RFR(M): 8027422: assert(_gvn.type(obj)->higher_equal(tjp)) failed: cast_up is no longer needed

[Roland Westrelin]

Hi Vladimir,

Here is a new webrev for this:
http://cr.openjdk.java.net/~roland/8027422/webrev.01/

I fixed the issues you mentioned in your review.
I added a call to remove_speculative() to the ConNode constructor. When a node becomes constant, its speculative part can be not null. The IGVN doesn’t kill ConNodes so without a call to remove_speculative() a ConNode with a speculative part can sneak past the call to Compile::remove_speculative_types().
I also added a verification method:
NodeHash::check_speculative_types()
to check that no TypeNode with a speculative type is still in the IGVN hash table after Compile::remove_speculative_types()

Roland.

On Nov 19, 2013, at 8:49 PM, Vladimir Kozlov <vladimir.kozlov at oracle.com> wrote:

> On 11/19/13 11:45 AM, Roland Westrelin wrote:
>> Thanks for reviewing this, Vladimir.
>> 
>> On Nov 19, 2013, at 1:34 AM, Vladimir Kozlov <vladimir.kozlov at oracle.com> wrote:
>> 
>>> Next 2 places in type.cpp pass 'true' to meet() unconditionally:
>>> 
>>> 1929     return TypeAry::make(_elem->meet(a->_elem, true),
>>> 
>>> 3812     const TypeAry *tary = _ary->meet(tap->_ary, true)->is_ary();
>>> 
>>> Should TypeAryPtr::remove_speculative() also clean _speculative in element's type?
>> 
>> You’re right. It probably should.
>> So I need to add a remove_speculative() method to TypeAry. Then the 2 places where true is passed to meet() for TypeAry don’t matter anymore because remove_speculative() is called from meet() and remove_speculative now has an effect on TypeAry, right?
> 
> Right, passing 'true' will work in all cases then.
> 
> Thanks,
> Vladimir
> 
>> 
>>> Could you make printing code with 'this_t' aligned again in Type::meet()?
>> 
>> Sure.
>> 
>> Roland.
>> 
>>> 
>>> thanks,
>>> Vladimir
>>> 
>>> On 11/18/13 1:15 PM, Roland Westrelin wrote:
>>>> The root of the problem is that during the null check when the type of obj is improved in GraphKit::cast_not_null():
>>>> const Type *t_not_null = t->join(TypePtr::NOTNULL, true);
>>>> The join with TypePtr::NOTNULL is not applied to the speculative part. In fact, no meet between a TypeOopPtr and a TypePtr modifies the speculative part. One way to fix it would be to apply the meet with a TypePtr to the speculative part as well as the standard part of the type which I tried: then we need to move the _speculative field up in TypePtr and modify all operations on TypePtr to operate on _speculative so that the type system remains symmetric.
>>>> In many places where we mix a TypePtr with a TypeOopPtr we actually don’t care about the speculative part. I changed the following operations on Type:
>>>> higher_equal()
>>>> meet()
>>>> join()
>>>> filter()
>>>> so that by default they don’t return a result that include the speculative part of the type. Where we need the speculative part of the type, we have to explicitly request it.
>>>> 
>>>> I also fixed a problem with Type nodes with a _type of TypeNarrowOop that wouldn’t drop the speculative part of the type during Compile::remove_speculative_types().
>>>> I included small clean ups that Mikael suggested privately (dropped the duplicate check for res->isa_oopptr() in TypeOopPtr::meet, make remove_speculative not go through the exercise of creating a new type if speculative is NULL).
>>>> 
>>>> http://cr.openjdk.java.net/~roland/8027422/webrev.00/
>>>> 
>>>> Roland.
>>>> 
>>