GNU bug report logs - #65491
[PATCH] Improve performance allocating vectors

Previous Next

Full log

Message #55 received at 65491 <at> debbugs.gnu.org (full text, mbox):

From: Mattias Engdegård <mattias.engdegard <at> gmail.com>
To: Stefan Monnier <monnier <at> iro.umontreal.ca>
Cc: 65491 <at> debbugs.gnu.org, Eli Zaretskii <eliz <at> gnu.org>
Subject: Re: bug#65491: [PATCH] Improve performance allocating vectors
Date: Sat, 16 Sep 2023 16:58:01 +0200

Results of some further experiments:

- Reducing the `vector_free_lists` array to its actually used first half does improve performance a bit, even with subsequent scanning speed-ups.

- Enlarging `struct vector_block` from 4 KiB does not necessarily speed things up but more measurement is needed.

- Using a bitmap for faster `vector_free_lists` scanning is definitely beneficial.

- The previously mentioned hack where scanning begins in the most recently added bucket is surprisingly effective, often even more so than a bitmap, but I'm wary of it being brittle. Need more measurements.

- Our special-casing of the bool-vector representation is annoying. It would be good if we didn't have to pay the price on 64-bit platforms.

I also pushed two small changes that were essential to working with the code so we might just as well have them in now: b1881d7dab (fix a static assertion) and 056c99a34c (better Lisp_Object untagging).

I'd like to try size-homogeneous blocks as well, but this requires some GC updates. Just a matter of taking the time, of course.

Previous Next

GNU bug tracking system
Copyright (C) 1999 Darren O. Benham, 1997,2003 nCipher Corporation Ltd, 1994-97 Ian Jackson.