https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103240
--- Comment #5 from frankhb1989 at gmail dot com ---
There are multiple issues.
1. The out-of-line definition and the inline definition of
std::type_info::before do different things. Specifically, only one name is
detected against to '*' in the former, but two in the latter. It seems the
latter is more correct. ("More", because it is still not quite correct, see
below.)
2. As stated in https://reviews.llvm.org/D100134, mixture of different methods
of comparisons breaks the ordering.
I've actually encountered the problem with std::type_index as the key type when
using flat_map::emplace (like
https://github.com/WG21-SG14/SG14/blob/master/SG14/flat_map.h, which uses
std::partition_point to find the position of insertion). The insertion suddenly
fails when std::partition_point meets two std::type_info objects x and y which
are unordered. It works with the workaround '-U__GXX_MERGED_TYPEINFO_NAMES
-D__GXX_MERGED_TYPEINFO_NAMES=1' in the compiler command line, but it seems not
enough in general without fixing the issue 2 here. (Although the same sequence
of key on std::map does not fail, occasionally, due to less comparisons are
made.)
3. Not sure the original change in r179236 is correct.
4. '||' in the condition of the inline definition of std::type_info::before
seems an overkill. '&&' may be enough. Assuming string comparison is more
expensive, this is a simple optimization. But once the issue 2 is fixed, the
change would look quite different.
