- Simplify mechanism for detecting traits of test allocators. There were
some portability issues, but rather than fix them I've just gone for a
simpler mechanism. Does mean that the relevant tests can't be run for
other allocators.
- Fix a couple of unnecessary_copy_tests, whose results were the wrong
way round.
- It appears that Visual C++ only implements RVO for implicitly defined
copy constructors in debug mode, so adjust a move_test to account for
the extra copies now that the copy constructors are explicitly
defined.
[SVN r73798]
It's pretty messy because I'm trying to avoid swapping allocators in
these cases. I'm also not sure of the exception requirements of
allocator swap and assignment.
[SVN r73756]
Seems to be better in some cases. For example, better forwarding if
rvalue references are supported but variadic template parameters aren't.
Also can use `boost::move` when inserting.
Older versions of gcc still fail this test. They perform more moves than
expected. It might be a limitation of Boost.Move, or maybe just poor
copy optimization.
[SVN r73539]
Store nodes in a single linked list, with hash values so that their
buckets can be found when needed. Iterators now only have to store a
pointer to the node and don't have to iterate over empty buckets to
reach the next node. This allows the container to meet the iterator
requirements - fixing the speed issues with `equal_range` and `erase`.
Also, define iterators in their own namespace, so that they don't
accidentally pull in detail functions via ADL.
I've simplified the code slightly by removing some of the special
cases for empty containers. Renamed a few things as well and other
minor changes that were made as I went along.
[SVN r71327]
In the draft standard the std::pair constructor from another pair is
only considered when overloading if both members are implicitly
constructible to their corresponding members. This breaks one of the
unordered tests which required an explicit conversion, so change it to
only require an implicit conversion.
[SVN r70028]
Borland is having some issues with the existing tests, since they call
inline functions before they're defined. It might be right to, although
all the other compilers are fine with it. But the test isn't really what
the forward headers are intended for. Try to make the tests better
represent the intent, and possibly work on Borland.
[SVN r66428]
It seems that the problem is calling sizeof on a dependent type when the
containers have only been used by reference. So by putting in these
dummy structures with the containers as members, it helps the compiler
instantiate the class to the level where sizeof works. I hope.
[SVN r58130]
Daniel James