Remove -Z time-passes and error with "-Z self-profile should be used instead".

[eddyb]

Remaining issues with -Z self-profile:

• at least as much granularity as -Z time-passes had pre-queries
• trigger on unsuccessful compilation / ICE / Ctrl+C, as well
• real-time progress indicator (Support a TUI "live view" version of -Z self-profile. #53630 perhaps?)

cc @steveklabnik @wesleywiser @alexcrichton @rust-lang/compiler

[nikomatsakis]

I'm in favor!

[Zoxc]

I'm against. I use -Z time-passes all the time when doing any performance work.

[bjorn3]

-Z self-profile provides a much less detailed output

[kennytm]

I'm mildly against if the replacement can't print some compilation progress at real time.

[eddyb]

@bjorn3 How so, does it not have a per-query breakdown? I'll shamefully admit I opened this issue away from my laptop when I remembered -Z self-profile landed, and I haven't played with it yet.

Either way, we should track the remaining problems with -Z self-profile in this issue, and remove -Z time-passes when they're resolved.

@Zoxc also mentioned on IRC that it doesn't work when there are errors.

@kennytm Ah, great point, I had forgotten. Would #53630 solve the "real time" problem for you?
I'd add that maybe we should catch Ctrl+C and also print the profile summary then too.

EDIT: I've added these to the issue description.

[kennytm]

@eddyb works for me 😃

[wesleywiser]

@eddyb We don't have per-query breakdown yet, just broad level categorization. Per-query is the next thing I plan on working on.

[nagisa]

I want some way to output various "events" pertaining to the query system (query X start, query X end, query X executes query Y, etc) in whatever format, as they occur. Preferably the format would target some tooling that already exists, but I cannot really advise on anything in particular at the moment. See this comment for examples of what I’m talking about.

This could possibly be the solution for the granularity problem. I can say from experience, that tracing such events is definitely not as expensive as it may sound and usually introduces overhead in proximity of tens to hundreds of ns per event.

[eddyb]

@nagisa I think we already have something similar because of dep graph tracking, but @michaelwoerister knows more than me here (I remember seeing event-collecting threads).

[nagisa]

Looking into webrender I noticed it uses thread_profiler, which outputs traces in Trace Event Format. Links lead to catapult, which upon the first blush seems similar to the tools I referenced before.

[nikomatsakis]

That is roughly what the old stuff did, iirc. I think we mostly ripped that out. I like the idea emitting events, at least in principle. I also like the idea of per-query profiling. I also like the idea of removing time-passes =)

It seems like we can fix things to emit a result in the case of error without great difficulty? Mostly a matter of using a Drop impl or something, no?

[nikomatsakis]

Ah I see @eddyb already made a check-list, 👍 from me. I'd be ok with removing -Ztime-passes earlier too but I don't mind waiting.

[FrankSD]

Anyone is working on this?

[wesleywiser]

Hi @FrankSD! The self-profile feature has a fair bit of churn going on right now. Before you decide to work on this, I'd recommend stopping by the t-compiler/wg-self-profile channel on Zulip.

[jyn514]

I don't think we should remove time-passes until self-profile measures memory usage: #81348

[john-h-k]

Is this still a desired thing?

[jyn514]

I would rather not make this change. -Z time-passes is nice because you can get feedback while compilation is still ongoing, without having to wait for it to finish. It's less accurate but it's also a lot nicer to use.

[fee1-dead]

I'm going to close this issue based on the comments above since it appears that no further input is likely.