[C++] Avoid one place of frontend calling into targetm

[Mark Mitchell]

Jan Hubicka wrote:

>> but, at least, the initializer only is called once.  Because this change
>> in behavior does not return to the original assembly before the first
>> patch, I tend to doubt that this patch fixes the regression correctly.
>> But closer!
> 
> Well, actually this was intended (in fact the code was produced before
> my patch too, just because of bug/misunderstanding in cgraph code it was
> always optimized out as unreachable and accidentally made everything
> working).  With -O1 the constructor will be already inlined.

FWIW, I think Jan's patch makes sense.  The only part I'm not sure about is:

!   /* It can be a static function as long as collect2 does not have
!      to scan the object file to find its ctor/dtor routine.  */
!   TREE_PUBLIC (current_function_decl) = ! targetm.have_ctors_dtors;

Jan, is there something in cgraph that will make sure the function is
public on targets that use collect2, now that you're unconditionally
making the function private in the C++ front end?  I can't read AIX
assembly well enough to be sure.

So, from my perspective, I think the patch is OK.  If it fixes the AIX
test failures, I think we should check it in, if that's OK with David.

However, Jan, it would be nice if you could arrange for these functions
to be always-inlined, even at -O0.  There's no reason for them to be
separate; the user doesn't know anything about them anyhow.  I don't
think that's a big deal, but it would be a bit tidier.

Thanks,

-- 
Mark Mitchell
CodeSourcery
mark@codesourcery.com
(650) 331-3385 x713