On Wed, Mar 21, 2012 at 10:56 AM, Kai Tietz <ktiet...@googlemail.com> wrote:
> 2012/3/21 Richard Guenther <richard.guent...@gmail.com>:
>> On Thu, Mar 15, 2012 at 3:45 PM, Kai Tietz <ktiet...@googlemail.com> wrote:
>>> 2012/3/15 Richard Guenther <richard.guent...@gmail.com>:
>>>> On Thu, Mar 15, 2012 at 2:46 PM, Kai Tietz <ktiet...@googlemail.com> wrote:
>>>>> 2012/3/15 Richard Guenther <richard.guent...@gmail.com>:
>>>>>> On Thu, Mar 15, 2012 at 2:09 PM, Kai Tietz <ktiet...@googlemail.com>
>>>>>> wrote:
>>>>>>> Hi,
>>>>>>>
>>>>>>> this is the second part of the patch for this problem. It adds some
>>>>>>> basic simplifications for ==/!=
>>>>>>> comparisons for eliminating redudant operands.
>>>>>>>
>>>>>>> It adds the following patterns:
>>>>>>> -X ==/!= Z - X -> Z ==/!= 0.
>>>>>>> ~X ==/!= Z ^ X -> Z ==/!= ~0
>>>>>>> X ==/!= X - Y -> Y == 0
>>>>>>> X ==/!= X + Y -> Y == 0
>>>>>>> X ==/!= X ^ Y -> Y == 0
>>>>>>> (X - Y) ==/!= (Z - Y) -> X ==/!= Z
>>>>>>> (Y - X) ==/!= (Y - Z) -> X ==/!= Z
>>>>>>> (X + Y) ==/!= (X + Z) -> Y ==/!= Z
>>>>>>> (X + Y) ==/!= (Z + X) -> Y ==/!= Z
>>>>>>> (X ^ Y) ==/!= (Z ^ X) -> Y ==/!= Z
>>>>>>
>>>>>> Can you re-base this patch to work without the previous one? Also
>>>>>> please coordinate with Andrew. Note that all of these(?) simplifications
>>>>>> are already done by fold_comparison which we could share if you'd split
>>>>>> out the EXPR_P op0/op1 cases with separated operands/code.
>>>>>>
>>>>>> Richard.
>>>>>
>>>>> Hmm, fold_comparison doesn't do the same thing as it checks for
>>>>> possible overflow. This is true for comparisons not being ==/!= or
>>>>> having operands of none-integral-type. But for ==/!= with integral
>>>>> typed arguments the overflow doesn't matter at all. And exactly this
>>>>> is what patch implements here.
>>>>
>>>> fold_comparison does not check for overflow for ==/!=.
>>>>
>>>>> This optimization of course is just desired in non-AST form, as we
>>>>> otherwise loose information in FE. Therefore I didn't added it to
>>>>> fold_const.
>>>>
>>>> Which pieces are not already in fold-const btw? forwprop already
>>>> re-constructs trees for the defs of the lhs/rhs of a comparison.
>>>>
>>>> Richard.
>>>
>>> I have tried to use here instead a call to fold_build2 instead, and I
>>> had to notice that it didn't optimized a single case (beside the - and
>>> ~ case on both sides).
>>>
>>> I see in fold const for example in the pattern 'X +- C1 CMP Y +- C2'
>>> to 'X CMP Y +- C2 +- C1' explicit the check for it.
>>>
>>> ...
>>> /* Transform comparisons of the form X +- C1 CMP Y +- C2 to
>>> X CMP Y +- C2 +- C1 for signed X, Y. This is valid if
>>> the resulting offset is smaller in absolute value than the
>>> original one. */
>>> if (TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (arg0))
>>> && (TREE_CODE (arg0) == PLUS_EXPR || TREE_CODE (arg0) == MINUS_EXPR)
>>> ...
>>
>> Because the transform is not valid if Y +- C2 +- C1 overflows. It is not
>> valid
>> because overflow is undefined, not because the comparison would do the
>> wrong thing. You'd have to change the addition to unsigned.
>>
>>> The same for pattern X +- C1 CMP C2 to X CMP C2 +- C1.
>>
>> Well, this is obviously just a missed optimization in fold-const.c then.
>> Mind
>> conditionalizing the overflow check to codes not NE_EXPR or EQ_EXPR?
>>
>>> The cases for '(X + Y) ==/!= (Z + X)' and co have the same issue or
>>> are simply not present.
>>
>> That's true. I suppose they were considered too special to worry about.
>> Did you see these cases in real code?
>>
>>> Sorry fold_const doesn't cover this at all.
>>
>> It covers part of it.
>>
>>> Kai
>
> Sure, the test code shown in this patch isn't that unusual.
> Especially in gimple (by using different statements) such construct
> are happening.
>
> Eg.:
>
> int f1 (int a, int b, int c)
> {
> if ((a + b) == (c + a))
> return 1;
> return 0;
> }
>
> int f2 (int a, int b, int c)
> {
> if ((a ^ b) == (a ^ c))
> return 1;
> return 0;
> }
>
>
> int f2 (int a, int b)
> {
> if (-a == (b - a))
> return 1;
> return 0;
> }
>
> In all those cases the use of variable should be optimized out.
> Instead we are producing pretty weak code for those cases.
True, I agree we should try to handle these. Did you talk to Andrew
with respect to the gimple-combining thing he is working on?
Richard.
> Kai
