[PATCH][RFC] bitwise CCP, 2nd try

[Paolo Bonzini]

On 08/04/2010 04:14 PM, Richard Guenther wrote:
> On Wed, 4 Aug 2010, Paolo Bonzini wrote:
>
>> On 08/04/2010 02:53 PM, Richard Guenther wrote:
>>>       gcc_assert (old_val->lattice_val<   new_val.lattice_val
>>>                   || (old_val->lattice_val == new_val.lattice_val
>>>     		  && ((!old_val->value&&   !new_val.value)
>>> + 		      /* Allow transitioning from&x to&x&   ~3.  */
>>> + 		      || (TREE_CODE (old_val->value) != INTEGER_CST
>>> + 			   && TREE_CODE (new_val.value) == INTEGER_CST)
>>> + 		      || (TREE_CODE (old_val->value) == INTEGER_CST
>>> + 			   && TREE_CODE (new_val.value) == INTEGER_CST
>>> + 			   && double_int_equal_p
>>> + 			       (double_int_and_not
>>> + 				  (tree_to_double_int (old_val->value),
>>> + 				   new_val.mask),
>>> + 				double_int_and_not
>>> + 				  (tree_to_double_int (new_val.value),
>>> + 				   new_val.mask)))
>>>     		      || operand_equal_p (old_val->value, new_val.value,
>>> 0))));
>>
>> Also, do you have a testcase for the CONSTANT/CONSTANT check?
>
> The &x -> &x & ~3 transition?  It happens during bootstrap - I yet
> have to create a set of (exhaustive) testcases.

That, and

+   /* We have to be careful to not go up the bitwise lattice
+      represented by the mask.
+      ???  This doesn't seem to be the best place to enforce this.  */

if they aren't the same thing.

Paolo