Added example to Unsupported phi use of const or let variable

[billhance]

I came across the "Unsupported phi use of const or let variable" bug while learning es6 and practicing a selection sort. I noticed my implementation was way slower than I expected so this might be a misunderstanding of some block scope gotcha in JS on my part.

Here is a fiddle that demonstrating the bug: https://jsfiddle.net/billhance/wfz9fcs6/1/

In console profiler, I saw an increase of 800% where input n = 1000 when this bug was present.

[leeoniya]

I've run into this before also and noticed the issue went away after transpiling down to ES5. i wonder if any of the recent ES6 improvements [1] cover this case.

[1] http://v8project.blogspot.com/2016/12/v8-release-56.html

[billhance]

@leeoniya I think you are right that is specific to ES6 because when I run the ES5 equivalent (specifically let to var) the bug goes away. Unfortunately, I don't have a beta build handy to test the new v8 changes.

[vhf]

You might want to try the snippet triggering this bailout on node nightly or on the latest version of v8. Both are available on https://mono.morph.ist

I'm happy to try it for you if you prefer.

[billhance]

Thanks for the suggestion @vhf, I've never used that tool before.
Not sure if I ran the test correctly, It seems like it's not fixed in the node nightly. https://mono.morph.ist/job/9johJCT9QWuC8NAbK

[vhf]

Thanks for the feedback @billhance . Yeah this tool is kinda new, I released it pretty recently and I know there are a few bugs in there. I'll work on them when I have time.

The results of your tests are quite surprising. TurboFan doesn't bail out on node 6.

I ran a similar test here https://mono.morph.ist/job/X7yqLKNJqTukaxu3g but it bails out on all known node versions it seems.

Also interesting is that using the destructuring variable swap trick doesn't bail out on nightly but bails out on node 6 : https://mono.morph.ist/job/pdCkANEcwDcLxEysD (see line 15)

[billhance]

Very cool tool 👍 . So am I to assume correctly that this is a recently introduced bug in v8 (sorry, a bit out of my depth)? If so, would https://bugs.chromium.org/p/v8/issues be the appropriate place to file the bug report? Thanks.

[vhf]

Well, it's really not a bug. The results you got depend on several things.

For instance, if transpiling to ES5 "solves" the issue, it doesn't mean it's a bug in later versions. The semantics are just different, and rightfully handled differently. The scoping of let and const have no ES5 equivalent. The whole TDZ (temporal dead zone) concept was introduced by ES2015 and seems not to be optimizable by Crankshaft. It also seems to me this bailout still happens in Turbofan under very specific conditions, I just cannot get which ones.

Also worth keeping in minde, when we observe some node versions optimizing this snippet via turbofan and some later node versions not optimizing it, it might simply mean that the later versions don't have turbofan enabled.

I don't work on V8 so I cannot really say whether opening an issue there would be the right move. I'm pretty sure it wouldn't be closed without someone giving some kind of explanation though. :)

[billhance]

I'm sure that team has enough on their plate for now to spend time indulging my curiosity so I'll hold off. Thanks so much for helping me understand this stuff better @vhf.

[dmonad]

This is already fixed in current Firefox and Chrome.

[vhf]

@dmonad Firefox never used v8 and Chrome does not use Crankshaft anymore. This repo serves as documentation even though it's legacy by now. ;)

[dmonad]

Ah okay. Thanks for the clarification!