[Mlir-commits] [mlir] [MLIR][Linalg] Use Top-Down traversal to safely optimize multi-use producer fusion (PR #172216)

[llvmlistbot at llvm.org]

MaheshRavishankar wrote:

Thanks for the PR. I think previously when I tried switching for the cases where we dont hit such issues (and I didnt hit this issue), the top-down or bottom-up traversal didnt make a difference. I think we need to look into compilation time impacts that we arent setup to do upstream.
Also you are only fixing a test here though.... Can you give a bit more context on how you are fixing this for your use case internally. Are you just updating the pattern application on your end. Could you update the documentation of `linalg::fuseElementwiseOps` ([here](https://github.com/llvm/llvm-project/blob/4ab36d8674f39725668722025097cd965dd26402/mlir/include/mlir/Dialect/Linalg/Transforms/Transforms.h#L565)) to indicate this.

The PR itself seems good to go for me.

https://github.com/llvm/llvm-project/pull/172216