[Mlir-commits] [mlir] [mlir] Optimize `isOpTriviallyDead` greedy rewriter checks for deeply nested IR (PR #231423)

[Ivan Butygin]

Hardcode84 wrote:

> I wonder about things where someone would tag something in the type system (like adding a "read-only" metadata on a custom memref) which would change the memory effect of every users. The cache wouldn't be correct here...

Not sure I fully follow. Changing value type involves changing op result/block arg type in place, which is already discouraged in rewrite patterns I think. We can potentially handle it by invalidating results/block args consumers from `modifyOpInPlace` but should we?

https://github.com/llvm/llvm-project/pull/231423