Re: [PATCH] riscv/atomic: Do proper sign extension also for unsigned in arch_cmpxchg

[Alexandre Ghiti]

On 30/01/2025 15:52, Andreas Schwab wrote:
> On Jan 30 2025, Alexandre Ghiti wrote:
>
> > That's a nice catch indeed. IIUC, we have the same issue here
> > https://elixir.bootlin.com/linux/v6.13/source/arch/riscv/include/asm/futex.h#L89
> > right?
> Indeed, though it doesn't result in wrong code currently.  This is
> because the compare value is passed unmodified as u32 to the asm and the
> compiler keeps the value sign extended in registers.  That would break
> if you would add a cast to long like in commit 6c58f25e6938 as that
> would erroneously zero extend it.

This is the disassembly I get:

ffffffff800fc540 <futex_atomic_cmpxchg_inatomic>:

...

ffffffff800fc566:       1605a8af                lr.w.aqrl a7,(a1)
ffffffff800fc56a:       00c89563                bne 
a7,a2,ffffffff800fc574 <futex_atomic_cmpxchg_inatomic+0x3
4>
ffffffff800fc56e:       1ed5a52f                sc.w.aqrl a0,a3,(a1)

a2 is used as it is passed by the calling function, so we can't be sure 
a2 is sign extended to me, what am I missing?