Re: [PATCH] powerpc: fix inline asm constraints for dcbz

[Segher Boessenkool]

On Fri, Aug 09, 2019 at 08:28:19PM +0200, Arnd Bergmann wrote:
> On Fri, Aug 9, 2019 at 8:21 PM 'Nick Desaulniers' via Clang Built
> Linux <clang-built-linux@xxxxxxxxxxxxxxxx> wrote:
> 
> >  static inline void dcbz(void *addr)
> >  {
> > -       __asm__ __volatile__ ("dcbz %y0" : : "Z"(*(u8 *)addr) : "memory");
> > +       __asm__ __volatile__ ("dcbz %y0" : "=Z"(*(u8 *)addr) :: "memory");
> >  }
> >
> >  static inline void dcbi(void *addr)
> >  {
> > -       __asm__ __volatile__ ("dcbi %y0" : : "Z"(*(u8 *)addr) : "memory");
> > +       __asm__ __volatile__ ("dcbi %y0" : "=Z"(*(u8 *)addr) :: "memory");
> >  }
> 
> I think the result of the discussion was that an output argument only kind-of
> makes sense for dcbz, but for the others it's really an input, and clang is
> wrong in the way it handles the "Z" constraint by making a copy, which it
> doesn't do for "m".

Yes.  And clang has probably miscompiled this in all kernels since we
have used "Z" for the first time, in 2008 (0f3d6bcd391b).

It is not necessarily fatal or at least not easily visible for the I/O
accessors: it "just" gets memory ordering wrong slightly (it looks like
it does the sync;tw;isync thing around an extra stack access, after it
has performed the actual I/O as any other memory load, without any
synchronisation).

> I'm not sure whether it's correct to use "m" instead of "Z" here, which
> would be a better workaround if that works. More importantly though,
> clang really needs to be fixed to handle "Z" correctly.

"m" allows offset addressing, which these insns do not.  That is the
same reason you need the "y" output modifier.  "m" is wrong here.

We have other memory constraints, but do those work with LLVM?


Segher