[Mlir-commits] [mlir] [MLIR][Affine] Avoid crash on non-affine memory effects (PR #205865)

[llvmlistbot at llvm.org]

mygitljf wrote:

Thanks for your time and review!
Yes, the operand fallback is intentionally conservative. For ops that do not model MemoryEffectOpInterface and are not recursively memory-effect-free, we cannot distinguish a memref operand that is merely carried from one that may be accessed, e.g. unknown/custom or call-like ops. This can introduce extra dependence edges and may prevent some fusion, but it keeps the analysis sound and avoids assuming such ops are memory-effect free.
I'll add a short comment to make that explicit.

https://github.com/llvm/llvm-project/pull/205865