`Vector2.clone()` performance hit for x86-64 and ARM64? [MOVED]

[JosefWN]

Hello! I'm experiencing a pretty heavy performance hit with Vector2.clone() on ARM64 (M3) and x86-64 in Dart 3.3.3 on macOS/Windows. This is a continuation of google/vector_math.dart#319, which was misplaced.

Example benchmark:

class Vector2CloneBenchmark extends BenchmarkBase {
  const Vector2CloneBenchmark() : super('CloneVector2');

  @override
  void run() {
    for (double i = -500; i <= 500; i += 0.75) {
      for (double j = -500; j <= 500; j += 0.75) {
        final _ = Vector2(j, i).clone();
      }
    }
  }
}

AOT:

CloneVector2(RunTime): 135596.2 us.

For comparison, the same benchmark loop with a custom PointDouble(i, j) + PointDouble(j, i) completes in 9000 us. Only creating and cloning a Vector2 is 15-20x slower than creating two PointDouble and adding them to form a third.

It could be completely unrelated, but the other day I observed another heavy performance hit for the double implementation of i.remainder(j) which seems to occur across different platforms: #55479.

[mraleph]

AOT compiler is rather conservative with its inlining heuristics (for code size reasons) and does not always correctly inline even profitable methods.

I would recommend that package:vector_math plasters @pragma('vm:prefer-inline') all over its APIs as a workaround.

I don't know if we can fix something short term here. /cc @alexmarkov

[mraleph]

/cc @johnmccutchan for visibility

[johnmccutchan]

Thanks for the @pragma('vm:prefer-inline') tip @mraleph

[a-siva]

Hopefully @pragma('vm:prefer-inline') helps to certain extent.