From: Alexander Grund (alexander.grund_at_[hidden])
Date: 2020-06-02 13:55:07
Am 02.06.20 um 15:38 schrieb Peter Dimov via Boost:
> Alexander Grund wrote:
>
>> With an API like `result<Foo> do() noexcept` instead of `Foo do()`
>> you won't get (N)RVO when you return a `Foo` (in the happy case) as
>> the standard only allows it when the types are exactly the same.
>
> This doesn't matter much nowadays. E.g. in
> https://godbolt.org/z/zC-Cah if you remove x.f(), x goes away.
Not sure if that really proves the point. I think you need to compare a
result<X> function to a function returning X directly. Like so:
https://godbolt.org/z/JFmn_j
And for that there is a difference: There are 4 main memory accesses in
the result<X> function and none in the "regular" one.
Writing and reading those values likely influences performance for the
worse. However it looks like for large objects (like Y above) there
seems to be no difference indeed and it very much looks like RVO is
happening. Didn't expect that.
So there is only a difference for small and trivial types which will be
passed through RAM instead of registers.