Skip to content

case217 — WEAK export loss demoted to a risk by a binding: weak rule

Field Value
Verdict 🔴 BREAKING
Category Breaking
Classification Rule
Platforms Linux
Flags ABI break
Detected ChangeKinds func_removed_elf_only
Source files catalog/cases/case217_export_loss_weak_binding_demoted/
Rule family export-loss-weak-binding-demoted
Subject ELF export and linker metadata

Verdict and consumer impact

NEW drops three WEAK exports that OLD had: a header-declared inline function (ns::hdr_inline), a WEAK export no public header declares (ns::weak_undeclared), and the ns::detail::P<float> constructor from an internal template namespace. Without a policy, abicheck reports them as BREAKING (func_visibility_changed, func_removed_elf_only) and exits 4.

These are copies the consumer's own translation units emit as well, so the loss is a risk, not a hard break. The policy in policy.yaml says so with a binding: weak selector scoped to the library's namespace. With that policy abicheck reports COMPATIBLE_WITH_RISK and exits 0.

This is the case that the binding: selector did not reach before PRs #1629 and #1630: the findings carried no symbol_binding, so the rule never matched.

Old/new diff

v1 (OLD) v2 (NEW)
hdr_inline() called from Pub::g(), emitted WEAK Pub::g() no longer calls it, not emitted
weak_undeclared() __attribute__((weak)), not in any header not defined
detail::P<float> ctor instantiated by Pub::q() Pub::q() returns nullptr, not instantiated
strong_fn() GLOBAL, declared GLOBAL, unchanged

abicheck command

Dump both sides from headers (L2), delete the OLD binary, then compare with the policy:

abicheck dump libv1.so -H v1.h -o v1.json
abicheck dump libv2.so -H v2.h -o v2.json
rm libv1.so
abicheck compare v1.json v2.json --policy policy.yaml

Expected abicheck finding

  • Verdict: COMPATIBLE_WITH_RISK, exit code 0.
  • Without --policy: BREAKING, exit code 4.
  • Kinds (both runs): func_visibility_changed on hdr_inline, func_removed_elf_only on weak_undeclared and the P<float> constructor. Each carries symbol_binding: "weak". Under the policy each is reclassified to risk.

Minimum evidence

min_evidence: L0: the expected kind (func_removed_elf_only) is detectable from the exported symbol table alone, and the gate derives the floor from it. func_visibility_changed on hdr_inline appears only when headers are supplied; the binary-only run reports the same symbol as func_removed_elf_only.

The walkthrough above still passes -H, because the verdict needs headers. Without -H the run also reports visibility_leak and public_surface_shrank, which the binding rule does not cover, so the verdict stays BREAKING there.

Why abicheck catches it

The ELF .dynsym tables of both sides give the removals and their WEAK binding (elf_metadata.py). The header AST (castxml) tells the detector that hdr_inline is still declared, so the removal is classified as a visibility change instead of an API removal. compare/undeclared_exports.py and the template detectors stamp the old-side binding on each finding (Change.symbol_binding); the binding: selector matches on that field (policy/selectors.py, EXPORT_LOSS_KINDS).

Sibling

case218_export_loss_strong_binding_stays_breaking uses the same policy.yaml and the same WEAK losses, plus the removal of a GLOBAL export. Its verdict stays BREAKING, which shows the selector does not demote strong exports.

Source files

  • policy.yaml
  • v1.cpp
  • v1.h
  • v2.cpp
  • v2.h

See also: Compatibility Catalog · All BREAKING cases · Category: Breaking · Rule: WEAK export loss demoted by a binding rule · Subject: ELF export and linker metadata.