gcse bugfix

[law at redhat dot com]

gcse can turn conditional jumps into nop-jumps  ie they can be turned into

(set (pc) (pc))


When this happens we properly decrement the LABEL_NUSES for the CODE_LABEL
associated with the jump.  However, we leave the JUMP_LABEL field alone.

This causes problems when cse (or some other pass) removes the nop jump
via delete_insn.  Why?  Because delete_insn decrements LABEL_NUSES for
any code label associated with a jump that's being deleted.    Thus we
can end up decrementing LABEL_NUSES twice.

The incorrect reference counting can potentially cause any number of
problems.  In my particular case it was causing cse.c to believe it
could extend a basic block, when in fact it should not have (we thought
we passed the only label reference, when in fact there were 2 label
references).  Extension of the block ultimately led to incorrect code as
we had incorrect equivalences in the hash table.

This has survived a bootstrap on PA32 hpux11 as well as x86-linux.  No
regressions in the testsuite on either platform -- and of course it fixes
a number of PA execution failures.


	* gcse.c (cprop_jump): Clear JUMP_LABEL field when we create
	a nop-jump.

Index: gcse.c
===================================================================
RCS file: /cvs/cvsfiles/devo/gcc/gcse.c,v
retrieving revision 1.125
diff -c -3 -p -r1.125 gcse.c
*** gcse.c	2001/10/22 08:35:07	1.125
--- gcse.c	2001/11/19 18:59:51
*************** cprop_jump (bb, insn, from, src)
*** 4256,4262 ****
        SET_SRC (set) = new;
  
        if (JUMP_LABEL (insn) != 0)
! 	--LABEL_NUSES (JUMP_LABEL (insn));
      }
  
    /* Otherwise, this must be a valid instruction.  */
--- 4256,4265 ----
        SET_SRC (set) = new;
  
        if (JUMP_LABEL (insn) != 0)
! 	{
! 	  --LABEL_NUSES (JUMP_LABEL (insn));
! 	  JUMP_LABEL (insn) = NULL_RTX;
! 	}
      }
  
    /* Otherwise, this must be a valid instruction.  */