dfaure added a comment.
I like the idea of enabling moves for KFileItem very much.
But here's a fun fact: your unittest passes even without the rest of the
patch.
PASS : KFileItemTest::testMove()
That's because std::move() doesn't move, it only makes the argument eligible
for e.g. a move ctor, but will call a copy ctor if there's no move ctor. So
that test is not really testing that move works ;)
That's always a bit tricky to test, I guess, because one can't really rely on
the state of the moved-from object to be anything in particular. And we want
=default, not to implement some counters in there.
Oh well, then maybe there's no real way to unittest that moving works.
Anyhow, the thing I'm wondering is the following: does the rule-of-zero lead
to more opportunities for optimizations by the compiler, who can see "from the
outside" that the 5 special members are the default generated ones, while your
patch "hides" the implementation, lowering the visibility for the compiler in
the rest of the code? Well, one could just move the 5 "=default" to the header
to fix that.
Of course in both cases (rule-of-zero or 5*=default in the header), it means
we are tied to the 5 defaults for all of KF5, for BIC reasons. But that seems
rather sensible in this case (the only member will always be the d pointer, and
it's unlikely we'll move away from refcounting...).
And in fact the other reason against rule-of-zero is that we can't just
remove the copy-ctor and operator=, that would be BIC (existing code links to
it).
In summary: this looks good to me, I'm just wondering if inlining the 5
"=default" wouldn't be better, for optimization purposes.
REPOSITORY
R241 KIO
REVISION DETAIL
https://phabricator.kde.org/D10341
To: markg, dfaure, mwolff
Cc: ngraham, #frameworks, michaelh
