PATCH: PR target/44588: Very inefficient 8bit mod/div

[Paolo Bonzini]

On 06/23/2010 04:26 PM, H.J. Lu wrote:
> On Tue, Jun 22, 2010 at 11:11 PM, Uros Bizjak<ubizjak@gmail.com>  wrote:
>> On Tue, Jun 22, 2010 at 8:57 PM, H.J. Lu<hjl.tools@gmail.com>  wrote:
>>
>>>>> If we use AX for quotient in 8bit divmod pattern, we have to make
>>>>> sure that AX is valid for quotient.  We have to extend AL with UNSPEC
>>>>> since AH isn't the part of quotient,.  Instead, I use AL for quotient and
>>>>> use UNSPEC_MOVQI_EXTZH to extract remainder from AL. Quotient
>>>>> access can be optimized very nicely. If remainder is used, we may have
>>>>> an extract move for UNSPEC_MOVQI_EXTZH. I think this is a reasonable
>>>>> comprise.
>>>>
>>>> Why we need to reinvent movqi_extzv_2 ?
>>>
>>> Because UNSPEC_MOVQI_EXTZH extracts AH from AL, not from
>>> AX, EAX, RAX.
>>
>> This is wrong, you can't extract AH out from AL...
>
> UNSPEC_MOVQI_EXTZH  is only used on AL from
> 8bit div where AH has remainder. Should I give it a different
> name, like UNSPEC_MOVQI_EXTZH_FROM_8BITDIV?
>
>>>
>>>> I guess that<u>divqi3 has to be implemented as multiple-set divmod
>>>> pattern using strict_low_part subregs to exactly describe in which
>>>> subreg quotient and remainder go.
>>>>
>>>
>>> I tried it and couldn't get it to work without adding a new
>>> constraint for the upper 8bit registers.
>>
>> "Q" with %h modifier ?
>>
>
> Although there are AX, AL, AH, ......, the rest of compiler only
> knows one hard register, AX_REG, which is register 0. We can't
> really access upper 8bit register,  like AH,  as a real register. We
> can only access them via sign_extract and zero_extract.  As far
> as the rest of gcc is concerned, there are no upper 8bit registers.
> I don't think we can describe where 8bit div remainder is to the
> rest of compiler.

Well, sure you can:

(set (reg:HI r)
    (ior:HI (lshift:HI
             (zero_extend:HI (umod:QI ...) (const_int 8)))
            (zero_extend:HI (udiv:QI ...) (const_int 8))))

(set (reg:QI s) ;; remainder
    (zero_extract (reg:HI r) (const_int 8) (const_int 8)))

(set (reg:QI t) ;; quotient
    (subreg (reg:HI r) 0))

The question is only whether fwprop, combine and regalloc are smart 
enough to reason about it.

Paolo