On Thu, 12 Sep 2024 11:30:12 GMT, Maurizio Cimadamore <[email protected]>
wrote:
>> Per Minborg has updated the pull request incrementally with two additional
>> commits since the last revision:
>>
>> - Fix yet another typo
>> - Fix typo
>
> src/java.base/share/classes/jdk/internal/foreign/SegmentBulkOperations.java
> line 204:
>
>> 202: // This gives about 20% performance increase for large values
>> of `length`.
>> 203: // On non-Aarch64 architectures, the unroll code will be
>> eliminated at compile time.
>> 204: if (Architecture.isAARCH64() && NATIVE_THRESHOLD_MISMATCH > 64)
>> {
>
> I'm a bit opposed to this - as it goes in the direction to add a lot of
> transient complexity when in reality the underlying issue is that aarch64
> mismatch intrinsics should be fixed. Tinkering with thresholds is borderline,
> but still acceptable - having different implementations one per platform
> starts to look "more wrong".
In other words, I don't think the goal of this (and related) PR is "improve
mismatch so that it blows other alternatives - like Unsafe, or array" out of
the water - as much as it is "make sure that using MemorySegment::mismatch is
competitive with other offerings".
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/20848#discussion_r1756678756
