mirror/guile
mirror/guile
1
Fork 0
mirror of https://git.savannah.gnu.org/git/guile.git synced 2025-05-19 19:20:23 +02:00
Code Activity
121 commits 61 branches 169 tags 66 MiB
Commit graph

121 commits

This branch
This branch
All branches
Author SHA1 Message Date
Andy Wingo
c824f17bd9 Rename gc-types.h to gc-api.h 2022-08-08 11:08:36 +02:00
Andy Wingo
67f9c89f2a Use fragmentation_low_threshold for venerable_threshold 
This way fragmentation from venerable blocks doesn't cause the collector
to keep evacuating.
 2022-08-04 11:32:06 +02:00
Andy Wingo
0450a282dd Skip mostly-tenured blocks during sweep/allocate after minor GC 2022-08-04 09:04:27 +02:00
Andy Wingo
0fe13e1cab Accelerate scanning of remembered set 2022-08-03 21:25:18 +02:00
Andy Wingo
47c07dd0eb Fix embarassing ctz issue 2022-08-03 16:40:34 +02:00
Andy Wingo
8f6a2692ab Update README 2022-08-03 12:13:25 +02:00
Andy Wingo
96b68095b7 Fix mark pattern updating for generational whippet 
After a minor collection, we were erroneously failing to sweep dead
objects with the survivor tag.
 2022-08-03 12:06:19 +02:00
Andy Wingo
0210a8caf0 Refactor out-of-memory detection 
Firstly, we add a priority evacuation reserve to prioritize having a few
evacuation blocks on hand.  Otherwise if we give them all to big
allocations first and we have a fragmented heap, we won't be able to
evacuate that fragmented heap to give more blocks to the large
allocations.

Secondly, we remove `enum gc_reason`.  The issue is that with multiple
mutator threads, the precise thread triggering GC does not provide much
information.  Instead we should make choices on how to collect based on
the state of the heap.

Finally, we move detection of out-of-memory inside the collector,
instead of the allocator.

