[External] : Re: MemorySegment.ofAddress(...).reinterpret(...)

[Jorn Vernee]

The fix for: https://bugs.openjdk.org/browse/JDK-8311594 would help with 
single accesses as well. For copies to/from Java arrays, I also found 
another issue: https://bugs.openjdk.org/browse/JDK-8311597 fixing both 
of those should help. If this is a benchmark of a whole application, you 
might also be measuring some of the extra warmup needed for the Panama 
version.

The Panama version is doing slightly more work, after all. Especially 
the bound checks on Java arrays is not something we can just drop. So, I 
think a slight regression is to be expected.

Jorn

On 07/07/2023 00:19, Brian S O'Neill wrote:
> When I use the "ALL" MemorySegment instead of allocating 
> MemorySegments on the fly for copies, the performance regression drops 
> to ~2%. When I revert to using unsafe native field accesses (getInt, 
> etc) and unsafe memory copies, the performance is back to what it was 
> before (+/- 0.0019%).
>
> The unsafe field access alone gives me back a bit more performance 
> than the unsafe copy, but that's because field accesses are more 
> common. The test runs for about nine minutes and performs ~20 million 
> field accesses per second and ~12 million copies per second. Many of 
> the copies are less than 10 bytes in length, and so bounds checks 
> would add significant overhead.
>
> On 2023-07-06 11:45 AM, Brian S O'Neill wrote:
>> Thanks for looking into this. I'm already doing unaligned accesses, 
>> so there's not much I can do at this point on my end. I'll try to 
>> identify which operations are the most expensive and report back.
>>