Re: [patch] libffi darwin structure implementation (PPC)

[Andreas Tobler <toa at pop dot agri dot ch>]

David Edelsohn wrote:

> 	This patch probably is algorithmically correct for long double,
> but not for structures in general.  Long doubles are word aligned, so
> there is no strangeness about padding.
>
> 	Currently GCC structure passing on Darwin does not conform to the
> AIX/Darwin ABI.  The recent patch to the AIX port makes it conform and a
> patch to the PPC64 Linux port makes it conform to its revised ABI (padded
> on the opposite end from AIX).  Darwin continues the backward-compatible
> alignment that depends on whether the object size is a power of 2.


You say that the things written in the OS-X runtime convention do not 
match the reality? Would not astonish me.
Ok, which patch? I don't see any recent patch for AIX. With the new 
testsuite I continously check if my tests pass or not.
The last build was from 'gcc version 3.4 20030828 (experimental)'

I'll happily rework my patch when I know on what to rely on.
Well, not so happy since I spent a lot of work into this and now you 
make me think it's somehow wasted. (only time aspect, not personal)


> 	Some technical comments about the patch:
>
> +      /* Don't align arguments for darwin since the first element in a 
> +       * struct is responsible for alignment. We take care about this in 
> +       * prep_cif_machdep.
> +       */
> +
> +#ifndef POWERPC_DARWIN
>        arg->size = ALIGN(arg->size, (*ptr)->alignment);
> +#endif
>        arg->size += (*ptr)->size;
>  
>        arg->alignment = (arg->alignment > (*ptr)->alignment) ? 
>
> This comment does not make sense and does not match the code.  The code is
> calculating the size, not the alignment.  The question is whether 
> arg->size is the raw value or the value rounded up to the alignment.

You're right. Sorry. Hopefully fixed.

> +	  if (ng < 8 ) {  
>                     ^--- minor formatting mistake.

Fixed, plus two more instances.

Thank you for the review time!

Regards,

Andreas