Re: PATCH [mainline] apple-ppc-darwin passing

[Fariborz Jahanian <fjahanian at apple dot com>]

On Jun 25, 2004, at 5:40 AM, Alan Modra wrote:

> On Thu, Jun 24, 2004 at 12:47:15PM -0700, Fariborz Jahanian wrote:
> > *************** emit_push_insn (rtx x, enum machine_mode
> > *** 3609,3614 ****
> > --- 3610,3627 ----
> >         int args_offset = INTVAL (args_so_far);
> >         int skip;
> >
> > +       if (mode == DFmode && reg && GET_CODE (reg) == PARALLEL)
> > +         {
> > + 	  /* Pattern for splitting up a DFmode part in GRP, part on stack
> > + 	     as well as fully in an FPR. */
>
> Your fix to emit_push_insn looks good to me, except that you really
> don't want the DFmode test here.  There's nothing special about DFmode.

No there isn't. But I just wanted to check the condition which I knew 
caused this
bug and protect against other cases of PARALLEL being gen'd. As you 
pointed
out below, there maybe hackery on other platforms that I wanted to 
guard against
and not breaking them. But thanks for your comments.

- fariborz


> It just happens to be one of the reg types, of size different to
> UNITS_PER_WORD, that you need to handle specially here.  I came up with
> with a similar patch, minus the DFmode test, when developing
> http://gcc.gnu.org/ml/gcc-patches/2004-05/msg00665.html.
>
> The reason I didn't post the patch at that stage was that I wanted the
> rs6000 backend fixes in first.  Also, you need to verify that any other
> backend emitting PARALLELs for arg passing doesn't have some hack in
> function_arg_partial_nregs to work around the emit_push_insn bug.  I
> didn't go so far as to do this verification..
>
> Oh, and I'm not qualified to review this patch, just commenting as
> someone who has fiddled enough with the function arg machinery to be
> reasonably familiar with this code.
>
> --
> Alan Modra
> IBM OzLabs - Linux Technology Centre