[llvm] AMDGPU: Add round-to-odd rounding during f64 to bf16 conversion (PR #133995)

[Changpeng Fang via llvm-commits]

changpeng wrote:

> > GCNPat<(v2bf16 (bf16_fpround v2f64:$src))
> 
> Right, we should remove these incorrect patterns. The work is done

Just realized that we should still need these patterns for unsafe fast math. Will take further look and may put them back. Thanks.

https://github.com/llvm/llvm-project/pull/133995