Quadratic behavior in kernel loading

[mkustermann]

The VM finalizes classes lazily when it needs them. When it decides to finalize a class it will call into the kernel loader to load all fields/functions/... of the (possibly top-level) class.

See runtime/vm/kernel_loader.cc:KernelLoader::FinishClassLoading:

void KernelLoader::FinishLoading(const Class& klass) {
  ...
  KernelLoader kernel_loader(...);
  ...

This constructor will execute the code in runtime/vm/kernel_loader.cc:KernelLoader::KernelLoader:

KernelLoader::KernelLoader(...) {
  ...
  const Array& scripts = Array::Handle(Z, kernel_program_info_.scripts());
  patch_classes_ = Array::New(scripts.Length(), Heap::kOld);
  ...
}

Now very often the entire application is represented as one KernelProgramInfo object.
The number of classes is always larger than the number of scripts / library since each script has it's own top-level class.

=> The complexity is O(#number-classes x #number-of-scripts)
=> For large number of libraries this can be very slow (finalization of each class can cause ~ 1MB allocation, causing very frequent GCs)

[mkustermann]

/cc @jensjoha Would you be interested in looking at this?

(We have internal customers with large apps which can make dev cycle really slow because of this)