[Mlir-commits] [mlir] [MLIR][OpenMP] Support basic materialization for `omp.private` ops (PR #81715)

[llvmlistbot at llvm.org]

NimishMishra wrote:

> If preferred, I can go back to doing that and then undoing this remapping after the conversion. This way the op is not cloned and at the same time is restored to its original state after lowering.

If this approach is taken, I think it implies the following. Correct me please if my understanding is incorrect.

1. If the remapped operation is short-lived (i.e if the number of transformations/checks we perform on the remapped operation are reasonably small), then the possibility of unforeseen problems arising with the remapped operation would be less. I think @skatrak had a comment on hard-to-debug problems with this.

2. By not explicitly cloning the parallel op, we would be avoiding perf overheads as @kiranchandramohan mentioned. I am not sure, but the perf overhead of remapping and undoing the same would be less than cloning? Is it the case?

   This is of course if the privCB is not called after outlining, as Kiran mentioned.

https://github.com/llvm/llvm-project/pull/81715