1344503 - fitzgen.github.io/bugzilla-todos/ results in hundreds of thousands of copies of the same string

Status: RESOLVED INCOMPLETE

(Core :: JavaScript Engine, defect, P3)

Description

[Jonathan Watt [:jwatt]]

Attachment: Snapshot from about:memory

After having Firefox open for several days, about:memory reported that the window using the most memory was https://fitzgen.github.io/bugzilla-todos/ .  I first clicked "Minimize memory usage" then "Measure" with "verbose" checked. What seems strange is that about:memory reports that there are hundreds of thousands of copies of the same strings. I'm not sure if the JS engine is expected to optimize these away, or if this is simply the fault of the Web site.

Comment 1

[Ben Kelly [:bkelly, not reviewing]]

My understanding is we de-dupe small strings (2 chars and less), but not all strings.

Comment 2

[Jan de Mooij [:jandem]]

(In reply to Jonathan Watt [:jwatt] from comment #0)
> I'm not sure if the JS
> engine is expected to optimize these away, or if this is simply the fault of
> the Web site.

It depends on how the website is creating these strings. Atoms are strings that have the invariant there's a single copy for the entire runtime, but there are lots of ways to create duplicate non-atomized strings. Sometimes we can avoid that but deduplicating all strings we allocate would add way too much overhead.

So we need to know what the website is doing and then we can figure out if there's a way to avoid the copies.

Comment 3

[Jonathan Watt [:jwatt]]

(In reply to Ben Kelly [not reviewing due to deadline][:bkelly] from comment #1)
> My understanding is we de-dupe small strings (2 chars and less), but not all
> strings.

From the attached about:memory snapshot, there is at least one length=2 string ("%d", with 61001 copies):

├──31,907,504 B (01.41%) -- strings
│  ├───3,677,216 B (00.16%) ── string(length=7, copies=153214, "getData")/gc-heap/latin1
│  ├───3,677,184 B (00.16%) ── string(length=5, copies=153216, "#TEXT")/gc-heap/two-byte
│  ├───3,677,136 B (00.16%) ── string(length=4, copies=153214, "span")/gc-heap/two-byte
│  ├───3,677,136 B (00.16%) ── string(length=8, copies=153214, "getData!")/gc-heap/latin1
│  ├───1,851,360 B (00.08%) -- string(length=4, copies=77140, "SPAN")/gc-heap
│  ├───1,838,592 B (00.08%) ── string(length=13, copies=76608, "[object Text]")/gc-heap/latin1
│  ├───1,838,592 B (00.08%) ── string(length=5, copies=76608, "#text")/gc-heap/two-byte
│  ├───1,838,592 B (00.08%) ── string(length=6, copies=76608, "SCRIPT")/gc-heap/latin1
│  ├───1,464,024 B (00.06%) ── string(length=2, copies=61001, "%d")/gc-heap/latin1

Comment 4

[Jon Coppeard (:jonco)]

Maybe we should de-duplicate strings when we do a shrinking GC.

Comment 5

[Matthew Gaudet (he/him) [:mgaudet]]

We do some deduplication these days.