On Tue, Jul 7, 2015 at 11:24 AM, Marc Glisse <marc.gli...@inria.fr> wrote:
> On Tue, 7 Jul 2015, Richard Biener wrote:
>
>> On Tue, Jul 7, 2015 at 8:06 AM, Marc Glisse <marc.gli...@inria.fr> wrote:
>>>
>>> On Tue, 7 Jul 2015, Hurugalawadi, Naveen wrote:
>>>
>>>> Please find attached the patch "PR25529.patch" that converts the
>>>> pattern:-
>>>> (unsigned * 2)/2 is into unsigned &0x7FFFFFFF
>>>
>>>
>>>
>>> +/* Simplify (unsigned t * 2)/2 -> unsigned t & 0x7FFFFFFF. */
>>> +(for div (trunc_div ceil_div floor_div round_div exact_div)
>>> + (simplify
>>> + (div (mult @0 INTEGER_CST@1) INTEGER_CST@1)
>>>
>>> You don't need to repeat INTEGER_CST, the second time @1 is enough.
>>>
>>> + (with { tree n2 = build_int_cst (TREE_TYPE (@0),
>>> + wi::exact_log2 (@1)); }
>>> + (if (TYPE_UNSIGNED (TREE_TYPE (@0)))
>>> + (bit_and @0 (rshift (lshift { build_minus_one_cst (TREE_TYPE (@0)); }
>>> + { n2; }) { n2; }))))))
>>>
>>> What happens if you write t*3/3?
>>
>>
>> Huh, and you posted this patch twice? See my reply to the other copy
>> for the correctness issues and better handling of exact_div
>
>
> They are not the same, one is for left shifts and the other one for right
> shifts. And that makes a big difference: in t*c/c, the division is always
> exact, so all divisions are equivalent. This is not the case for t/c*c.
Ah, sorry. Still the same comment for computing the constant and
placing of the 'with' applies. For signed types with TYPE_OVERFLOW_UNDEFINED
you can simply cancel the operation (even for non-power-of-two multipliers).
In fold-const.c extract_muldiv contains magic to handle this kind of cases.
Otherwise for signed division (only the sign of the division matters, so you
can probably ignore sign-changing conversions of the multiplication result)
you can simplify it to a sign-extension from bit precision - log2 with
the proposed introduction of a SEXT_EXPR (see other thread about
type promotion).
Richard.
> --
> Marc Glisse
