RFR: JDK-8198285: More consistent Access API for arraycopy

Roman Kennke rkennke at redhat.com
Wed Apr 11 20:18:09 UTC 2018


I just realized we use ptrdiff_t for offset in the rest of the API. I am
not sure that it's really a good fit. ptrdiff_t is meant to be used in
pointer arithmetic and is relative to the size of the pointer type. We
use it to give an offset in bytes from object start to the field, which
is always positive and can be unsigned (ptrdiff_t is signed) and is
always in bytes. Alas, I changed the API to use ptrdiff_t for now to
make it consistent. If we want to change this, we shall change them all.

Please review the updated full webrev:
http://cr.openjdk.java.net/~rkennke/JDK-8198285/webrev.01/

Thanks, Roman

>  Currently, the arraycopy API in access.hpp gets the src and dst oops,
> plus the src and dst addresses. In order to be most useful to garbage
> collectors, it should receive the src and dst oops together with the src
> and dst offsets instead, and let the Access API / GC calculate the src
> and dst addresses.
> 
> For example, Shenandoah needs to resolve the src and dst objects for
> arraycopy, and then apply the corresponding offsets. With the current
> API (obj+ptr) it would calculate the ptr-diff from obj to ptr, then
> resolve obj, then re-add the calculate ptr-diff. This is fragile because
> we also may resolve obj in the runtime before calculating ptr (e.g. via
> arrayOop::base()). If we then pass in the original obj and a ptr
> calculated from another copy of the same obj, the above resolution logic
> would not work. This is currently the case for obj-arraycopy.
> 
> I propose to change the API to accept obj+offset, in addition to ptr for
> both src and dst. Only one or the other should be used. Heap accesses
> should use obj+offset and pass NULL for raw-ptr, off-heap accesses (or
> heap accesses that are already resolved.. use with care) should pass
> NULL+0 for obj+offset and the raw-ptr. Notice that this also allows the
> API to be used for Java<->native array bulk transfers.
> 
> An alternative would be to break the API up into 4 variants:
> 
> Java->Java transfer:
> arraycopy(oop src, size_t src_offs, oop dst, size_t dst_offs, size_t len)
> 
> Java->Native transfer:
> arraycopy(oop src, size_t src_offs, D* raw_dst, size_t len)
> 
> Native->Java transfer:
> arraycopy(S* src_raw, oop dst, size_t dst_offs, size_t len)
> 
> 'Unsafe' transfer:
> arraycopy(S* src_raw, D* dst_raw, size_t len)
> 
> 
> But that seemed to be too much boilerplate copy+pasting for my taste.
> (See how having this overly complicated template layer hurts us?)
> 
> Plus, I had a better idea: instead of accepting oop+offset OR T* for
> almost every Access API, we may want to abstract that and take an
> Address type argument, which would be either HeapAddress(obj, offset) or
> RawAddress(T* ptr). GCs may then just call addr->address() to get the
> actual address, or specialize for HeapAddress variants and resolve the
> objs and then resolve the address. This would also allow us to get rid
> of almost half of the API (all the *_at variants would go) and some
> other simplifications. However, this seemed to explode the scope of this
> RFE, and would be better handled in another RFE.
> 
> This changes makes both typeArrayKlass and objArrayKlass use the changed
> API, plus I identified all (hopefully) places where we do bulk
> Java<->native array transfers and make them use the API too. Gets us rid
> of a bunch of memcpy calls :-)
> 
> Please review the change:
> http://cr.openjdk.java.net/~rkennke/JDK-8198285/webrev.00/
> 
> Thanks, Roman
> 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: OpenPGP digital signature
URL: <https://mail.openjdk.org/pipermail/hotspot-gc-dev/attachments/20180411/dd3c778a/signature.asc>


More information about the hotspot-gc-dev mailing list