Re: [dpdk-dev] [PATCH v5 2/4] examples: add lthread subsystem for performance-thread

[Bruce Richardson <bruce.richardson@intel.com>]

On Thu, Dec 03, 2015 at 08:31:39AM -0800, Stephen Hemminger wrote:
> On Thu,  3 Dec 2015 09:28:23 +0000
> ibetts <ian.betts@intel.com> wrote:
> 
> > +/*
> > + * Atomically set a value and return the old value
> > + */
> > +static inline uint64_t
> > +atomic64_xchg(uint64_t *ptr, uint64_t val) __attribute__ ((always_inline));
> > +static inline uint64_t
> > +atomic64_xchg(uint64_t *ptr, uint64_t val)
> 
> You don't need a forward declaration for this.
> Instead do:
> 
> static inline uint64_t __attribute__((always_inline))
> atomic_xchg64(uint64_t *ptr, uint64_t val)
> 
> Really should be in rte_atomic.h as a primitive
> and the assembly macro is missing change to ptr so Gcc might optmize it away.
> 
> Something like this mayb?
> 
> static inline uint64_t __attribute__ ((always_inline));
> rte_atomic64_xchg(uint64_t *ptr, uint64_t val)
> {
> 	asm volatile (
> 			MPLOCKED
> 			"xchgq %[ptr],%[val];"
> 			: [val] "=r" (val)
>                           [ptr] "=m" (*ptr)
> 			: [ptr] "m" (*ptr),
> 			  "a" (val)
> 			: "memory");
> 
> 	return val;
> }

Rather than using assembly, I believe the gcc builtin __sync_lock_test_and_set
is actually an xchg op, so can be used here.

/Bruce