[PATCH] test: only test for double alignment on 64 bit. (#36986)

[Peter Hutterer]

On Thu, May 19, 2011 at 09:58:00PM +0200, Julien Cristau wrote:
> On Thu, May 19, 2011 at 14:54:28 +1000, Peter Hutterer wrote:
> 
> > X.Org Bug 36986 <http://bugs.freedesktop.org/show_bug.cgi?id=36986>
> > 
> > Signed-off-by: Peter Hutterer <peter.hutterer at who-t.net>
> > ---
> > 
> > On Wed, May 18, 2011 at 12:45:05AM -0700, Keith Packard wrote:
> > > On Wed, 18 May 2011 13:30:44 +1000, Peter Hutterer
> > > <peter.hutterer at who-t.net> wrote:
> > > 
> > > > +    const int sz_dbl = sizeof(double);
> > > >      /* force alignment with double */
> > > > -    union align_u { ValuatorClassRec valc; double d; } *align;
> > > > +    union align_u {
> > > > +        ValuatorClassRec valc;
> > > > +        double d[(sizeof(ValuatorClassRec) + sz_dbl - 1)/sz_dbl];
> > > > +    } *align;
> > > 
> > > This doesn't make sense to me. The requirement for the union is that you
> > > be able to allocate an array of them and store into each element:
> > > 
> > >         union align_u   foo[12];
> > > 
> > >         foo[0].d = 0.0;
> > >         foo[1].d = 1.0;
> > > 
> > > This should require double alignment for the entire union, even though
> > > 'd' is far smaller than 'valc'.
> > > 
> > > Of course, on a 32-bit x86 machine, doubles can be stored without
> > > penalty on 4-byte boundaries, so the union is only aligned to 52 bytes.
> > > 
> > > Can you explain why this alignment isn't acceptable in this context?
> > 
> > because I naïvely thought that we needed double alignment on 32 bit too and
> > the test was written that way. Oh well, how about this patch instead then.
> > it just disables the alignment check on 32 bit.
> > 
> Maybe check for __alignof__(double) instead of sizeof(double)?

__alignof__(double) on my i686 install claims 8, even though that doesn't
seem to be true.

Cheers,
  Peter