memory is not free

[Mike Stump]

> Date: Wed, 1 Dec 1999 19:10:59 -0800
> From: Richard Henderson <rth@cygnus.com>

> On Wed, Dec 01, 1999 at 12:09:50AM -0800, Alex Samuel wrote:
> > ! 	    case 'e': case 'u':
> > ! 	      exp = XEXP (r, i);
> > ! 	      if (ggc_test_and_set_mark (exp))
> > ! 		{ 
> > ! 		  if (next_rtx == NULL) 
> > ! 		    next_rtx = exp; 
> > ! 		  else 
> > ! 		    ggc_mark_rtx_children (exp);
> > ! 		} 

> We might want to special-case the instruction rtxes.  I would imagine
> that we could keep the nesting depth significantly lower by forcing
> NEXT_INSN to be the rtx we optimize for the tail call.

:-) My code was exceptionally clear.  You knew that 1 was the tail
recursive case.  His code behaves _exactly_ like mine (in all the ways
that matter), though I had to think about it for a while before I
convinced myself this _was_ the case.  The reason is that 0 is an int,
1 is first one that hit this code.  1 is the important one, as start
marking with the return insn (artifact of other parts of the GC code),
and work backwards (recursing).  So while NEXT_INSN _seems_ like the
right one, it is not.  PREV_INSN is.  Anyway, his code is more robust
than my version, as if one day we need to walk the other direction, his
code will just work.

Welcome to gcc.