Canonicalize X u< X to UNORDERED_EXPR

Hello,

this case seemed to be missing in the various X cmp X transformations. It 
does not change the generated code in the testcase.

The missing :c is rather trivial. I can commit it separately if you 
prefer.

Bootstrap+regtest on powerpc64le-unknown-linux-gnu.

2016-05-02  Marc Glisse  <marc.glisse@inria.fr>

gcc/
 	* match.pd ((A & B) OP (C & B)): Mark '&' as commutative.
 	(X u< X, X u> X): New transformations

gcc/testsuite/
 	* gcc.dg/tree-ssa/unord.c: New testcase.

On Sat, Apr 30, 2016 at 8:44 PM, Marc Glisse <marc.glisse@inria.fr> wrote:
> Hello,
>
> this case seemed to be missing in the various X cmp X transformations. It
> does not change the generated code in the testcase.
>
> The missing :c is rather trivial. I can commit it separately if you prefer.

I think it's not missing.  Commutating the first one is enough to eventually
make the @1s match up.  I think you should get a diagnostic on a duplicate
pattern when adding another :c (hmm, no, it's indeed "different" patterns but
still redundant).

> Bootstrap+regtest on powerpc64le-unknown-linux-gnu.

Ok for the new pattern.

Thanks,
Richard.

> 2016-05-02  Marc Glisse  <marc.glisse@inria.fr>
>
> gcc/
>         * match.pd ((A & B) OP (C & B)): Mark '&' as commutative.
>         (X u< X, X u> X): New transformations
>
> gcc/testsuite/
>         * gcc.dg/tree-ssa/unord.c: New testcase.
>
> --
> Marc Glisse
> Index: trunk/gcc/match.pd
> ===================================================================
> --- trunk/gcc/match.pd  (revision 235654)
> +++ trunk/gcc/match.pd  (working copy)
> @@ -783,21 +783,21 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>    @0)
>   /* (~x | y) & x -> x & y */
>   /* (~x & y) | x -> x | y */
>   (simplify
>    (bitop:c (rbitop:c (bit_not @0) @1) @0)
>    (bitop @0 @1)))
>
>  /* Simplify (A & B) OP0 (C & B) to (A OP0 C) & B. */
>  (for bitop (bit_and bit_ior bit_xor)
>   (simplify
> -  (bitop (bit_and:c @0 @1) (bit_and @2 @1))
> +  (bitop (bit_and:c @0 @1) (bit_and:c @2 @1))
>    (bit_and (bitop @0 @2) @1)))
>
>  /* (x | CST1) & CST2 -> (x & CST2) | (CST1 & CST2) */
>  (simplify
>    (bit_and (bit_ior @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
>    (bit_ior (bit_and @0 @2) (bit_and @1 @2)))
>
>  /* Combine successive equal operations with constants.  */
>  (for bitop (bit_and bit_ior bit_xor)
>   (simplify
> @@ -1914,20 +1914,24 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>   (simplify
>    (cmp @0 @0)
>    (if (cmp != NE_EXPR
>         || ! FLOAT_TYPE_P (TREE_TYPE (@0))
>         || ! HONOR_NANS (@0))
>     { constant_boolean_node (false, type); })))
>  (for cmp (unle unge uneq)
>   (simplify
>    (cmp @0 @0)
>    { constant_boolean_node (true, type); }))
> +(for cmp (unlt ungt)
> + (simplify
> +  (cmp @0 @0)
> +  (unordered @0 @0)))
>  (simplify
>   (ltgt @0 @0)
>   (if (!flag_trapping_math)
>    { constant_boolean_node (false, type); }))
>
>  /* Fold ~X op ~Y as Y op X.  */
>  (for cmp (simple_comparison)
>   (simplify
>    (cmp (bit_not@2 @0) (bit_not@3 @1))
>    (if (single_use (@2) && single_use (@3))
> Index: trunk/gcc/testsuite/gcc.dg/tree-ssa/unord.c
> ===================================================================
> --- trunk/gcc/testsuite/gcc.dg/tree-ssa/unord.c (revision 0)
> +++ trunk/gcc/testsuite/gcc.dg/tree-ssa/unord.c (working copy)
> @@ -0,0 +1,7 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O -fdump-tree-optimized" } */
> +
> +int f(double a){double b=a;return !__builtin_islessequal(a,b);}
> +int g(double a){double b=a;return !__builtin_isgreaterequal(a,b);}
> +
> +/* { dg-final { scan-tree-dump-times " unord " 2 "optimized" } } */
>

On Mon, 2 May 2016, Richard Biener wrote:

> On Sat, Apr 30, 2016 at 8:44 PM, Marc Glisse <marc.glisse@inria.fr> wrote:
>> Hello,
>>
>> this case seemed to be missing in the various X cmp X transformations. It
>> does not change the generated code in the testcase.
>>
>> The missing :c is rather trivial. I can commit it separately if you prefer.
>
> I think it's not missing.  Commutating the first one is enough to eventually
> make the @1s match up.  I think you should get a diagnostic on a duplicate
> pattern when adding another :c (hmm, no, it's indeed "different" patterns but
> still redundant).

