kaivalnp commented on PR #15272:
URL: https://github.com/apache/lucene/pull/15272#issuecomment-3358029323
On printing some JVM inlining internals, I came across entries like:
```
@ 35
org.apache.lucene.internal.vectorization.PanamaVectorUtilSupport$ArrayLoader::length
(6 bytes) inline (hot) callee changed to
org.apache.lucene.internal.vectorization.PanamaVectorUtilSupport$MemorySegmentLoader::length
(13 bytes) inline (hot) callee changed to
org.apache.lucene.internal.vectorization.PanamaVectorUtilSupport::dotProductBody
(250 bytes) \-> TypeProfile (1028/43516 counts) =
org/apache/lucene/internal/vectorization/PanamaVectorUtilSupport$MemorySegmentLoader
callee changed to
org.apache.lucene.internal.vectorization.PanamaVectorUtilSupport::dotProductBody
(250 bytes) \-> TypeProfile (42488/43516 counts) =
org/apache/lucene/internal/vectorization/PanamaVectorUtilSupport$ArrayLoader
```
..so I turned off JVM method inlining for those [internal
classes](https://github.com/apache/lucene/blob/cd88a97a63a97dc857d692ec3eb73b229bb5f372/lucene/core/src/java25/org/apache/lucene/internal/vectorization/PanamaVectorUtilSupport.java#L323-L359)
using some compiler directives (so that their callers are compiled separately,
into more type-appropriate methods)
Here are the results:
```
Benchmark (size)
Mode Cnt Score Error Units
VectorScorerBenchmark.binaryDotProductMemSeg 1024
thrpt 15 6.374 ± 0.026 ops/us
VectorScorerBenchmark.binaryDotProductMemSegWithVectorDirectives 1024
thrpt 15 7.401 ± 0.005 ops/us
```
Looks like we can regain most of the performance drop from call site
pollution!
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
