RFR(S): 8218201: Failures when vmIntrinsics::_getClass is not inlined
tobias.hartmann at oracle.com
Fri Mar 8 08:25:33 UTC 2019
Vladimir, Dean, thanks for the review.
My first fix version had '_return_allocated = false' but then I incorrectly assumed that the
behavior would be the same as if we don't special case at all. I missed that the receiver is of
course also involved and you are right that we should not treat it as "global escape".
On 07.03.19 20:20, dean.long at oracle.com wrote:
> I agree. We don't want obj.getClass() to treat obj as "global escape".
> PS - Original change goes back to JDK-6488063.
> On 3/7/19 10:46 AM, Vladimir Kozlov wrote:
>> _getClass is special case because it is native call and we can't do bytecode analysis.
>> But we know what it does (it loads klass from object and mirror from klass) - no allocations, no
>> locals, no arguments returned.
>> I think it is simple missing _return_allocated = false setting in addition to _return_local =
>> false. _return_allocated is set to true by default optimistically.
>> On 3/7/19 4:54 AM, Tobias Hartmann wrote:
>>> please review the following patch:
>>> When intrinsification is disabled, the BCEscapeAnalyzer marks the return value of the (native)
>>> method Object::getClass as "return allocated value" which means that "only newly allocated unescaped
>>> objects are returned". The OptimizePtrCompare optimization then uses this information to incorrectly
>>> fold 'obj.getClass() == Object.class' (see TestGetClass.java:39) to always false.
>>> This is a very old issue and I can't trace back why a special case for the _getClass intrinsic has
>>> been added to the BCEscapeAnalyzer. Since I don't think we should make any assumptions about the
>>> returned Object, I've removed the special case.
More information about the hotspot-compiler-dev