Re: RFA: autoincrement patches for gcc 4 - updated patch

[Joern RENNECKE <joern dot rennecke at st dot com>]

Ian Lance Taylor wrote:

> Joern RENNECKE <joern.rennecke@st.com> writes:
>
>  
>
> > gen_add3_insn (rtx r0, rtx r1, rtx c)
> > {
> >   int icode = (int) add_optab->handlers[(int) GET_MODE (r0)].insn_code;
> > +  int mcode;
> > +  rtx s;
> >
> >   if (icode == CODE_FOR_nothing
> >       || !(insn_data[icode].operand[0].predicate
> > -	   (r0, insn_data[icode].operand[0].mode))
> > -      || !(insn_data[icode].operand[1].predicate
> > +	   (r0, insn_data[icode].operand[0].mode)))
> > +    return NULL_RTX;
> > +
> > +  if ((insn_data[icode].operand[1].predicate
> > 	   (r1, insn_data[icode].operand[1].mode))
> > -      || !(insn_data[icode].operand[2].predicate
> > +      && (insn_data[icode].operand[2].predicate
> > 	   (c, insn_data[icode].operand[2].mode)))
> > +    return GEN_FCN (icode) (r0, r1, c);
> > +  
> > +  mcode = (int) mov_optab->handlers[(int) GET_MODE (r0)].insn_code;
> > +  if (REGNO (r0) == REGNO (r1)
> > +      || !(insn_data[icode].operand[1].predicate
> > +	   (r0, insn_data[icode].operand[1].mode))
> > +      || !(insn_data[icode].operand[2].predicate
> > +	   (r1, insn_data[icode].operand[2].mode))
> > +      || !(insn_data[mcode].operand[0].predicate
> > +	   (r0, insn_data[mcode].operand[0].mode))
> > +      || !(insn_data[mcode].operand[1].predicate
> > +	   (c, insn_data[mcode].operand[1].mode)))
> >     return NULL_RTX;
> >
> > -  return GEN_FCN (icode) (r0, r1, c);
> > +  start_sequence ();
> > +  emit_insn (GEN_FCN (mcode) (r0, c));
> > +  emit_insn (GEN_FCN (icode) (r0, r0, r1));
> > +  s = get_insns ();
> > +  end_sequence ();
> > +  return s;
> > }
> >    
>
> I think we need a clearer description of what gen_add3_insn is for.
> In general this function is hard to use safely and easily, as it
> requires checking for a NULL return even though most of the callers do
> not actually check.  I note that the MIPS backend uses it, and the
> changes above don't seem quite right for the MIPS backend.
>  
Could you explain in more detail how this is not right for the MIPS backend?
I don't see and code in mips.md that relies on gen_add3_insn to return 
NULL_RTX
when a constant is not valid as an input to the add insn.