[llvm] [X86] Remove LowerFCanonicalize and use generic expansion (PR #147877)

via llvm-commits llvm-commits at lists.llvm.org
Wed Jul 16 03:27:02 PDT 2025


woruyu wrote:

> I am not sure I understant this PR correct.
> 
> While just define a function `expandFCANONICALIZE`, and it do like
> 
> ```
> if (isLegal(STRICT_FMUL, TYPE))
>    DAG.getNode(ISD::STRICT_FMUL ....);
> else
>   return SDValue()
> ```

Thanks, I’ll give that a try.

That said, my concern is that this might not work out cleanly — in my mind, the expand logic here (used by AMD and SystemZ) is designed to scalarize for optimization, whereas the vector-based STRICT_FMUL expansion I’m targeting is fundamentally a different case. I’ll experiment and report back.

https://github.com/llvm/llvm-project/pull/147877


More information about the llvm-commits mailing list