org/D84959#2188060>, @aeubanks wrote:
>
>> I had a random thought, maybe CVP itself shouldn't be invalidating LVI, but rather it should be explicitly invalidated in `PassBuilder::buildFunctionSimplificationPipeline()` with a `InvalidateAnalysisPass<LazyValueAnalysis>`. Since the description mentions that this is based on the knowledge that no passes after CVP will use LVI. ???
>
> I'm not super familiar with the details, but if LVI should really always created on demand then immediately invalidated after CVP, then this lgtm.

Yeah I was pondering the same thing (since it looks like CVP is invoked twice per function), but presumably intermediate passes may hit the update compile time issue described in D70376 <https://reviews.llvm.org/D70376>. So I think for now this is a good approach to address the regression?

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D84959/new/

https://reviews.llvm.org/D84959