(10) RFR (actually S) 8169881: Remove implicit Handle conversions oop->Handle

[David Holmes]

Hi Coleen,

On 14/02/2017 1:53 PM, Coleen Phillimore wrote:
>
>
> On 2/13/17 8:44 PM, David Holmes wrote:
>> Hi Coleen,
>>
>> "actually S"? :)
>
> Well, maybe not but it shouldn't have a lot of merge conflicts for
> backports.
>
>>
>> On 14/02/2017 6:10 AM, Coleen Phillimore wrote:
>>> Summary: Pass THREAD to Handle as argument instead of implicit
>>> Thread::current() call.
>>
>> Well there's a bit more to it than that as you also change parameter
>> types so that we (sometimes) avoid:
>>
>> - oop -> Handle -> oop
>> - Handle -> oop -> Handle
>
> Yes, I put that in the bug but not in the RFR.  The recommendation is to
> Handle the oop as far up as possible in the call stack and pass Handle
> around.

I hear what you are saying but it is hard to actually see that reflected 
in the code.

>> across method calls. This leads to some puzzling differences eg:
>>
>> src/share/vm/aot/aotCodeHeap.cpp
>> src/share/vm/classfile/javaClasses.hpp
>>
>> The change in aotCodeHeap.cpp reverses the use of Handle and oop, but
>> it is unclear why Throwable::print and Throwable::print_stack_trace
>> have different signatures in that regard. They used to both take a
>> Handle but now one takes an oop instead. ??
>
> Yes, you found the exceptions in Throwable.  A lot of the
> java_lang_Throwable (and other javaClasses.hpp classes) pass oop instead
> of Handle and I changed print to oop to make it consistent, and because
> they it being called from other functions in Throwable with oop.
> Throwable::print_stack_trace actually goes to a safepoint so has to be
> Handle.

I think in javaClasses.hpp the rule should be oop unless Handle is 
definitely needed.

>
>>
>> More generally it is unclear why you made the changes you did in
>> places - see below for some specifics. I'm left questioning how to
>> know when to take an oop and when to take a Handle. Of course that
>> question already exists, but these changes didn't make it any clearer
>> for me.
>
> Pass a Handle early and often is still the rule.  For all but GC code.
> There are still some functions that are very small and don't safepoint
> and take oop.  Some of these have migrated to taking Handle, but some
> haven't.  I didn't want to make this change larger for now.
>
>>
>> I understand why the oop to Handle implicit conversion could be
>> problematic due to the Thread::current() calls, but I don't understand
>> why you couldn't keep (? add?) the Handle to oop implicit conversions ??
>
> There weren't Handle -> oop implicit conversions.  The conversion back
> to oop uses the operator() which remains.

Okay then it seems an implicit Handle -> oop would avoid the need to 
change the caller from foo to foo() if the underlying API switched from 
oop to Handle.

> The oop -> Handle conversion has cost, both at runtime and also to the
> GC because it increases the number of Handles on the Thread->_handle_area.

Sure.

> I did this same experiment years ago when Thread::current() was at the
> top of the profiling call stack but back then I needed to add too many
> explicit Thread::current() calls.  There aren't many in this change, so
> it's worth doing.  Also, metadata handles have the same problem but
> there are still too many of them to remove the implicit conversion.
> And {instance}KlassHandle needs to be removed first.  They do nothing
> now but hide the type.
>
>>
>> ---
>>
>> src/share/vm/c1/c1_Runtime1.cpp
>>
>> Aside: I'm a little surprised that there was not a Handle assignment
>> operator or copy constructor such that given:
>>
>> 863   Handle appendix(THREAD, NULL);
>> 970         appendix = cpce->appendix_if_resolved(pool);
>>
>> that it simply updated the NULL to the desired value, while keeping
>> the THREAD intact.
>
> Handle doesn't save the THREAD so a NULL handle is just oop* NULL. The
> RHS of the expression has to be converted to Handle and the default
> assignment operator copies it to appendix.

I had assumed Handle saved the Thread.

>>
>> ---
>>
>> src/share/vm/ci/ciInstance.cpp
>>
>> Not at all obvious that replacing Handle with raw oop is correct. I'm
>> not saying it isn't, just that it isn't obvious - and I'll wonder why
>> the Handle was used in the first place.
>
> It is safe because it's immediately unhandled in all the case
> statements, and I wanted to avoid a Thread::current call.
>>
>> ---
>>
>> src/share/vm/classfile/javaClasses.cpp
>>
>> Why change java_lang_String::as_symbol_or_null to take a Handle
>> instead of an oop if you are simply going to unwrap the Handle to get
>> the oop back. ??? Particular when a caller like
>> src/share/vm/prims/methodHandles.cpp has to create the Handle from oop
>> in the first place.
>
> I changed it to be the same as java_lang_String::as_symbol(Handle
> java_string ...) which took a handle. java_lang_String::as_symbol() was
> the same - it immediately unhandles the argument to pass to the rest of
> the functions.  I changed java_lang_String::as_symbol() to take an oop
> to match, and there are few scattered changes from that.

