[Mlir-commits] [mlir] [MLIR][Math] add canonicalize-f32-promotion pass (PR #92482)

[Ivy Zhang]

crazydemo wrote:

> Could you make this an edit to `legalize-to-f32` instead? I think reasonable people can agree that (as seen over with the X86 backend changes) that removing redundant implicit rounding is fine.

Thanks for your comment. I add a `use-canonicalize-f32-promotion` option  to `legalize-to-f32` instread.

https://github.com/llvm/llvm-project/pull/92482