Re: How aggressive is CSE about removing redundant CC sets?

[Ulrich Weigand <weigand at i1 dot informatik dot uni-erlangen dot de>]

Hi Kazu,

> > (Optionally, it may be beneficial to provide a third pattern that
> > has *only* the SET CC, and a (clobber (match_scratch)) to do away
> > with the main output; this will be chosen if you're performing the
> > operation only to look at the CC.)
> 
> I'm sorry, I don't quite understand.  Could you give me an example
> pattern?  Something like this?

Most arithmetic instructions in the s390 backend follow the scheme of 
having those three patterns, called XXX, XXX_cc, and XXX_cconly.
As a prototypical example here's the xordi3 patterns (reg 33 is
the condition code; the s390_match_ccmode stuff handles the fact
that more than one CCmode is acceptable for these insns):

This is the expander (just clobbers CC):
(define_insn "xordi3"
  [(set (match_operand:DI 0 "register_operand" "=d,d")
        (xor:DI (match_operand:DI 1 "nonimmediate_operand" "%0,0")
                (match_operand:DI 2 "general_operand" "d,m")))
   (clobber (reg:CC 33))]
  "TARGET_64BIT"
  "@
   xgr\t%0,%2
   xg\t%0,%2"
  [(set_attr "op_type"  "RRE,RXY")])

The full pattern with two SETs:
(define_insn "*xordi3_cc"
  [(set (reg 33)
        (compare (xor:DI (match_operand:DI 1 "nonimmediate_operand" "%0,0")
                         (match_operand:DI 2 "general_operand" "d,m"))
                 (const_int 0)))
   (set (match_operand:DI 0 "register_operand" "=d,d")
        (xor:DI (match_dup 1) (match_dup 2)))]
  "s390_match_ccmode(insn, CCTmode) && TARGET_64BIT"
  "@
   xgr\t%0,%2
   xg\t%0,%2"
  [(set_attr "op_type"  "RRE,RXY")])

And the 'cconly' pattern that has a SET for CC and a CLOBBER
for the output of the arithmetical operation:
(define_insn "*xordi3_cconly"
  [(set (reg 33)
        (compare (xor:DI (match_operand:DI 1 "nonimmediate_operand" "%0,0")
                         (match_operand:DI 2 "general_operand" "d,m"))
                 (const_int 0)))
   (clobber (match_scratch:DI 0 "=d,d"))]
  "s390_match_ccmode(insn, CCTmode) && TARGET_64BIT"
  "@
   xgr\t%0,%2
   xr\t%0,%2"
  [(set_attr "op_type"  "RRE,RXY")])

The advantage of the last is that it, like the first, is accepted
by single_set, and thus takes part in many optimization steps
the second pattern doesn't.  It will be chosen by combine if only
the result of the comparison is needed.

Bye,
Ulrich

-- 
  Dr. Ulrich Weigand
  weigand@informatik.uni-erlangen.de