Severe performance regression with object spread operator introduced in v12.18.1 and v14.5.0

[uwinkelvos]

• Version: v12.18.1, v12.18.2, v14.5.0
• Platform: Linux lalala-nb 5.7.7-arch1-1 deps: update openssl to 1.0.1j #1 SMP PREEMPT Wed, 01 Jul 2020 14:53:16 +0000 x86_64 GNU/Linux
• Subsystem: v8

What steps will reproduce the bug?

When using the object spread operator to "join" two objects, the performance has decreased by ~ factor 20. This only happens with objects that have at least 50 keys and the keys of the first one need to be non trivial.

see:
https://github.com/uwinkelvos/node_object-spread_preformance-regression

If i had to guess it's related to:
af95bd7#diff-ff06109b32824fc20fec697f584a42e3

How often does it reproduce? Is there a required condition?

Always reproducible on aforementioned platforms.

What is the expected behavior?

performance level like 12.18.0

What do you see instead?

factor 20 performance loss.

Additional information

https://bugs.chromium.org/p/v8/issues/detail?id=10763

[addaleax]

You might want to report this at https://bugs.chromium.org/p/v8/issues/list as well.

[nelsonsvo]

agreed

[uwinkelvos]

Hey! I was out for two weeks, but you are right of course. Created a v8 issue: https://bugs.chromium.org/p/v8/issues/detail?id=10763

[mcollina]

Confirmed it's caused by 78eb420.

[mcollina]

78eb420 was introduced to fix #32049... however it caused regressions of its own. What should we do?

cc @nodejs/tsc @nodejs/v8

[mmarchini]

The overhead here seems worse than #32049, correct? I wonder which regression our users are more likely to hit (although we have no way of knowing it for sure). The best thing would be to cherry-pick a fix as soon as one is available upstream, but until then we might want to consider reverting 78eb420.

[uwinkelvos]

Btw. there has been some progress on the upstream bug which is alread on main branch. I have not tested the changes though.

[camillobruni]

I'm working on this on the side, currently these are just cleanup CLs to pave the way for a proper fix.

[uwinkelvos]

Ah, alright! Thanks for the work and the update!

[alexcfyung]

Acmeair benchmark have ~40% degradation on my end because of 78eb420, you can also see the effect here, and it affect v14.5 onward as well.

[19shubham11]

Is this still an issue? I'm trying to update to v14.17 and was wondering if I should take this into account? 🤔