On Apr 23, 2020, at 1:33 PM, Maurizio Cimadamore
<maurizio.cimadam...@oracle.com> wrote:
Hi,
time has come for another round of foreign memory access API incubation (see JEP 383
[3]). This iteration aims at polishing some of the rough edges of the API, and adds some
of the functionalities that developers have been asking for during this first round of
incubation. The revised API tightens the thread-confinement constraints (by removing the
MemorySegment::acquire method) and instead provides more targeted support for parallel
computation via a segment spliterator. The API also adds a way to create a custom native
segment; this is, essentially, an unsafe API point, very similar in spirit to the JNI
NewDirectByteBuffer functionality [1]. By using this bit of API, power-users will be
able to add support, via MemorySegment, to *their own memory sources* (e.g. think of a
custom allocator written in C/C++). For now, this API point is called off as
"restricted" and a special read-only JDK property will have to be set on the
command line for calls to this method to succeed. We are aware there's no precedent for
something like this in the Java SE API - but if Project Panama is to remain true about
its ultimate goal of replacing bits of JNI code with (low level) Java code, stuff like
this has to be *possible*. We anticipate that, at some point, this property will become a
true launcher flag, and that the foreign restricted machinery will be integrated more
neatly into the module system.
A list of the API, implementation and test changes is provided below. If you
have any questions, or need more detailed explanations, I (and the rest of the
Panama team) will be happy to point at existing discussions, and/or to provide
the feedback required.
Thanks
Maurizio
Webrev:
http://cr.openjdk.java.net/~mcimadamore/8243491_v1/webrev
Javadoc:
http://cr.openjdk.java.net/~mcimadamore/8243491_v1/javadoc
Specdiff:
http://cr.openjdk.java.net/~mcimadamore/8243491_v1/specdiff/overview-summary.html
CSR:
https://bugs.openjdk.java.net/browse/JDK-8243496
API changes
===========
* MemorySegment
- drop support for acquire() method - in its place now you can obtain a
spliterator from a segment, which supports divide-and-conquer
- revamped support for views - e.g. isReadOnly - now segments have access
modes
- added API to do serial confinement hand-off
(MemorySegment::withOwnerThread)
- added unsafe factory to construct a native segment out of an existing address; this
API is "restricted" and only available if the program is executed using the
-Dforeign.unsafe=permit flag.
- the MemorySegment::mapFromPath now returns a MappedMemorySegment
* MappedMemorySegment
- small sub-interface which provides extra capabilities for mapped segments
(load(), unload() and force())
* MemoryAddress
- added distinction between *checked* and *unchecked* addresses; *unchecked*
addresses do not have a segment, so they cannot be dereferenced
- added NULL memory address (it's an unchecked address)
- added factory to construct MemoryAddress from long value (result is also
an unchecked address)
- added API point to get raw address value (where possible - e.g. if this is
not an address pointing to a heap segment)
* MemoryLayout
- Added support for layout "attributes" - e.g. store metadata inside
MemoryLayouts
- Added MemoryLayout::isPadding predicate
- Added helper function to SequenceLayout to rehape/flatten sequence layouts
(a la NDArray [4])
* MemoryHandles
- add support for general VarHandle combinators (similar to MH combinators)
- add a combinator to turn a long-VH into a MemoryAddress VH (the resulting
MemoryAddress is also *unchecked* and cannot be dereferenced)
Implementation changes
======================
* add support for VarHandle combinators (e.g. IndirectVH)
The idea here is simple: a VarHandle can almost be thought of as a set of method handles
(one for each access mode supported by the var handle) that are lazily linked. This gives
us a relatively simple idea upon which to build support for custom var handle adapters:
we could create a VarHandle by passing an existing var handle and also specify the set of
adaptations that should be applied to the method handle for a given access mode in the
original var handle. The result is a new VarHandle which might support a different
carrier type and more, or less coordinate types. Adding this support was relatively easy
- and it only required one low-level surgery of the lambda forms generated for adapted
var handle (this is required so that the "right" var handle receiver can be
used for dispatching the access mode call).
