Simplified LXR GC by wenyuzhao · Pull Request #1508 · mmtk/mmtk-core

wenyuzhao · 2026-06-05T03:50:14Z

DRAFT: We may merge this PR as long as the LXR plan is working correctly, even if it still contains code that needs refactoring. This PR should also do its best to keep other parts of MMTk clean. We need to ensure the following issues are resolved before this PR is merged.

Some modules/types/functions/etc. are made public. We need to reduce their visibility to what's necessary.
Some metadata are made public. We need to reduce their visibility to what's necessary.
Plans other than LXR should not suffer functionality loss, including:
- They should still support reference processing and finalization.
- They should not assume there are only two spaces.
- They should not assume the raw address of object reference == object start.
The performance of plans other than LXR should not regress.

This pull request introduces a simplified version of the LXR collector. There are many features not included in this version of LXR, such as reference processing, and they will be ported to the master branch in the future.

In addition to the LXR plan, this PR also makes the following changes to mmtk-core

Field unlog bits metadata and field-logging barrier. They can be used by generational and concurrent Immix plans to implement remembered set and SATB.
Pause::RefCount: a new pause kind for reference counting.
RootsWorkFactory::create_process_roots_work_with_root_kind: an experimental API with RootKind.
Many changes to the ImmixSpace and LargeObjectSpace to support reference counting-based Immix plans.
Additional object metadata to support LXR.
Additional scheduler stages to support LXR.
Each work bucket now has two queues to support deferred work packets.
Renamed pause-start and pause-end hooks to merge the two mechanisms from the master branch and the LXR branch to make the names consistent.- UnstraddlableRegion to prevent the invocation of region.contains on regions which objects can straddle, specifically Line.
Merged the implementations of ObjectReference::iterate_fields.
The definition of "chunk index" is changed.
Map32: 0 is now a valid entry of the linked list, and invalid entries are represented as -1.

Related PRs:

mmtk-openjdk JDK-21 PR: Simplified LXR GC mmtk-openjdk#358

…et is open

qinsoon · 2026-07-31T01:51:31Z

-                Line::MARK_TABLE
-                    .bzero_metadata(block.start(), crate::policy::immix::block::Block::BYTES);
-                // mark objects if concurrent marking is active
-                if self.immix_space().should_allocate_as_live() {


I am not sure why it is removed. I didn't find an alternative approach in the PR to mark new allocated objects as live.

This is added back for non-LXR plans.

qinsoon · 2026-07-31T03:34:04Z

+        // is a defrag source) for every Immix-family plan, not just plans with reference
+        // counting enabled. So this table must always be mapped, regardless of which plan
+        // requests it.
+        ret.push(crate::policy::immix::block::Block::DEFRAG_STATE_TABLE);


This is added by Claude during fixing mmtk-core test/build. It is clearly wrong. We need to revert this.

We let ConcurrentImmix use `add_deferred` for the `Concurrent` bucket like LXR. This simplifies the design so that the `Concurrent` bucket is always flipped during the end of a pause.

This reverts commit 3027b79.

qinsoon · 2026-08-03T02:02:42Z

+        // as if drained, even though unprocessed concurrent-marking work is still sitting in the
+        // now-inactive queue. Keep this the same enable/disable-only mechanism as master unless
+        // the plan is LXR.
+        let is_lxr = mmtk.get_plan().downcast_ref::<LXR<VM>>().is_some();


This is a workaround for now. Should fix it before merging the PR.

…only skipped for LXR.

qinsoon · 2026-08-03T05:17:03Z

-        self.reusable_blocks.flush_all();
+        // FIXME: Do we need this for LXR? We observed this to cause fails on conix.
+        if !self.rc_enabled {
+            self.reusable_blocks.flush_all();


This line was removed from this PR. Now it has been added back for non-LXR plans.

Currently the VO bit is not precise enough for conservative stack scanning because dead objects in lines that also contain live objects will still have VO bits. It is enough for asserting that INC and DEC never operate on dead objects.

We remove methods of `Block` and `Line` if they are already implemented by the `Region` trait. We split `Line::contains` into `Line::contains_obj_start` and `Line::contains_obj_ref` to disambiguate the different intentions.

Fixed build errors. Both master and lxr-x/simplified introduced concurrent GC options. We use the option introduced in master recently. Both master and lxr-x/simplified introduced pause start/end hooks. We combine them and rename them `on_pause_start` and `on_pause_end` to be consistent with the methods in `GCTriggerPolicy`.

qinsoon · 2026-08-07T04:18:40Z

binding-refs
OPENJDK21_BINDING_REPO=wenyuzhao/mmtk-openjdk
OPENJDK21_BINDING_REF=lxr-x/jdk21-simplified

We introduce the `UnstraddlableRegion` and move the method `Region::containing(object: ObjectReference)` into it. The primary intention is to prevent the invocation of `Line::containing`. An object may straddle multiple lines. In this case, the line that contains the *starting address* of the object and the line that contains the *raw address* of `ObjectReference` will be different. Removing `Line::containing` will eliminate unintended misuse.

…' into lxr-x/simplified

There is no "relaxed fence"

The SlotIterator was introduced with ConcurrentImmix, and was inspired by the SlotIterator in the LXR branch. Observing from the use case, we find that it is usually used as a method of `ObjectReference`, and is not specific to tracing. We move its implementation to `ObjectReference::iterate_fields`.

