Severe performance regression in Dart VM after switching to Dart 2.0

[mkustermann]

The example at the bottom of this issue (which was reduced from a real world example, without needing --no-background-compilation --enable-inlining-annotations) had orders of magnitude regression:

% out/ReleaseX64/dart --enable-inlining-annotations --no-background-compilation test.dart
Took 0:00:16.288135
% out/ReleaseX64/dart --no-preview-dart-2 --enable-inlining-annotations --no-background-compilation test.dart
Took 0:00:00.090875

This is a slowdown of 180 x!

main() {
  final closure = outer<int, int>();
  final sw = new Stopwatch()..start();
  for (int i = 0; i < 10000000; ++i) {
    fooClosure(closure);
  }
  print('Took ${sw.elapsed}');
}

final dynamic fooClosure = foo;

@NeverInline
foo(FI arg) { }

B Function(A) outer<A, B>() {
  B inner(A a) => null;
  return inner;
}

typedef FI = int Function(int);
const NeverInline = 'NeverInline';

The cause for this slowdown is that we have no fast cases for type tests where the instance tested against is a closure which has a direct or indirect parent function with type parameters.

This is most likely a problem for all users of package:stack_trace on the Dart VM - which is a real problem!

While looking at this I've also discovered a bug in the implementation of function subtyping implementation -- see #34051 .

I've started working on this now.

[mkustermann]

@isoos @sortie

This performance issue is a reproduction of the issue the pub site has on Dart 2 and was fixed in 575a8f8. So you could try to switch to a Dart 2 sdk again and do a load test. That being said, JIT in Dart 2.0 is still slower than Dart 1.