Traverse performance

[JLHwung]

Q	A
Minor: Bug Fix?	👍
Major: Breaking Change?	No if one is using public API.
Tests Added + Pass?	Yes
License	MIT

In this PR we improve the performance of babel-traverse by revising the memory layout of NodePath. It results to 10% performance gain compared to the control branch on the following test cases.

const babel = require("/path/to/local/babel-core");
const code = require("fs").readFileSync("./fixtures/es5/babylon-dist.js");
babel.transform(code, { preset: ["@babel/env"] });

I can observe this branch outperforms control branch consistently.

┌─────────────────────┬──────────────────────────────┐
│ fixture             │ devTransform                 │
├─────────────────────┼──────────────────────────────┤
│ es5/babylon-dist.js │ 1.02 ops/sec ±31.84% (978ms) │
└─────────────────────┴──────────────────────────────┘
┌─────────────────────┬───────────────────────────────┐
│ fixture             │ baseTransform                 │
├─────────────────────┼───────────────────────────────┤
│ es5/babylon-dist.js │ 0.86 ops/sec ±20.52% (1167ms) │
└─────────────────────┴───────────────────────────────┘

And the memory usage improvements due to compressing traverse flags and adding accessor methods for inList and parentPath.

┌─────────────────────┬──────────────────┐
│ fixture             │ devTransform x5  │
├─────────────────────┼──────────────────┤
│ es5/babylon-dist.js │ heap: 239.37 MiB │
└─────────────────────┴──────────────────┘
┌─────────────────────┬──────────────────┐
│ fixture             │ baseTransform x5 │
├─────────────────────┼──────────────────┤
│ es5/babylon-dist.js │ heap: 305.47 MiB │
└─────────────────────┴──────────────────┘

The following changes are implemented:

• Removed NodePath.inList data property, a get accessor is provided for backward compatibility (partially).
• Removed NodePath.parentKey data property, a get accessor is provided for backward compatibility (partially).
• Removed NodePath.typeAnnotation data property, it is introduced in Initialize properties to avoid hidden class thrashing. #1752 but since then it is not used by any routines. I believe it is a mistype of Node.typeAnnotation.
• Compressed NodePath.shouldSkip/shouldStop/removed into _traverseFlags bit array, both get accessor and set accessor is provided for backward compatibility.
• Lazily assign this.skipKeys in setContext.
• Lazily initialize this.data in NodePath.constructor.

In practice inList and parentKey should be treated as read-only since these two properties are rewritten in setContext. The current behavior accepts one assigning these two properties.

The shouldSkip/shouldStop/removed traverse flags are handy and turbofan would inline these simple accessors we don't introduce performance overhead here. By compressing these states into bit array we optimize the performance of setContext.

[nicolo-ribaudo]

Regarding the breaking changes, I think that any plugin relying on the old behavior was actually broken: if someone wanted to change the parent node by doing path.parentPath = ... it wouldn't have worked.

[babel-bot]

Build successful! You can test your changes in the REPL here: https://babeljs.io/repl/build/11690/

[nicolo-ribaudo]

I wish https://github.com/tc39/proposal-logical-assignment wasn't just stage 1 😛

[JLHwung]

Offtopic: we still need a PR to add nullish coalescing and optional chaining.

[hzoo]

We can always use new stage features if you are willing to move off of them if they change (I think that one is small enough that it's fine) 😁

[JLHwung]

@nicolo-ribaudo I am sorry I make a typo before. It should be the parentKey instead of parentPath, but your point is still valid on parentKey.

I think that any plugin relying on the old behavior was actually broken: if someone wanted to change the parent node by doing path.parentKey = ... it wouldn't have worked.

I agree with you that setting inList and/or parentKey never works in practice. Instead of failing silently, babel will now report can not assign to an accessor. In this way it seems more like a bug fix, though. 😁

[buildsize]

File name	Previous Size	New Size	Change
babel-preset-env.js	2.69 MB	2.74 MB	52.46 KB (2%)
babel-preset-env.min.js	1.63 MB	1.65 MB	24.77 KB (1%)
babel.js	2.9 MB	2.95 MB	52.47 KB (2%)
babel.min.js	1.6 MB	1.63 MB	24.78 KB (2%)

[hzoo]

Some great work doing this research!

I would just like to note that using bitwise flags and operators is for sure going to make contributing and reading the coding that much more difficult (for me, and I will assume some others). I think I mentioned it earlier on Slack and it's a hard call but trading off readability/performance. I don't know what else we can do other than writing comments to explain things though