Together, these changes let mt-gcbench (with fragmentation) operate in
smaller heaps.
 2022-08-03 10:10:33 +02:00
Andy Wingo
1358d99abc Fix yield calculation after evacuating collections 2022-08-02 22:15:13 +02:00
Andy Wingo
a4e1f55f37 Implement generational collection 
Not really battle-tested but it seems to work.  Need to implement
heuristics for when to do generational vs full-heap GC.
 2022-08-02 15:37:02 +02:00
Andy Wingo
13b3bb5b24 Update barrier functions to also have the object being written 
Also remove read barriers, as they were unused, and we have no plans to
use them.
 2022-08-02 15:37:02 +02:00
Andy Wingo
7f405c929e Initial live mask does not include young allocations 
After rotation, the young bit wasn't being included anyway.  This just
improves the first collection.
 2022-08-02 15:37:02 +02:00
Andy Wingo
1781c5aed4 Fix evacuation allocator to clear any holes 2022-08-02 15:36:59 +02:00
Andy Wingo
22a9cc87a0 Update TODO 2022-07-20 14:40:47 +02:00
Andy Wingo
279309b821 mt-gcbench allocates garbage between live data 
This obviously invalidates previous benchmark results; perhaps we should
make this optional.
 2022-07-20 14:40:47 +02:00
Andy Wingo
d106f3ca71 Mutator collects evacuation target blocks 2022-07-20 14:40:47 +02:00
Andy Wingo
92b05a6310 Add implementation of parallel evacuation 2022-07-20 14:40:47 +02:00
Andy Wingo
4a9908bc4d Refactor evacuation vs pinning support 
Marking conservative roots in place effectively prohibits them from
being moved, and we need to trace the roots anyway to discover
conservative roots.  No need therefore for a pin bit.
 2022-07-20 14:40:47 +02:00
Andy Wingo
a16bb1833c Add logic to compute evacuation candidate blocks 2022-07-20 14:40:47 +02:00
Andy Wingo
c7c8fa2d32 Refactor to add "block_list" type 2022-07-20 14:40:47 +02:00
Andy Wingo
a8214af467 Whippet reserves a bit in object kind for forwarding 
Tags without the bit are forwarding addresses.
 2022-07-20 14:40:47 +02:00
Andy Wingo
8409383ee1 Refactor post-collection for mark space 2022-07-20 14:40:47 +02:00
Andy Wingo
69caead182 Add heuristics to choose when to compact or mark in place 
We can choose to compact (evacuate) or mark in place.  What we choose
has some effects on how we mark.
 2022-07-20 14:40:47 +02:00
Andy Wingo
09d2df1626 Compute GC yield as fraction of total heap size 2022-07-20 14:40:47 +02:00
Andy Wingo
c998f1cd5c Measure fragmentation as fraction of total heap size 
This allows a relatively more fragmented mark space if the majority of
the heap is taken up by lospace.
 2022-07-20 14:40:47 +02:00
Andy Wingo
7af8bb6bd0 Add machinery to disable ragged-stop marking 
We'll need to disable the optimization that mutators mark their own
stacks once we support evacuation.
 2022-07-20 14:40:47 +02:00
Andy Wingo
e4342f6c45 Add helper for yielding in a spinlock 2022-07-20 14:40:47 +02:00
Andy Wingo
52166fe286 Add gc_edge data structure 
Less casting in user programs, and it's a step on the way to evacuation
in whippet.
 2022-07-20 14:40:47 +02:00
Andy Wingo
808d365f4b We identify empty blocks lazily now 2022-07-20 14:40:47 +02:00
Andy Wingo
bc73c5ad02 Whitespace fix 2022-07-20 14:40:47 +02:00
Andy Wingo
157d40466b mark_space_reacquire_memory updates pending_unavailable_bytes 2022-07-20 14:40:47 +02:00
Andy Wingo
33a3af2c73 Large object space properly acquires blocks from mark space 
If the mutator finds completely empty blocks, it puts them on the side.
The large object space acquires empty blocks, sweeping if needed, and
causes them to be unmapped, possibly causing GC.
 2022-07-20 14:40:47 +02:00
Andy Wingo
71b656bca4 When sweeping, return empty blocks to global freelist 
This will facilitate management of overhead for defragmentation as well
as blocks to unmap, for compensating large object allocations.
 2022-07-20 14:40:47 +02:00
Andy Wingo
8f06b914b0 Refactor to allow "next" pointer embedded in block summary 2022-07-20 14:40:47 +02:00
Andy Wingo
7d80d45c79 Rename mark-sweep.h to whippet.h 2022-07-20 14:40:44 +02:00
Andy Wingo
061d92d125 Update README 2022-05-15 22:06:41 +02:00
Andy Wingo
69d7ff83dd More wording 2022-05-11 22:29:37 +02:00
Andy Wingo
c39e26159d Some README updates 2022-05-11 22:25:09 +02:00
Andy Wingo
7ac0b5bb4b More precise heap size control 
No longer clamped to 4 MB boundaries.  Not important in production but
very important for comparing against other collectors.
 2022-05-11 21:19:26 +02:00
Andy Wingo
fa3b7bd1b3 Add global yield and fragmentation computation 2022-05-09 22:03:50 +02:00
Andy Wingo
3bc81b1654 Collect per-block statistics 
This will let us compute fragmentation.
 2022-05-09 21:46:27 +02:00
Andy Wingo
7461b2d5c3 Be more permissive with heap multiplier 
Also if there's an error, print the right argument
 2022-05-06 15:08:24 +02:00
Andy Wingo
815f206e28 Optimize sweeping 
Use uint64 instead of uintptr when bulk-reading metadata bytes.  Assume
that live objects come in plugs rather than each object being separated
by a hole.  Always bulk-load metadata bytes when measuring holes, and be
less branchy.  Lazily clear hole bytes as we allocate.  Add a place to
record lost space due to fragmentation.
 2022-05-06 15:07:43 +02:00
Andy Wingo
0d0d684952 Mark-sweep does bump-pointer allocation into holes 
Instead of freelists, have mark-sweep use the metadata byte array to
identify holes, and bump-pointer allocate into those holes.
 2022-05-01 17:07:30 +02:00
Andy Wingo
f51e969730 Use atomics when sweeping 
Otherwise, there is a race with concurrent marking, though possibly just
during the ragged stop.
 2022-05-01 16:23:10 +02:00
Andy Wingo
2a68dadf22 Accelerate sweeping 
Read a word at a time from the mark byte array.  If the mark word
doesn't correspond to live data there will be no contention and we can
clear it with one write.
 2022-05-01 16:09:20 +02:00
Andy Wingo
ce69e9ed4c Record object sizes in metadata byte array 
This will let us avoid paging in objects when sweeping.
 2022-05-01 15:19:13 +02:00
Andy Wingo
3a04078044 mark-sweep uses all the metadata bits 
Don't require that mark bytes be cleared; instead we have rotating
colors.  Beginnings of support for concurrent marking, pinning,
conservative roots, and generational collection.
 2022-05-01 15:04:21 +02:00
Andy Wingo
f97906421e Sweep by block, not by slab 
This lets mutators run in parallel.  There is a bug currently however
with a race between stopping mutators marking their roots and other
mutators still sweeping.  Will fix in a followup.
 2022-05-01 14:46:36 +02:00
Andy Wingo
83bf1d8cf3 Fix bug ensuring zeroed memory 
If the granule size is bigger than a pointer, we were leaving the first
granule uncleared.
 2022-05-01 14:45:25 +02:00