Repository navigation
Link: with alternatives as plain edges, fernflower's StructClass.create exceeds the work ceiling through a callee #89
Description
With alternatives as plain edges ("alternatives are not aliases", #72), fernflower's StructClass.create(DataInputFullStream, boolean, LazyLoader) exceeds the work ceiling, is degraded to a shallow summary, and so becomes nonModifying=false. Without the change it is analysed in full and non-modifying.
create marks no alternatives in its own statements: the retry without them (LinkComputerImpl.doMethod, which saves DomHelper.findGeneralStatement) does not trigger. So the extra cost comes in through a callee whose summary changed under the change, and makes create's own link graph heavier.
What is known
- A/B on laser1, PARALLEL=1,
NOALTSOURCES NOLIBRARYRETURNagainst neither: the only fernflower FPDUMP difference, and the only degraded method that differs (17 without, 18 with). - Every attempt is a work ceiling (
[work ceiling]in the "Cycle protection tripped" line), in every iteration. - A class isolate of
StructClass, and ofDomHelpertogether with its reached types, doesn't reproduce such blowups: the callees there are stubs, or too close (maddi#180 for the isolate's own defects).
Next step
Find the callee: log LINKWORK for create's callees on both sides (a run with the link computer's debug logging limited to the org.jetbrains.java.decompiler.struct package) and compare their summaries. Then decide between a rule that keeps that shape grouped and a retry that also covers a method whose callees used alternatives.
Activity
bnaudts commentedon Nov 3, 2026
Found and fixed on
ws/maddi-null-89(9d4c768), on top of the #72 stack rebased onto devel (ws/maddi-null-72stack). Not merged yet; guava is being measured.The callee
StructClass.createmarks no alternatives itself, and without them it already uses 53–60% of the work ceiling. One callee's summary changed:GenericMain.parseClassSignature, throughGenericType.parse. Logging every recipient of alternatives (a probe, not committed) showed two kinds marked there, neither of them aliasing the method can observe:type = getType(c)(one of CodeConstants' ints),wildcard = switch (..) { .. };descriptor.superclass = GenericType.parse(superCl)takes one of fournew GenericType(..)(oc:122, oc:126, oc:129, oc:133). As plain edges the result's fields went from≺ oc(field of) to≈ oc(shares fields with), both ways, andcreate's closure grew from 9.1k to 12.9k edges.The fix
Both keep their group (
Graph.aliasingIsMoot, gatesNOALTPRIMITIVE,NOALTFRESH). guava's TypeResolver case has the caller's own locals as sources and is unaffected (TestCallResults.alternativeSources).StructClass.createWith both rules, fernflower's verdicts equal those with alternatives off.
Still open
bnaudts commentedon Nov 3, 2026
Guava check, KOTLIN-pass nullability on guava with every verdict and local dumped (PARALLEL=1, Mac):
Of 69,598 verdict and local lines, 2 differ:
TypeResolver.resolveParameterizedType:resolvedOwneris nullable with the fix (A: not null),resolvedRawTypestays not null. The case the alternatives were introduced for holds under the fix.ImmutableMap.copyOf(Iterable): the elements ofentryArraymove from nullable (A) to not null (B).From the earlier laser1 run, at modification level (FPDUMP): the fix changes nothing compared with the change before it (R vs B, 0 lines). Alternatives themselves change 252 lines against A, e.g.
ImmutableSet.SetBuilderImpl.build()and threeContiguousSet.*Implmethods become non-modifying. SWITCHFORK=1 changes 44 more lines (B vs S).The 19 trips are the same in both legs, so they are not caused by alternatives.
bnaudts commentedon Nov 5, 2026
Merged into devel (fix 9d4c768, gates NOALTPRIMITIVE and NOALTFRESH). The remaining work-ceiling retries are #92.