[Mlir-commits] [mlir] [mlir][bufferization] Hoist allocations through exactly-once regions (PR #224921)

[purnima shrivastava]

purnima-nlp wrote:

> I'm writing this on my phone, but how does this fix relate to this issue #223944? Does it resolve both?

I tested the #223944 reproducer with this patch, and it does not resolve that issue.

I think the reason is this PR, the missed hoist happens because the placement walk stops at a non-loop region (`scf.execute_region`). The fix allows traversal through regions that are guaranteed to execute exactly once.

For #223944, `scf.while` is already recognized as a loop, and the allocation is still rejected by the existing alias/back-edge check because the loop exit result contributes to the alias dominator. So that issue still needs the separate exit-edge/alias-set refinement discussed there.


https://github.com/llvm/llvm-project/pull/224921