RFR: 8199739: Use HeapAccess when loading oops from static fields in javaClasses.cpp

[Roman Kennke]

Am 20.03.2018 um 11:07 schrieb Erik Österlund:
> Hi Roman,
> 
> On 2018-03-19 21:11, Roman Kennke wrote:
>> Am 19.03.2018 um 20:35 schrieb coleen.phillimore at oracle.com:
>>>
>>> On 3/19/18 3:15 PM, Stefan Karlsson wrote:
>>>> On 2018-03-19 20:00, coleen.phillimore at oracle.com wrote:
>>>>> I like Roman's version with static_field_base() the best.  The reason
>>>>> I wanted to keep static_field_addr and not have static_oop_addr was
>>>>> so there is one function to find static fields and this would work
>>>>> with the jvmci classes and with loading/storing primitives also.  So
>>>>> I like the consistent change that Roman has.
>>>> That's OK with me. This RFE grew in scope of what I first intended, so
>>>> I'm fine with Roman taking over this.
>>>>
>>>>> There's a subtlety that I haven't quite figured out here.
>>>>> static_field_addr gets an address mirror+offset, so needs a load
>>>>> barrier on this offset, then needs a load barrier on the offset of
>>>>> the additional load (?)
>>>> There are two barriers in this piece of code:
>>>> 1) Shenandoah needs a barrier to be able to read fields out of the
>>>> java mirror
>>>> 2) ZGC and UseCompressedOops needs a barrier when loading oop fields
>>>> in the java mirror.
>>>>
>>>> Is that what you are referring to?
>>> I had to read this thread over again, and am still foggy, but it was
>>> because your original change didn't work for shenandoah, ie Kim's last
>>> response.
>>>
>>> The brooks pointer has to be applied to get the mirror address as well
>>> as reading fields out of the mirror, if I understand correctly.
>>>
>>> OopHandle::resolve() which is what java_mirror() is not accessorized but
>>> should be for shenandoah.  I think.  I guess that was my question
>>> before.
>> The family of _at() functions in Access, those which accept oop+offset,
>> do the chasing of the forwarding pointer in Shenandoah, then they apply
>> the offset, load the memory field and return the value in the right
>> type. They also do the load-barrier in ZGC (haven't checked, but that's
>> just logical).
>>
>> There is also oop Access::resolve(oop) which is a bit of a hack. It has
>> been introduced because of arraycopy and java <-> native bulk copy stuff
>> that uses typeArrayOop::*_at_addr() family of methods. In those
>> situations we still need to 1. chase the fwd ptr (for reads) or 2. maybe
>> evacuate the object (for writes), where #2 is stronger than #1 (i.e. if
>> we do #2, then we don't need to do #1). In order to keep things simple,
>> we decided to make Access::resolve(oop) do #2, and have it cover all
>> those cases, and put it in arrayOopDesc::base(). This does the right
>> thing for all cases, but it is a bit broad, for example, it may lead to
>> double-copying a potentially large array (resolve-copy src array from
>> from-space to to-space, then copy it again to the dst array). For those
>> reasons, it is advisable to think twice before using _at_addr() or
>> in-fact Access::resolve() if there's a better/cleaner way to do it.
> 
> Are we certain that it is indeed only arraycopy that requires stable
> accesses until the next thread transition?
> I seem to recall that last time we discussed this, you thought that
> there was more than arraycopy code that needed this. For example
> printing and string encoding/decoding logic.
> 
> If we are going to make changes based on the assumption that we will be
> able to get rid of the resolve() barrier, then we should be fairly
> certain that we can indeed get rid of it. So have the other previously
> discussed roadblocks other than arraycopy disappeared?

No, I don't think that resolve() can go away. If you look at:

http://mail.openjdk.java.net/pipermail/hotspot-gc-dev/2018-March/021464.html

You'll see all kinds of uses of _at_addr() that cannot be covered by
some sort of arraycopy, e.g. the string conversions stuff.

The above patch proposes to split resolve() to resolve_for_read() and
resolve_for_write(), and I don't think it is unreasonable to distinguish
those. Besides being better for Shenandoah (reduced latency on read-only
accesses), there are conceivable GC algorithms that require that
distinction too, e.g. transactional memory based GC or copy-on-write
based GCs. But let's probably continue this discussion in the thread
mentioned above?

Thanks, Roman