On Friday, June 17, 2016 at 3:57:01 PM UTC+1, Gerald Squelart wrote: > On Friday, June 17, 2016 at 2:31:15 PM UTC+1, smaug wrote: > > On 06/16/2016 06:40 PM, smaug wrote: > > > On 05/24/2016 08:33 AM, [email protected] wrote: > > >> For > > >> > > >> RefPtr<T> GetFoo() { > > >> RefPtr<T> foo; > > >> // ... > > >> } > > >> > > >> should we: > > >> > > >> 1. return foo and expect RVO to kick in to eliminate additional > > >> AddRef/Release > > >> 2. return foo.forget() > > >> 3. return Move(foo) > > >> > > >> Which one is preferred? > > >> > > >> ps: I find gcc is able to apply RVO even with "-O0". Not sure if it is > > >> also true for other compilers. > > >> > > > > > > Can we somehow guarantee that RVO kicks in? It feels quite error prone to > > > rely on compiler optimization when dealing with > > > possibly slow Addref/Release calls. > > > https://en.wikipedia.org/wiki/Return_value_optimization#Compiler_support > > > hints that there are cases when RVO might not happen. > > > > > > > I've been told we can't guarantee RVO to kick in, which means, I think, > > that we should assume we don't have RVO ever and need to > > rely on other coding patterns which are guaranteed to optimize out extra > > addref/release. > > Currently having already_AddRefed as the return type is one such pattern. > > > > What are the other options? > > > > > > At least we can't start using RefPtr<T> as return type before we know it > > won't be causing performance regressions. > > So I would recommend using already_AddRefed for now. > > From what *I* understand, RVO is guaranteed (or at least supported > everywhere?) when there is only one stack variable that is returned, e.g.: > C foo() > { > C rv; > // ... (put stuff in rv) > return rv; > } > In this case, the caller function stack can host 'rv' directly, no copies > needed. > > RVO won't happen when there may be multiple variables in the function, only > one of which will be returned, e.g.: > C foo() > { > C rv1, rv2; bool b; > // ... (assign a value to b) > return b ? rv1 : rv2; > } > The caller cannot know which of rv1 or rv2 will be returned, so no return > value can be lifted. > > (Any actual expert who knows the truth?) (sorry for the spam)
If I'm correct, then the question becomes: Can we trust developers to always make the right choice? Or better: Can we write static analyzers that will catch RVO-hostile code? _______________________________________________ dev-platform mailing list [email protected] https://lists.mozilla.org/listinfo/dev-platform
