Re: [PATCH 4/5] atomic: Introduce atomic_try_cmpxchg()

[Peter Zijlstra]

On Mon, Feb 06, 2017 at 12:24:28PM +0800, Boqun Feng wrote:
> On Fri, Feb 03, 2017 at 02:26:02PM +0100, Peter Zijlstra wrote:

> > 
> > 	for (;;) {
> > 		new = val $op $imm;
> > 		if (try_cmpxchg(ptr, &val, new))
> > 			break;
> > 	}
> > 
> > while also generating better code (GCC6 and onwards).
> > 
> 
> But switching to try_cmpxchg() will make @val a memory location, which
> could not be put in a register. And this will generate unnecessary
> memory accesses on archs having enough registers(PPC, e.g.).

GCC was perfectly capable of making @val a register in the code I was
looking at.

> > +#ifndef atomic_try_cmpxchg
> > +
> > +#define __atomic_try_cmpxchg(type, _p, _po, _n)				\
> > +({									\
> > +	typeof(_po) __po = (_po);					\
> > +	typeof(*(_po)) __o = *__po;					\
> > +	bool success = (atomic_cmpxchg##type((_p), __o, (_n)) == __o);	\
> > +	*__po = __o;							\
> 
> Besides, is this part correct? atomic_cmpxchg_*() wouldn't change the
> value of __o, so *__po wouldn't be changed.. IOW, in case of failure,
> *ptr wouldn't be updated to a new value.
> 
> Maybe this should be:
> 
> 	bool success;
> 	*__po = atomic_cmpxchg##type((_p), __o, (_n));
> 	sucess = (*__po == _o);
> 
> , right?

Yes, botched that. Don't think I even compiled it to be honest :/