[foreign-memaccess+abi] RFR: 8278151: Heap segments should handle alignment constraints in a deterministic fashion [v2]
Maurizio Cimadamore
mcimadamore at openjdk.java.net
Thu Dec 2 14:31:45 UTC 2021
On Thu, 2 Dec 2021 14:27:17 GMT, Maurizio Cimadamore <mcimadamore at openjdk.org> wrote:
>> This patch fixes the issue with alignment checks on heap segment leading to failures which depend on VM implementation details and layout choices.
>>
>> This is discussed in more details here:
>> https://mail.openjdk.java.net/pipermail/panama-dev/2021-November/015852.html
>>
>> The fix consists in adding a *max alignment* mask to all segments. Native segments have mask set at 0, whereas heap segment have the mask set at the element size of their backing array (e.g. 1 for `byte[]`, 2 for `short[]`/`char[]` and so forth).
>>
>> The alignment check is the computed as follows:
>>
>>
>> if (((address | maxAlignMask) & alignmentMask) != 0) {
>> throw ...
>> }
>>
>>
>> Where `address` is the long address being dereferenced, `maxAlignMask` is the max alignment associated with the segment being dereferenced, and `alignmentMask` is the alignment associated with the layout used during a dereference operation.
>> This patch also applies alignment checks to bulk copy operations as well, and it consolidates the way in which alignment checks are carried out (by adding a couple of helper methods, in `Utils` and `AbstractMemorySegment`, respectively).
>>
>> This patch also changes alignment errors to uniformly throw IllegalArgumentException, in all cases.
>>
>> Finally, this patch clarifies that during bulk copy and spliterator calls, the element layout alignment cannot be bigger than its size (as the size is used as a stride).
>>
>> A new test `TestHeapAlignment` has been added to check that alignment of heap segment is correctly enforced.
>>
>> Performance-wise, this patch doesn't change anything; that's because the value layout constants in `ValueLayout` are all unaligned. For code using aligned layout, some slight regression when using heap segment is possible (whereas native segment should have same performance as before). Eventually, these performance issues should disappear once [1] is fixed.
>>
>> [1] - https://bugs.openjdk.java.net/browse/JDK-8277850
>
> Maurizio Cimadamore has updated the pull request incrementally with one additional commit since the last revision:
>
> Add benchmarks for aligned access
src/jdk.incubator.foreign/share/classes/jdk/internal/foreign/AbstractMemorySegmentImpl.java line 127:
> 125: throw new IllegalArgumentException("Element layout size cannot be zero");
> 126: }
> 127: if (elementLayout.byteAlignment() > elementLayout.byteSize()) {
I wonder if we should add a similar check on the various `MemorySegment/MemoryAddress::[get/set]atIndex` ? I think accessing hyper-aligned elements is always going to fail, for certain offsets - in fact when deriving a var handle from a layout path, it is never possible to have an array var handle for an hyper-aligned element layout.
-------------
PR: https://git.openjdk.java.net/panama-foreign/pull/622
More information about the panama-dev
mailing list