Re: [PATCH] Simplify floating point comparisons

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

On Mon, Apr 30, 2018 at 8:49 PM, Marc Glisse <marc.glisse@inria.fr> wrote:
> On Fri, 12 Jan 2018, Wilco Dijkstra wrote:
>
>> Hi,
>>
>> Here is the updated version:
>>
>> This patch implements some of the optimizations discussed in
>> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71026.
>>
>> Simplify (C / x >= 0.0) into x >= 0.0 with -fno-signed-zeros
>> and -ffinite-math-only.  If C is negative the comparison is reversed.
>> Only handle >= and <= for now since C / x can underflow if C is small.
>>
>>
>> Simplify (x * C1) > C2 into x > (C2 / C1) with
>> -funsafe-math-optimizations.
>> If C1 is negative the comparison is reversed.
>>
>> OK for commit?
>>
>> ChangeLog
>> 2018-01-10  Wilco Dijkstra  <wdijkstr@arm.com>
>>             Jackson Woodruff  <jackson.woodruff@arm.com>
>>
>>    gcc/
>>         PR 71026/tree-optimization
>>         * match.pd: Simplify floating point comparisons.
>>
>>    gcc/testsuite/
>>         PR 71026/tree-optimization
>>         * gcc.dg/div-cmp-1.c: New test.
>>         * gcc.dg/div-cmp-2.c: New test.
>> --
>>
>> diff --git a/gcc/match.pd b/gcc/match.pd
>> index
>> 435125a317275527661fba011a9d26e507d293a6..8a6fee906de6a750201362119862f8326868f26b
>> 100644
>> --- a/gcc/match.pd
>> +++ b/gcc/match.pd
>> @@ -376,6 +376,21 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>>  (rdiv @0 (negate @1))
>>  (rdiv (negate @0) @1))
>>
>> +/* Simplify (C / x op 0.0) to x op 0.0 for C != 0, C != Inf/Nan.
>> +   Only handle >= and <= since C / x may underflow to zero.  */
>> +(for op (le ge)
>> +     res_op (lt ge)
>> +     neg_op (ge lt)
>> + (simplify
>> +  (op (rdiv REAL_CST@0 @1) real_zerop@2)
>> +  (if (!HONOR_SIGNED_ZEROS (@1) && !HONOR_INFINITIES (@1))
>> +   (switch
>> +    (if (real_less (&dconst0, TREE_REAL_CST_PTR (@0)))
>> +     (res_op @1 @2))
>> +    /* For C < 0, use the inverted operator.  */
>> +    (if (real_less (TREE_REAL_CST_PTR (@0), &dconst0))
>> +     (neg_op @1 @2))))))
>
>
> Let's try with C = DBL_MIN and x = 婊BL_MAX. I don't believe it involves
> signed zeros or infinities, just an underflow. First, the result depends on
> the rounding mode. And in the default round-to-nearest, both divisions give
> 0, and thus compare the same with 0, but we replace that with a sign test on
> x, where they clearly give opposite answers.
>
> What would be the proper flag to test to check if we care about underflow?

We have none specific so this makes it flag_unsafe_math_optimizations.

Richard.

> --
> Marc Glisse