c/9700: va_start gets __offset wrong in some situations

[Aron Griffis]

bangerth@dealii.org wrote:[Fri Feb 14 2003, 09:10:12PM EST]
>     Admittedly I only looked at the program given in the first
>     report. So, let me ask again: I think that that program
>     does exactly what should be expected, is this right? If so,
>     the comment in that program about expectations is wrong.

Ah, this is where we're not communicating.  The comment is accurate.  I
can rephrase it as follows, perhaps this is more clear:

    /*
     * This is the first invocation of va_arg.  The argument list is as
     * follows:
     *
     * 		a = 0
     * 		b = 0
     * 		c = 0
     * 		d = 0
     * 		e = 0
     * 		ap = { undef, undef }
     * 		first_custom = "one"
     * 		... = "two", "three"
     *
     * Therefore va_arg should return "two".  However in this example it
     * erroneously returns "one", which is first_custom.
     */

>     Second question: if the program in the first report doesn't
>     show the problem at all, is it demonstrated in the second?
>     If so, I'll be happy to take a look at that one on Monday.

It is demonstrated in both, but most clearly in the second, where __base
is clearly pointing 8-bytes prior to the correct location.  It seems
like the alignment of __base in the single instance is being corrected
downward instead of upward.

Note you will need an Alpha to build the second example program since
the definition of va_list is different on x86.  However it might not be
necessary since I've posted the output.

Aron