Re: [Patch] Altivec ABI as default for powerpc64 linux

[Hartmut Penner <HPENNER at de dot ibm dot com>]

Geoff wrote:
> > > > For starters, try this with powerpc-linux-gnu and -mabi=altivec
(but
> > > > no -maltivec), and fix it :).
> > > >
> > > >         typedef int vec __attribute__((mode(V4SI)));
> > > >         vec poo, bar;
> > > >         extern void funk(vec);
> > > >         f(){
> > > >                 funk(poo);
> > > >         }
> > >
> > > How can this work?  "funk" expects its parameter in v1, but without
> > > -maltivec there are no instructions that can put it there.
> >
> > Exactly.  It won't.  If they want one ABI, they need to fix all these
> > places.  I haven't looked in this particular case, but the test should
> > probably be TARGET_ALTIVEC_ABI *and* TARGET_ALTIVEC, otherwise stuff
> > should go in GPRs.

> So how does "funk" know whether or not it was called by something that
> was compiled with -maltivec?

Yes, there is no way to solve this. I would propose an error message,
if somebody wants pass vector register values, but has none available.
To have abi=altivec as default prevents that altivec applications and
non altivec shared libraries don't work together. To have a clean error
message rather than a ICE helps developer to find out, if their
executable only run on altivec hardware or not.
      Hartmut


Index: rs6000.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/config/rs6000/rs6000.c,v
retrieving revision 1.587
diff -u -p -r1.587 rs6000.c
--- rs6000.c      2 Feb 2004 06:31:05 -0000     1.587
+++ rs6000.c      3 Feb 2004 08:06:32 -0000
@@ -840,6 +840,13 @@ rs6000_override_options (const char *def
      rs6000_long_double_type_size = size;
     }

+  /* Set Altivec ABI as default for powerpc64 linux.  */
+  if (TARGET_ELF && TARGET_64BIT)
+    {
+      rs6000_altivec_abi = 1;
+      rs6000_altivec_vrsave = 1;
+    }
+
   /* Handle -mabi= options.  */
   rs6000_parse_abi_options ();

@@ -3980,6 +3987,8 @@ function_arg_advance (CUMULATIVE_ARGS *c
     {
       if (USE_ALTIVEC_FOR_ARG_P (cum, mode, type, named))
      cum->vregno++;
+      if (!TARGET_ALTIVEC)
+     error ("Sorry, no altivec registers available to pass argument");

       /* PowerPC64 Linux and AIX allocates GPRs for a vector argument
       even if it is going to be passed in a vector register.
@@ -4296,6 +4305,10 @@ function_arg (CUMULATIVE_ARGS *cum, enum
           and also in GPRs or memory, in absence of prototype.  */
        int align_words;
        rtx slot;
+
+       if (!TARGET_ALTIVEC)
+      error ("Sorry, no altivec registers available to pass argument");
+
        align_words = (cum->words + 1) & ~1;

        if (align_words >= GP_ARG_NUM_REG)