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

[Hans-Peter Nilsson]

On Thu, 12 Feb 2015, Richard Henderson wrote:
> When we fixed PR54005,

Hm, there's confusion.  When was this fixed?  (Not fixed
AFAICT.)  Maybe you mean PR54004, but there was no note there
either.  Or maybe there's a typo and you meant some other PR and
that PR54005 is supposedly fixed by this patch (committed as
r221701)

...but it doesn't seem right: you use a specific object when
deducing the alignment for the fake-pointer, so it's used anyway
and is_lock_free must not be object-specific (despite its name)
and only type-specific as mandated by the standard (see PR).

To wit, deduce from the known-alignment of _Tp, not
known-alignment of _M_i. Or is this what happens; they're the
same?  Why then use __alignof(_M_i) (the object-alignment)
instead of _S_alignment (the deduced alas insufficiently
increased type-alignment)?

> making sure that atomic_is_lock_free returns the same
> value for all objects of a given type,

(That would work but it doesn't seem to be the case.)

> we probably should have changed the
> interface so that we would pass size and alignment rather than size and object
> pointer.
>
> Instead, we decided that passing null for the object pointer would be
> sufficient.  But as this PR shows, we really do need to take alignment into
> account.

Regarding what's actually needed, alignment of an atomic type
should always be *forced to be at least the natural alignment of
of the object* (with non-power-of-two sized-objects rounded up)
and until then atomic types won't work for targets where the
non-atomic equivalents have less alignment (as straddling a
page-boundary won't be lock-less-atomic anywhere where
straddling a page-boundary may cause a non-atomic-access...) So,
not target-specific except for targets that require even
higher-than-natural alignment.

> The following patch constructs a fake object pointer that is maximally
> misaligned.

(With regards to the known object alignment of the _M_i object.)

>  This allows the interface to both the builtin and to libatomic to
> remain unchanged.  Which probably makes this back-portable to maintenance
> releases as well.
>
> I believe that for all of our current systems, size_t == uintptr_t, so the
> reinterpret_cast ought not generate warnings.
>
> The test case is problematic, as there's currently no good place to put it.
> The libstdc++ testsuite doesn't have the libatomic library path configured, and
> the libatomic testsuite doesn't have the libstdc++ include paths configured.
> Yet another example where we really need an install tree for testing.  Thoughts?

brgds, H-P