Skip to content

A header-only default template-argument change reads NO_CHANGE at L2 on castxml, and castxml records no default (2026-10-21)

Status (2026-10-25): PARTIAL — every modelled backend and template family either reports the change or names the gap; clang alias and variable templates are a named gap, not a finding, and the textual scan still misses macro-generated or indirectly-included ones. Function-template defaults (template<class T = int> void f();) are recorded on the clang pattern (schema v64; see the snapshot format) and read as the same template_param_default_changed kind. A clang function template whose default cannot be spelled (for example sizeof(T)) now records the defaults as UNSUPPORTED, and the assurance rollup counts function-template patterns too, so that case reads partial with template-parameter defaults (a default could not be spelled) (fixed 2026-10-24; before, it read complete). castxml cannot do better than the gap: probed with castxml 0.7.0, its XML has no element for an uninstantiated class, function, alias or variable template, and an instantiation keeps only the spelling the code wrote (Box<char>), so no default can be recovered (option of extracting it from castxml rejected). The clang backend models neither alias templates (template<class T, class U = T> using P = ...) nor variable templates. Since extractor generation 23 (2026-10-25) a clang run scans its public headers for an alias or variable template with a default template argument and names template-parameter defaults (the clang backend does not model alias or variable templates), so such a change reads partial instead of complete NO_CHANGE. Residual: no finding is reported for it on clang (no model entity carries the default, so an unchanged default reads partial too), and the textual scan misses a macro-generated alias or variable template. castxml's header scan names the gap for them as well. tests/test_template_default_assurance_gap.py pins both sides. ADR-085 (Accepted) landed D1 and D2: the clang backend records the defaults (schema v62; see the snapshot format), and a header-only default change is the API_BREAK kind template_param_default_changed on clang. The castxml side reports the gap but not a finding: castxml never emits the template pattern, so its verdict reads NO_CHANGE. The ADR-085 D3 header-text scan (schema v63) names the gap, and the run reads partial. The assurance rollup also reports the gap for a pattern that exists with unrecorded defaults (a pre-v62 snapshot or a non-type parameter). The ABI break is caught at L2 whenever the default-substituted type reaches an exported instantiation or a public signature. The one shape L2 does not report is a default change on a template that nothing public names and nothing exports.

Repro (header only, no .so):

// v1.h                                  // v2.h
template <typename T, typename U = T>    template <typename T, typename U = int>
struct Box { T t; U u; };                struct Box { T t; U u; };

abicheck dump -H v1.h / -H v2.h then abicheck compare reports API_BREAK (template_param_default_changed) with compile.frontend: clang, and NO_CHANGE with compile.frontend: castxml, whose analysis is partial with the gap note template-parameter defaults (castxml cannot record them) on each side whose header declares a default (ADR-085 D3; for case224 that is the new side only). The L4 source replay reports public_dependency_role_changed (risk, default_template_arg → field), not default_template_arg_changed.

Why each backend cannot report it:

  • CastXML emits no record at all for an uninstantiated class template, so there is no entry to carry a default argument. tests/test_default_template_arg_l2_matrix.py pins this.
  • Clang emits the template pattern (is_template_pattern) and its TemplateTypeParmDecl.defaultArg. As of ADR-085 D1 the clang backend records it (RecordType.template_param_defaults_fact, schema v62), so this shape is no longer a clang gap.

An earlier design, a clang-only BREAKING detector, was rejected: a header-only edit has no binary evidence, and castxml could not reproduce the verdict. The shipped design is API_BREAK, which is the source-compatibility verdict ADR-084 defines for a header-level break. castxml now reports the same shape as a coverage gap, so neither backend is silent for it. Their verdicts still differ for the same header.

Fix (decided in ADR-085, Accepted): the default is a Fact on the template pattern (schema v62). The header-only kind is TEMPLATE_PARAM_DEFAULT_CHANGED with verdict API_BREAK (not DEFAULT_TEMPLATE_ARG_CHANGED, which stays BREAKING for exported instantiations). The detector fires only when both sides recorded the default. Decided (ADR-085 Remaining item 1, 2026-10-23): castxml's verdict stays NO_CHANGE for this shape and its status reads partial with the gap named. Weaker evidence narrows the conclusion and is reported; it does not fabricate a finding.

Index: Known gaps