[clang] [llvm] [AMDGPU] Use 1-iteration Newton-Raphson refinement for FP32 reciprocal. (PR #194716)

[Jay Foad via cfe-commits]

jayfoad wrote:

> > Is it possible to handle subnormal arguments with scaling, just like you handle large normal arguments with scaling?
> 
> I added it in the latest version of this patch. It slows down the best case from 50% to 27% due to the extra complexity of dealing with subnormals (extra selects).

Oh I see you added the select to handle subnormals, but you still have the conditional branch to select the new or old path. Why do you need that? Are there still some inputs that the new path does not handle correctly?

https://github.com/llvm/llvm-project/pull/194716