Re: [Patch, AVR]: Fix PR46779

[Georg-Johann Lay <avr@gjlay.de>]

Denis Chertykov wrote:
> 2011/6/26 Georg-Johann Lay <avr@gjlay.de>:
>> Denis Chertykov schrieb:
>>> 2011/6/24 Richard Henderson <rth@redhat.com>:
>>>
>>>> On 06/23/2011 01:15 PM, Denis Chertykov wrote:
>>>>
>>>>>>  text    data     bss     dec     hex filename
>>>>>> 10032      25       0   10057    2749 bld-avr-orig/gcc/z.o
>>>>>>  5816      25       0    5841    16d1 bld-avr-new/gcc/z.o
>>>>> Richard, can you send me this z.c file ?
>>>>> Right now I'm notice that new code is worse.
>>>> That's gcc.c-torture/compile/950612-1.c.
>>> I have founded that postreload optimizations can't handle results of
>>> new L_R_A code.
>>> I think that it's can be handled by CSE (postreload).
>> Did you try to add constraint alternative to *addhi3?
>> Like "*!d,d,n" or even "*!r,r,n"
>>
>> I saw some code improvement with that alternative.
> 
> I'm trying:
> 
> (define_insn "*addhi3"
>   [(set (match_operand:HI 0 "register_operand" "=r,!w,!w,d,r,r,!d")
>  	(plus:HI
>  	 (match_operand:HI 1 "register_operand" "%0,0,0,0,0,0,!r")
>  	 (match_operand:HI 2 "nonmemory_operand" "r,I,J,i,P,N,!ri")))]
>   ""
>   "@
>  	add %A0,%A2\;adc %B0,%B2
>  	adiw %A0,%2
>  	sbiw %A0,%n2
>  	subi %A0,lo8(-(%2))\;sbci %B0,hi8(-(%2))
>  	sec\;adc %A0,__zero_reg__\;adc %B0,__zero_reg__
>  	sec\;sbc %A0,__zero_reg__\;sbc %B0,__zero_reg__
>         #"
>   [(set_attr "length" "2,1,1,2,3,3,4")
>    (set_attr "cc" "set_n,set_czn,set_czn,set_czn,set_n,set_n,set_n")])
> 

That split will split always:

> ;; Special split three addressing addhi3
> ;; to make postreload optimization possible
> (define_split ; addhi3 !d,!r,!ri
>   [(set (match_operand:HI 0 "d_register_operand" "")
> 	(plus:HI (match_operand:HI 1 "register_operand" "")
> 		 (match_operand:HI 2 "nonmemory_operand" "")))]
>   "reload_completed"
     && REGNO(operands[0]) != REGNO(operands[1])"
>   [(set (match_dup 0) (match_dup 2))
>    (set (match_dup 0) (plus:HI (match_dup 0) (match_dup 1)))]
>   "")
Maybe it can also restrict to const_int_operand in #2 and then it's
best to
    (set (match_dup 0)
         (match_dup 1))
    (set (match_dup 0)
         (plus:HI (match_dup 0)
                  (match_dup 2)))
> 
> The main problem for me is that the new addressing mode produce a
> worse code in many tests.

You have an example source?

> Although, results for gcc.c-torture/compile/950612-1.c is
> significantly better with new addressing.

That testcase is pathologic for AVR...

> 
> Denis.
> 

Johann