Re: [libstdc++/65033] Give alignment info to libatomic

[Jonathan Wakely <jwakely@redhat.com>]

On 07/04/15 06:51 -0400, Hans-Peter Nilsson wrote:
>On Tue, 7 Apr 2015, Jonathan Wakely wrote:
>> On 05/04/15 21:07 -0400, Hans-Peter Nilsson wrote:
>> > On Fri, 3 Apr 2015, Jonathan Wakely wrote:
>> >
>> > > On 03/04/15 05:24 -0400, Hans-Peter Nilsson wrote:
>> > > > On Thu, 2 Apr 2015, Hans-Peter Nilsson wrote:
>> > > > > Why then use __alignof(_M_i) (the object-alignment)
>> > > > > instead of _S_alignment (the deduced alas insufficiently
>> > > > > increased type-alignment)?
>> > >
>> > > Isn't the object aligned to _S_alignment?
>> >
>> > We did specify that with the alignas.  Is the alignof always
>> > exactly the same as an alignas, if one is specified?  (And will
>> > that not change in a future amendment, standard and/or
>> > implementation?)  Either way, is there a test-case to guard all
>> > this?
>>
>> The language guarantees that's what alignas() does, if the argument is
>> a valid alignment (which it must be if we derive it from some other
>> type's alignment).
>
>I'm more worried about alignof reporting a higher value for a
>specific object than alignas to be wrong.

That shouldn't be possible because the C++ standard says it's an error
to use alignas with a less strict alignment than would be used if it
was omitted, i.e. an error to use alignas with a value less than the
result alignof would give.  However, G++ doesn't reject it (PR65685).

It still won't be possible here, because the alignas value we use is
not less than alignof(_Tp).

>Your question quoted just below seems to indicate a similar
>worry.

I was thinking about cases like this:

  struct __attribute__((packed)) Bad {
    char c;
    std::atomic<long long> a;
  };

But G++ ignores the packed attribute here, which is good (Clang
doesn't seem to ignore it, and mis-aligns the atomic).

>> > > Or is it different if a std::atomic<T> is included in some other
>> > > struct and the user forces a different alignment on it? I don't think
>> > > we really need to support that, users shouldn't be doing that.
>> >
>> > Why do we even need to ask those questions, when the patch takes
>> > care of the per-type business without doubt?
>>
>> Well if we know the object is guaranteed to be correctly aligned we
>> might not even need a fake, minimally aligned pointer. We could go
>> back to passing &_M_i or just a null pointer to __atomic_is_lock_free.
>
>The target has a say in __atomic_is_lock_free and could impose
>some crazy pointer-value-specific condition like "only in the
>first half of a page" (anything can happen to work around chip
>errata) so I suggest staying with an alignment-generated
>pointer.  No, I only dreamt that up, shoot it down if we only
>care for current targets.
>
>> The whole point of alignas() is to fix the alignment to a known value.
>
>And alignof and __alignof to report *exactly* that, I hope?

Yes, alignas sets the alignment requirement, alignof gets it.