Re: ICE in `find_reloads', at reload.c:4106

[Jeffrey A Law <law at cygnus dot com>]

  In message <200006151747.NAA27641@hiauly1.hia.nrc.ca>you write:
  > > Though I wonder if the no input reloads restriction could be further
  > > refined -- the idea behind that restriction is we can't guarantee
  > > that cc0 will always have the right value after emitting the reload
  > > insns.
  > 
  > I am currently testing.
  > 
  > 2000-06-15  J. David Anglin  <dave@hiauly1.hia.nrc.ca>
  > 
  > 	* reload.c (find_reloads): Don't reload a nonoffsettable memory address
  > 	into an offsettable base register when no_input_reloads is true.
  > 
  > --- reload.c.orig	Thu Jun  1 10:50:59 2000
  > +++ reload.c	Wed Jun 14 20:20:03 2000
  > @@ -3560,7 +3560,8 @@
  >  	   so we don't bother with it.  It may not be worth doing.  */
  >  	else if (goal_alternative_matched[i] == -1
  >  		 && goal_alternative_offmemok[i]
  > -		 && GET_CODE (recog_data.operand[i]) == MEM)
  > +		 && GET_CODE (recog_data.operand[i]) == MEM
  > +		 && !no_input_reloads)
  >  	  {
  >  	    operand_reloadnum[i]
  >  	      = push_reload (XEXP (recog_data.operand[i], 0), NULL_RTX,
  > 
  > It results in roughly the following code for the line that previously
  > aborted:
  > 
  >         movzwl $387,-(sp)
  > 	cmpl _maxtype,*_typevec
  > 	clrl r1
  > 	adwc $0,r1
  > 	pushl r1
  > 	calls $2,_my_friendly_assert
  > 
  > My guess is that "clrl r1; adwc $0,r1; pushl r1" is as efficient if not
  > more so than "clrl -(sp); adwc $0,(sp)".  It is clear that the compiler
  > wants to generate sCODE instructions with auto inc and dec's.  This is
  > not allowed by the vax sltu and sgeu instruction constraints, and thus
  > the reload attempt.  This doesn't happen very often, thus I wonder if
  > it is worth trying to find a way to handle this case.
  > 
  > There are not many cc0 friendly instructions on the vax, except for a
  > few that don't muck with the carry bit C.  It might be tricky to try
  > to further optimise the no_input_reload constraint.
IMHO, the vax machine description is in error.  This insn uses cc0 and
therefore must never need reloading.

You may be able to fix the problem by writing a tighter predicate, another
would be to actually include the set of cc0 within that pattern, then
split it up into multiple patterns after reload.  ie

(define_insn "sltu"
  [(set (match_operand:SI 0 "general_operand" "=ro")
        (ltu (compare (op1, op2)) (const_int 0)))]
  ""
  "#"

Then have a splitter which emits a cmpsi insn followed by an insn that looks
like the current sltu.  
jeff