Severe performance penalty found in VSCode rendering loop

https://github.com/microsoft/vscode/issues/272155

by anticensor • 9 months ago

63 45 9 months ago

github.com

nawgz2• 9 months ago

A bit sloppy but easily resolved - surprised it took so long to notice, or maybe it was new?

minitech2• 9 months ago

It’s been around since the root commit in 2015: https://github.com/microsoft/vscode/blob/8f35cc4768393b25468...

anticensor4• 9 months ago

Yeah, it was a bit surprising to me as well.

muglug2• 9 months ago

Given that the issue already gives a before-and-after metric it's extremely odd there's no POC PR attached.

This just seems like an AI slop GitHub issue from beginning to end.

And I'd be very surprised if VS Code performance could be boosted that much by a supposedly trivial fix.

duskwuff2• 9 months ago

Even if it is a real performance issue, the reasonable fix would be to move the sort call out of the loop - implementing a new data structure in JS is absolutely not the way to fix this.

muglug2• 9 months ago

Right, and also this would show up in the profiler if it were a time sink — and I'm 100% certain this code has been profiled in the 10 years it's been in the codebase.

oe• 9 months ago

Adding a new data structure just for this feels like such an AI thing. I've added to our agents.md a rule to prefer using existing libraries and types, otherwise Gemini will just happily generate things like this.

nneonneo• 9 months ago

There’s clearly functionality to push more work to the current window’s queue, so I would not be surprised if the data structure needs to be continually kept sorted.

(Somewhere in the pile of VSCode dependencies you’d think there’d be a generic heap data structure though)

gigatexal• 9 months ago

I’ve already moved from VSCode to Zed. It’s native. Faster. Has most of the functionality I had before. I’m a huge fan.

sillythrowawy9• 9 months ago

OP’s account also seems automated. This certainly feel like automated post to social media for PR clout

anticensor4• 9 months ago

Not really, I read HN more than I post to it, but I found this one interesting.

ollin• 9 months ago

Yeah the issue reads as if someone asked Claude Code "find the most serious performance issue in the VSCode rendering loop" and then copied the response directly into GitHub (without profiling or testing anything).

a-dub• 9 months ago

i see emojis in the comments.

also no discussion of measured runtimes for the rendering code. (if it saves ~1.3ms that sounds cool, but how many ms is that from going over the supposed 16ms budget.)

hyperhello2• 9 months ago

Real-world impact: With 50+ view parts (text, cursors, minimap, scrollbar, widgets, decorations, etc.), this wastes 1-2ms per frame

Good thing to find...

blharr2• 9 months ago

How does it possibly take 1-2ms to sort... 50 items? I'd expect that to happen in an order of microseconds

klodolph• 9 months ago

It’s being sorted not once per frame, but once per item.

If you have 50 items in the list, then the list gets sorted 50 times. If you have 200 items in the list, the list is sorted 200 times.

This is unnecessary. The obvious alternative is a binary heap… which is what the fix does. Although it would also be obvious to reuse an existing binary heap implementation, rather than inventing your own.

hdjfjkremmr• 9 months ago

it's sorting 50 times a list going from 50 to 0 items.

retsibsi• 9 months ago

The issue seems to be a direct copy-paste from an LLM response, so I suspect "this wastes 1-2ms per frame" is estimated/made up.

webprofusion• 9 months ago

Looks interesting.

I see they also contributed a fix to the OnlyFans notification robot. Clearly doing the important work that the internet needs.

brokencode• 9 months ago

This is what I want to do when I retire. Maybe not OnlyFans fixes specifically, but just go around fixing random stuff.

Like if Batman turned out to be bad at fighting criminals so had to fight null pointer exceptions instead.

gnarlouse• 9 months ago

bugman

OrderlyTiamat• 9 months ago

"Fear not the bugs citizen! For in my utility belt, I have REGEX and VIM!"

nurettin• 9 months ago

Maybe hack into facilities, optimize their scripts and deployment, then leave without a trace confusing the IT department.

anticensor4• 9 months ago

That notification robot codebase is actually generic, Zara Darcy just used OnlyFans branding to boost her follower base.

adwn2• 9 months ago

I'm confused: Does top.execute() modify currentQueue in some way, like pushing new elements to it? If it doesn't, then why not simply move the sort out of the loop? This is simpler and faster than maintaining a binary heap.

adwn2• 9 months ago

One more thing: Nowadays sort() functions ary usually heavily optimized and recognize already sorted subsequences. If currentQueue isn't modified during the loop, then the sort() call should run in O(n) after the first iteration, instead of O(n * log n). Still worse than not having it inside the loop at all, of course.

nateb20222• 9 months ago

If it doesn't, then why not simply move the sort out of the loop?

Yup, they should definitely move the sort outside of the loop. Shifting is O(N) so overall complexity would be O(N^2) but they could avoid shifting by reverse-sorting outside the loop and then iterating backwards using pop()

jgoldshlag• 9 months ago

This seems like a nonsense issue. Sorting 50 things takes 1-2 ms? What sort of potato was that timed on.

tylerhou• 9 months ago

No, sorting 50/2ish things 50 times allegedly takes 1-2ms. Which is only slightly more believable.

jojobas• 9 months ago

The penalty is called "Electron".

geokon• 9 months ago

I feel with Valgrind (in C++land) or VisualVM (JVMland) stuff like this is very easy to zero in on.

I don't work in JS-land.. but are Electron apps difficult to do performance profiling on?

nateb20222• 9 months ago

https://www.electronjs.org/docs/latest/tutorial/performance

nawgz2• 9 months ago

No. Browser dev tools are available, and make it pretty easy to do performance profiling, and get a flamegraph etc..

Just seems like the reality of things is that the number of extensions or widgets or whatever has remained low enough that this extra sorting isn't actually that punitive in most real-world use cases. As a long-time developer working mainly in VSCode, I notice no difference between performance/snappiness in VSCode compared to JetBrains Rider, which is the main other IDE I have meaningful experience with these days.

rockorager• 9 months ago

If you work with LLM agents, you will immediately be able to tell this issue is written by one. The time cost of this sort is almost certainly not real, as others have pointed out.

I’ve had agents find similar “performance bottlenecks” that are indeed BS.

znpy• 9 months ago

as one of the commenters to the issue wrote, arguing whether the text is ai-generated or not is essentially useless.

the important question is: is this an actual performance bug?

cgriswald• 9 months ago

The question is whether credence, and therefore time, should be given to the claim that the performance bug exists.

ec1096852• 9 months ago

I hate ai sometimes — an AI generated pull request (really some rando found a way of shaving 12% off the run loop?) responded to by an ai comment bot:

This feature request is now a candidate for our backlog. The community has 60 days to upvote the issue. If it receives 20 upvotes we will move it to our backlog. If not, we will close it. To learn more about how we handle feature requests, please see our documentation.

dkdcio• 9 months ago

reasonably confident that’s just an automated response bot, not AI…

also it’s an issue, not a PR

ec1096852• 9 months ago

Even worse (issue part).

flowerthoughts• 9 months ago

I would have expected V8 sort() to be optimized for runs of presorted input, like other implementations nowadays. So O(n²) seems more likely than O(n² log n). Not that it matters much.

But then again, probably AI slop with "performance gain" numbers taken out of thin air. Who knows if the number 50 and 1-2ms are based on fantasy novels or not.

Like when I used Claude to build a door video intercom sytem, and first asked it to create a plan. It inserted how many weeks each milestone would take, and it was an order of magnitude off. But I guess milestone documents have time estimates, so that's how it's supposed to look, information accuracy be damned.

Severe performance penalty found in VSCode rendering loop

Related Stories