I have to disagree with this one. AFAICS there is no reason this:

517 Symbol* java_lang_String::as_symbol(Handle java_string, TRAPS) {

needs to take a Handle in the first place. The java_string is unwrapped 
to get the oop and the oop is only passed to a few other String methods 
and then not touched. Looking at the API's in javaClass.hpp the norm 
should be to take oop (as the bulk of methods do) with Handle only used 
when actually required.

>
>>
>> ---
>>
>> src/share/vm/jvmci/jvmciCompilerToVM.hpp
>>
>> It is not at all obvious that JavaArgumentUnboxers are
>> thread-confined! I'm also concerned by API's (unfortunately a number
>> pre-existing) that take a Thread/JavaThread argument but really
>> require it to be the current thread. To that end I'd rather see
>> _thread as _cur_thread and an assertion when it is set.
>
> Since it's a ResourceObj and not StackObj (setting aside debate about
> that), I agree with you.  I changed it to use Thread::current() where it
> handles the object, which it explicitly did before.  In this case, it is
> a stack allocated thing but it's not guaranteed to be that.

Ok.

>>
>> ---
>>
>> src/share/vm/oops/cpCache.cpp
>>
>> Nit: objArrayHandle resolved_references (Thread::current(), ...
>>
>> Please remove space before (
>
> fixed.
>
>>
>> ---
>>
>> src/share/vm/prims/jvm.cpp
>>
>> Nit: need to fix indent on StackWalk::walk call second line.
>
> fixed.
>
>>
>> ---
>>
>> src/share/vm/prims/jvmtiGetLoadedClasses.cpp
>>
>> Ditto re _thread -> _cur_thread
>>
>
> This one is a StackObj (a Closure which should be used that way). So I
> changed _thread to _cur_thread and added an assert that it ==
> Thread::current() in the constructor.

Thanks.

>> ---
>>
>> src/share/vm/prims/jvmtiImpl.cpp
>>
>> Nit: space before ( again (on handle constructor)
>
> fixed.
>>
>>> It's a very small change to a number of files.  Notably the new JVMCI
>>> files have a lot of implicit conversions from oop to handle. I want to
>>> get this change in early for jdk10 to stop this usage.  I also added a
>>> few HandleMarks where I found that the HandlesArea was growing large.
>>
>> The addition and removal of the HandleMarks seems problematic because
>> there are no obvious rules being applied. It is hard to know where and
>> when a HandleMark is needed or not. Looking at
>> src/share/vm/jvmci/jvmciCompilerToVM.cpp do we really want to
>> create/destroy the HandleMark on each loop iteration? That would only
>> seem necessary if we know the number of iterations is so high that we
>> might consume all our handle space.
>
> Nobody knows when and where to put HandleMarks.   It's a dark art. I

So a trade-off between too much cleanup overhead if we have too many, 
and too much GC overhead if we have too few. And no real way to know 
either way - Ouch! :(

> also was testing with a JVM that restored the assert if HandleArea got
> over 200 handles and the numbers got very high in these functions.  The
> GC has to walk these HandleAreas and I am trying to reduce work for them.
>
> I tried to write about this in the bug.   I'd like to see if changing
> Handle to a proper constructor/destructor like methodHandles, so that
> only the Handles that are active need to be collected.   Then we don't
> need HandleMark and their associated mystery.   There are several
> HandleMarks in code that obviously doesn't allocate any Handles.  I took
> out several but restored them so that this patch was smaller.   To
> change Handle to a scoped object would require changing passing Handle
> as a const reference like I did with methodHandle which is too big of a
> change for early-jdk10.

Ok. This sounds like a way forward though. In combination with an easy 
way to detect when handles are required, this would allow very precise 
management of handle lifetimes.

> The reason I worked on this RFE was because I was curious to see how
> many oops/Handles we're passing around and using from within the JVM
> these days.  It's more than I thought.
>
> Thank you for reviewing this.  I'm going to rerun some tests with the
> changes above and send another webrev.

Ok. Thanks.

David
-----

> Coleen
>
>>
>> Thanks,
>> David
>> ------
>>
>>> See bug for more on the motivation of this change.
>>>
>>> open webrev at http://cr.openjdk.java.net/~coleenp/8169881.01/webrev
>>> bug link https://bugs.openjdk.java.net/browse/JDK-8169881
>>>
>>> Tested by running all hotspot jtreg tests with -XX:+CheckUnhandledOops.
>>> There weren't any unhandled oops amazingly.
>>>
>>> Thanks,
>>> Coleen
>