Re: [PATCH 00/11] ARC atomics update

[Peter Zijlstra]

On Thu, Aug 05, 2021 at 04:18:29PM +0000, Vineet Gupta wrote:
> On 8/5/21 2:02 AM, Peter Zijlstra wrote:
> > On Wed, Aug 04, 2021 at 12:15:43PM -0700, Vineet Gupta wrote:
> > 
> >> Vineet Gupta (10):
> >>    ARC: atomics: disintegrate header
> >>    ARC: atomic: !LLSC: remove hack in atomic_set() for for UP
> >>    ARC: atomic: !LLSC: use int data type consistently
> >>    ARC: atomic64: LLSC: elide unused atomic_{and,or,xor,andnot}_return
> >>    ARC: atomics: implement relaxed variants
> >>    ARC: bitops: fls/ffs to take int (vs long) per asm-generic defines
> >>    ARC: xchg: !LLSC: remove UP micro-optimization/hack
> >>    ARC: cmpxchg/xchg: rewrite as macros to make type safe
> >>    ARC: cmpxchg/xchg: implement relaxed variants (LLSC config only)
> >>    ARC: atomic_cmpxchg/atomic_xchg: implement relaxed variants
> >>
> >> Will Deacon (1):
> >>    ARC: switch to generic bitops
> > 
> > Didn't see any weird things:
> > 
> > Acked-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> 
> Thx Peter. A lot of this is your code anyways ;-)
> 
> Any initial thoughts/comments on patch 06/11 - is there an obvious 
> reason that generic bitops take signed @nr or the hurdle is need to be 
> done consistently cross-arch.

That does indeed seem daft and ready for a cleanup. Will any
recollection from when you touched this?

AFAICT bitops/atomic.h is consistently 'unsigned int nr', but
bitops/non-atomic.h is 'int nr' while bitops/instrumented-non-atomic.h
is consistently 'long nr'.

I'm thinking 'unsigned int nr' is the most sensible allround, but I've
not gone through all the cases.