Re: [PATCH 1/5] std::latch: reduce internal implementation from ptrdiff_t to int

[Jonathan Wakely <jwakely@redhat.com>]

On 03/03/21 14:56 +0000, Jonathan Wakely wrote:
>On 01/03/21 09:56 +0100, Richard Biener via Libstdc++ wrote:
>>On Sun, Feb 28, 2021 at 10:53 PM Hans-Peter Nilsson <hp@bitrange.com> wrote:
>>>
>>>
>>>
>>>On Fri, 26 Feb 2021, Thiago Macieira via Gcc-patches wrote:
>>>
>>>> On Friday, 26 February 2021 11:31:00 PST Andreas Schwab wrote:
>>>> > On Feb 26 2021, Thiago Macieira wrote:
>>>> > > On Friday, 26 February 2021 10:14:42 PST Andreas Schwab wrote:
>>>> > >> On Feb 26 2021, Thiago Macieira via Gcc-patches wrote:
>>>> > >> > -    alignas(__alignof__(ptrdiff_t)) ptrdiff_t _M_a;
>>>> > >> > +    alignas(__alignof__(int)) int _M_a;
>>>> > >>
>>>> > >> Futexes must be aligned to 4 bytes.
>>>> > >
>>>> > > Agreed, but doesn't this accomplish that?
>>>> >
>>>> > No.  It uses whatever alignment the type already has, and is an
>>>> > elaborate no-op.
>>>>
>>>> I thought so too when I read the original line. But I expected it was written
>>>> like that for a reason, especially since the same pattern appears in other
>>>> places.
>>>>
>>>> I can change to "alignas(4)" (which is a GCC extension, I believe). Is that
>>>> the correct solution?

It's not a GCC extensions. You're thinking of alignas(obj) which is a
GCC extension.

>>>IMNSHO make use of the corresponding atomic type.  Then there'd
>>>be no need for separate what's-the-right-align-curse games.
>
>That won't work though, because we need direct access to the integer
>object, not to a std::atomic<int> which contains it.
>
>>Or align as the corresponding atomic type (in case using an actual
>>std::atomic<int> is undesirable).  OTOH the proposed code isn't
>>any more bogus than the previous ...
>
>The previous code accounts for the fact that ptrdiff_t is a typedef
>for an unspecified type, and that some ABIs allow struct members to have
>weaker alignment than they would have otherwise.
>
>e.g. __alignof__(long long) != alignof(long long) on x86.
>
>Yes, I know ptrdiff_t isn't long long on x86, but since it's a typedef
>for some target-specific type, it's still possible that
>alignof != __alignof__. So alignas(__alignof__(T)) is not necessarily
>a no-op. So not bogus.
>
>For int, there shouldn't be any need to force the alignment. I don't
>think any ABI supported by GCC allows int members to be aligned to
>less than __alignof__(int). Users could break it by compiling with
>#pragma pack, but I have no sympathy for such silliness.

Jakub said on IRC that m68k might have alignof(int) == 2, so we'd need
to increase that alignment to use it as a futex.

N.B. std::atomic and std::atomic_ref don't use alignas(__alignof__(T))
they do this instead:

       static_assert(is_trivially_copyable_v<_Tp>);

       // 1/2/4/8/16-byte types must be aligned to at least their size.
       static constexpr int _S_min_alignment
	= (sizeof(_Tp) & (sizeof(_Tp) - 1)) || sizeof(_Tp) > 16
	? 0 : sizeof(_Tp);

     public:

       static constexpr size_t required_alignment
	= _S_min_alignment > alignof(_Tp) ? _S_min_alignment : alignof(_Tp);

Using std::atomic_ref<T>::required_alignment would give that value.

For something being used as a futex we should just use alignas(4)
though, since that's what the kernel requires.