Reuse of mlir::AliasAnalysis in flang

[Renaud-K]

For context, there have been a couple of dormant discussions on alias analysis so, as a heads-up, I was encouraged to start a new one.

Alias Analysis Framework in MLIR

I am going to add an aliasing analysis for FIR (Alias analysis in LLVM Flang - #16 by Renaud-K) I was going try reuse the AliasAnalysis infrastructure but I believe I see an issue The LocalAliasAnalysis is always installed and is the first AliasAnalysis::AliasAnalysis(Operation *op) { addAnalysisImplementation(LocalAliasAnalysis()); } It ignores Allocate and Free effects and starts with the optimistic result noModRef. So it would come out returning noModRef for an operation performing eith…

Alias analysis in LLVM Flang

I would like to add some initial coarse aliasing analysis based on FIR. It will heavily rely on FIR types, attributes and operations. It’s goal would be to help reduce the creation of temporary arrays be used as a building block for other analyses such as global code motion help in pushing aliasing information down to the LLVM level https://reviews.llvm.org/D135901

Flang is planning on reusing the infrastructure in place for alias analysis.
I believe what is there is a good start and we will attempt to combine alias analyses with mlir::AliasAnalysis::addAnalysisImplementation.

I currently have this up for review ⚙ D136889 Infrastructure for alias analysis in flang, which touches MLIR. It does not make any substantial changes though other than move code for reuse.

I also have in mind a few changes that are more substantial, that I would like to discuss at some point, once we adopt this change. But that would be a different topic.

-Renaud