http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51643
--- Comment #14 from Richard Earnshaw <rearnsha at gcc dot gnu.org> 2011-12-22 23:27:29 UTC --- Because the ABI says it only works for bare metal. On a system with shared libraries, you can't tell at static link time if a weak symbol will be resolved by a shared library, so it has to left up to the dynamic linker which will fill in a PLT stub. Once you have those, it's pretty hard to make the call become a stub (and even if it were, it wouldn't be a NOP (the linker would have to turn the call into something that simply returned as the PLT sequence can't be patched). Tail-calling PLT sequences is perfectly safe, so there's no reason not to perform the optimization.
