Re: [PATCH 1/5] riscv: Optimize fmaximum_num and fminimum_num
Jeffrey Law <[email protected]>
| Newsgroups | gmane.comp.lib.glibc.alpha |
|---|---|
| Message-ID | <[email protected]> |
On 8/6/2026 12:34 PM, Adhemerval Zanella Netto wrote: > > On 06/08/26 07:25, Julian Zhu wrote: >> Implement the C23 maximumNumber and minimumNumber operations with the RISC-V fmax and fmin instructions for float and double. These instructions provide the required NaN and signed-zero semantics using the base F and D extensions. >> >> Signed-off-by: Julian Zhu <[email protected]> > The version 2.2 of the F extension requirement really makes this messy to keep > deployment generic. Ideally it would be better to implement it with builtin and > let compiler handle it, and then we gate it through USE_*_BUILTIN. If deployers > want to support pre 2.1 hardware, using the generic implementation is the price > to pay. I think that's the key thing I'd relay as well. Let's use a builtin and let the compiler figure it out. Jeff