On 04/19/2011 02:52 AM, Denis Chertykov wrote:
> 2011/4/19 Georg-Johann Lay <a...@gjlay.de>:
>> How can add, sub etc. be split? This would need an explicit
>> representation of carry.
>
> Yes.
>
> Look at http://gcc.gnu.org/ml/gcc/2005-03/msg00871.html
Well, sort-of, but not really.
It gets a tad ugly, but have a look at the adddi3* patterns in
the mn10300 and rx ports.
In particular note how both the inputs and outputs to the insn
(not the expander) are all SImode, allowing for lower_subreg to
do its job. The patterns are split post-reload -- some of that
is for scheduling, some of that simply makes computing the
individual insn lengths significantly easier -- but you wouldn't
really have to do that for AVR.
For AVR things would become even trickier. You might consider
(define_predicate "concat_operator"
(match_code "concat"))
(define_insn "addsi3_qqqq"
[(set (match_operand:QI 0 "register_operand" "=r")
(truncate:QI
(plus:SI
(match_operator:SI 12 "concat_operator"
[(match_operand:QI 4 "register_operand" "0")
(match_operand:QI 5 "register_operand" "1")
(match_operand:QI 6 "register_operand" "2")
(match_operand:QI 7 "register_operand" "3")])
(match_operator:SI 13 "concat_operator"
[(match_operand:QI 8 "reg_or_0_operand" "rL")
(match_operand:QI 9 "reg_or_0_operand" "rL")
(match_operand:QI 10 "reg_or_0_operand" "rL")
(match_operand:QI 11 "reg_or_0_operand" "rL")])))
(set (match_operand:QI 1 "register_operand" "=r")
(truncate:QI
(lshiftrt:SI
(plus:SI (match_dup 24) (match_dup 25))
(const_int 8))))
(set (match_operand:QI 2 "register_operand" "=r")
(truncate:QI
(lshiftrt:SI
(plus:SI (match_dup 24) (match_dup 25))
(const_int 16))))
(set (match_operand:QI 3 "register_operand" "=r")
(truncate:QI
(lshiftrt:SI
(plus:SI (match_dup 24) (match_dup 25))
(const_int 24))))]
""
"add %0,%Z8\;adc %1,%Z9\;adc %2,%Z10\;adc %3,%Z11"
[(set_attr "length" "4")]
)
This may require a little bit of code inside combine to handle
CONCAT in a reasonable way, but that should be fairly minimal.
It may also want some more special case patterns and/or peep2s
to more efficiently handle constants, particularly considering
adiw and subic. But I think it's at least worth investigating.
r~
