[PATCH 4/7] riscv: Implement xchg8/16() using Zabha

[Alexandre Ghiti]

Hi Andrea,

On Tue, May 28, 2024 at 8:00 PM Andrea Parri <parri.andrea at gmail.com> wrote:
>
> > -#define __arch_xchg_masked(prepend, append, r, p, n)                 \
> > +#define __arch_xchg_masked(swap_sfx, prepend, append, r, p, n)               \
>
> This actually indicates a problem in the current (aka, no Zabha)
> implementation: without your series, xchg16() gets mapped to
>
>   lr.w     a2,(a3)
>   and      a1,a2,a5
>   or       a1,a1,a4
>   sc.w     a1,a1,(a3)
>   bnez     a1,43c <.L0^B1>
>
> which is clearly wrong... (other "fully-ordered LR/SC sequences"
> instead follow the mapping
>
>   lr.w     a2,(a3)
>   and      a1,a2,a5
>   or       a1,a1,a4
>   sc.w.rl  a1,a1,(a3)
>   bnez     a1,43c <.L0^B1>
>   fence    rw,rw  )
>
> A similar consideration for xchg8().

Ok, I will fix that separately and make it merged in -rc2.

Thanks,

Akex

>
>   Andrea