Let's see. The current pattern is
   (bitop (bit_and:c @0 @1) (bit_and @2 @1))

This matches:
(X & Y) ^ (Z & Y)
(Y & X) ^ (Z & Y)

If I have for instance (Y & X) ^ (Y & Z), I don't see how that is going to 
match, and indeed we don't simplify that. On the other hand, if I have 
bit_ior instead of bit_xor, then we have another very similar 
transformation a 100 lines up in match.pd

(for op (bit_and bit_ior)
      rop (bit_ior bit_and)
  (simplify
   (op (convert? (rop:c @0 @1)) (convert? (rop @0 @2)))

That one also commutes only one, but starting from a different match. We 
should probably reorganize them a bit.

>> Bootstrap+regtest on powerpc64le-unknown-linux-gnu.
>
> Ok for the new pattern.
>
> Thanks,
> Richard.
>
>> 2016-05-02  Marc Glisse  <marc.glisse@inria.fr>
>>
>> gcc/
>>         * match.pd ((A & B) OP (C & B)): Mark '&' as commutative.
>>         (X u< X, X u> X): New transformations
>>
>> gcc/testsuite/
>>         * gcc.dg/tree-ssa/unord.c: New testcase.
>>
>> --
>> Marc Glisse
>> Index: trunk/gcc/match.pd
>> ===================================================================
>> --- trunk/gcc/match.pd  (revision 235654)
>> +++ trunk/gcc/match.pd  (working copy)
>> @@ -783,21 +783,21 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>>    @0)
>>   /* (~x | y) & x -> x & y */
>>   /* (~x & y) | x -> x | y */
>>   (simplify
>>    (bitop:c (rbitop:c (bit_not @0) @1) @0)
>>    (bitop @0 @1)))
>>
>>  /* Simplify (A & B) OP0 (C & B) to (A OP0 C) & B. */
>>  (for bitop (bit_and bit_ior bit_xor)
>>   (simplify
>> -  (bitop (bit_and:c @0 @1) (bit_and @2 @1))
>> +  (bitop (bit_and:c @0 @1) (bit_and:c @2 @1))
>>    (bit_and (bitop @0 @2) @1)))
>>
>>  /* (x | CST1) & CST2 -> (x & CST2) | (CST1 & CST2) */
>>  (simplify
>>    (bit_and (bit_ior @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
>>    (bit_ior (bit_and @0 @2) (bit_and @1 @2)))
>>
>>  /* Combine successive equal operations with constants.  */
>>  (for bitop (bit_and bit_ior bit_xor)
>>   (simplify
>> @@ -1914,20 +1914,24 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>>   (simplify
>>    (cmp @0 @0)
>>    (if (cmp != NE_EXPR
>>         || ! FLOAT_TYPE_P (TREE_TYPE (@0))
>>         || ! HONOR_NANS (@0))
>>     { constant_boolean_node (false, type); })))
>>  (for cmp (unle unge uneq)
>>   (simplify
>>    (cmp @0 @0)
>>    { constant_boolean_node (true, type); }))
>> +(for cmp (unlt ungt)
>> + (simplify
>> +  (cmp @0 @0)
>> +  (unordered @0 @0)))
>>  (simplify
>>   (ltgt @0 @0)
>>   (if (!flag_trapping_math)
>>    { constant_boolean_node (false, type); }))
>>
>>  /* Fold ~X op ~Y as Y op X.  */
>>  (for cmp (simple_comparison)
>>   (simplify
>>    (cmp (bit_not@2 @0) (bit_not@3 @1))
>>    (if (single_use (@2) && single_use (@3))
>> Index: trunk/gcc/testsuite/gcc.dg/tree-ssa/unord.c
>> ===================================================================
>> --- trunk/gcc/testsuite/gcc.dg/tree-ssa/unord.c (revision 0)
>> +++ trunk/gcc/testsuite/gcc.dg/tree-ssa/unord.c (working copy)
>> @@ -0,0 +1,7 @@
>> +/* { dg-do compile } */
>> +/* { dg-options "-O -fdump-tree-optimized" } */
>> +
>> +int f(double a){double b=a;return !__builtin_islessequal(a,b);}
>> +int g(double a){double b=a;return !__builtin_isgreaterequal(a,b);}
>> +
>> +/* { dg-final { scan-tree-dump-times " unord " 2 "optimized" } } */

On Mon, May 2, 2016 at 11:18 AM, Marc Glisse <marc.glisse@inria.fr> wrote:
> On Mon, 2 May 2016, Richard Biener wrote:
>
>> On Sat, Apr 30, 2016 at 8:44 PM, Marc Glisse <marc.glisse@inria.fr> wrote:
>>>
>>> Hello,
>>>
>>> this case seemed to be missing in the various X cmp X transformations. It
>>> does not change the generated code in the testcase.
>>>
>>> The missing :c is rather trivial. I can commit it separately if you
>>> prefer.
>>
>>
>> I think it's not missing.  Commutating the first one is enough to
>> eventually
>> make the @1s match up.  I think you should get a diagnostic on a duplicate
>> pattern when adding another :c (hmm, no, it's indeed "different" patterns
>> but
>> still redundant).
>
>
> Let's see. The current pattern is
>   (bitop (bit_and:c @0 @1) (bit_and @2 @1))
>
> This matches:
> (X & Y) ^ (Z & Y)
> (Y & X) ^ (Z & Y)
>
> If I have for instance (Y & X) ^ (Y & Z), I don't see how that is going to
> match, and indeed we don't simplify that.

