Query glibc version only once to speed up JSTransformer on Linux

[pastelsky]

↪️ Pull Request

While profiling Parcel as it ran on a large app at Atlassian, I noticed that calls to process.report.getReport() stood out in the trace —

This function is in fact called once per transformation, to get the glibc version on linux and on an average takes 10ms per call. However, the version of glibc on a system isn't likely to change during a parcel run. Moving the call out so it is queried only once led to up to a ~35% improvement in overall build times on Linux.

This was tested on a 15 Core c6i AWS box, on Node 16.15.0, with both the SWC and Babel transformers enabled. The % improvement might vary based on individual setups, however, it is still likely to be significant.

[devongovett]

hmm it's surprising to me that this is running more than once per thread given this check here. any idea why that's happening?

[pastelsky]

isLoadedOnMainThread is set to true only if the legacy glibc version condition is met —

parcel/packages/transformers/js/src/JSTransformer.js

Lines 950 to 957 in db3bcae

	if (glibcVersionRuntime && parseFloat(glibcVersionRuntime) <= 2.17) {
	let api = WorkerFarm.getWorkerApi();
	await api.callMaster({
	location: __dirname + '/loadNative.js',
	args: [],
	});
	isLoadedOnMainThread = true;

Else, its called once per transform —

parcel/packages/transformers/js/src/JSTransformer.js

Lines 308 to 314 in db3bcae

	async transform({asset, config, options, logger}) {
	let [code, originalMap] = await Promise.all([
	asset.getBuffer(),
	asset.getMap(),
	init,
	loadOnMainThreadIfNeeded(),
	]);

I think moving the check outside might be better in that case?

    if (glibcVersionRuntime && parseFloat(glibcVersionRuntime) <= 2.17) {
      let api = WorkerFarm.getWorkerApi();
      await api.callMaster({
        location: __dirname + '/loadNative.js',
        args: [],
      });
    }

   isLoadedOnMainThread = true;

[devongovett]

Ah right

[pastelsky]

Moved the isLoadedOnMainThread = true out instead, which should be a neater way to do this.

[pastelsky]

@devongovett do you mind approving the CI workflow whenever you can? Curious to see the effect this has on the parcel benchmark suite.

[pastelsky]

Looks like the benchmark pipeline has been broken, and performance hasn't been published for a while according to —

parcel-bundler/parcel-benchmark-action#93 (comment)

So nevermind, feel free to merge this if all looks good.