https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119386
--- Comment #54 from Alexander Monakov <amonakov at gcc dot gnu.org> --- I think the x86_64 behavior is simply copied as-is from i386. On i386, there was a time when Glibc wouldn't preserve eax+ecx+edx in the PLT trampoline, but preserving those became necessary when GCC exposed -mregparm and the corresponding function attribute. https://sourceware.org/cgit/glibc/diff/sysdeps/i386/dl-machine.h?id=831372e7c1bb907f9f2c3d78909b15717b8ac095 It doesn't really explain to me why PLT avoidance for mcount was necessary (cross-DSO regparm calls with unfixed Glibc would be buggy for everything, not just mcount), and apparently the avoidance goes back even further than that, to a.out times. So the relevance to the matter at hand is unclear. On amd64, PLT trampolines may clobber r11 (and lazy resolution trampolines always do). So, yeah, looks like inherited PLT avoidance plays some role now (even if by accident).
