Re: egcs-971127: results on powerpc-ibm-aix4.2.1.0

[Franz Sirl <Franz.Sirl-kernel@lauterbach.com>]

At 22:26 Uhr -0000 29.11.1997, Jeffrey A Law wrote:
>  In message < 34801C0C.A67C5FAB@msu.ru >you write:
>  > Why there is a regression in gcc testsuite
>  > (gcc.c-torture/execute/va-arg-5.c)?
>Bug in Meissner's rework of some of Geoffrey Keating's patches;
>
>This should fix the problem (and will be included in the release).
>
>Index: rs6000.c
>===================================================================
>RCS file: /cvs/cvsfiles/egcs/gcc/config/rs6000/rs6000.c,v
>retrieving revision 1.3.2.4
>retrieving revision 1.3.2.5
>diff -c -3 -p -r1.3.2.4 -r1.3.2.5
>*** rs6000.c	1997/11/27 00:06:51	1.3.2.4
>--- rs6000.c	1997/11/29 21:37:41	1.3.2.5
>*************** function_arg_partial_nregs (cum, mode, t
>*** 1387,1395 ****
>  	return 0;
>      }
>
>-   if (GET_MODE_CLASS (mode) == MODE_FLOAT && TARGET_HARD_FLOAT)
>-     return 0;
>-
>    if (cum->words < GP_ARG_NUM_REG
>        && GP_ARG_NUM_REG < (cum->words + RS6000_ARG_SIZE (mode, type,
>named)))
>      {
>--- 1387,1392 ----

Hmm, this also solves a sig11 while compiling glibc-2.1 for me, but it
didn't change the test behaviour on powerpc-unknown-linux-gnulibc1. The
strange thing is, after getting the sig11 I started applying Geoff's
remaining patches step by step and the following one also prohibited the
sig11. Is Michaels rework really buggy?

BTW, while looking at *.s files generated by -save-temps I noticed that I
can't switch on -mmultiple, it is always directly followed by
-mno-multiple. I tried with various combinations of -mpowerpc,
-mcpu=powerpc, etc. and was never able to get rid of the -mno-multiple.
According to the docs this should be possible. Is there any other option
that forces -mno-multiple?

Anyway, egcs-971127 is now quite fit for release from the Linux/PPC
viewpoint ;-). Thanks all the helpful guys involved in this!

Bye,
Franz.

Patch by Geoff Keating, originally for an older version of egcs, so it
applies with offset:

* PIC code breaks when nested functions are used; for instance, the following
  compiled with 'cc1 -O2 -fpic -Wall'.  (The -Wall gives a nice visible crash,
  rather than quietly generating invalid code.)

extern void x1(int), x2(int);
extern int c1, c2;
void x3 (int p) {
  static inline void i1 (void) { x1(c1); }
  static inline void i2 (void) { i1(); }
  if (p) i2();
  x2(c2);
}

--- config/rs6000/rs6000.c~     Thu Aug 28 00:39:28 1997
+++ config/rs6000/rs6000.c      Fri Aug 29 01:16:00 1997
@@ -2292,6 +2292,12 @@

   free (machine);
   p->machine = (struct machine_function *)0;
+
+  /* If the nested function used the GOT, pic_offset_table_rtx would
+     have been set to the hard register that the function was using as
+     a pointer into the GOT table.  Reset it to NULL so that if this
+     function uses the GOT it will get GOT_TOC_REGNUM.  */
+  pic_offset_table_rtx = NULL;
 }

 /* Do anything needed before RTL is emitted for each function.  */