Re: allocsa module update

[Ralf Wildenhues]

Hello Bruno,

A while ago you applied this go gnulib:

* Bruno Haible wrote on Sat, Jul 22, 2006 at 03:53:42PM CEST:
> Merge from gettext:

> 2005-02-10  Bruno Haible  <address@hidden>
> 
>       * allocsa.h (sa_alignof): Define differently with AIX xlc, to avoid
>       a bug of this compiler on AIX 3.2.5 dealing with enums.

> *** allocsa.h 14 May 2005 06:03:57 -0000      1.4
> --- allocsa.h 22 Jul 2006 13:55:31 -0000

> ***************
> *** 81,86 ****
> --- 92,101 ----
>     /* Work around a HP-UX 10.20 cc bug with enums constants defined as 
> offsetof
>        values.  */
>   # define sa_alignof(type) (sizeof (type) <= 4 ? 4 : 8)
> + #elif defined _AIX
> +   /* Work around an AIX 3.2.5 xlc bug with enums constants defined as 
> offsetof
> +      values.  */
> + # define sa_alignof(type) 4
>   #else
>   # define sa_alignof(type) offsetof (struct { char __slot1; type __slot2; }, 
> __slot2)
>   #endif

It looks rather scary to me, for two reasons: First, the alignment may
be altered by the -qalign compiler option; granted, this is not so much
of a concern in practice as it creates binary-incompatible software.
But second, newer AIX versions may align 'long long' on an 8 byte
boundary.  So, if sa_alignof needs only mention an upper bound (which I
don't know, as I don't know this code), then that should be 8.  But
maybe it's either not worth catering for AIX 3.2.5 any more (hey, AIX
4.3.3 has reached end of service life for a while), or you could at
least check for something like
  #elif defined _AIX && !defined _AIX41

to limit impact for newer versions (AFAIK all newer versions define
the macro _AIX41), given that this is fixed there.

What was the original bug report that prompted this, by the way?  It
should affect some Autoconf proper code as well.

Cheers,
Ralf