On Mon, Jul 17, 2017 at 3:02 AM, Richard Biener
<richard.guenther@gmail.com> wrote:
> On Thu, Jul 13, 2017 at 6:18 AM, Andrew Pinski <pinskia@gmail.com> wrote:
>> On Wed, Jul 12, 2017 at 9:10 PM, Marc Glisse <marc.glisse@inria.fr> wrote:
>>> On Wed, 12 Jul 2017, Andrew Pinski wrote:
>>>
>>>> Hi,
>>>>  Unlike most other expressions, BIT_INSERT_EXPR has an implicit
>>>> operand of the precision/size of the second operand.  This means if we
>>>> have an integer constant for the second operand and that compares to
>>>> the same constant value, vn_nary_op_eq would return that these two
>>>> expressions are the same.  But in the case I was looking into the
>>>> integer constants had different types, one with 1 bit precision and
>>>> the other with 2 bit precision which means the BIT_INSERT_EXPR were
>>>> not equal at all.
>>>>
>>>> This patches the problem by checking to see if BIT_INSERT_EXPR's
>>>> operand 1's (second operand) type  has different precision to return
>>>> false.
>>>>
>>>> Is this the correct location or should we be checking for this
>>>> differently?  If this is the correct location, is the patch ok?
>>>> Bootstrapped and tested on aarch64-linux-gnu with no regressions (and
>>>> also tested with a few extra patches to expose BIT_INSERT_EXPR).
>>>>
>>>> Thanks,
>>>> Andrew Pinski
>>>>
>>>> ChangeLog:
>>>> * tree-ssa-sccvn.c (vn_nary_op_eq): Check BIT_INSERT_EXPR's operand 1
>>>> to see if the types precision matches.
>>>
>>>
>>> Hello,
>>>
>>> since BIT_INSERT_EXPR is implicitly an expression with 4 arguments, it makes
>>> sense that we may need a few such special cases. But shouldn't the hash
>>> function be in sync with the equality comparator? Does operand_equal_p need
>>> the same?
>>
>> The hash function does not need to be exactly the same.  The only
>> requirement there is if vn_nary_op_eq returns true then the hash has
>> to be the same.  Now we could improve the hash by using the precision
>> which will allow us not to compare as much in some cases.
>>
>> Yes operand_equal_p needs the same handling; I did not notice that
>> until you mention it..
>> Right now it does:
>>         case BIT_INSERT_EXPR:
>>           return OP_SAME (0) && OP_SAME (1) && OP_SAME (2);
>
> Aww.  The issue is that operand_equal_p treats INTEGER_CSTs of different
> type/precision but the same value as equal.
>
> Revisiting that, while a good idea, shouldn't block a fix here.  So ...
>
> Index: tree-ssa-sccvn.c
> ===================================================================
> --- tree-ssa-sccvn.c    (revision 250159)
> +++ tree-ssa-sccvn.c    (working copy)
> @@ -2636,6 +2636,14 @@ vn_nary_op_eq (const_vn_nary_op_t const
>      if (!expressions_equal_p (vno1->op[i], vno2->op[i]))
>        return false;
>
> +  /* BIT_INSERT_EXPR has an implict operand as the type precision
> +     of op1.  Need to check to make sure they are the same.  */
> +  if (vno1->opcode == BIT_INSERT_EXPR)
> +    if (INTEGRAL_TYPE_P (TREE_TYPE (vno1->op[0]))
> +       && TYPE_PRECISION (TREE_TYPE (vno1->op[1]))
> +           != TYPE_PRECISION (TREE_TYPE (vno2->op[1])))
> +      return false;
> +
>
> the case can be restricted to INTEGER_CST vno1->op[0] I think:
>
>   if (vno1->opcode == BIT_INSERT_EXPR
>       && TREE_CODE (vno1->op[0]) == INTEGER_CST
>       && TYPE_PRECISION (....
>
> and yes, operand_equal_p needs a similar fix.  Can you re-post with that added?

Here is that with the changes you requested too.

> Do you have a testcase?

I don't have one which fails with the trunk.  With lowering of
bit-fields accesses (which I hope to submit soon; just getting in the
required patches first), many testcases fail (bootstrap fails for the
same reason too).

OK? Bootstrapped and tested on aarch64-linux-gnu with no regressions.

Thanks,
Andrew

ChangeLog:
* tree-ssa-sccvn.c (vn_nary_op_eq): Check BIT_INSERT_EXPR's operand 1
to see if the types precision matches.
* fold-const.c (operand_equal_p): Likewise,


>
> Thanks,
> Richard.
>
>> Thanks,
>> Andrew Pinski
>>
>>>
>>> --
>>> Marc Glisse