Case 95: Allocator Nested-Typedef Removed¶

Field	Value
Verdict	🔴 BREAKING
Category	Breaking
Platforms	Linux, macOS, Windows
Flags	API break
Detected `ChangeKind`s	`typedef_removed`
Source files	browse on GitHub

Category: Source API contract | Verdict: 🔴 BREAKING

What breaks¶

An allocator-style class drops its historical nested typedef set (value_type, pointer, reference, size_type, difference_type). The exported member functions keep their mangled names, so the .so symbol table is unchanged and previously-linked binaries still load. Every consumer source TU that wrote typename Alloc::value_type (or participates in STL-style generic code that does) fails to compile against v2 headers.

Why this fires "noise cannon"-style¶

DWARF stores nested typedefs by their bare name (value_type), not qualified by the containing class. A diff therefore emits TYPEDEF_REMOVED once per removed alias name across the entire library. For a real allocator + container stack, removing the same five aliases from allocator<T>, vector<T>::allocator_type, and friends shows up as many findings — even though it's a single coherent design change.

How abicheck catches it¶

The existing TYPEDEF_REMOVED detector fires for each removed alias. That detection is correct on its own; the noise problem is policy.

The new member_name suppression selector (added in this batch) lets maintainers control that noise with a single rule:

version: 1
suppressions:
  - member_name: "(value_type|pointer|reference|size_type|difference_type)"
    change_kind: typedef_removed
    reason: "Modernization — STL nested-typedef set removed."
    expires: 2026-12-01

member_name fullmatches the last ::-segment of change.symbol. Combined with change_kind: typedef_removed, it scopes the rule tightly: non-typedef changes are not affected, and only the listed aliases are suppressed.

Combine with type_pattern for even tighter scoping when the host class contributes the qualifier (e.g. when typedef symbols are emitted as Allocator::value_type rather than the bare alias).

Code diff¶

v1	v2
`typedef int value_type;` and four more nested aliases	(removed)
`int* allocate(size_type n);`	`int* allocate(unsigned long n);`

How to fix (as a library maintainer)¶

If consumers depend on STL-style allocator traits, keep the typedef set even when modernizing the public API.
If you really want them gone, ship a deprecation cycle: in release N–1, alias the old names to the new types with [[deprecated]]; in release N, remove them.

Real failure demo¶

# v1 header, v1 .so:
g++ -std=c++17 -I. app.cpp -L. -lmylib -o app   # compiles and runs

# v2 header, v2 .so: app.cpp's `my_allocator::value_type` is undefined.
g++ -std=c++17 -I. app.cpp -L. -lmylib -o app
# → error: 'value_type' is not a member of 'mylib::my_allocator'

References¶

C++ allocator traits — historical nested types
See tests/test_suppression.py::test_member_name_* for the selector's test coverage.