Performance regression in v12 caused by primordials

[joyeecheung]

I have heard about performance regressions caused by transition to primordials in v12, but couldn't find a tracking issue for the regression, so opening this to make sure we are tracking it and can work with the upstream to get this handled, on way or another (feel free to close this if there is already an issue opened here).

The regressions seem to come from two types of code patterns:

1. Calling functions through Function.prototype.{call, apply} instead of just calling them directly. There is an issue opened in the upstream by @bmeck
   https://bugs.chromium.org/p/v8/issues/detail?id=9702
2. Looking up properties from fronzen objects - objects from our primordials namespace are frozen so lookups like const { Reflect } = primordials; Reflect.apply(...); is slower than just Reflect.apply when Reflect comes from the global object. This can be mitigated by caching the lookup results upfront, e.g. event: improve performance of EventEmitter.emit #29633 by @mcollina There is also a fairly odd tracking issue for this in the upstream: https://bugs.chromium.org/p/v8/issues/detail?id=6831

cc @MylesBorins (https://twitter.com/MylesBorins/status/1173390304742785024)

[mcollina]

From my tests, it seems to be related to inlining.

[joyeecheung]

I just noticed that things like primordials.Reflect are not frozen, but instead the slowdown probably comes from Object.create(null) which turns the objects into dictionary mode.

[devsnek]

Can we have V8 migrate everything captured in the snapshot to fast properties?

_{And if all else fails... try { eval('%ToFastProperties(primordials.Reflect)') } catch {} :)}

[joyeecheung]

cc @nodejs/v8

[joyeecheung]

For reference the primordials are baked in https://github.com/nodejs/node/blob/6ce87c0/lib/internal/per_context/primordials.js

[joyeecheung]

One idea about the way forward: put the primordials beind a configure-time flag, and figure out how to fix the performance regression before turning it back on.

[joyeecheung]

@nodejs/process ^ any opinions on the idea in #29766 (comment) ?

[addaleax]

How much would we gain by not freezing the objects, for now? That could be put behind a flag pretty easily, right?

[joyeecheung]

@addaleax I think we need to identify a suitable benchmark to compare the impact of removing certain bits in the primorials.js...any suggestions? (from the tweet I think @mcollina @MylesBorins @bmeck have experience on this)

[mcollina]

I just did it manually, building two node versions and run some of our microbenchmarks.

[mcollina]

The way how I fixed it is by just doing a simple assignment / destructuring:

node/lib/events.js

Line 24 in ed5eaa0

const { Math, Object, Reflect } = primordials;

I would just do that everywhere, and we would be good I think.

[joyeecheung]

@mcollina Thanks, that would probably be good for a code & learn task or a beginner task. I'll see if I can get this done through nodejs/code-and-learn#97 and if not, I'll spin off a separate issue about this specific task.