[Mlir-commits] [mlir] [mlir][scf] Revisit affected loops during range folding (PR #230961)

[llvmlistbot at llvm.org]

Michael-Chen-NJU wrote:

> Thanks for the experiment. Can you address the `walk` comment and try again, looking at `applyOpPatternsGreedily`, `findCommonAncestor` can go quadratic if list on nested ops are passed.

Thanks! I've removed the explicit `walk` and switched to `applyPatternsGreedily` on the root op (`applyOpPatternsGreedily` with only the root doesn't visit nested `ForOp`s). All tests pass.

Unfortunately, the performance regression remains: the region-based greedy version took ~42s on the synthetic 1000-level nested-loop case, compared to ~0.2s with the targeted worklist.

Profiling the previous greedy version showed that `findCommonAncestor` took only ~4ms, while `isOpTriviallyDead` accounted for ~35s (~92% of runtime), due to repeated DCE checks on nested loops.

The region-based driver also removes unrelated dead operations. Would you prefer keeping the greedy approach despite these trade-offs, or reverting to the targeted worklist? Or is there any other ways to solve this regression?

https://github.com/llvm/llvm-project/pull/230961