Hi Andrej,
Sorry for the delay getting back to you.
On 29/05/2014 10:24 PM, Andrej Golovnin wrote:
Hi David,
The valueOf calls may also allocate a new object so you can't just
delete the JvmtiExport::post_vm_object_alloc call. Unfortunately you
can't tell whether a new object was allocated or not. It is only for the
smaller primitive types that any kind of Object caching is mandated.
It is only for the smaller values (-128 to +127) of the integer primitives
types (plus boolean) that caching is mandated. Float.valueOf and Double.valueOf
always create objects.
You are right, that #valueOf call may allocate an object.
But as far as I understand currently the JvmtiExport::post_vm_object_alloc call
is only needed, because today the native code itself allocates an object
by calling java_lang_boxing_object::create(type, value, CHECK_NULL);.
Right, sorry - I was misunderstanding the purpose of the
post_vm_object_alloc call:
http://docs.oracle.com/javase/8/docs/platform/jvmti/jvmti.html#VMObjectAlloc
So from the perspective that you are diverting this back to Java code
the hotspot changes look okay to me.
The more general question, for the core-libs folk, is whether changing
everything to use valueOf is overkill (given the limits of the required
caching mechanisms) or good to do from a consistency perspective. I'm
slightly on the overkill side of things but not enough to reject things.
On the performance/benefit side, if I read things correctly you only see
the 9GB of Boolean objects because you disable reflection-inflation - is
that right? In that case, as Joel states, the gains are not really
general, but on the other hand I don't see anything wrong with trying to
improve the general efficiency here even if the greatest benefit comes
from a "non-mainstream" usecase.
David
-----
My code changes this behavior and delegates object allocation back to Java
by calling
JavaCalls::call_static(&boxed_value,
klass_handle,
vmSymbols::valueOf_name(),
valueOf_signature,
&args,
THREAD);
But maybe I misunderstood the implementation of JavaCalls.
Best regards,
Andrej Golovnin
