1.0.7 (2026-10-14)

Changes since 1.0.6. The public API and ABI are unchanged: existing code compiles as before, and abidiff against 1.0.6 reports no removed or changed functions. A few fixes do change behaviour that a program may have relied on; they are listed first.

Behaviour changes

String key order on x86. dict_str_cmp now compares bytes as unsigned char, exactly as strcmp() does. On x86, x86_64 and i386, keys containing bytes >= 0x80 (for example UTF-8 text) therefore sort differently than before: "é" now sorts after "a". This affects iteration order, dict_select() positions and the neighbours found by the near searches; lookups are unaffected. Nothing changes on ARM, PowerPC or s390x, which already behaved this way.
hb_itor_search() is an exact search. It used to find the first key greater than or equal to the target. dict_itor_search() on an hb_tree was already exact; use hb_itor_search_ge() for the old behaviour.
Hash table iterators detect invalidation. After an insert that grows the table, an explicit resize, or dict_clear() (and, for hashtable2, after any removal), every iterator over the table reports itself invalid: dict_itor_valid() is false and dict_itor_next() returns false. Loops that modify a hash table while iterating now stop instead of silently skipping or repeating elements, or reading freed memory. Reposition with dict_itor_first() or dict_itor_search(). Inserting another key without a resize, and (for hashtable) removing another element, keep iterators valid as before. The rules for all containers are documented under "Iterator invalidation".
Inserting an existing key never grows a hash table. Previously the resize check ran before the lookup.
vtable slots hold internal wrapper functions. Calling container functions through incompatible function pointer types was undefined behaviour. As a result, comparing a vtable slot with a public function (for example dct->_vtable->insert == hashtable_insert) is now always false; compare the _vtable pointer with that of a dict of the known type instead.

Fixed

dict_traverse() on an hb_tree with two or more keys crashed.
dict_select() with a NULL key or datum argument crashed on sp_tree and tr_tree.
skiplist_new() with max_link 0 or 1 lost every element and leaked it; max_link is now raised to 2.
wb_tree rebalancing broke above about 6 million elements (32-bit overflow).
A degenerate splay tree or treap (about 175,000+ nodes in a chain) crashed the *_path_length functions and sp_tree_verify/tr_tree_verify with a stack overflow; they are now iterative.
On 32-bit targets, creating or resizing a hash table with about 2^30 slots or more allocated too little memory and wrote past it; it now fails.
hb_tree and rb_tree now abort with a message, also in release builds, if the allocator returns misaligned memory, instead of corrupting the tree.
Undefined behaviour: __builtin_ctz(0) in the skip list, relational comparison of unrelated pointers in dict_ptr_cmp, and comparisons of pointers to freed nodes in several trees.
dict_verify() now catches corruption it used to accept: wrong black height in rb_tree, a subtree on the wrong side of an ancestor (all trees), a wrong element count, and out-of-order skip list links. It prints the failed check to stderr.
dict_itor_compare() on a hash table evaluates to 0 instead of calling a NULL function pointer; hashtable2_itor_valid() no longer aborts after a removal.
sp_tree now splays when inserting an existing key, preserving its amortized O(log n) bound; the rotation_count statistics of sp_tree and pr_tree count single and double rotations consistently.
demo, anagram and benchmark: crashes on a bare search command and on blank or NUL-led input lines, an overlapping strcpy, negative <ctype.h> arguments, printf format mismatches, unchecked allocations, and memory leaks. benchmark now skips blank input lines; anagram skips words with ten or more of one letter.
util.h could not be included by two translation units.

Documentation

Every function of the public headers is documented, and the published Doxygen documentation now includes it (header comments were not extracted before) with the README as its main page.
The README described dict_insert_result.inserted backwards; it is true when the key was inserted.
New: iterator invalidation rules for every container, the comparison function contract, treap priority direction (max-heap), and usage notes in the README. Many wrong comments were corrected.

Build and tests

The GNU makefile works again (its recipe tabs had been replaced by spaces) and builds warning-free with clang -Weverything.
The test suite grew from 13 to 41 tests, including checks of every container for traversal, selection and iterator behaviour. Slow tests (a 6-million-element tree) run only with LIBDICT_SLOW_TESTS=1.

1.0.6 and earlier

See the GitHub releases and the git history.