Closed gahaas closed 3 months ago
@gahaas we have been benchmarking alternatives to v8::FastApiTypedArray
and they are all significantly slower.¨
Refs: https://github.com/nodejs/node/pull/54103 Refs: https://github.com/nodejs/node/pull/54087
@ronag What exactly did you measure? If I remember correctly, the performance regression of removing v8::FastApiTypedArray
was 6 CPU cycles per call. In a micro-benchmark that calls an empty API function, this regression was indeed clearly measurable, but as soon as I added reasonable workload, I could not measure a difference anymore.
memmove 64 bytes.
Pleas see the references
Thank you for making me aware of this issue. Could you please file a bug in the V8 bug tracker.
Please add information on how I can reproduce the issue locally.
Does node compile with PGO and LTO? I could imagine that the issue is that some API function calls get inlined in Chrome but not in node, as that would explain the different performance behavior.
Does node compile with PGO and LTO? I could imagine that the issue is that some API function calls get inlined in Chrome but not in node, as that would explain the different performance behavior.
@targos Do you know someone that can answer this?
In the examples where
v8::FastApiTypedArray
was used, there is actually no advantage of using V8's fast API calls instead of regular API calls. Therefore this CL just removes the fast API call targets instead of adjusting them.