There is a fairly comprehensive overview of what compilers do / don't do for RVO + NRVO at: http://www.byte-physics.de/index.php?eID=tx_nawsecuredl&u=0&g=0&t=1427845187&hash=d85e8402f1e5b835e3ff4f266a0a7256516ea57e&file=fileadmin/user_upload/kundenbereich/copy-elision-revisited-v1.0.pdf
Move constructors you'll likely still have to copy some of the data (The size of the vector for instance), but you save copying anything that is pointed in (the data in the vector). Definitely it makes it a lot cheaper for a lot of things, and things smaller than 1024 bytes most likely the copy is cheaper than any sort of reference scheme on a modern processor. On Mon, Mar 30, 2015 at 3:31 PM, Benjamin Mahler <[email protected]> wrote: > IIRC, it was when we changed Option:::get to return a const ref. Some > callers of 'Option<V> hashmap<K,V>::get()' were grabbing a const ref to the > temporary coming from the option (case 1 in your markdown), and it then > became a temporary to a temporary's member (case 2 in your markdown): > > E.g.: https://reviews.apache.org/r/18386/diff/#6 > > Re: optimizations > > Do you know what the state of the art is for RVO and NRVO? When is the > compiler unable to perform these optimizations? > I noticed you didn't mention move constructors. It looks like move > constructors guarantee we elide the copy into the caller's object, which > seems to be the whole motivation behind capturing a const ref? > > On Mon, Mar 30, 2015 at 12:15 PM, Jie Yu <[email protected]> wrote: > > > I am a +1 on this too. We had a similar bug long time ago when we changed > > the return value of Try::get() to be const ref. > > > > On Mon, Mar 30, 2015 at 11:52 AM, Niklas Nielsen <[email protected]> > > wrote: > > > > > Big +1; You could also mention the recent bug in the Two fold > > > authentication patch that you guys discovered with this bug. > > > > > > Niklas > > > > > > On 30 March 2015 at 10:49, Joris Van Remoortere <[email protected]> > > > wrote: > > > > > > > I would like to propose we disallow capturing temporaries using a > > > constant > > > > reference: > > > > const T& val = f(); > > > > > > > > *The reasons, and examples for this are outlined in the proposal > here:* > > > > [Review Board]: https://reviews.apache.org/r/32630/ > > > > [original markdown version]: > > > > https://gist.github.com/jmlvanre/8a3de53ae88c2d19b375 > > > > > > > > > > > > Please feel free to comment here on the dev-list, though I also urge > > you > > > to > > > > comment on the Review Request as putting proposal markdowns on RB is > > > > something new we're trying! > > > > > > > > Joris > > > > > > > > > >
