Skip to content

fix(agent-feed): objects published through an array or updater update function are ordered (#1336, part) - #1448

Merged
PIsberg merged 6 commits into
mainfrom
fix/p4-reference-slot-forms
Nov 21, 2026
Merged

PIsberg merged 6 commits into
mainfrom
fix/p4-reference-slot-forms

Conversation

@PIsberg

@PIsberg PIsberg commented Nov 20, 2026 •

Copy link
Copy Markdown
Owner

Part of #1336. Stacked on #1444, #1446 and #1447, review the last two commits.

Why

With the agent attached, an AtomicReferenceArray element replaced by a successor built inside updateAndGet (or getAndUpdate, accumulateAndGet, getAndAccumulate, compareAndExchange, weakCompareAndSetVolatile), or a field replaced through an AtomicReferenceFieldUpdater's update or accumulate forms, was reported by AtomicityValidator at PROMPT. That happened when another thread read the successor's state through the slot. These forms store with a volatile write that released nothing in the model; #1334 fixed only AtomicReference's.

The issue was ranked P4 as unmeasured. It is measured now: both new must-stay-silent corpus idiom rows were reported before this change.

Change

  • 10 hooks in TelemetryRegistry, each running the form as the woven get and the woven compareAndSet of its slot kind, the loop the JDK runs inside it. The function's argument is acquired, the stored reference is published, and a lost swap publishes nothing.
  • FieldAccessWeaver: rows in REFERENCE_ARRAY_FORMS (6) and REFERENCE_UPDATER_FORMS (4). weakCompareAndSetPlain and the deprecated weakCompareAndSet stay unwoven, and SpinLockHookTableTest pins that.
  • The AsyncTestAgent @AICore note names the new substitutions, and the guardrail files are regenerated from it (tools/guardrail-drift.sh exits 0).

Still open on #1336: the acquire and release flavours on every slot kind, which need a decision on what a one-sided swap publishes (the VarHandle volatile forms are in the second commit, below).

Verified

  • Corpus idioms lane: idiom_arrayUpdateAndGet_buildsTheSuccessorInTheFunction and idiom_updaterUpdateAndGet_buildsTheSuccessorInTheFunction were both reported before the hooks. With them the lane passes 224/224, and their plain-field twins fire. The full corpus (6 lanes) is green after a fresh install.
  • TelemetryRegistryHooksMatchTheJdkTest compares every new hook with the JDK call. SpinLockHookTableTest checks that the forms are substituted and the plain ones are not.
  • Library -P fast: 4339 tests green. Agent -P e2e: 610 tests green. Static gates, the javadoc jar and the guardrail drift check are clean.

Docs

  • docs/agent/attaching.md: which array and updater forms order what, and what is still not woven.
  • docs/CHANGELOG.md.
  • The AsyncTestAgent guardrail note, and through it the generated CLAUDE.md, GEMINI.md and rule files.
  • Javadoc on each hook and on both weaver tables.

Second commit (VarHandle forms)

A reference VarHandle's compareAndExchange and weakCompareAndSet (volatile on a handle) now publish, on an instance field, a static field and an array element. There are six hooks, each run as the woven compareAndSet plus the woven getVolatile witness. referenceHandleHook matches them, and the hook descriptor returns Object (narrowed back by the existing CHECKCAST) or boolean. weakCompareAndSetPlain and the acquire and release flavours stay unwoven, and SpinLockHookTableTest pins both sides.

Verified failing first: the new must-stay-silent row idiom_varHandleCompareAndExchange_retriesFromTheWitness was reported against the jar without these hooks and passes with them. Idioms lane 226/226, full corpus green, agent -P e2e 610, library -P fast green. The JDK-equivalence rows cover all six hooks, and the guardrail drift check exits 0.

Still open on #1336 after both commits: only the acquire and release flavours on every slot kind.

