bpo-47189: What's New in 3.11: Faster CPython

[Fidget-Spinner]

The static objects section's credits go to Kumar Aditya.

TODO:

• Lazy __dict__ optimizations
• [ ] dict memory saving optimizations by Inada-san, but I have no clue where to start for those

https://bugs.python.org/issue47189

[Fidget-Spinner]

CC @gvanrossum @markshannon @brandtbucher @iritkatriel @sweeneyde @corona10 @kumaraditya303 @JelleZijlstra. For your opinions and review please (especially on the parts y'all contributed).

[corona10]

Thank you, It is a great honor to see my name on the list.

[markshannon]

Thank you, It is a great honor to see my name on the list.

You deserve it.

[markshannon]

Thanks for doing this. It is much appreciated.
I've left a few comments.

[markshannon]

I would rather claim a range, say 10% to 60%.

Once the latest PRs are merged it will be about 25%, if you must have a single number.

https://gist.github.com/markshannon/bf1b99177290db30dd0ab8f56446d0ea

[Fidget-Spinner]

I will update this number when the benchmark link updates its values/ when those PRs are merged into main.

[Fidget-Spinner]

@pablogsal are you OK with this going into 3.11 whats new? Or would you prefer I host this somewhere else and link to it in the whatsnew?

[pablogsal]

@pablogsal are you OK with this going into 3.11 whats new? Or would you prefer I host this somewhere else and link to it in the whatsnew?

For the time leave it here in its own section as in this way I will be able to easily do the final edits. Removing or moving information is much easier to add it as that way I don't need to consider what may be missing :)

[gvanrossum]

I put my nit-picking hat on while reviewing this. :-) Overall great work, and can't wait to see this show up on docs.python.org!

[gvanrossum]

Is this an oblique reference to PRECALL_NO_KW_METHOD_DESCRIPTOR_O? Or does it refer to something else?

[Fidget-Spinner]

No, it just means that CALL_FUNCTION_EX for things like f(1, 2, *args, **kwargs) aren't inlined.

[gvanrossum]

Hm. It's rather subtle what you intend to say then. The text linked under inline-calls already describes Python-to-Python calls. But they are described here again because specialization adds... what again? It feels like claiming credit for the same optimization in two places. I do realize that there are two separate code changes, but to the readers of a high-level document like this that seems a bit too subtle.

Also, it's not clear for what operation you are claiming a 170% speedup (so a factor 2.7). The table layout makes it look like C(arg). But surely the speedup for class constructors is not a single extremely large number?

[gvanrossum]

Bunch of one-character nits. :-)

[gvanrossum]

Honestly this PR is so large and comments are flying so fast we should probably land it ASAP and then iterate.

[gvanrossum]

Hm. It's rather subtle what you intend to say then. The text linked under inline-calls already describes Python-to-Python calls. But they are described here again because specialization adds... what again? It feels like claiming credit for the same optimization in two places. I do realize that there are two separate code changes, but to the readers of a high-level document like this that seems a bit too subtle.

Also, it's not clear for what operation you are claiming a 170% speedup (so a factor 2.7). The table layout makes it look like C(arg). But surely the speedup for class constructors is not a single extremely large number?

[Fidget-Spinner]

Also, it's not clear for what operation you are claiming a 170% speedup (so a factor 2.7). The table layout makes it look like C(arg). But surely the speedup for class constructors is not a single extremely large number?

Oof the table layout worked against me. The 170% speedup is tied to "Calls to certain python functions are inlined", not C(arg). I have since removed that line altogether.

[gvanrossum]

Thanks, this looks good now. Thanks very much for all your work on this!