Bug 292147 – REGRESSION (291398@main): Slower String.prototype.split with result length < 100

• Status: RESOLVED
• Resolution: FIXED
• Component: JavaScriptCore

[Dylan Conway]

Created attachment 475061 [details]
out.js

The recent `Array.prototype.indexOf` (e1b9447) optimization seems to cause a performance regression in `String.prototype.split`. This only applies to result lengths less than 100 because the extra work involved for the optimization (WeakGCMap.ensureValue) happens when it is less than 100.

out.js is a bundled script that uses the `convert-csv-to-json` package to read a csv (37 cols, 130453 rows) into json. Comparing main and a build with `atomStringsArrayLimit` set to 20 to avoid the optimization, the results are:

```
> hyperfine -n "main" "../webkit-main/Tools/Scripts/run-jsc out.js" -n "atomStringArrayLimit=20" "./Tools/Scripts/run-jsc out.js" --warmup 5
Benchmark 1: main
  Time (mean ± σ):     571.2 ms ±   4.7 ms    [User: 511.0 ms, System: 78.2 ms]
  Range (min … max):   561.9 ms … 576.7 ms    10 runs

Benchmark 2: atomStringArrayLimit=20
  Time (mean ± σ):     258.7 ms ±   2.8 ms    [User: 219.7 ms, System: 57.2 ms]
  Range (min … max):   255.3 ms … 263.4 ms    11 runs

Summary
  atomStringArrayLimit=20 ran
    2.21 ± 0.03 times faster than main
```

another example, splitting the csv by line then comma:

```
// split-csv-37-cols.js
const csv = readFile("input-37-cols.csv");

csv.split("\n").forEach(line => {
    line.split(",");
});
```

results:

```
> hyperfine -n "main" "../webkit-main/Tools/Scripts/run-jsc split-csv-37-cols.js" -n "atomStringArrayLimit=20" "./Tools/Scripts/run-jsc split-csv-37-cols.js" --warmup 5
Benchmark 1: main
  Time (mean ± σ):     458.1 ms ±   5.8 ms    [User: 380.8 ms, System: 68.9 ms]
  Range (min … max):   451.5 ms … 468.0 ms    10 runs

Benchmark 2: atomStringArrayLimit=20
  Time (mean ± σ):     168.3 ms ±   2.2 ms    [User: 107.3 ms, System: 47.4 ms]
  Range (min … max):   166.7 ms … 176.2 ms    17 runs

Summary
  atomStringArrayLimit=20 ran
    2.72 ± 0.05 times faster than main
```

input csv: https://github.com/Schlumberger/hackathon/blob/master/backend/dataset/data-large.csv

Originally reported here https://github.com/oven-sh/bun/issues/17373

[Radar WebKit Bug Importer]

<rdar://problem/150668072>

[Vassili Bykov]

Pull request: https://github.com/WebKit/WebKit/pull/45069

[EWS]

Committed 294689@main (d84f22e0d2e4): <https://commits.webkit.org/294689@main>

Reviewed commits have been landed. Closing PR #45069 and removing active labels.