The method `Mutator::assert_is_flushed` is never used, and the method `Barrier::assert_is_flushed` was removed long ago.

Move the constants `LOGGED_VALUE` and `UNLOGGED_VALUE` into the `log_bit` module.

We keep the signature of `RootsWorkFactory::create_process_roots_work` unchanged while adding a `create_process_roots_work_with_root_kind` method for LXR and OpenJDK. Currently the `RootKind` contains OpenJDK-specific concerns, and should be made VM-agnostic.

wenyuzhao added 30 commits November 28, 2023 17:51

Add regression test results

9351018

Fix potential leak and virtual memory OOM

2dc90ac

minor

be2fdb5

lxr_precise_incs_counter

466ffd3

minor

9e9f916

minor

110b5fd

minor

cf5bf17

Flush barrier-generated mark queue before FinishConcurrentWork buck…

5d6c0fa

…et is open

Fix SATB without lazy decrements

4c45354

Fix lazy sweep counters

a56e00b

Improve SATB performance

cf1faf2

wip: simple SATB trigger

7d83dae

wip: limit number of concurrent workers

af39191

wip

2f9c481

Minor SATB optimizations

fda36aa

Trigger final mark pause right after SATB is finished

598228e

wip

9dda971

working

896011f

wip

f51a58a

cleanup

9ec3b58

cleanup

2f11502

minor

432c273

Concurrent marking optimizations

6ca0a55

Fix space check optimization

281aed8

measure trace rate

2dedda9

lxr_no_mature_defrag

43d912e

minor fix

7502428

wip

b094de6

Optimize defrag check

9ed9544

Fix CM array slice enqueuing

da40c1b

wks added 2 commits July 29, 2026 13:49

Revert changes to Block::sweep

d72e64b

Use -1 as invalid chunk

8f3c129

qinsoon reviewed Jul 31, 2026

View reviewed changes

wks and others added 5 commits July 31, 2026 15:39

GC status transition logging

b147e1e

Use add_deferred for concurrent bucket.

3027b79

We let ConcurrentImmix use `add_deferred` for the `Concurrent` bucket like LXR. This simplifies the design so that the `Concurrent` bucket is always flipped during the end of a pause.

Revert "Use add_deferred for concurrent bucket."

b37aba1

This reverts commit 3027b79.

Don't flip queues for non-lxr plans

e04e7ec

Merge branch 'master' into lxr-x/simplified

9997259

qinsoon reviewed Aug 3, 2026

View reviewed changes

Gate two places: reusable block flush and allocate as live. They are …

4f608cc

…only skipped for LXR.

qinsoon reviewed Aug 3, 2026

View reviewed changes

wks and others added 8 commits August 3, 2026 14:47

Log LXR GC scheduling

f824670

cargo fmt

6d9956f

Enable VO bits for LXR

087bab9

Currently the VO bit is not precise enough for conservative stack scanning because dead objects in lines that also contain live objects will still have VO bits. It is enough for asserting that INC and DEC never operate on dead objects.

Remove methods already implemented by Region

e8edabd

We remove methods of `Block` and `Line` if they are already implemented by the `Region` trait. We split `Line::contains` into `Line::contains_obj_start` and `Line::contains_obj_ref` to disambiguate the different intentions.

Merge branch 'master' into lxr-x/simplified

824bbfe

Merge branch 'master' into lxr-x/simplified

7bc24da

Add missing cfg(feature = "vo_bit")

7609f50

wks added 9 commits August 7, 2026 13:18

Merge remote-tracking branch 'refs/remotes/wenyuzhao/lxr-x/simplified…

e64a162

…' into lxr-x/simplified

Cleanup

41d579d

There is no "relaxed fence"

Fix clippy warning

ddb2f2e

Restore unreachable! in unreachable_prepare_func

ff05c9d

Remove assert_is_flushed

7925414

The method `Mutator::assert_is_flushed` is never used, and the method `Barrier::assert_is_flushed` was removed long ago.

Move LOGGED_VALUE and UNLOGGED_VALUE to log_bit

3d69aa9

Move the constants `LOGGED_VALUE` and `UNLOGGED_VALUE` into the `log_bit` module.

Navigation Menu

Uh oh!

Simplified LXR GC - #1508

Simplified LXR GC#1508
wenyuzhao wants to merge 1754 commits into
mmtk:masterfrom
wenyuzhao:lxr-x/simplified

wenyuzhao commented Jun 5, 2026 •

edited by wks

Loading

Uh oh!

qinsoon Jul 31, 2026

Uh oh!

qinsoon Aug 3, 2026

Uh oh!

qinsoon Jul 31, 2026

Uh oh!

qinsoon Aug 3, 2026

Uh oh!

qinsoon Aug 3, 2026

Uh oh!

qinsoon commented Aug 7, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

wenyuzhao commented Jun 5, 2026 • edited by wks Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

qinsoon Jul 31, 2026

Choose a reason for hiding this comment

Uh oh!

qinsoon Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

qinsoon Jul 31, 2026

Choose a reason for hiding this comment

Uh oh!

qinsoon Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

qinsoon Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

qinsoon commented Aug 7, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

wenyuzhao commented Jun 5, 2026 •

edited by wks

Loading