Re: [RFC PATCH v2] target/loongarch: Fix incorrect rounding in fnm{add,sub} under certain modes

[WANG Rui <wangrui@loongson.cn>]

Hi Richard,

On Wed, May 7, 2025 at 11:04 PM Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> On 5/7/25 02:14, WANG Rui wrote:
> > This patch fixes incorrect results for `[xv]fnm{add,sub}.{s,d}`
> > instructions when rounding toward {zero, positive, negative}.
> >
> > According to the LoongArch ISA specification, the result of an
> > instruction like `FNMSUB.D` is computed as:
> >
> >    FR[fd] = -FP64_fusedMultiplyAdd(FR[fj], FR[fk], -FR[fa])
> >
> > Here, `FP64_fusedMultiplyAdd()` performs a fused multiply-add operation
> > compliant with IEEE 754-2008. The negation is applied to the fully
> > rounded result of the fused operation - not to any intermediate value.
> > This behavior is specifiec to LoongArch and differs from other arches,
> > which is why the existing `float_muladd_negate_result` flag does not
> > model it correctly.
>
> Loongarch does not differ from other arches; we got it wrong for everyone.
> There's no need for a new flag.

I'm not familiar with PowerPC and s390x. The official PowerPC docs[^1]
doesn't clearly specify the order of negation and rounding operations.
I also don't have access to the hardware to run experiments. However,
I did find the Linux kernel's emulation code[^2][^3] for the PowerPC
`fnmsub` instruction, which seems to suggest that negation occurs
before rounding - though it's possible that interpretation is
incorrect.

[^1]: https://www.ibm.com/docs/en/aix/7.2?topic=set-fnmsub-fnms-floating-negative-multiply-subtract-instruction
[^2]: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/powerpc/math-emu/fnmsub.c?h=v6.15-rc5#n49
[^3}: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/powerpc/math-emu/fnmsub.c?h=v6.15-rc5#n55

Regards,
-hev

>
>
> r~