http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59626
--- Comment #18 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #5)
> But the function doesn't inline itself, that is why it uses the asm alias
> and GCC shouldn't be looking through that and merging the two decls because
> of that.
>
> Breaking this breaks pretty much all of glibc -D_FORTIFY_SOURCE, and there
> really isn't any other way how to write it.
>
> To show the reason for the inline:
> __fortify_function __nonnull ((1, 2)) __wur ssize_t
> __NTH (readlink (const char *__restrict __path, char *__restrict __buf,
> size_t __len))
> {
> if (__bos (__buf) != (size_t) -1)
> {
> if (!__builtin_constant_p (__len))
> return __readlink_chk (__path, __buf, __len, __bos (__buf));
>
> if ( __len > __bos (__buf))
> return __readlink_chk_warn (__path, __buf, __len, __bos (__buf));
> }
> return __readlink_alias (__path, __buf, __len);
> }
>
> The inline must be always_inline, because it is a security matter if it is
> inlined or not, and we want it to expand as a call to __readlink_chk if
> it needs runtime verification, otherwise as a call to the original function,
> not inlined. So, this is really a LTO/IPA bug.
Note that __readlink_alias was _not_ declared always-inline and thus at least
the call edge (even if eventually re-directing to readlink itself) should
not be considered always-inline. But we always check the always-inline flag
on the callee decl, even if it was substituted from an alias. As we throw
away the callgraph edges it's also hard to maintain this edge property there.