Will unaligned memory access work on all platforms?

[Manuel Bleichenbacher]

Thanks for the clarification.

It would certainly make sense to mention it in the documentation as well.

I did an experiment with CAS similar to the one in your answer as I had
seen the restriction mentioned in the MemorySegment documentation.
Strangely enough, I couldn't find any mention in the VarHandle
documentation. Possibly, it's documented somewhere else. It seems
additional restrictions apply to CAS, e.g. it's not supported for "short"
(2 byte) access and likely several other value types. But that's not really
relevant for my work...

Regards
Manuel

Am Di., 2. Jan. 2024 um 10:35 Uhr schrieb Per-Ake Minborg <
per-ake.minborg at oracle.com>:

> Hi Manuel!
>
> Thanks for your feedback!
>
> In short, var handle operations with normal memory semantics (such as
> get() and set()) will always work with unaligned access whereas atomic
> operations, such as compare-and-set, will always fail with an
> IllegalArgumentException if attempted on a physical memory address that is
> not properly aligned. MemorySegment get and set operations will just
> delegate to the var handle operations and consequently, they will always
> work (as they always use normal memory semantics).
>
> The behavior is the same regardless of whether it is a heap or a native
> segment and it is not possible to crash the JVM using unaligned access.
>
> I think the documentation can be improved in this area. I will bring it up
> with the team.
>
> Here is an example of how to reproduce unaligned CAS-operation alignment
> behavior:
>
> jshell --enable-preview
> |  Welcome to JShell -- Version 21
> |  For an introduction type: /help intro
>
>
> jshell> var segment = java.lang.foreign.Arena.ofAuto().allocate(16)
> segment ==> MemorySegment{ heapBase: Optional.empty
> address:105553116302944 limit: 16 }
>
> jshell> var h =
> java.lang.foreign.ValueLayout.JAVA_INT_UNALIGNED.varHandle();
> h ==> VarHandle[varType=int, coord=[interface
> java.lang.foreign.MemorySegment]]
>
> jshell> h.compareAndSet(segment, 0, 1);
> $4 ==> true
>
> jshell> var slice = segment.asSlice(1)
> slice ==> MemorySegment{ heapBase: Optional.empty address:105553116302945
> limit: 15 }
>
> jshell> h.compareAndSet(slice, 0, 1);
> |  Exception java.lang.IllegalArgumentException: Misaligned access at
> address: 105553116302945
> |        at
> VarHandleSegmentViewBase.newIllegalArgumentExceptionForMisalignedAccess
> (VarHandleSegmentViewBase.java:57)
> |        at VarHandleSegmentAsInts.offset (VarHandleSegmentAsInts.java:89)
> |        at VarHandleSegmentAsInts.compareAndSet
> (VarHandleSegmentAsInts.java:192)
> |        at (#6:1)
>
> Note that in the final API (in Java 22), there is an extra coordinate for
> an offset value for the VarHandle above.
>
> Best, Per Minborg
> ------------------------------
> *From:* panama-dev <panama-dev-retn at openjdk.org> on behalf of Manuel
> Bleichenbacher <manuel.bleichenbacher at gmail.com>
> *Sent:* Thursday, December 28, 2023 4:52 PM
> *To:* panama-dev at openjdk.org <panama-dev at openjdk.org>
> *Subject:* Will unaligned memory access work on all platforms?
>
> Hi,
>
> Instances of MemoryLayout support specifying alignment constraints.
> However, if the alignment is set to 1 byte for all layouts, will unaligned
> access work on all platforms?
>
> In a short test on a Mac with Apple Silicon, unaligned access works with
> both native and heap segments, across 8-byte boundaries using
> MemorySegment.get(), MemorySegment.set(), VarHandle.get() and
> VarHandle.set(). And the result will likely be the same on x86 machine,
> independent of the operating system. But is it guaranteed to work on any
> operating system and any CPU architecture, even on those ones where machine
> instructions may not read from or write to unaligned addresses?
>
> If it does not work on all platform, what would happen instead? Will it
> throw an exception? Could it crash the JVM? Is the behavior the same for
> native and heap segments?
>
> While researching this topic, I kept wondering where to look for
> documentation about it. The only description I've found is in documentation
> of the MemorySegment class. It has quite a long chapter about alignment,
> yet it doesn't really tell why you would want to use alignment constraints
> in the first place and what happens if an unaligned access is made anyway,
> with or without alignment constraint on the layout.
>
> Regards
> Manuel Bleichenbacher
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/panama-dev/attachments/20240102/eb7b3e1a/attachment-0001.htm>