🤖 Generated with Claude Code

PIsberg and others added 5 commits November 20, 2026 18:59
…#1317, part)

#1315 made set, getAndSet, compareAndSet and the arithmetic forms of
AtomicBoolean/AtomicInteger/AtomicLong release and get acquire. A field
published through lazySet, setRelease, updateAndGet/getAndUpdate/
accumulateAndGet/getAndAccumulate, a won compareAndExchange(Release) or
a won weakCompareAndSetVolatile/Release, and read after a get that saw
the value, was still reported by AtomicityValidator as an unordered
publication on correct code. And a reader whose read was getAndSet's
previous value, a compareAndExchange witness, or what an add or update
returned acquired nothing.

Those hooks now call primitiveSlotStoring (before the store; after a
won swap for the compare forms) and HappensBefore.acquireVolatile on
the value read. The update and accumulate forms run as a CAS loop like
addPublishing, so the value to be stored is known before the store.
setPlain, setOpaque and weakCompareAndSetPlain stay unordered, as the
JMM says; AtomicFlagPublicationWeavingTest pins setPlain.

Verified failing first: 23 new AtomicFlagPublicationWeavingTest rows
(one per form) failed before the hook change and pass after; the
setPlain negative passed throughout. Library atomic, spinlock,
happens-before and atomicity tests plus MinTrustGateTest under -P e2e,
the full agent suite under -P e2e and the corpus eval are green;
static gates and the javadoc jar are clean.

Still open on #1317: getAcquire, intValue and longValue reads, the
AtomicLong forms the weaver does not substitute, and the updater,
VarHandle and array forms.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The previous commit made the value an increment, add, update or
accumulate returns acquire. The corpus idioms lane then filled a 2 GB
heap within seconds on CI (Java 26 OOM; 21 and 25 hung past 25 min),
reproduced locally under -Xmx2g. A heap histogram of the dead fork
showed 1.38 GB of long[] and 690 MB of int[] averaging 32 KB and 14 KB:
HappensBefore stamps near MAX_ENTRIES (4096). The add releases before
its store, so a counter every thread bumps (CorpusRecorder's static
BODY_EXECUTIONS, bumped once per body execution) collects every
bumper's clock; acquiring on the read handed that whole clock to each
bumper, and every access it recorded afterwards pinned a copy. A user
suite with a global id counter would hit the same OOM.

The add and update forms now release only, as on main before #1317.
getAndSet and the compareAndExchange witness still acquire, and a get
of the counter still acquires what it published.

Verified failing first: AtomicCounterClockTest (11 forms, 200 earlier
bumpers) failed with a clock of 201 threads on every form and passes
after. AtomicFlagPublicationWeavingTest's two add/update-result rows
are inverted to assertRaced, pinning the decision. Idioms lane under
-Xmx2g: OOM before, 220/220 in 37 s after; full corpus (6 lanes)
green; library -P fast 4336 tests green; static gates and javadoc jar
clean.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The woven addAll released every element before the call and, outside
queues, withdrew nothing. An element a concurrent set refused (it held
an equal one) or one a call throwing part way never added stayed
published, so a thread that reached it another way was ordered after
the producer by a hand-off that never happened, hiding a race.
HappensBefore.retract could not be reused: it keeps only the thread's
last release and forgets it at the next stamped access, which the
element's own equals/hashCode makes inside addAll.

HappensBefore.releaseWithdrawable hands each release back and withdraw
takes it back exactly (CAS of the element's clock from what the release
installed to what it held before; a later release by another thread
keeps it). collectionAddAll withdraws every element's release when the
call returns false, and those of elements the collection does not hold
by identity when it throws.

Verified failing first: AgentHappensBeforeFeedTest's two new tests
(newKeySet already holding equal tags; newKeySet.addAll stopping at a
null) failed before and pass after; the twin (the element that went in
stays published) passes throughout. Library -P fast 4338 green, full
corpus green after a fresh install, static gates and javadoc jar clean.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…se (#1316, part)

The woven increment, add and update forms of AtomicInteger/AtomicLong
release the value they are about to store before each compare-and-set
attempt (#1315). An attempt that lost to another thread storing the
same value left its release as the atomic's latest, so a get that saw
the winner's value was ordered after the loser's earlier writes too:
an edge the program does not have, which can hide a race on them.

