[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[PATCH] A few simplifications/maintainability improvements of the GC

From: Peter Bex
Subject: [PATCH] A few simplifications/maintainability improvements of the GC
Date: Sun, 19 Apr 2020 08:47:20 +0200
User-agent: Mutt/1.10.1 (2018-07-13)

Hi all,

After the recent patches to the GC, I was looking over that code again
and decided it would be quite easy to make that code a lot more readable
and maintainable, with less if statements and getting rid of a few gotos,
as well.  I haven't yet managed to get rid of i_like_spaghetti, though :)

Attached are a bunch of patches which incrementally simplify the GC.
I realised that the basic functionality of the three GC modes (minor, major
and realloc) are mostly the same:

- Mark live objects from saved stack and mutation stack
- Mark live heap objects (not in minor mode)
- Run Cheney's algorithm to mark nested objects

There's some other stuff that actually differs per mode (like finalizers
and locatives), but that's stuff that can stay the same.  The marking of
live objects and then nested objects is copy/pasted three times: There's
an if inside really_mark to distinguish between major/minor mode and then
there's of course really_remark.

But the copying is exactly the same!  The only thing that differs are
the addresses that we use to copy to, and the forwarding pointer
handling.  So, these patches change really_mark to accept the three
heap pointers (start, top and limit) as arguments, then pass in the
corresponding pointers to [C_]fromspace{start,top,limit},
tospace{start,top,limit} or new_tospace_{start,top,limit}, depending
on the mode we're in.

This is done by patches 1 and 2.  Patch 3 is a trivial change to drop
a few unused functions.  Patch 4 moves the marking of live objects into
a separate function.  Patch 5 moves marking of nested objects into a
separate function.

Finally, patch 6 is simply something to fix an annoyance during
development of this: Emacs insists on indenting, so fixing the basic
offset in .dir-locals.el.

I know it's a lot of patches, and they're somewhat largish, but it's
mostly just moving code around.  I didn't really make any fundamental
changes to the code.

This work is also available in the simplify-gc branch.  I've ran
Salmonella against it, and it doesn't break anything:

And I also ran the benchmarks against this, and the performance stays
the same (modulo some noise differences, as usual).


Attachment: 0001-Simplify-really_mark-to-be-a-bit-less-branchy.patch
Description: Text Data

Attachment: 0002-Also-convert-remark-into-really_mark-calls.patch
Description: Text Data

Attachment: 0003-Remove-a-few-unused-variables-from-rereclaim.patch
Description: Text Data

Attachment: 0004-Extract-out-the-marking-of-live-objects-common-to-al.patch
Description: Text Data

Attachment: 0005-Move-Cheney-algorithm-into-its-own-function.patch
Description: Text Data

Attachment: 0006-Fix-dir-locals-for-c-mode-indentation-depth.patch
Description: Text Data

Attachment: signature.asc
Description: PGP signature

reply via email to

[Prev in Thread] Current Thread [Next in Thread]