[Development] What's the status of a moved-from object?

[Konstantin Ritt]

пн, 27 мая 2019 г. в 11:46, Edward Welbourne <edward.welbourne at qt.io>:

> сб, 25 мая 2019 г., 12:30 Mutz, Marc via Development
> >>>> Repeat after me: default ctors do _not_ establish a valid value.
>
> As an aged C programmer I can respect that ;^>
>
> On 2019-05-25 17:24, Konstantin Ritt wrote:
> >>> Perhaps you mean "trivial type's default ctors do _not_ establish a
> >>> valid value"?
>
> сб, 25 мая 2019 г., 20:42 Mutz, Marc via Development <
> development at qt-project.org<mailto:development at qt-project.org>>:
> >> No, I actually meant default ctor. What should the default value be?
> >> _Obviously_ zero-initialisation. But oh, no, QSize is different. So
> >> maybe it's not _quite_ so obvious what the default value is. Why is
> >> pen's default value a black solid line instead of NoPen? Why is
> >> QBrush's default ctor not solid black, but NoBrush? See? It's just
> >> nonsense to try to pick a default value, so this is to say: Don't
> >> try.
> >>
> >> There's exactly one exception: containers. _Clearly_ no-one would
> >> _ever_ expect a container to start out anything but empty.
>
> Indeed.
>
> Konstantin Ritt (26 May 2019 10:51) replied:
> > Sure we should force a no-op default ctor where possible, and provide
> > some helper code that returns a valid default/specialized value of
> > that type where {0} isn't enough or non-intuitive
> > (i.e. QQuaternion::identity(), QRect::empty(), QSize::invalid()).
>
> (I can't help but wonder whether QRect::empty() should take a point at
> which to place its empty rectangle.  An animated rectangle that shrinks
> and expands repeatedly [0] shouldn't forget its location at the moment
> when it is transiently empty, so empty rectangles at distinct locations
> are distinct.  But I'm unfamiliar with QRect, so maybe not.)
>
> [0] example (in SVG), remembering position and orientation while empty:
>     http://www.chaos.org.uk/~eddy/img/geom/transpyth.svg
>
> > There could [be] more exceptions when we're talking about non-trivial
> > types, at least some of them could be resolved by providing
> > convenience/helper initializers.
>
> Many types need an invalid value (this is the current default for
> QDate, QTime, QDateTime and QTimeZone).  But not all.
>
> For arithmetic types, up to and including QQuaternion and matrices,
> you'll need a zero-value; many (including square matrices) shall also
> need a unit value.  While zero is a natural default, being explicit is
> good; and standard integral/floating types don't default to it.
>
> Various others are (at present) constructed with default values.  In the
> case of QLocale, that can be controlled by calling a static method,
> QLocale::setDefault(), and we'd need to add (as someone tried to do a
> little while ago) a QLocale::defaultLocale() or QLocale::getDefault() to
> get that (note that we can't use the name QLocale::default() as it's a
> reserved word) if we stopped making the default constructor produce it.
> Note that QLocale's default isn't the same thing as QLocale::system(),
> although it starts out so, until setDefault() is called.
>
> We're not particularly consistent about what default construction
> produces; sometimes undefined, sometimes defined but invalid, sometimes
> a zero, sometimes a "reasonable" default.  Which, as Marc tripped over,
> is a source of confusion for authors of client code.  Having static
> methods to return invalid, zero, unit, default values could be made a
> consistent pattern, with consistent naming, to reduce confusion.
>
> We can, of course, do that without changing what the default-constructed
> objects are, but changing that also to a consistent pattern - the same
> as the moved-from state - would make matters clearer (for any choice of
> what moved-from state to consistently use).  But it'd be SiC.
>
>         Eddy.
>

Completely agreed.


Konstantin
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.qt-project.org/pipermail/development/attachments/20190527/8b31b01a/attachment.html>