[Mlir-commits] [flang] [mlir] [mlir][AliasAnalysis] Add opt-in query caching; use it in flang LICM (PR #207030)

[Kareem Ergawy]

https://github.com/ergawy commented:

Thanks Slava, I have a question about the design. Can we make the cache itself RAII-based and not just its scope?

The main advantages are:
1. We decouple the cache from AA's API and do not have to extend MLIR's AA for it.                                   
2. We already had to add a RAII scoping component: QueryCacheScope, so I think moving the resource management to the cache itself away from the analysis makes sense.
3. Maybe in the future we can implement different caches (for different query results) all implementing a shared interface: `AAResultCache` and can augment AA with customized caches based on a particular client's needs.
 
Does that make sense in your opinion?

https://github.com/llvm/llvm-project/pull/207030