[llvm] AMDGPU][True16][CodeGen] FP_Round f64 to f16 in true16 (PR #128911)

[Brox Chen via llvm-commits]

broxigarchen wrote:

> I don't understand this change. You shouldn't have to make any of this code conditional on true16. Are you just avoiding the FP_TO_FP16? We can and should do that independently of true16

Hi Matt. Sorry for the delay. Just checked this PR (I am not the original author). We basically do FP64 lowering differently between true16 and fake16 mode.

The unsafe mode "F64 to F16" is lowered differently between true16/fake16 flow. Fake16 mode lowered to FP_TO_FP16  while in true16 mode it get lowered to two FP_ROUND. (This patch https://reviews.llvm.org/D154528 stops from combining these two FP_ROUND back)

The post lowering operation is also changed as we replace fptrunc to bitcast in true16 mode



https://github.com/llvm/llvm-project/pull/128911