On Thursday, April 28, 2016 at 11:00:12 AM UTC+10, Gerald Squelart wrote:
> On Thursday, April 28, 2016 at 10:41:21 AM UTC+10, Ehsan Akhgari wrote:
> > On 2016-04-28 8:00 AM, Gerald Squelart wrote:
> > > On Thursday, April 28, 2016 at 9:35:56 AM UTC+10, Kyle Huey wrote:
> > >> Can we catch this pattern with a compiler somehow?
> > >>
> > >> Foo foo;
> > >> foo.x = thing;
> > >> DoBar(mozilla::Move(foo));
> > >> if (foo.x) { /* do stuff */ }
> >
> > I think so. We already have an analysis which would detect whether the
> > return value of a function is used somewhere or not. We should be able
> > to reuse that to find the call to DoBar(), and then look for future
> > occurrences of foo used as an rvalue in the rest of the function. Once
> > we detect a use of "foo" as an lvalue, further usages of it as an rvalue
> > in the same function should be OK and not trigger the error. File a bug?
> >
> > > Definitely something that would be nice.
> > >
> > > But if we have/implement such a catcher, I'd like to have an annotation
> > > to say "yep I really want to reuse this moved-from object".
> > > Because sometimes the function will choose not to actually move from an
> > > rvalue-ref, or the object knows to revert to a fully-reusable state, etc.
> >
> > What you're describing sounds like a violation of move semantics, right?
> > The first case should only happen if DoBar doesn't accept an rvalue
> > reference, in which case the code above is definitely doing something
> > that the author did not expect, given that they have used Move(). The
> > latter case sounds completely broken, and if there is an actual good use
> > case for it, the C++ move semantics sound like the wrong tool to achieve
> > that goal to me.
> >
> > If you feel like I'm missing something or you can make a strong argument
> > on why breaking move semantics is OK in some cases, please let me know. :-)
> >
> > Cheers,
> > Ehsan
>
> std::move and mozilla:Move are just casts that make an l-value object *look*
> like an r-value, so that when the compiler considers which 'DoBar' to use,
> one that takes an r-value reference will be picked first.
>
> "Move" is probably not the best name because it gives this impression that an
> actual move happens, but that's what we're stuck with in the standard.
>
> I don't see a "violation of move semantics" in there, could you please
> elaborate on what exact move semantics are violated?
> I'd say it's probably more a "perversion of the move spirit". :-)
>
> In any case, a moved-from object *must* stay valid after that call, because
> at the minimum it will be destroyed at the end of its enclosing scope, by
> invoking its normal destructor, no magic or special path there.
>
> Now what to do with a moved-from object, is I think more a philosophical
> question! Some argue that we should do nothing (except the inevitable implied
> destruction). Others think it should be fine to also allow re-assignment
> (i.e. reuse the variable for something completely different). And yet others
> would allow total reuse.
>
> My position is that 'Move(x)' *usually* means we give the value away and
> don't want to use it again, and therefore a compiler warning/error would help
> catch unexpected reuses; but also that some situations call for reuse (e.g.
> the function doesn't always steal the object's contents, based on other
> factors) and a programmer in the know should be allowed to annotate this
> special case.
Note that we talked a bit about this situation in:
https://groups.google.com/d/topic/mozilla.dev.platform/VLtl2yL_TlA/discussion
Referring to:
http://mxr.mozilla.org/mozilla-central/source/mfbt/UniquePtr.h#183
Which talks about conditionally moving from a UniquePtr.
_______________________________________________
dev-platform mailing list
dev-platform@lists.mozilla.org
https://lists.mozilla.org/listinfo/dev-platform
