1245213 - Servo docs hang for >20s in buildIndex() when Gecko profiler is enabled

Status: RESOLVED WORKSFORME

(Core :: JavaScript Engine, defect)

Description

[Bobby Holley (:bholley)]

STR:
(1) Install the gecko profiler addon in Nightly: https://github.com/bgirard/Gecko-Profiler-Addon/blob/master/geckoprofiler-signed.xpi?raw=true
(2) Make sure the icon is enabled (not grey)
(3) Visit http://doc.servo.org/string_cache/macro.atom!.html
(4) Very long hang

Comment 1

[Bobby Holley (:bholley)]

Looks like we spend forever in buildIndex - the profiler may be screwing up the JIT somehow. Jan said he'd take a look.

Comment 2

[Jan de Mooij [:jandem]]

(In reply to Bobby Holley (busy) from comment #1)
> Looks like we spend forever in buildIndex - the profiler may be screwing up
> the JIT somehow. Jan said he'd take a look.

Hm this also happens with the JITs disabled. According to a profile, we spend an insane amount of time under PCToLineNumber (called by the sampler thread).

I'll do some digging to see if this is a large script or something.

Comment 3

[Jan de Mooij [:jandem]]

(In reply to Jan de Mooij [:jandem] from comment #2)
> I'll do some digging to see if this is a large script or something.

Bingo. We're loading this huge, 15.21 (!!) MB file: http://doc.servo.org/search-index.js

And PCToLineNumber is choking on that...

Comment 4

[Jan de Mooij [:jandem]]

Filed a rustdoc bug: https://github.com/rust-lang/rust/issues/31387

FWIW, PCToLineNumber being slow is nothing new (bug 929950).

Comment 5

[Jan de Mooij [:jandem]]

Bug 929950 is on my TODO list now, unfortunately I have some other things to do first.

Comment 6

[Bobby Holley (:bholley)]

(In reply to Jan de Mooij [:jandem] from comment #5)
> Bug 929950 is on my TODO list now, unfortunately I have some other things to
> do first.

Awesome - thanks for digging in Jan!

Comment 7

[Simon Sapin (:SimonSapin)]

The file has very few lines for its size:

     $ wc search-index.js 
     167   316606 15954940 search-index.js

Splitting lines some more help? Or is the time spent just proportional to the total file size?

Comment 8

[Jan de Mooij [:jandem]]

(In reply to Simon Sapin (:SimonSapin) from comment #7)
> Splitting lines some more help? Or is the time spent just proportional to
> the total file size?

Yes it's the total file size. We emit source notes for line and column numbers, and we walk all source notes from the start until we reach our bytecode pc. Splitting lines will result in more source notes, so it may actually make us slower (probably not noticeable though).

Comment 11

[Xidorn Quan [:xidorn] UTC+10]

Although it's not badly slow when profiler is disabled, it still feels slower than ideal in cases like the page of Iterator trait.