All the new adapters in the MemoryHandles API (which are really defined inside
VarHandles) are really just a bunch of MH adapters that are stitched together
into a brand new VH. The only caveat is that, we could have a checked exception
mismatch: the VarHandle API methods are specified not to throw any checked
exception, whereas method handles can throw any throwable. This means that,
potentially, calling get() on an adapted VarHandle could result in a checked
exception being thrown; to solve this gnarly issue, we decided to scan all the
filter functions passed to the VH combinators and look for direct method
handles which throw checked exceptions. If such MHs are found (these can be
deeply nested, since the MHs can be adapted on their own), adaptation of the
target VH fails fast.
* More ByteBuffer implementation changes
Some more changes to ByteBuffer support were necessary here. First, we have added support for
retrieval of "mapped" properties associated with a ByteBuffer (e.g. the file descriptor,
etc.). This is crucial if we want to be able to turn an existing byte buffer into the "right
kind" of memory segment.
Conversely, we also have to allow creation of mapped byte buffers given
existing parameters - which is needed when going from (mapped) segment to a
buffer. These two pieces together allow us to go from segment to buffer and
back w/o losing any information about the underlying memory mapping (which was
an issue in the previous implementation).
Lastly, to support the new MappedMemorySegment abstraction, all the memory
mapped supporting functionalities have been moved into a common helper class so
that MappedMemorySegmentImpl can reuse that (e.g. for
MappedMemorySegment::force).
* Rewritten memory segment hierarchy
The old implementation had a monomorphic memory segment class. In this round we aimed at splitting
the various implementation classes so that we have a class for heap segments
(HeapMemorySegmentImpl), one for native segments (NativeMemorySegmentImpl) and one for memory
mapped segments (MappedMemorySegmentImpl, which extends from NativeMemorySegmentImpl). Not much to
see here - although one important point is that, by doing this, we have been able to speed up
performances quite a bit, since now e.g. native/mapped segments are _guaranteed_ to have a null
"base". We have also done few tricks to make sure that the "base" accessor for
heap segment is sharply typed and also NPE checked, which allows C2 to speculate more and hoist.
With these changes _all_ segment types have comparable performances and hoisting guarantees (unlike
in the old implementation).
* Add workarounds in MemoryAddressProxy, AbstractMemorySegmentImpl to special case
"small segments" so that VM can apply bound check elimination
This is another important piece which allows to get very good performances out
of indexes memory access var handles; as you might know, the JIT compiler has
troubles in optimizing loops where the loop variable is a long [2]. To make up
for that, in this round we add an optimization which allows the API to detect
whether a segment is *small* or *large*. For small segments, the API realizes
that there's no need to perform long computation (e.g. to perform bound checks,
or offset additions), so it falls back to integer logic, which in turns allows
bound check elimination.
* renaming of the various var handle classes to conform to "memory access var
handle" terminology
This is mostly stylistic, nothing to see here.
Tests changes
=============
In addition to the tests for the new API changes, we've also added some stress tests for
var handle combinators - e.g. there's a flag that can be enabled which turns on some
"dummy" var handle adaptations on all var handles created by the runtime. We've
used this flag on existing tests to make sure that things work as expected.
To sanity test the new memory segment spliterator, we have wired the new
segment spliterator with the existing spliterator test harness.
We have also added several micro benchmarks for the memory segment API (and
made some changes to the build script so that native libraries would be handled
correctly).
[1] -
https://docs.oracle.com/en/java/javase/14/docs/specs/jni/functions.html#newdirectbytebuffer
[2] - https://bugs.openjdk.java.net/browse/JDK-8223051
[3] - https://openjdk.java.net/jeps/383
[4] -
https://docs.scipy.org/doc/numpy/reference/generated/numpy.reshape.html#numpy.reshape
