Repository navigation
Link: work-ceiling retries without alternatives cost ~6 min on fernflower #92
Description
Follow-up to #89 (fix 9d4c768). With alternatives on, about a dozen methods per iteration reach the work ceiling in LinkComputerImpl.doMethod and are analysed a second time without alternatives. On fernflower (PARALLEL=1) that comes to:
| setting | run time | ceiling retries |
|---|---|---|
| alternatives off | 29 min | 0 |
| alternatives + both #89 rules | 35 min | 117 |
| fresh-object rule only | 37 min | 137 |
The retries recur every iteration for the same methods: each time they spend the full ceiling and then fall back.
Ideas:
- remember per method that the alternatives run hit the ceiling and go straight to the run without alternatives in later iterations (memo keyed on the method; reset if its inputs change shape);
- find which recipients still multiply the closure in those methods; the Link: with alternatives as plain edges, fernflower's StructClass.create exceeds the work ceiling through a callee #89 rules removed primitive and fresh-object recipients, and the rest may share a pattern.
On guava there are no retries (0 in both legs), so this is a fernflower-shaped cost.
Activity
bnaudts commentedon Nov 6, 2026
Probably the same cause as #97.
DomHelper.findGeneralStatementsits at 87% of the work ceiling with 15,302∩facts out of about 16,950 (90%). About half are materialized from dying temporaries, and most of the rest is∩copied along aliases. Details and the proposed direction in #97.Link engine: an orphan that a surviving middle vertex re-derives is n…
Link engine: lazy materialization for '∩' orphans only; LAZYMATANY fo…
Link: gate NOOVERLAP drops '∩' from the closure (#92 measurement)
Link: '∩' computed on demand instead of derived (#92), WIP
Link: implicit '∩' search continues with the stored label of a pair t…
Link: implicit '∩' as union-find classes; removed temporaries leave f…
Link: implicit '∩' — class partners by default (OVERLAPSEARCH diagnos…
Link: MLVDUMP=<fqn substring> prints computed summaries (diagnostic, #92
Link: OVLSTATS timing and gate OVLRELEVANT for the implicit '∩' (diag…
Link: implicit '∩' opt-in (gate IMPLICIT) while #92 is open; old deri…
Link: OVLSTATS counts what triggers the implicit-'∩' freezes and wher…
Link: implicit '∩' freeze skips the after-partition when no class can…
Link: implicit '∩' remembers each hub's creations while the rows they…
Revert "Link: implicit '∩' remembers each hub's creations while the r…
Link: gate NOFREEZESKIP and OVLSTATS rebuild classification for the i…
Revert the implicit-'∩' freeze shortcut and its A/B gate (#92)
bnaudts commentedon Nov 9, 2026
Findings, 2026-11-08/09: computing
∩on demand (class model)Where things stand. The class model is opt-in (
IMPLICIT=1). The old derivation is the default again (5b8e32f), because making the class model the default failed 10 tests in the maddi gate.Who reads
∩LinkToFielddeliberately skips∩.FieldAnalyzerImpl.computeIndependent: the catch-all for non-decoration links.field ∩ localbecomesfield ~ p, a nature that is judged.ShadowModificationPass.followsReturnLink: any link between two real variables carries modification.∩.WriteLinksAndModification: a mutualrv ∩ Ybecomes a§m ≡.RedundantLinks: groups∩with≤/≥, so a∩can drop a≤/≥.So
∩moves verdicts mostly by being converted into something else.Semantics
∩is absorbing and∩∘∩ = ∩. The complete fixpoint is therefore a set of classes, and the class model computes exactly that. The old engine's set is incomplete and order-dependent, so any complete algorithm differs from it somewhere.∩only for summary variables (OVLRELEVANT)∩was answered only for parameters, the return value,thisand fields.Fernflower's
EcjSwitchCandidatefields become@Dependent.GenericType.parseand fourElementAwareLinkedListTestmethods become modifying. The analyzer suite loses two expected local∩s (TestInstanceOf,TestModificationBasics.test3). So∩between locals does reach verdicts, and restricting it to the summary is not an option.Cost of the class model
On elasticsearch: 625 s, against 108 s for the old engine. The time goes to full rebuilds of the classes after every closure change, and to the before/after rebuilds of the freeze at every vertex removal.
Look-ups are cheap: 60–70% are cache hits. The worst method is
DesiredBalanceComputer.compute, at 33 s per pass, 12 s of it in∩.Tried, measured and reverted. Both were exact;
OVLCHECKfound no mismatch and the corpus verdicts were identical:What precedes each rebuild (langchain4j): 28% follow only new facts, and the classes only grew in all of them. 72% follow a removal or a label change, mostly temporaries leaving the graph.
Keeping the classes up to date incrementally therefore saves at most ~30–40% of the
∩bookkeeping. That might bring elasticsearch to about 550 s, still far from 108 s. Not built.Open
∩once per statement instead of per change. This is the only lever that cuts rebuilds by a large factor, but it changes results relative to the class model and would need its own verdict comparison.ChatRequestOptions.Builder.addListenerAttributecontext: the old engine deriveskey ∩ value, and an isolated probe does not reproduce it.ExprProcessor.ShallowMethodAnalyzercascade in langchain4j (EmbeddingModel).Diagnostic gates left in place
All are off by default:
IMPLICIT,OVLSTATS(timing counters),OVLRELEVANT,OVERLAPSEARCH,NOCHAIN,NOOVERLAP.OVLCHECKwent with the reverted shortcuts.Link: gate OVLSNAPSHOT computes the implicit '∩' once per statement (…
Link: OVLSTATS counts all link computation time; OVLMETHODMS sets the…
Link: OVLSTATS counts statements without reuse, reused and extracted …
Link: OVLSTATS prints every 500 methods (diagnostic, #92)
Link: gate OVLLAST shows the implicit '∩' only to the last statement'…