expr_not_equal_to: use value_range API

[Aldy Hernandez]

On 11/8/18 9:43 AM, Richard Biener wrote:
> On Thu, Nov 8, 2018 at 3:27 PM Aldy Hernandez <aldyh@redhat.com> wrote:
>>
>>
>>
>> On 11/8/18 9:21 AM, Richard Biener wrote:
>>> On Thu, Nov 8, 2018 at 1:09 PM Aldy Hernandez <aldyh@redhat.com> wrote:
>>>>
>>>> All this nonsense:
>>>>
>>>> -      rtype = get_range_info (t, &min, &max);
>>>> -      if (rtype == VR_RANGE)
>>>> -       {
>>>> -         if (wi::lt_p (max, w, TYPE_SIGN (TREE_TYPE (t))))
>>>> -           return true;
>>>> -         if (wi::lt_p (w, min, TYPE_SIGN (TREE_TYPE (t))))
>>>> -           return true;
>>>> -       }
>>>> -      else if (rtype == VR_ANTI_RANGE
>>>> -              && wi::le_p (min, w, TYPE_SIGN (TREE_TYPE (t)))
>>>> -              && wi::le_p (w, max, TYPE_SIGN (TREE_TYPE (t))))
>>>>
>>>> Replaced by an API like Kutulu intended.
>>>>
>>>> +      get_range_info (t, vr);
>>>> +      if (!vr.may_contain_p (wide_int_to_tree (TREE_TYPE (t), w)))
>>>>
>>>> Ain't it grand?
>>>
>>> Well.  The not-so-grand thing is that you possibly ggc-allocate
>>> three INTEGER_CST nodes here.
>>
>> Hmmm... I'd really prefer to use a simple API call, instead of having to
>> twiddle with the extremes manually.  Ideally no one should be looking
>> inside of a value_range.
>>
>> Do recommend another way of implementing may_contain_p ?
> 
> I think many places dealing with get_range_info () should instead
> work on the (to be created...) 1:1 copy of value_range ontop of
> wide-int-range instead.

I'd prefer to not expose that we're going to use wide_int or any other 
implementation to the users of get_range_info().

> 
> So - can you add a wide_int_range class to wide-int-range.h
> that implements the same (but with wide-ints rather than trees
> obviously) API as value-range?

Hmmm, I don't have time for this release cycle.  Perhaps something to be 
entertained for GCC+1?

Again, I prefer my patch as is.  I cleans up the code, and keeps us from 
introducing problematic bugs.  Anything dealing with anti ranges is 
fraught with peril, as my cleanups to tree-vrp revealed.

If using these INTEGER_CST's causes any measurable performance 
difference, I'd be happy to look into it.

Aldy