Re: GDB 4.17 Patch for stack aligned i386 code

[nsd@cygnus.com]

>Are you handling non-leaf functions even when the stack pointer is changing?

Yes.  It works by emulating instructions to figure out the stack pointer
on entry to or exit from the function.  The opcodes library, with a few
minor changes, provides the instruction information in an
architecture-independent format.

>Does your patch handle functions where the prologue uses "andl" to align
>the frame?

I don't know.  Do you happen to have an example of how to generate such
code?  I'd be interested in trying it.

Nick
From geoffk@cygnus.com Fri Mar 24 16:08:00 2000
From: Geoff Keating <geoffk@cygnus.com>
To: gdb-patches@sourceware.cygnus.com
Subject: patch to psim for denormal values
Date: Fri, 24 Mar 2000 16:08:00 -0000
Message-id: <200003250008.QAA31192@localhost.cygnus.com>
X-SW-Source: 2000-03/msg00573.html
Content-length: 1148

gcc/testsuite/gcc.c-torture/execute/ieee/20000320-1.c was sending the
sim into a very long (perhaps infinite) loop, because of this typo.

My copy of the ppc 601 user's manual correctly uses '+' here (on page
F-2).

OK to commit?

-- 
- Geoffrey Keating <geoffk@cygnus.com>

===File ~/patches/cygnus/sim-ppc-denormal.patch=============
2000-03-23  Geoff Keating  <geoffk@cygnus.com>

	* ppc-instructions (Disabled_Exponent_Underflow): Increment
	the exponent when denormalizing.

Index: ppc-instructions
===================================================================
RCS file: /cvs/cvsfiles/devo/sim/ppc/ppc-instructions,v
retrieving revision 1.46
diff -p -u -u -p -r1.46 ppc-instructions
--- ppc-instructions	2000/03/02 09:24:10	1.46
+++ ppc-instructions	2000/03/24 01:53:30
@@ -4282,7 +4282,7 @@ void::function::invalid_zero_divide_oper
 	  }
 	    /* G|R|X == zero from above */
 	    while (exp < -126) {
-	      exp = exp - 1;
+	      exp = exp + 1;
 	      frac_grx = (INSERTED64(EXTRACTED64(frac_grx, 0, 54), 1, 55)
 	                  | MASKED64(frac_grx, 55, 55));
 	    }
============================================================
From john@feith.com Fri Mar 24 16:30:00 2000
From: John Wehle <john@feith.com>
To: nsd@bosbc.com
Cc: cgf@cygnus.com, eliz@delorie.com, gdb-patches@sourceware.cygnus.com, jimb@cygnus.com, msnyder@cygnus.com
Subject: Re: GDB 4.17 Patch for stack aligned i386 code
Date: Fri, 24 Mar 2000 16:30:00 -0000
Message-id: <200003250030.TAA14328@jwlab.FEITH.COM>
X-SW-Source: 2000-03/msg00574.html
Content-length: 1281

>>Are you handling non-leaf functions even when the stack pointer is changing?
>
> Yes.  It works by emulating instructions to figure out the stack pointer
> on entry to or exit from the function.  The opcodes library, with a few
> minor changes, provides the instruction information in an
> architecture-independent format.

Cool.

>>Does your patch handle functions where the prologue uses "andl" to align
>>the frame?
>
> I don't know.  Do you happen to have an example of how to generate such
> code?  I'd be interested in trying it.

I'm not sure what I still have laying around other than a very old patch
for GCC.  The idea is to handle a prologue which looks like:

func:
        pushl %ebp
        movl %esp,%ebp
        andl $-15,%esp
        subl $32,%esp

function arguments are relative to %ebp and local variables are relative to
%esp.  Extra alignment has been applied to the stack pointer in order to
ensure that local variables don't cross a cache boundary.

-- John
-------------------------------------------------------------------------
|   Feith Systems  |   Voice: 1-215-646-8000  |  Email: john@feith.com  |
|    John Wehle    |     Fax: 1-215-540-5495  |                         |
-------------------------------------------------------------------------