[Mlir-commits] [mlir] [mlir][arith.constant]Fix element type of the dense attributes in target attributes to be consistent with result type in LLVM::detail::oneToOneRewrite() (PR #149787)

[Diego Caballero]

dcaballe wrote:

> We could probably change the behavior for these unsupported floats to do the bitcasts during convert-to-llvm, but I've got the rough sense that this has a readability impact ... though I wouldn't be too opposed to that change (which is, IIRC, this PR)

Thanks for elaborating. I didn't know about the translation bit. This change sounds reasonable to me. I think we would have to change this PR so that the element type conversion happens for any `ShapedType`, including tensor types. We should also add a test for that if we don't have any.

I'd also suggest that we have the PR strengthening the element type verification of the `llvm.mlir.constant` op handy before landing this... to make sure we are not missing anything...

Pinging also @ftynse and @joker-eph in case they have any feedback.

https://github.com/llvm/llvm-project/pull/149787