This tracker is for issues related to:

• Dart VM

Description

I was doing some performance testing on iOS and noticed that 7% of CPU time was spent in GetCurrentThreadCPUMicrosForTimeline. It is the second hottest symbol I'm executing beyond memcpy (I'm copying thousands of buffers that are a megabyte big for microbenchmarks). Compared to the other symbols, some that are also doing system calls, it's shocking how much CPU time it is taking.

Steps to reproduce

1. Check out https://github.com/flutter/flutter/tree/master/dev/benchmarks/platform_channels_benchmarks
2. Make the following code change so channels use compute (which will call Dart_TimelineEvent which calls GetCurrentThreadCPUMicrosForTimeline)\

--- a/packages/flutter/lib/src/services/platform_channel.dart
+++ b/packages/flutter/lib/src/services/platform_channel.dart
@@ -57,7 +57,10 @@ class BasicMessageChannel<T> {
   /// Returns a [Future] which completes to the received response, which may
   /// be null.
   Future<T?> send(T message) async {
-    return codec.decodeMessage(await binaryMessenger.send(name, codec.encodeMessage(message)));
+    final ByteData? encodedResult =
+      await binaryMessenger.send(name, codec.encodeMessage(message));
+    return compute<ByteData?, T?>(codec.decodeMessage, encodedResult);    
+//    return codec.decodeMessage(await binaryMessenger.send(name, codec.encodeMessage(message)));
   }

3. Run the platform channels benchmarks with flutter run --profile with instruments running a Time Profiler probe.

Profile

All of the time is in the thread_info call.

Version

Dart version 2.16.0 (build 2.16.0-56.0.dev)

Evaluation

Can't we just use the system clock for Dart_TimelineEvent, why does it have to be the clock from thread_info?