[Mlir-commits] [mlir] [MLIR] Add single definition multiple regions for mem2reg (PR #89107)

[Christian Ulmann]

Dinistro wrote:

> Right. I think for mem2reg liveness and dominance are somewhat similar, at least in the scope of what this mem2reg infrastructure was designed for. If we successfully model dominance in regions, this should provide enough information to compute a useful amount of variable tracking for mem2reg.

While this is a non-trivial task, it might be possible to somehow generalize dominance for this. Unfortunately, one will also require to design interfaces that allow adding results to regions in some kind of generic way. I do not have any concrete idea about how to achieve this goal at the moment, so such a Mem2Reg generalization is not something we will see in due time.

@fanfuqiang are you lowering back to LLVM dialect at some point? If so, can you not just apply mem2reg on that level for now?

https://github.com/llvm/llvm-project/pull/89107