<wchar.h> / <wctype.h> were declarations with no implementation

• Type: bug — Track B (library / crtl), tag compat (gcc/glibc parity)
• Status: done
• Resolved: 2026-08-05 in 570e72de3 (verified on origin/master after the rebase)
• Opened: 2026-08-05
• Found by: sweeping all 317 buildable test/c*.c for a spurious DT_NEEDED while fixing [[bug-cfront-spurious-dt-needed-libc-with-no-imports]] — test/cwide_string_literal was importing wcslen.

Symptom

lib/crtl/include/wchar.h and wctype.h declared wcslen, towlower, towupper and the twelve isw* predicates. There was no src/wchar.c and no src/wctype.c, so nothing was ever pulled, every prototype stayed external, and each call became a glibc dynamic import.

Same shape as the socket bug found alongside it, and it hides the same way: on a glibc host the program works, so the only visible symptom is that the binary is dynamically linked — which nothing was checking.

The interesting part is what happens above 127

crtl is C-locale-only, and it would be easy to treat that as a limitation to apologise for. It is not: glibc's C locale answers FALSE for every one of these predicates on every value above 127, and towlower/towupper are the identity there. Read off a gcc build over the whole range before writing a line:

So the ASCII range is not a subset being settled for — it is the answer. An implementation that tried to be locale-aware above 127 would be wrong here.

Fix

• lib/crtl/src/wchar.c — wcslen (a plain scan; wchar_t is 32-bit, so no encoding is involved), towlower, towupper.
• lib/crtl/src/wctype.c — the twelve predicates, plus wctype()/iswctype(). The wctype_t tags are private: C promises only that a wctype() result is meaningful to iswctype(), so they need not match glibc's and the test compares behaviour rather than the number.

No header changes were needed: <wctype.h> already includes <wchar.h> first, so src/wchar.c is pulled before src/wctype.c and there is none of the guard cycle that the socket veneer ran into.

Test

test/cwctype.c, in lib-test, with no recorded expectations — the whole output is diffed against a gcc build of the same file. It covers WEOF through 255 against all twelve predicates plus both case functions, and four wide values a locale-aware libc would classify. Identical to gcc on all 262 lines, and on i386, arm32, aarch64 and riscv32 as well.

lib-test also asserts the linkage, because the output diff passes either way on a glibc host — exactly the blind spot that let this sit.

Sweep result

All 318 buildable test/c*.c now build statically except two, and neither is a defect: crtl_libc_oracle links libc by design (it is the oracle this surface is measured against) and cquickjs_prereq needs --threadsafe ([[bug-c-pthread-without-threadsafe-builds-then-dies-at-load]]).

Gate

tools/gate.sh lib GREEN; c-conformance i386 219 pass / 0 fail (1 known VLA skip) — the cross re-run any crtl change owes, since crtl compiles for every target while lib-test is x86-64 only.