https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102589
Jonathan Wakely <redi at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- Resolution|--- |INVALID Status|UNCONFIRMED |RESOLVED --- Comment #1 from Jonathan Wakely <redi at gcc dot gnu.org> --- (In reply to Gregor Jasny from comment #0) > I stumbled over this issue when porting a code base from C++17 to C++20. > Both: GNU libstdc++ as well as MSVC STL show the unexpected behaviour. > Libc++ seems to do the right thing. Probably because it doesn't implement <=> support in the standard library at all. Libstdc++ and MSVC do. > Describe the bug > > Compiling the test case with C++17 works as expected whereas compiling it > with C++20 makes the test assumption fail. The assumption is wrong. > The reason is that for C++ 20 builds the specialised operator<(const FooPtr& > left, const FooPtr& right) is not picked up. Instead the default operator<=> > for std::shared_ptr is used which only compares the raw pointers and not the > content they point to. This is the correct behaviour. > Expected behavior > > I'd expect that with C++17 and C++20 the STL would prefer the specialised > operator<(const FooPtr& left, const FooPtr& right) over the synthesised <=> > one. That would also align with the strong backwards compatibilities C++ > strives for. No. Comparing the two std::map objects has to use <=> because there is no < for std::map in C++20, so you get a synthesized < from the <=> for std::map. And obviously <=> for std::map is going to try to use <=> for its elements. So it uses <=> for std::pair which uses <=> for its members, so it uses <=> for std::shared_ptr. The only < comparison that happens is the one for std::map, after that everything is a three-way comparison.