On 17/11/2023 04:09, Dmitry Gutov wrote:
> I can evaluate individual joaot/with-benchmark-group forms, and they 
> print -- how'd you call it -- minified Lisp data to messages, but it's 
> not formatted the same way as in your results.txt, nor is it easy to get 
> out of the -nw session because the clipboard is naturally not shared. Is 
> -nw needed? I expected it'd be used to print something to stdout.

OK, I used just that, as well as 'benchmark-run-compiled' inside 
'joaot/bench' instead of the complex expression which chose whether to 
compile a form.

Here are some results.

0) "mine" - just running the tests on your branch.

1) "mine-proper" - is me splitting of the typecheck into a separate 
method. Not an optimization, just better code.

2) "mine-proper+astoff" - including Augusto's optimization as well. 
Notably, seq-some gets ahead on big lists by about 1.2x. Not sure how 
reliable this number is, though.

3) "astoff" - just Augusto's patch. In some cases it's faster than 2 
because of no extra dispatch (or multimorphic methods, I guess), but the 
order-of-magnitude difference is still there in many set-difference 
examples.

Loading m6sparse.el (using (load (expand-file-name "../m6sparse.el" 
load-file-name)) at the top) didn't seem to affect the results of (2) 
too much, though it brought -- predictably -- the performance of 
seq-difference (the original version) further down. Though the results 
fluctuate a lot, so I might have missed some.