[Performance regression] Performance regression in Typescript 5.1 on large files

[moufmouf]

🔎 Search Terms

performance regression, Typescript 5.1

🕗 Version & Regression Information

• This changed between versions 5.0.4 and 5.1
• This changed in commit 511921e (detected via every-ts)
• This is the behavior in every version I tried after 5.1, including the latest 5.6 version

⏯ Playground Link

https://github.com/moufmouf/typescript-perf-issue

💻 Code

See Github repo for the code sample (the issue is that very big files take wayyyyy longer to be analyzed)

🙁 Actual behavior

The issue is that the compilation time of some very large files (22k lines) has increased significantly between v5.0 and v5.1 (and is still very long in the latest 5.6 release)

The file I'm trying to compile is 22k lines long.

Compilation time in v5.0: 6s.
Compilation time in v5.1: 110s.

This is a 18x increase.

I generated trace files but I don't know how to interpret them.

Tracefile for 5.0: traceDir_5.0.log
Tracefile for 5.1: traceDir_5.1.log

🙂 Expected behavior

The analysis time would not greatly differ between 5.0 and recent releases.

Additional information about the issue

The file that is suddenly taking very long to analyze was generated using ts-proto.

I could perfectly disable the analysis for this file (since it was autogenerated) but there is probably an underlying issue.

For the record, the original file comes from the WorkAdventure project (https://github.com/workadventure/workadventure)

[s3xysteak]

After updating to [email protected], the Typescript IntelliSense is so slow😥. I'm not sure is that related to this issue.

[jakebailey]

@s3xysteak TypeScript 5.1 was ages ago. If things got slower when updating to this month's VS Code version, your issue is definitely not this issue.

[RyanCavanaugh]

Bisects to #53435

[RyanCavanaugh]

We're spending huge amounts of time analyzing functions like this one https://github.com/moufmouf/typescript-perf-issue/blob/master/src/messages.ts#L6122 which have extremely deep control flow graphs with possible narrowing effects due to the conditional assignments.

If this is generated code, it's a huge waste of time to be typechecking these bodies (which would almost certainly be written differently by a human, i.e. in a way that is more amenable to typechecking). It should really be using any in all function bodies, since any is effectively free from a perf perspective. I would also question why it's not using switch, e.g. here it would be much shorter and also efficient both in terms of runtime and typechecking perf

    if (message.message?.$case === "batchMessage") {
      obj.batchMessage = BatchMessage.toJSON(message.message.batchMessage);
    }
    if (message.message?.$case === "errorMessage") {
      obj.errorMessage = ErrorMessage.toJSON(message.message.errorMessage);
    }
    if (message.message?.$case === "roomJoinedMessage") {
      obj.roomJoinedMessage = RoomJoinedMessage.toJSON(message.message.roomJoinedMessage);
    }
    if (message.message?.$case === "webRtcStartMessage") {
      obj.webRtcStartMessage = WebRtcStartMessage.toJSON(message.message.webRtcStartMessage);
    }
    if (message.message?.$case === "webRtcSignalToClientMessage") {
      obj.webRtcSignalToClientMessage = WebRtcSignalToClientMessage.toJSON(message.message.webRtcSignalToClientMessage);
    }

It's possible there's a fix here but it's always going to be possible (but unlikely for a human) to write over-deep CFA graphs like this.

I'm open to a PR on this but overall I would say this is taking an expectable amount of time for the amount of CFA graph complexity the code is presenting; the 5.0 behavior was fast because it was also wrong in many cases.

[typescript-bot]

This issue has been marked as "Not a Defect" and has seen no recent activity. It has been automatically closed for house-keeping purposes.