Es6.array.from and Es6.array.iterator do something with Float32Array/Array? which makes it 100x slower on Chrome

[vlajos]

Demo scripts:
http://lavela.hu/bug/test-array.from.html
http://lavela.hu/bug/test-array.iterator.html
http://lavela.hu/bug/test-array.others.html

Speed results on my machine:

• test-array.from.html 1236.399999999776 ms
• test-array.iterator.html 1345.0000000002547 ms
• test-array.others.html 11.300000000119326 ms

Other notes:

• The third one is just a reference that the other ES6.array.* modules does not seem to cause any trouble. (The speed is the same without core-js.)
• It seems that the bug does something with the chrome tab which affects the later executions in the same tab. (Run first a slow one and then try to run the array.others one. It will be also slow.)
• The test script: http://lavela.hu/bug/test.js But practically the new Float32Array(arr) operation is getting slower.
• es6.promise seems to affected as well. (I guess through a shared internal dependency.)
• Firefox does not seem to be affected.
• My chrome version: Version 64.0.3282.119 (Developer Build) built on Debian buster/sid, running on Debian buster/sid (64-bit) (But I had the same results on 63.x on Windows.)
• Core.js version: 2.5.3
• All the test details are here: http://lavela.hu/bug.tar.gz (built es6 modules, script to build them, test htmls, test js, test results, flame graphs, performance profiles.)

Related flamegraphs from chrome:

(The story started originally here: pixijs/pixijs#4662 and here: babel/babel#7362 But I think I copied everything relevant here.)

[vlajos]

Can I help anything with this issue?

[zloirock]

@vlajos now I have no time for exploring this issue. Feel free explore it and propose changes.

[vlajos]

@zloirock I tried to dig a bit deeper.
It seems that this line is the problematic one:
https://github.com/zloirock/core-js/blob/master/modules/_iter-detect.js#L18
At least if I extract the same logic, it seems that this script can cause the same trouble:

<script>
var ITERATOR = Symbol.iterator;
var arr = [7];
var iter = arr[ITERATOR]();
iter.next = function () { return { done: safe = true }; };
arr[ITERATOR] = function () { return iter; };
</script>
<script src="test.js"></script>

And if I comment out that line, the problem disappears. (Or if I return after/before that line in the original function.)
Unfortunately I am a little lost what is happening here...

I tried to revert arr[ITERATOR] to the original in a later step, but it did not make any difference.
Any idea?

A little shorter example:

<script>
var arr = [7];
arr[Symbol.iterator] = function () {};
</script>
<script src="test.js"></script>

(I'll carry on, just wanted to share these results first...)

[zloirock]

Funny. Seems somehow any array with a custom iterator somehow deoptimise array iteration (which also used in typed array constructors) in V8.

[vlajos]

@zloirock
Would it worth reporting the v8 devs?
What does this iter_detect tries to achieve? Is not it possible to do it without that assignment somehow?

[zloirock]

You could report it.

It's detection of support iterables in methods / constructors which should accept it.

It's possible to do it without that assignment, but IIRC that caused some problems in some old Promise.{all, race} implementations which accept only arrays and throws uncatchable errors to dev tools.

[zloirock]

I added a workaround to v3 branch. Feel free to test it. If it will cause some problems, I will revert it.

[vlajos]

@zloirock Thanks! I have just tried it, but there is no difference.
I checked out the v3 branch, then npm i;npm run build
I used this test page in chrome:

<script src="core.js"></script>
<script src="test.js"></script>

Just do double check if the new code is applied:

$ sed '2q;d' core.js 
 * core-js 3.0.0-alpha.1

grep -c ITERATION_SUPPORT core.js 
3

But the speed of the test is still the same.
So it seems that it still deoptimises that thing. :-(

[zloirock]

So seems the problem not (or not only) in this code.

[zloirock]

I reworked typed arrays and iterators polyfills, found some cases which can theoretically somehow deoptimise it and now, after calling your test, seems it's fixed in v3 branch. If it still not fixed, please, report me.

[vlajos]

@zloirock Thank you very much. The test is definitly quick now. Do you think I could test our "full" app against the v3 branch? Is it stable enough?

[zloirock]

@vlajos yes, now it's enough stable, today I finished testing on all supported engines. Seems actual version of v3 branch will be published as the first public alpha version of core-js@3 without any serious changes.

[zloirock]

core-js@3.0.0-alpha.1 is available.

[vlajos]

@zloirock Thanks. I checked earlier and i can confirm that our app works properly with core-js@3.0.0-alpha.1.
I see there are some more related commits, should I retest it with them?

[zloirock]

@vlajos nope, it's duplicates after rebasing.