Re: Questions about comparison insns in mips.md

[Eric Fisher <joefoxreal at gmail dot com>]

08 Oct 2005 10:40:40 -0700, Ian Lance Taylor <ian@airs.com>:
> Eric Fisher <joefoxreal@gmail.com> writes:
>
> >     Who can explains for me about the comparison insns in mips.md.
>
> Look at the documentation for the bCOND instruction patterns.
>
> >     One question, why there is only define_expand patterns but no
> > define_insn.
>
> Because the gcc model for comparison instructions is that they set
> condition codes, but MIPS processors do not have condition codes.
>
> > Supposed that I have a comparison instruction say 'cmp
> > %1, %2', should I implement define_insn patterns for comparison insns?
>
> It depends on where the result goes.
>
> >     The second question. As for SETTING A REGISTER FROM A COMPARISON,
> > must I implement them?
>
> You do not have to implement them, although you should if your target
> has that sort of instruction.  If you don't implement them, gcc will
> use conditional branches to get the result of a comparison.
>
> > Here is a pattern from mips.md,
> > (define_expand "seq"
> >   [(set (match_operand:SI 0 "register_operand" "=d")
> >       (lt:SI (match_dup 1)
> >              (match_dup 2 )))]
> >   ""
> > {
> >    /* Set up operands from compare.  */
> >   operands[1] = branch_cmp[0];
> >   operands[2] = branch_cmp[1];
> > })
> > Why here use match_dup not match_operand? And why to set up operands
> > from compare?
>
> Because that is how gcc invokes the patterns.  See the documentation
> for the sCOND instructions.
>
> Ian
>


Thanks for the help. Since ARM has condition codes, I'm wondering why it uses
a psuedo register internally to represent things that use or set the
condition codes?
I think it should use CPSR directly.

Eric.