RFR: 8287788: reuse intermediate segments allocated during FFM stub invocations [v3]

[Jorn Vernee]

On Mon, 20 Jan 2025 16:20:20 GMT, Matthias Ernst <duke at openjdk.org> wrote:

>> Certain signatures for foreign function calls (e.g. HVA return by value) require allocation of an intermediate buffer to adapt the FFM's to the native stub's calling convention. In the current implementation, this buffer is malloced and freed on every FFM invocation, a non-negligible overhead.
>> 
>> Sample stack trace:
>> 
>>    java.lang.Thread.State: RUNNABLE
>> 	at jdk.internal.misc.Unsafe.allocateMemory0(java.base at 25-ea/Native Method)
>> ...
>> 	at jdk.internal.foreign.abi.SharedUtils.newBoundedArena(java.base at 25-ea/SharedUtils.java:386)
>> 	at jdk.internal.foreign.abi.DowncallStub/0x000001f001084c00.invoke(java.base at 25-ea/Unknown Source)
>> ...
>> 	at java.lang.invoke.Invokers$Holder.invokeExact_MT(java.base at 25-ea/Invokers$Holder)
>> 
>> 
>> To alleviate this, this PR remembers and reuses up to two small intermediate buffers per carrier-thread in subsequent calls.
>> 
>> Performance (MBA M3):
>> 
>> 
>> master at 764d70b7df18e288582e616c62b0d7078f1ff3aa
>> Benchmark                         Mode  Cnt    Score    Error  Units
>> PointsAlloc.circle_by_ptr         avgt   30    9.197 ?  0.037  ns/op
>> PointsAlloc.circle_by_value       avgt   30   42.195 ?  0.088  ns/op   <= #######
>> PointsAlloc.jni_ByteBuffer_alloc  avgt   30  226.127 ? 35.378  ns/op
>> PointsAlloc.jni_long_alloc        avgt   30   25.297 ?  2.457  ns/op
>> PointsAlloc.panama_alloc          avgt   30   27.053 ?  1.915  ns/op
>> 
>> After:
>> Benchmark                         Mode  Cnt    Score    Error  Units
>> PointsAlloc.circle_by_ptr         avgt   30    9.156 ?  0.021  ns/op
>> PointsAlloc.circle_by_value       avgt   30   11.995 ?  0.051  ns/op   <= #######
>> PointsAlloc.jni_ByteBuffer_alloc  avgt   30  211.161 ? 23.284  ns/op
>> PointsAlloc.jni_long_alloc        avgt   30   24.885 ?  2.461  ns/op
>> PointsAlloc.panama_alloc          avgt   30   26.905 ?  1.935  ns/op
>> 
>> 
>> `-prof gc` also shows that the new call path is fully scalar-replaced vs 160 byte/call before.
>
> Matthias Ernst has updated the pull request incrementally with three additional commits since the last revision:
> 
>  - shift api boundary
>  - move bench
>  - revert formatting

src/java.base/share/classes/jdk/internal/foreign/abi/CallBufferCache.java line 41:

> 39:         // Storing addresses, not MemorySegments turns out to be slightly faster (write barrier?).
> 40:         private long address1;
> 41:         private long address2;

Can we store just `MemorySegment`s here? It would avoid having to call `ofAddress` in `acquireOrAllocate`, and would slightly reduce the footprint when using compressed oops.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/23142#discussion_r1922704992