Substitute all T::Struct usage for plain Objects

[vinistock]

Motivation

I wanted to get a more complete understanding of the LSP's performance on a long running session, so I added the following to server.rb to generate profile data.

def start
  # ...
  when "initialized"
    StackProf.start(mode: :cpu, raw: true)

  when "shutdown"
    StackProf.stop
    data = StackProf.results
    File.write("profile.json", JSON.pretty_generate(data))
  # ...
end

I edited and saved files for 2 minutes to triggered as many requests as possible. To my surprise, of the 120 seconds we spent 30 of them just initializing T::Struct objects (specifically, running the construct with defaults thing and performing runtime typechecks with is_a?).

Other than the desire to have very compact class definitions, we really have no other requirement for using T::Struct and it doesn't seem like a good trade off for the performance penalty.

Using benchmark IPS to check the difference on a single request, not counting objects created for queueing, there's a 7% performance difference.

Warming up --------------------------------------
                 new    15.350k i/100ms
Calculating -------------------------------------
                 new    154.727k (± 1.1%) i/s -    782.850k in   5.060214s

Comparison:
                 new:   154727.1 i/s
                 old:   144077.7 i/s - 1.07x  (± 0.00) slower

This also benefits YJIT, which in my local machine goes from 1.24 -> 1.32 interp / yjit ratio.

Implementation

Switched all T::Struct to regular classes.

[paracycle]

Can we figure out how to improve T::Struct as well? Maybe we should always be shipping our own simplified version of T::Struct at runtime, so that we never have to do this work, nor worry about performance.

[Morriar]

It seems there are some improvements we could be shipping Sorbet's upstream either on T::Struct itself or its runtime configuration.

[vinistock]

Indeed, there might be an opportunity. I think @bitwise-aiden has done some exploration on the performance of instantiating T::Struct.

[bitwise-aiden]

Can we figure out how to improve T::Struct as well? Maybe we should always be shipping our own simplified version of T::Struct at runtime, so that we never have to do this work, nor worry about performance.

+1 to this. Off the back of the investigation into the constructor, I think that it might be worth replacing T::Struct for PORO at runtime. I will have a bit of a look into that this week before going on break.