RFR(XL): 6312651: Compiler should only use verified interface types for optimization

[Vladimir Ivanov]

Roland,

> Fixing CheckCastPPNode::Value() exposes a bug in the C2 type checking
> logic: if a type check is proved to always fail during optimizations,
> the CheckCastPPNode becomes top, but the actual type checking logic in
> Phase::gen_subtype_check() doesn't always optimize out so the data path
> dies but not the control path. To fix that, I added a
> PartialSubtypeCheckNode::Value() method that triggers on subtype check
> failures consistently with CheckCastPPNode::Value. Problem is
> PartialSubtypeCheckNode is not checked first in
> Phase::gen_subtype_check(). So I added a duplicate test on the
> PartialSubtypeCheckNode result first in that logic, under an Opaque4Node
> so the actual is never compiled in the final code.

I have been working on JDK-8222075 [1] to fix a performance regression 
(JDK-8220708 [2]) caused by interface subtype checks.

I checked that your patch fixes it.

What I came up with is a solution based on late expansion of subtype 
checks [3]. It makes IR shape simpler (guard + deopt) and it works well 
with predicate hoisting.

Do you see any value in representing subtype checks as macro nodes once 
C2 type system handles interfaces better?

Best regards,
Vladimir Ivanov

[1] https://bugs.openjdk.java.net/browse/JDK-8222075

[2] https://bugs.openjdk.java.net/browse/JDK-8220708

[3] http://cr.openjdk.java.net/~vlivanov/8222075/webrev.00/