Repository navigation
codegen: perry_method_<Class>__<name> emission order is HashMap-nondeterministic when a method name is shared across classes (#7038/#7039 family) #7172
Description
Found during PR #7171's IR-neutrality verification (macOS arm64, same-compiler control runs): emitted LLVM IR is nondeterministic when a method name is shared across classes — perry_method_<Class>__<name> symbols are emitted in HashMap iteration order, so two runs of the same compiler on the same source produce differently-ordered IR.
Observed on real dependency files: vite's module-runner.js and rollup's shared/index.js (the two corpus files excluded as control-failed in #7171's 47/0/2 IR comparison — reported honestly as control failures, not counted as passing).
Class
Same family as #7038/#7039 (closure-source HashMap iteration made @.str.N numbering a per-process permutation on macOS — fixed by sorting) and #7131 (Linux temp names — fixed). This is the method-emission sibling: cross-class shared method names iterate an unordered map.
Why it matters
- Any object-hash or IR-diff A/B silently excludes affected files (or worse, reads phantom diffs) — two agents have now had to build control-run machinery to detect it.
- Fix shape is the fix(codegen): sort closure-source emission so the same input compiles to the same IR (#7038) #7039 precedent: sort the emission order by a stable key (class id, then name).
Repro
Compile vite/dist/node/module-runner.js (or rollup shared/index.js) twice with the same compiler, --trace llvm or --no-link, diff the IR: method-symbol ordering differs run-to-run.
Activity
proggeramlug commentedon Aug 9, 2026
Closing — fixed by
c413af072(PR #7625, "make emitted IR run-to-run deterministic").It sorts the
ctx.class_idsiteration inlower_call/property_get/dynamic_dispatch.rs(dispatch_roots.sort_unstable_by(|a,b| a.1.cmp(&b.1).then_with(|| a.0.cmp(&b.0)))) and sortsmethod_registry.rs'sclass_tablewalk that builds the(class, method) → perry_method_*symbol map — the exact map this issue names.The commit reproduces this issue's symptom verbatim: "the same source compiled twice by the same binary emitted the same arms naming DIFFERENT
perry_method_*callees per position."Backed by sabotage-tested coverage in
codegen/emission_order_tests.rsand a 126-program zero-diff A/B inchangelog.d/7625-codegen-determinism.md— which matters here more than usual, since nondeterminism is precisely the class of bug that a single passing run cannot rule out.