[Mlir-commits] [mlir] [MLIR][Transforms] add eliminate-explicit-rounding pass (PR #93443)
Ivy Zhang
llvmlistbot at llvm.org
Thu Jun 6 00:50:21 PDT 2024
crazydemo wrote:
> I think it makes sense to modify the PR slightly so that it only removes the extf/truncf pair if they have the "fastmath=contract" attribute, and followed by future PRs which mark these attributes for extf/truncf op introduced by various promotion passes. This is exactly what suggested by [#93443 (comment)](https://github.com/llvm/llvm-project/pull/93443#issuecomment-2138154222).
>
> This mean that we all agree to extend the "fastmath=contract" semantics to MLIR extf/truncf op and allow MLIR optimization to remove the rounding effects associated with these two ops.
>
> We also need to consider whether and how user specified compiler options on "fastmath" impacts the promotion passes to set "fastmath=contract" for extf/truncf. But I think it is a global issue and need to discussed in separate thread (or please let us know if there is already such one!).
I agree with your comments, and I will update the approach.
https://github.com/llvm/llvm-project/pull/93443
More information about the Mlir-commits
mailing list