[Bug libstdc++/65147] alignment of std::atomic object is not correct

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65147

--- Comment #10 from Jonathan Wakely <redi at gcc dot gnu.org> ---
(In reply to Alexey Lapshin from comment #7)
> It looks like this fix makes alignment of atomic object to be the same as
> alignment of integral non-atomic object of the same size.

Actually it only did that for non-integral atomic objects, e.g. I didn't do
anything to change std::atomic<long long>.

> The gcc behavior is different it makes alignment of atomic objects of sizes
> 1,2,4,8,16 to match with size :

That's not strictly true, there is a target hook (atomic_align_for_mode) which
specifies the alignment for 1/2/4/8/16-byte objects, and the result is not
necessarily the same as the size. Or so I'm told. That's why I used the nested
conditional expressions with alignof(integral type) instead of just using
alignas(sizeof(T)).

> $g++ -latomic -std=c++11 -m32 all.cc 
> $./a.out
> 
>  sizeof(ac) 1 alignof(ac) 1
>  sizeof(as) 2 alignof(as) 2
>  sizeof(al) 4 alignof(al) 4
>  sizeof(all) 8 alignof(all) 4
>  sizeof(a16) 16 alignof(a16) 1

There are two problems here.

The first is that alignof(std::atomic<long long>) is less than alignof(long
long), and my recent changes didn't address that. That is easy to fix:

--- a/libstdc++-v3/include/bits/atomic_base.h
+++ b/libstdc++-v3/include/bits/atomic_base.h
@@ -235,7 +235,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
   // 8 bytes, since that is what GCC built-in functions for atomic
   // memory access expect.
   template<typename _ITp>
-    struct __atomic_base
+    struct alignas(_ITp) __atomic_base
     {
     private:
       typedef _ITp     __int_type;


The second problem is that alignof(struct S16) is not increased. That's because
libstdc++ doesn't support __int128 on x86, so this bit of code doesn't do
anything:

#ifdef _GLIBCXX_USE_INT128
    : sizeof(_Tp) == sizeof(__int128)  ? alignof(__int128)
#endif

I'm not sure how to fix this. Maybe we should just bodge it like this and hope
it is valid for all important targets:

#ifdef _GLIBCXX_USE_INT128
    : sizeof(_Tp) == sizeof(__int128)  ? alignof(__int128)
#else
    : sizeof(_Tp) == 16  ? 16
#endif

(The real solution is a new attribute that uses the target hook, so we can
guarantee the same result as the C front end, but it's too late to do that for
GCC 5).