[Bug c/71255] Implement #pragma may_alias

["fw at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71255

--- Comment #16 from Florian Weimer <fw at gcc dot gnu.org> ---
(In reply to Marek Polacek from comment #15)
> Yeah, only the C++ side was changed.  I think it's wrong that we reject the
> testcase in Comment 14 in C (I have a fix for that).

Good.

> But even with that fixed we still need the new #pragma because of the second
> testcase in Comment 13, right?

I'm surprised this is even valid C.  The test case as-is does not seem relevant
to the glibc usage scenario because all completions would use the may_alias
attribute.  Only some forward declarations would not.  This could be a relevant
fringe case we need to support:

struct S s;
struct __attribute__((may_alias)) S { int i; };

Although it can only arise if there is a definition *before* including the
relevant header file.

I think the real question is whether it matters anywhere if a pointer to an
incomplete struct has the may_alias attribute or not.