RFC [1/3] divmod transform v2

[Jeff Law]

On 10/24/2016 01:28 AM, Richard Biener wrote:
[ big snip ]

>>>>>
>>>>>> +
>>>>>> +  /* Update all statements in stmts.
>>>>>> +     if stmt is lhs = op1 TRUNC_DIV_EXPR op2, change to lhs =
>>>>>> REALPART_EXPR<divmod_tmp>
>>>>>> +     if stmt is lhs = op1 TRUNC_MOD_EXPR op2, change to lhs =
>>>>>> IMAGPART_EXPR<divmod_tmp>.  */
>>>>> I'd just emit a copy from RES to the appropriate lhs operand just after
>>>>> the
>>>>> divmod and delete the now unnecessary TRUNC_DIV_EXPR and TRUNC_MOD_EXPR
>>>>> statements.
>>>>
>>>> That sounds like a good idea.
>>> Um sorry, not sure if I understood this part.
>>> For eg:
>>> t1 = x / y;
>>> t2 = x % y;
>>>
>>> do we want to transform it to:
>>> complex_tmp = DIVMOD (x, y);
>>> div_tmp = REALPART_EXPR<complex_tmp>
>>> mod_tmp = IMAGPART_EXPR<complex_tmp>
>>
>> complex_tmp = DIVMOD (x, y)
>> t1 = REALPART_EXPR (complex_tmp)
>> t2 = IMAGPART_EXPR (compex_tmp)
>>
>> Then remove the
>> t1 = x/y
>> t2 = x%y
>>
>>  statements
>
> OTOH implementation-wise that's more complicated.
If so, then I wouldn't bother.  I only mention it because I've found 
that model is sometimes easier on the implementation side.  I don't 
consider it a big deal.

jeff