Re: [PATCH] remove split condition from *fix_truncdi_1, *fix_truncsi_1

[Uros Bizjak <uros at kss-loka dot si>]

Hello!

> > > > -  ""
> > > > +  "&& 1"
> > > >        
> > 2004-11-12  Uros Bizjak  <uros@kss-loka.si>
> >
> >     * configure/i386/i386.md: (*fix_truncdi_1, *fix_truncsi_1):
> >     Remove unneeded "&& 1" from split constraint.
> >     (two unnamed define_peephole2): Remove unneeded "&& 1" from
> >     constraints.
> >    
Zack Weinberg wrote:

> Looks good to me; check it in.
>
> zw
>  
Richard Henderson wrote:

> NOT ok.  See my previous reply.
>
> r~
>  
From the documentation, define_insn_and_split is equal to define_insn 
and define_split of the same pattern.

This instruction (without "&& 1" in its split constraint):

(define_insn_and_split "*fix_truncdi_1"
 [(set (match_operand:DI 0 "nonimmediate_operand" "=m,?r")
   (fix:DI (match_operand 1 "register_operand" "f,f")))
  (clobber (reg:CC FLAGS_REG))]
 "TARGET_80387 && FLOAT_MODE_P (GET_MODE (operands[1]))
  && !reload_completed && !reload_in_progress
  && (!SSE_FLOAT_MODE_P (GET_MODE (operands[1])) || !TARGET_64BIT)"
 "#"
 ""
 [(const_int 0)]
{
...
}
 [(set_attr "type" "fistp")
  (set_attr "i387_cw" "trunc")
  (set_attr "mode" "DI")])

is then equal to:

(define_insn "*fix_truncdi_1"
 [(set (match_operand:DI 0 "nonimmediate_operand" "=m,?r")
   (fix:DI (match_operand 1 "register_operand" "f,f")))
  (clobber (reg:CC FLAGS_REG))]
 "TARGET_80387 && FLOAT_MODE_P (GET_MODE (operands[1]))
  && !reload_completed && !reload_in_progress
  && (!SSE_FLOAT_MODE_P (GET_MODE (operands[1])) || !TARGET_64BIT)"
 "#"
 [(set_attr "type" "fistp")
  (set_attr "i387_cw" "trunc")
  (set_attr "mode" "DI")])

with

(define_split
 [(set (match_operand:DI 0 "nonimmediate_operand" "")
   (fix:DI (match_operand 1 "register_operand" "")))
  (clobber (reg:CC FLAGS_REG))]
 ""
 [(const_int 0)]
{
...
})

The instruction will only be generated, when its condition will match, 
and will be splitted afterwards.  However,  there is a small (-> zero?) 
possibility that combiner produces an exact pattern that match the split 
pattern. In this case, instruction will be split without any split 
constraints. To be safe, should  "&& 1" be present in split condition?

Thanks,
Uros.