[10] RFR(S): 8188785: CCP sets invalid type for java mirror load

Tobias Hartmann tobias.hartmann at oracle.com
Fri Oct 20 08:04:04 UTC 2017


please review the following patch:

Since 8186777 [1], we require two loads to retrieve the java mirror from a klass oop:

LoadP(LoadP(AddP(klass_oop, java_mirror_offset)))

The problem is that now the type of the outermost LoadP does not depend on the inner LoadP (which has a raw pointer 
type) but on the type of the AddP which is one level up. CPP only propagates the types downwards to the direct users and 
as a result, the mirror LoadP ends up with an incorrect (too narrow/optimistic) type.

I've verified the fix with the failing test and also verified that 8188835 [2] is a duplicate.

Gory details:
During CCP, we compute the type of a Phi that merges oops of type A and B where B is a subtype of A. Since the type of 
the A input was not computed yet (it was initialized to TOP at the beginning of CCP), the Phi temporarily ends up with 
type B (i.e. with a type that is too narrow/optimistic). This type is propagated downwards and is being used to optimize 
a java mirror load from the klass oop:


The mirror load is then folded to TypeInstPtr::make(B) which is not correct because the oop can be of type A at runtime.


[1] https://bugs.openjdk.java.net/browse/JDK-8186777
[2] https://bugs.openjdk.java.net/browse/JDK-8188835

More information about the hotspot-compiler-dev mailing list