+1.
I noticed some slight performance degradation in Richards and other
octane benchmarks that seems to be caused by the ScriptFunction fix. But
it after some in-depth checks there is no additional invalidation going
on, so it must be the slightly heavier guards which are actually
necessary to fix the object wrapping bug we already had.
Hannes
Am 2013-08-08 09:04, schrieb A. Sundararajan:
enhanced :-)
http://cr.openjdk.java.net/~sundar/8022524/webrev.03/
-Sundar
On Thursday 08 August 2013 11:58 AM, Attila Szegedi wrote:
+1 on your webrev.02;
You could still enhance few things, such as:
- you could just write the code for getInvokeByName and
getDynamicInvoker once if you wrote a generic version as private
static T getLazilyCreatedValue(final Object key, final Callable<T>
creator, final ConcurrentMap<Object, T>) (that's basically equivalent
to computeIfAbsent)
- you could inline the single-arg createIteratorCallbackInvoker into
the two-arg version, since it's no longer used from elsewhere
but those are minor things.
Attila.
On Aug 8, 2013, at 7:24 AM, A. Sundararajan
<[email protected]> wrote:
Thanks. I got it. Basically InvokeByName/MethodHandle may be created
by Callable atmost twice -- and the second one will be thrown away.
I'll make these changes.
-Sundar
On Thursday 08 August 2013 10:41 AM, Attila Szegedi wrote:
On Aug 8, 2013, at 5:45 AM, A. Sundararajan
<[email protected]> wrote:
I'd like to lazily initialize InvokeByName and dynamic method
handles in global instance. I am not sure of the refactoring in
NativeArray that you suggested.
Are you talking about these?
private final MethodHandle someInvoker =
getSOME_CALLBACK_INVOKER();
If I pass key to a refactored method, I've to do if..else on
object identity check.. Or am I missing something?
This is what I had in mind:
private static MethodHandle getEVERY_CALLBACK_INVOKER() {
return createIteratorCallbackInvoker(EVERY_CALLBACK_INVOKER,
boolean.class);
}
private static MethodHandle getSOME_CALLBACK_INVOKER() {
return createIteratorCallbackInvoker(SOME_CALLBACK_INVOKER,
boolean.class);
}
...
private static createIteratorCallbackInvoker(final Object key,
final Class<?> rtype) {
return Global.instance().getDynamicInvoker(key,
new Callable<MethodHandle>() {
@Override
public MethodHandle call() {
return createIteratorCallbackInvoker(rtype);
}
});
}
Also, on avoiding synchronized in Global.java. If we'd like to
avoid jdk8 specific API/syntax in nashorn, I can't use
computeIfAbsent. But putIfAbsent forces computing the value...
Again, am I missing something?
private final ConcurrentMap<Object, InvokeByName> namedInvokers =
new ConcurrentHashMap<>();
@Override
public InvokeByName getInvokeByName(final Object key, final
Callable<InvokeByName> creator) {
final InvokeByName invoke = namedInvokers.get(key);
if(invoke != null) {
return invoke;
}
final InvokeByName newInvoke = creator.call();
final InvokeByName existingInvoke =
namedInvokers.putIfAbsent(key, newInvoke);
return existingInvoke != null ? existingInvoke : newInvoke;
}
thanks,
-Sundar
On Thursday 08 August 2013 08:56 AM, A. Sundararajan wrote:
On Thursday 08 August 2013 02:29 AM, Attila Szegedi wrote:
- CompileUnit: While making fields non-final and nulling out
fields is certainly a solution, I don't like it as it feels
fragile - you end up with an object that has a member nulled
out, and what if something later would want to depend on it etc.
As an example, consider CompileUnit, which now has its
ClassEmitter nulled out. Seems like architecturally it's a
better idea is to remove the field from the CompileUnit
altogether, and use a composite object being a tuple of
(CompileUnit, ClassEmitter) in the compiler, and only pass down
the CompileUnit part of the tuple to things in the IR package
that require it.
While code can be refactored for a longer term, as of now, it
does leak memory. Moment class is loaded, we don't need lots of
info maintained by ASM's ClassEmitter. I suggest we go with short
term solution and revisit refactoring changes to
FunctionNode/CompileUnit/Compiler later.
- Another issue I have is with synchronization in the Global
object; I'd rather use a ConcurrentMap and the (new for Java 8)
computeIfAbsent() method.
<http://download.java.net/jdk8/docs/api/java/util/Map.html#computeIfAbsent(K,
java.util.function.Function)>. If you don't want to rely on
computeIfAbsent() (but I don't see why wouldn't you, frankly),
you could still use a composition of get() and putIfAbsent().
We still don't use any jdk8 specific API in nashorn codebase yet
(I believe). I'll restructure this with older API.
- In NativeArray, you could factor out the pattern of getting an
invoker for an iterator callback repeated across 4 methods into
a method taking a key and a return type.
Will do.
- Ostensibly, NativeObject could just use Global.TO_STRING
instead of having its own now. Not too convinced about this, as
these things sort-of represent call sites, so maybe it's okay as
it is.
Yes - it is a different callsite (although I doubt how much
InvokeByName and dynamic invokers help now!)
- We still keep GlobalObject interface around?
Yes - we do. That calls for more refactorings. As I said, I'd
like to keep it minimal (as much as possible) for now.
- Why does RecompilableScriptFunctionData.ensureHasAllocator
have to be synchronized? If we absolutely need atomic updates to
the allocator field, I'd consider using an AtomicReference for
it instead. Having synchronization in path of every "new
SomeClass()" bothers me. Even if it's completely unsynced and
the field is not volatile, we only "risk" creating the method
handle multiple times; shouldn't be a big deal as we're (a)
rarely multithreaded and (b) it's idempotent. So, I'd rather
choose a bit of a statistical redundancy than a certain
performance hit.
- Why does ensureCodeGenerated have to be synchronized? Can the
modifications of fields possibly occur on multiple threads? I
mean, functionNode.canSpecialize() will be determined at first
execution and fields nulled out. Also, wouldn't a second call to
ensureCodeGenerated() after functionNode was nulled out (if
that's possible) result in a NPE on functionNode.isLazy(), or is
this guarded by !code.isEmpty()? At least this synchronization
only happens once on every linking event and not on every
invocation, unlike allocate() but I still don't really see the
necessity.
I'll check again.
-Sundar
Attila.
On Aug 7, 2013, at 6:56 PM, A. Sundararajan
<[email protected]> wrote:
Please review http://cr.openjdk.java.net/~sundar/8022524/
Thanks
-Sundar
