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

[David Edelsohn]

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

From: David Edelsohn <dje@watson.ibm.com>
To: Geoff Keating <geoffk@redhat.com>
Cc: amodra@bigpond.net.au, d.mueller@elsoft.ch, gcc-gnats@gcc.gnu.org,
   gcc-patches@gcc.gnu.org
Subject: Re: other/7114: ICE building strcoll.op from glibc-2.2.5 
Date: Wed, 17 Jul 2002 16:28:02 -0400

 >>>>> Geoff Keating writes:
 
 >> PUSH/POP cannot work on PowerPC.  On AIX PUSH/POP were corrupting the
 >> stack.
 
 Geoff> Because they were implemented wrongly?
 
 Geoff> Clearly, push/pop can work, because procedures push and pop call
 Geoff> frames all the time.  It's just necessary to do it the right way. 
 
 	We are discussing PUSH/POP in the context of the macros to save
 and restore individual registers used when generating a profiler call in
 final.c, so GCC's ability to correctly allocate call frames is irrelevant.
 
 	The discussion about the PUSH/POP macros in March 1999 provides
 more background about the problem: the macros were not respecting the
 ABI-defined location of dynamic allocation on the stack.  PUSH/POP ideally
 should work like alloca and open a hole in the stack frame at the alloca
 location.  The PUSH/POP macros modify the stack pointer without updating
 the compiler's internal knowledge about the stack layout and assume that
 the alloca area is adjacent to the top of the stack, so adjusting the
 stack pointer will not have any bad consequences, e.g., if a function call
 is invoked.
 
 	On AIX, things went haywire because the PUSH/POP macros only
 copied the backchain pointer, not all of the ABI-specified area between
 the backchain pointer and the alloca area.  When GCC called the profiler
 function, the call frame had random garbage in ABI-specified locations,
 causing the application to run off the rails when an ABI stack location
 was accessed in the called function.  Copying the backchain *and* CR *and*
 LR may be sufficient for the AIX case for this particular use.
 
 	Because SVR4 PowerPC invokes the profiler before the prologue and
 has a slightly different stack layout, the simplistic definition of
 PUSH/POP may work correctly.
 
 David