patch for update_flow_info REG_WAS_0

patch for update_flow_info REG_WAS_0

[Joern Rennecke]

update_flow_info tried in vain to find the new setter for a (mem:DI ...
value after this was split into two memory accesses...

Fri Feb 13 00:25:02 1998  J"orn Rennecke <amylaar@cygnus.co.uk>

	* sched.c (update_flow_info, REG_WAS_0): Ignore if setting insn
	was deleted.

Index: sched.c
===================================================================
RCS file: /cvs/cvsfiles/devo/gcc/sched.c,v
retrieving revision 1.120
diff -p -r1.120 sched.c
*** sched.c	1998/02/06 08:35:42	1.120
--- sched.c	1998/02/13 00:28:12
***************
*** 1,5 ****
  /* Instruction scheduling pass.
!    Copyright (C) 1992, 93-96, 1997 Free Software Foundation, Inc.
     Contributed by Michael Tiemann (tiemann@cygnus.com)
     Enhanced by, and currently maintained by, Jim Wilson (wilson@cygnus.com)
  
--- 1,5 ----
  /* Instruction scheduling pass.
!    Copyright (C) 1992, 93-97, 1998 Free Software Foundation, Inc.
     Contributed by Michael Tiemann (tiemann@cygnus.com)
     Enhanced by, and currently maintained by, Jim Wilson (wilson@cygnus.com)
  
*************** update_flow_info (notes, first, last, or
*** 3854,3859 ****
--- 3854,3867 ----
  	  break;
  
  	case REG_WAS_0:
+ 	  /* If the insn that set the register to 0 was deleted, this
+ 	     note cannot be relied on any longer.  The destination might
+ 	     even have been moved to memory.
+              This was observed for SH4 with execute/920501-6.c compilation,
+ 	     -O2 -fomit-frame-pointer -finline-functions .  */
+ 	  if (GET_CODE (XEXP (note, 0)) == NOTE
+ 	      || INSN_DELETED_P (XEXP (note, 0)))
+ 	    break;
  	  /* This note applies to the dest of the original insn.  Find the
  	     first new insn that now has the same dest, and move the note
  	     there.  */

Re: patch for update_flow_info REG_WAS_0

[Jeffrey A Law]

  In message < 199802130030.AAA24740@phal.cygnus.co.uk >you write:
  > update_flow_info tried in vain to find the new setter for a (mem:DI ...
  > value after this was split into two memory accesses...
  > 
  > Fri Feb 13 00:25:02 1998  J"orn Rennecke <amylaar@cygnus.co.uk>
  > 
  > 	* sched.c (update_flow_info, REG_WAS_0): Ignore if setting insn
  > 	was deleted.
One might argue that the note should have been removed when the
insn was deleted.  Can you look into that?

jeff

Re: patch for update_flow_info REG_WAS_0

[Joern Rennecke]

> One might argue that the note should have been removed when the
> insn was deleted.  Can you look into that?

I first looked at the REG_WAS_0 description in rtl.texi.  It appears that
is a known fact that the insn might be deleted.  So I presume that that's
OK, and we have to deal with it.

@findex REG_WAS_0
@item REG_WAS_0
The single output of this insn contained zero before this insn.
@var{op} is the insn that set it to zero.  You can rely on this note if
it is present and @var{op} has not been deleted or turned into a @code{note};
its absence implies nothing.
@end table

These notes describe linkages between insns.  They occur in pairs: one
insn has one of a pair of notes that points to a second insn, which has
the inverse note pointing back to the first insn.

Re: patch for update_flow_info REG_WAS_0

[Jeffrey A Law]

  In message < 199802131014.KAA10625@phal.cygnus.co.uk >you write:
  > > One might argue that the note should have been removed when the
  > > insn was deleted.  Can you look into that?
  > 
  > I first looked at the REG_WAS_0 description in rtl.texi.  It appears that
  > is a known fact that the insn might be deleted.  So I presume that that's
  > OK, and we have to deal with it.
Ok.  Thanks.

In that case your patch is the right way to go.  Please install
it into sched.c and haifa-sched.c.

jeff