Re: other/7114: ICE building strcoll.op from glibc-2.2.5

[Geoff Keating <geoffk@geoffk.org>]

The following reply was made to PR other/7114; it has been noted by GNATS.

From: Geoff Keating <geoffk@geoffk.org>
To: amodra@bigpond.net.au
Cc: d.mueller@elsoft.ch, gcc-gnats@gcc.gnu.org, gcc-patches@gcc.gnu.org,
   dje@watson.ibm.com
Subject: Re: other/7114: ICE building strcoll.op from glibc-2.2.5
Date: Wed, 17 Jul 2002 00:07:11 -0700

 > Date: Wed, 17 Jul 2002 11:17:39 +0930
 > From: Alan Modra <amodra@bigpond.net.au>
 
 > On Tue, Jul 16, 2002 at 10:48:14AM -0700, Geoff Keating wrote:
 > > The profiling function isn't allowed to clobber r30.  It never has
 > > been, so this should be no surprise.
 > 
 > I'm not quite sure what to make of this response.  We're talking about
 > this code from rs6000.c:10479
 > 
 >       if (current_function_needs_context)
 > 	asm_fprintf (file, "\tmr %s,%s\n",
 > 		     reg_names[30], reg_names[STATIC_CHAIN_REGNUM]);
 >       fprintf (file, "\tbl %s\n", RS6000_MCOUNT);
 >       if (current_function_needs_context)
 > 	asm_fprintf (file, "\tmr %s,%s\n",
 > 		     reg_names[STATIC_CHAIN_REGNUM], reg_names[30]);
 
 I see, I was confused.  I thought we were already using r30 for
 STATIC_CHAIN_REGNUM.  The ABI specifies r31, but I see we can't use
 that if we want trampolines to be efficient.
 
 > This is currently emitted _before_ the prologue in the nested function,
 > thus trashes r30.  I was considering the idea of adding a clobber of
 > r30 to CALL_INSN_FUNCTION_USAGE when calling a nested function.  That's
 > a workable solution, but means you need to zap r30 on all calls via
 > function pointers too.
 
 From David's mail message when he put the code in:
 
      I have ripped out all of the stack PUSH/POP stuff that was
      causing ABI problems and replaced it with explicit moves to a
      temporary register.  This includes having the SVR4 ABI act more
      like AIX using a register instead of the dangerous stack
      save/restore game.  I could not test the SVR4 changes, so I would
      appreciate if the LinuxPPC testers would make sure that I have
      not broken anything when profiling is enabled.
 
 So, thanks for testing it!  We now know this doesn't work. :-)
 
 So, why don't we go back to the push/pop implementation, but this time
 do it properly?  We'd only need to push/pop in the (rare)
 nested-function case.
 
 -- 
 - Geoffrey Keating <geoffk@geoffk.org> <geoffk@redhat.com>