Re: [RFC] optimize x - y cmp 0 with undefined overflow

[Richard Biener <richard.guenther@gmail.com>]

On Fri, Jun 6, 2014 at 12:52 PM, Eric Botcazou <ebotcazou@adacore.com> wrote:
>> So when computing a range for z in
>>
>>  z = y - x;
>>
>> with x = [-INF, y - 1] and y = [x + 1, +INF] (deduced from !(x >= y)) we
>> fail to do sth sensible with [y, y] - [-INF, y - 1] or [x + 1, +INF] - [x,
>> x] but we do sth with [x + 1, +INF] - [-INF, x]?  That seems odd to me.
>>
>> With the patch we compute z to [1, +INF(OVF)]
>
> Right, and note the overflow.
>
>> Going the [x + 1, +INF] - [x,x] path first we obtain
>>
>>   [1, -x + INF]
>>
>> which fails the sanity checking
>>
>>   cmp = compare_values (min, max);
>>   if (cmp == -2 || cmp == 1)
>>     {
>>       /* If the new range has its limits swapped around (MIN > MAX),
>>          then the operation caused one of them to wrap around, mark
>>          the new range VARYING.  */
>>       set_value_range_to_varying (vr);
>>     }
>>   else
>>     set_value_range (vr, type, min, max, NULL);
>>
>> but clearly the same reasoning you can apply that makes trying
>> with [-INF, x] valid (it's just enlarging the range) can be applied
>> here, too, when computing +INF - x for the upper bound.  We can
>> safely increase that to +INF making the range valid for the above
>> test.
>
> I don't think we can enlarge to +INF because -x + INF can overflow, we can
> only enlarge to +INF(OVF).
>
>> But I wonder what code path in the routine still relies on that sanity
>> checking to produce a valid result (so I'd rather try removing it, or
>> taking uncomparable bounds as a valid range).
>>
>> Simplest would be to simply do
>>
>>           set_value_range (vr, type, min, max, NULL);
>>           return;
>>
>> and be done with that in the plus/minus handling.  With that the
>> testcase optimizes ok for me.
>
> With [1, -x + INF] as the resulting range?  But it can be bogus if x is itself
> equal to +INF (unlike the input range [x + 1, +INF] which is always correct)

Hmm, indeed.

> so this doesn't look valid to me.  I don't see how we can get away without a
> +INF(OVF) here, but I can compute it in extract_range_from_binary_expr_1 if
> you prefer and try only [op0,op0] and [op1,op1].

Yeah, I'd prefer that.

Thanks,
Richard.

> --
> Eric Botcazou