On Fri, Jun 28, 2024 at 01:33:06AM +0100, Jiaxun Yang wrote:
>
>
> 在2024年6月27日六月 下午8:54，Maciej W. Rozycki写道：
> > On Thu, 27 Jun 2024, Jiaxun Yang wrote:
> >
> >> >> @@ -318,6 +318,10 @@ void mips_set_personality_nan(struct arch_elf_state *state)
> >> >> t->thread.fpu.fcr31 = c->fpu_csr31;
> >> >> switch (state->nan_2008) {
> >> >> case 0:
> >> >> + if (!(c->fpu_msk31 & FPU_CSR_NAN2008))
> >> >> + t->thread.fpu.fcr31 &= ~FPU_CSR_NAN2008;
> >> >> + if (!(c->fpu_msk31 & FPU_CSR_ABS2008))
> >> >> + t->thread.fpu.fcr31 &= ~FPU_CSR_ABS2008;
> >> >
> >> > why is this needed?
> >>
> >> Because t->thread.fpu.fcr31 comes from c->fpu_csr31, in this case we the default
> >> value of c->fpu_csr31 is read from hardware and we don't know what would that be.
> >
> > But it has always been like this. What has changed with your patch that
> > you need to mask the bit out now?
>
> After this patch kernel's copy of t->thread.fpu.fcr31 can disagree with hardware.
> When disagree happens, we trigger emulation.
>
> Before that patch for nan legacy binary running on nan2008 CPU t->thread.fpu.fcr31
> will still be nan2008 (for ieee754=relaxed) so that's not relevant.
I'm considering to apply your patch, how much testing/verification did
this patch see ? Do have some test binaries ?
Thomas.
--
Crap can work. Given enough thrust pigs will fly, but it's not necessarily a
good idea. [ RFC1925, 2.3 ]