HappensBefore.releaseVolatile now remembers, per thread and without
allocating, the field clock and the release it installed;
withdrawVolatile puts the field back to the release before it while it
is still the latest. addPublishing (int, long) and updatePublishing
call it after a lost compareAndSet. When the winner released after the
loser, its cumulative release already holds the loser's clock and
nothing is withdrawn; #1316 stays open for that order.

Verified failing first: AgentHappensBeforeFeedTest
.aLostAttemptWithdrawsItsRelease (the update function lets another
thread win the first swap, then a reader gets the winner's value) was
red before and is green after; its twin (the attempt stored the value)
stays silent throughout. Library -P fast 4339, agent
AtomicFlagPublicationWeavingTest 35/35 under -P e2e, full corpus after
a fresh install, static gates and javadoc jar all green.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… function are ordered (#1336, part)

With the agent attached, an AtomicReferenceArray element replaced by a
successor built in updateAndGet (or getAndUpdate, accumulateAndGet,
getAndAccumulate, compareAndExchange, weakCompareAndSetVolatile), or a
field replaced through an AtomicReferenceFieldUpdater's update or
accumulate forms, was reported by AtomicityValidator at PROMPT when
another thread read the successor's state through the slot. These forms
stored with a volatile write that released nothing; #1334 fixed only
AtomicReference's.

Each form now has a hook that runs it as the woven get and the woven
compareAndSet of its slot kind, the loop the JDK runs inside it, and
FieldAccessWeaver's REFERENCE_ARRAY_FORMS and REFERENCE_UPDATER_FORMS
substitute them. weakCompareAndSetPlain and the deprecated
weakCompareAndSet stay unwoven (SpinLockHookTableTest pins it). The
AsyncTestAgent @aicore note names the new substitutions; the guardrail
files are regenerated from it.

Verified failing first: the two new must-stay-silent corpus idiom rows
(idiom_arrayUpdateAndGet_/idiom_updaterUpdateAndGet_buildsTheSuccessor
InTheFunction) were both reported before the hooks; with them the lane
passes 224/224 and the full corpus is green. TelemetryRegistryHooksMatch
TheJdkTest compares every new hook with the JDK call. Library -P fast
4339, agent -P e2e 610, static gates green.

Still open on #1336: the reference VarHandle compareAndExchange and
weak swaps, and the acquire/release flavours on every slot kind.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@codecov

codecov Bot commented Nov 20, 2026 •

Copy link
Copy Markdown

…ompareAndSet publish (#1336, part)

With the agent attached, an object published through a reference
VarHandle compareAndExchange loop on a volatile field was reported by
AtomicityValidator at PROMPT when another thread read its state through
the handle: compareAndExchange and weakCompareAndSet (volatile on a
handle) stored with a volatile write that released nothing.

Each now has a hook for an instance field, a static field and an array
element, run as the woven compareAndSet and, for a lost exchange, the
woven getVolatile witness. FieldAccessWeaver.referenceHandleHook matches
them and the hook descriptor returns Object (narrowed back by the
existing CHECKCAST) or boolean. weakCompareAndSetPlain and the acquire
and release flavours stay unwoven; SpinLockHookTableTest pins both
sides. The AsyncTestAgent @aicore note and the generated guardrails
name the new substitutions.

Verified failing first: the new must-stay-silent idiom row
idiom_varHandleCompareAndExchange_retriesFromTheWitness was reported
against the jar without these hooks (lane 226, 1 failure) and passes
with them; the full corpus is green. TelemetryRegistryHooksMatchTheJdk
Test compares each hook with the handle's own call. Agent -P e2e 610,
library -P fast (JavadocDescribesRatherThanRestates caught two
restating @PARAM lines, fixed), static gates and javadoc jar green.

Still open on #1336: the acquire and release flavours on every slot
kind, which need a decision on what a one-sided swap publishes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant