I'm trying to write a configure check for whether the __sync atomic builtin functions are present -- and it never fails! For some reason, now whenever I tell g++ to compile an example .cpp file that uses the __sync instructions, it compiles without errors, even if I don't provide the march flag. Also, my CFLAGS, CPPFLAGS, and CXXFLAGS are empty, so it should be using the defaults. In fact, if I deliberately specify -march=i386, it STILL compiles fine, when it should fail since the 386 had no lock instructions. Indeed, I have seen __sync using software fail to compile until CXXFLAGS="-march=i686" was specified, but for some reason invoking g++ myself outside of any build tool doesn't trigger this behavior. It seems to be using my native architecture even if I deliberately tell it not to, which makes it hard to verify my configure script works correctly in the failure case. Is this a bug or a feature? Either way it's undesirable for me in this particular case.... My GCC version is 4.2.2 on Solaris.
