On Wed, Dec 4, 2013 at 8:07 PM, Jakub Jelinek <ja...@redhat.com> wrote:
>> > @@ -8617,6 +8740,49 @@
>> > [(set_attr "type" "negnot")
>> > (set_attr "mode" "SI")])
>> >
>> > +;; Negate with jump on overflow.
>> > +(define_expand "negv<mode>3"
>> > + [(parallel [(set (reg:CCO FLAGS_REG)
>> > + (ne:CCO (match_operand:SWI 1 "register_operand")
>> > + (const_int 0)))
>> > + (set (match_operand:SWI 0 "register_operand")
>> > + (neg:SWI (match_dup 1)))])
>> > + (set (pc) (if_then_else
>> > + (eq (reg:CCO FLAGS_REG) (const_int 0))
>> > + (label_ref (match_operand 2))
>> > + (pc)))]
>> > + ""
>> > +{
>> > + rtx minv = GEN_INT (HOST_WIDE_INT_M1U
>> > + << (GET_MODE_BITSIZE (<MODE>mode) - 1));
>> > + emit_insn (gen_negv<mode>3_1 (operands[0], operands[1], minv,
>> > operands[2]));
>> > + DONE;
>> > +})
>>
>> No, please use
>>
>> "operands[3] = GEN_INT (....);"
>>
>> and use (match_dup 3) in the pattern. The pattern below is not needed then.
>
> My memory is fuzzy about that, but I think that was my first version which
> didn't work, because with match_dup then it requires on the internal-fn.c
> side to pass 4 arguments instead of just 3. I can try again though.
I believe it should work, please see for example expNcorexf3 expander
and many of its (match_dup X) expressions.
>> BTW: can we use
>>
>> gen_int_mode (1 << (GET_MODE_BITSIZE (mode) - 1), mode)
>>
>> instead?
>
> With HOST_WIDE_INT_1U instead of 1 and s/mode/<MODE>mode/g perhaps.
gen_int_mode calls trunc_int_for_mode that is introduced by the comment:
/* Truncate and perhaps sign-extend C as appropriate for MODE. */
But, admittedly, I didn't test it...
Uros.
