[9] RFR(S): 8150804: C2 Compilation fails with assert(_base >= OopPtr && _base <= AryPtr) failed: Not a Java pointer

[Tobias Hartmann]

Hi Vladimir,

thanks for the review!

On 14.03.2016 02:28, Vladimir Kozlov wrote:
> It is strange that next subgraph did not collapse - CheckCastPPNode::Identity() should remove second CheckCastPP:
> 
>>         CheckCastPP
>>          A:NotNull
>>          /       \
>> CheckCastPP     |
>>   A:NotNull      |
>>           \     /
>>             Phi
>>              A

When CheckCastPPNode::Identity() is first invoked, the graph looks like this

 92	CheckCastPP	===  90  32  [[]]  #A:NotNull *  Oop:A:NotNull * !jvms: TestPhiElimination::test @ bci:8
 135	CheckCastPP	===  133  92  [[]]  #A:NotNull * (speculative=A:NotNull:exact * (inline_depth=2))  Oop:A:NotNull * (speculative=A:NotNull:exact * (inline_depth=2)) !jvms: A::get @ bci:9 TestPhiElimination::test @ bci:11

Because 135 has a speculative part, the types are not equal and 135 is not replaced by 92.

> I think next code in PhiNode::Ideal() should check can_reshape too since during parsing graph is incomplete:
> 
>   if (uin == NULL) {
>     uncasted = true;
>     uin = unique_input(phase, true);
>   }

Yes, that's a better solution. I verified that it works (Phi is folded after parsing and has correct A:NotNull type).

Here is the new webrev:
http://cr.openjdk.java.net/~thartmann/8150804/webrev.01/

Thanks,
Tobias

> 
> Thanks,
> Vladimir
> 
> On 3/11/16 7:40 AM, Tobias Hartmann wrote:
>> Hi,
>>
>> please review the following patch.
>>
>> https://bugs.openjdk.java.net/browse/JDK-8150804
>> http://cr.openjdk.java.net/~thartmann/8150804/webrev.00/
>>
>> We fail in Compile::Process_OopMap_Node() while processing monitors of a safepoint node because the monitor object is TOP. The crash is rare but reproduces with my regression test. The problem is the elimination of Phi nodes with a unique input which was broken by the fixes for JDK-8139771 [1] and JDK-8146999 [2].
>>
>> Here are the details (for context, see 'TestPhiElimination.java'):
>> A::get() is inlined into test(obj) producing the following graph:
>>
>>          Parm (obj)
>>       TestPhiElimination
>>              |
>>            CastPP
>>   TestPhiElimination:NotNull
>>              |
>>         CheckCastPP
>>          A:NotNull
>>          /       \
>> CheckCastPP     |
>>   A:NotNull      |
>>           \     /
>>             Phi
>>              A
>>              |
>>          Safepoint
>>
>> https://bugs.openjdk.java.net/secure/attachment/57820/before_ideal.png
>>
>> PhiNode::ideal() then replaces the Phi by a CheckCastPP because it has a unique input (see PhiNode::unique_input()):
>>
>>          Parm (obj)
>>       TestPhiElimination
>>              |
>>         CheckCastPP
>>              A
>>              |
>>          Safepoint
>>
>> https://bugs.openjdk.java.net/secure/attachment/57821/after_ideal.png
>>
>> We completely lose the NotNull information provided by the CastPP. Therefore, we cannot prove that obj != null when accessing a field of obj and add an uncommon trap. Obj is also used as a monitor (A::get() is synchronized) and set to TOP in the uncommon trap branch. We are never able to prove that the null branch is not reachable and later fail when emitting code in Process_OopMap_Node because the monitor object is still TOP.
>>
>> Before the fix for JDK-8139771, we had a check to verify that the type of the unique (uncasted) input is "at least as good" as the type of the PhiNode:
>>
>>   phase->type(uncasted_input)->higher_equal(type()))
>>
>> http://hg.openjdk.java.net/jdk9/hs-comp/hotspot/rev/9e17d9e4b59f#l4.79
>>
>> Re-adding this check, fixes the problem. However, I'm concerned that this check is not strong enough. For example, in the case where the type of the PhiNode is Object:
>>
>>          Parm (obj)
>>       TestPhiElimination
>>              |
>>            CastPP
>>   TestPhiElimination:NotNull
>>              |
>>         CheckCastPP
>>          A:NotNull
>>          /       \
>> CheckCastPP     |
>>   A:NotNull      |
>>           \     /
>>             Phi
>>            Object
>>
>> We would still replace the Phi because TestPhiElimination->higher_equal(Object) and again lose the NotNull information. I therefore added a slightly stronger check that also checks the types in-between. I had to remove the assert that Roland added.
>>
>> What do you think?
>>
>> Thanks,
>> Tobias
>>
>> [1] https://bugs.openjdk.java.net/browse/JDK-8139771
>> [1] https://bugs.openjdk.java.net/browse/JDK-8146999
>>