Hmm, indeed.  I might have wrongly ommitted some :c then in other places
as well.  So the original patch is ok.

> On the other hand, if I have
> bit_ior instead of bit_xor, then we have another very similar transformation
> a 100 lines up in match.pd
>
> (for op (bit_and bit_ior)
>      rop (bit_ior bit_and)
>  (simplify
>   (op (convert? (rop:c @0 @1)) (convert? (rop @0 @2)))
>
> That one also commutes only one, but starting from a different match. We
> should probably reorganize them a bit.

Yeah.

Richard.

>
>>> Bootstrap+regtest on powerpc64le-unknown-linux-gnu.
>>
>>
>> Ok for the new pattern.
>>
>> Thanks,
>> Richard.
>>
>>> 2016-05-02  Marc Glisse  <marc.glisse@inria.fr>
>>>
>>> gcc/
>>>         * match.pd ((A & B) OP (C & B)): Mark '&' as commutative.
>>>         (X u< X, X u> X): New transformations
>>>
>>> gcc/testsuite/
>>>         * gcc.dg/tree-ssa/unord.c: New testcase.
>>>
>>> --
>>> Marc Glisse
>>> Index: trunk/gcc/match.pd
>>> ===================================================================
>>> --- trunk/gcc/match.pd  (revision 235654)
>>> +++ trunk/gcc/match.pd  (working copy)
>>> @@ -783,21 +783,21 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>>>    @0)
>>>   /* (~x | y) & x -> x & y */
>>>   /* (~x & y) | x -> x | y */
>>>   (simplify
>>>    (bitop:c (rbitop:c (bit_not @0) @1) @0)
>>>    (bitop @0 @1)))
>>>
>>>  /* Simplify (A & B) OP0 (C & B) to (A OP0 C) & B. */
>>>  (for bitop (bit_and bit_ior bit_xor)
>>>   (simplify
>>> -  (bitop (bit_and:c @0 @1) (bit_and @2 @1))
>>> +  (bitop (bit_and:c @0 @1) (bit_and:c @2 @1))
>>>    (bit_and (bitop @0 @2) @1)))
>>>
>>>  /* (x | CST1) & CST2 -> (x & CST2) | (CST1 & CST2) */
>>>  (simplify
>>>    (bit_and (bit_ior @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
>>>    (bit_ior (bit_and @0 @2) (bit_and @1 @2)))
>>>
>>>  /* Combine successive equal operations with constants.  */
>>>  (for bitop (bit_and bit_ior bit_xor)
>>>   (simplify
>>> @@ -1914,20 +1914,24 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>>>   (simplify
>>>    (cmp @0 @0)
>>>    (if (cmp != NE_EXPR
>>>         || ! FLOAT_TYPE_P (TREE_TYPE (@0))
>>>         || ! HONOR_NANS (@0))
>>>     { constant_boolean_node (false, type); })))
>>>  (for cmp (unle unge uneq)
>>>   (simplify
>>>    (cmp @0 @0)
>>>    { constant_boolean_node (true, type); }))
>>> +(for cmp (unlt ungt)
>>> + (simplify
>>> +  (cmp @0 @0)
>>> +  (unordered @0 @0)))
>>>  (simplify
>>>   (ltgt @0 @0)
>>>   (if (!flag_trapping_math)
>>>    { constant_boolean_node (false, type); }))
>>>
>>>  /* Fold ~X op ~Y as Y op X.  */
>>>  (for cmp (simple_comparison)
>>>   (simplify
>>>    (cmp (bit_not@2 @0) (bit_not@3 @1))
>>>    (if (single_use (@2) && single_use (@3))
>>> Index: trunk/gcc/testsuite/gcc.dg/tree-ssa/unord.c
>>> ===================================================================
>>> --- trunk/gcc/testsuite/gcc.dg/tree-ssa/unord.c (revision 0)
>>> +++ trunk/gcc/testsuite/gcc.dg/tree-ssa/unord.c (working copy)
>>> @@ -0,0 +1,7 @@
>>> +/* { dg-do compile } */
>>> +/* { dg-options "-O -fdump-tree-optimized" } */
>>> +
>>> +int f(double a){double b=a;return !__builtin_islessequal(a,b);}
>>> +int g(double a){double b=a;return !__builtin_isgreaterequal(a,b);}
>>> +
>>> +/* { dg-final { scan-tree-dump-times " unord " 2 "optimized" } } */
>
>
> --
> Marc Glisse