[Mlir-commits] [mlir] [mlir][Transforms] Dialect Conversion: Fix folder implementation (PR #150775)

[Matthias Springer]

matthias-springer wrote:

I am missing some context here. Is the folder-undo modifying the module? What is the race condition here? And does the verifier run during the dialect conversion? That seems a bit unusual.

> However, is it not allowed to access a parent op (ModuleOp) in the verification of a nested op?

What's not allowed: Accessing the parent op (ModuleOp) in a pass that operates on a nested op (function, etc.). That can cause race conditions because the pass processes the functions in parallel. (Maybe just reading is fine, but there seems to be both read + write here.) I'm not quite sure what that means for the design of a verifier.

`SymbolUserOpInterface` sounds like the right approach to me. For reference, see https://reviews.llvm.org/D89512: one of the reason why this interface was added was to avoid race conditions.


https://github.com/llvm/llvm-project/pull/150775