On 08/04/2012 03:26 PM, Marc Glisse wrote:
On Sat, 4 Aug 2012, Paolo Carlini wrote:
.. note anyway, that only the new testcase was failing, no
regressions on pre existing testcases.
What I am seeing is a different testcase (with the same name but in a
different directory) failing, because:
typedef std::pair<const rvalstruct,rvalstruct> V;
static_assert(std::is_constructible<V, V&&>::value,"too bad");
and it makes sense, since you end up having to construct a rvalstruct
from a rvalstruct const&&.
Oops, you are right, sorry. To be clear, the testcase which was failing
with the patch applied is (just check testresults, many examples) is:
23_containers/map/element_access/2.cc
make_pair constructs a pair without const, from which a pair with
const is constructible, though I am surprised it doesn't fail
somewhere further. I don't know what the right solution is, maybe
something emplace-like. In any case, make_pair is unlikely to be right.
I'm not sure to understand which specific testcase you are discussing,
but for sure we don't want regressions. I agree that we should assume a
priori that the standard is right, but correcting the make_pair should
not lead to failures elsewhere (unless a proper analysis establishes
that the existing testcases are wrong)
Paolo.
