multidict 7.0.0 on Python PyPI

7.0.0 is a major release. It makes istr final, adds a
public C API for third-party extensions, and makes the C extension
substantially faster across the board.

Breaking change. istr can no longer be subclassed,
on either backend. Code that subclassed it has to wrap or convert instead.
The C extension also started rejecting an argument passed both
positionally and by name with TypeError, where it used to silently
misbehave.

Public C and Cython API. Other C extensions and Cython modules can now
create, read and mutate multidicts through a capsule, without going through
the Python-level API. The capsule comes with multidict_capi.h, a
cimport-able multidict/__init__.pxd and
multidict.get_include(). The API includes MultiDict_ForEach() and
a watchers API modeled on CPython's dict watchers. See the C API and
Cython API references.

Performance. On CodSpeed's GIL-build benchmarks against 6.9.1, 82
C-extension benchmarks got faster and none got slower. Rough figures:

CIMultiDict keyed by plain str: construction,
add(), extend(), update() and item assignment got 2 to 3.8
times faster, and lookups about 2 times faster.
CIMultiDict keyed by istr, and
case-sensitive MultiDict: lookups, insertion and
deletion got 10% to 70% faster.
getall() and iterating
items() got about 35% faster, and the view set
operations 10% to 85% faster.
popitem() on a whole mapping went from
quadratic to linear time.
On the free-threaded build, measured in instructions against 6.9.1,
lookups got 24% to 33% cheaper on MultiDict and 4 to 5
times cheaper on CIMultiDict with lowercase str
keys, and item assignment 15% to 19% cheaper. Construction and deletion on
MultiDict cost 2% to 7% more. With several threads
mutating at once, a d[key] = value got about four times faster.

The pure-Python backend is unchanged in speed.

Robustness. This release fixed several crashes and leaks in the C
extension, including use-after-free bugs when a key's lower() or a
value's __eq__ mutated a multidict, a crash at interpreter shutdown,
and table and reference leaks on the free-threaded build.

Bug fixes

Fixed a segmentation fault in the C extension when a finalizer let the
same multidict be resized while del md[key] or
popall() was still scanning the table, and a
spurious RuntimeError from getall()
when a garbage collection triggered while building its result mutated
the multidict -- by @asvetlov.
Related issues and pull requests on GitHub:
#1492, #1512.
Fixed a reference leak of the key and value in the C extension when
growing the hash table fails with MemoryError during
add(), update(),
merge(), setdefault()
or item assignment -- by @asvetlov.
Related issues and pull requests on GitHub:
#1507.
Fixed a data race on the free-threaded build of the C extension where a
MultiDict's own version counter was read by
getversion() and written on every mutation with a plain,
non-atomic load and store. Both sides switched to a relaxed atomic
load/store, matching the pattern already used for the object's item
count -- by @asvetlov.
Related issues and pull requests on GitHub:
#1514.
Fixed the C extension discarding a MemoryError or
KeyboardInterrupt raised by a positional argument's __len__
while the size of an update was estimated. extend(), update(),
merge() and subclass construction started propagating it, matching the
pure-Python backend; only the TypeError from an unusable
__length_hint__ stayed ignored -- by @asvetlov.
Related issues and pull requests on GitHub:
#1517.
Ensured that canonical keys in pure-Python MultiDict and
CIMultiDict are converted to exact str instances
without invoking object.__str__() overrides, matching the C extension --
by @agustin18.
Related issues and pull requests on GitHub:
#1535.
Fixed the argument counts reported by the C extension when a method is
called with too many positional arguments; methods with an optional
second argument claimed to take exactly one, and methods requiring two
claimed to take a range. A method called with none of its two required
arguments now reports both as missing -- by @asvetlov.
Related issues and pull requests on GitHub:
#1545.
Fixed the C extension accepting an argument passed both positionally and
by name instead of raising TypeError. md.add("a", key="k", value="v") stored a wrong value, and md.get("a", "b", key="c")
silently ignored the surplus arguments -- by @asvetlov.
Related issues and pull requests on GitHub:
#1545.
Fixed a memory leak in the free-threaded build of the C extension.
clear() hands its hash table to a drain that
frees it once no lock-free reader is in flight, and a drain that finds
one leaves the table for the next drain to free. Deallocation ran no
drain of its own, so a multidict dropped after such a clear, with no
operation in between, never freed that table nor released the references
its entries still held. Builds with the GIL enabled emit byte-identical
code and are unaffected -- by @asvetlov.
Related issues and pull requests on GitHub:
#1555, #1562.
Fixed a crash during interpreter shutdown in the C extension: a
multidict started keeping a reference to its own module, so the module
state it reads while being torn down could not be freed first
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1563.
Fixed retired hash tables lingering in the free-threaded build of the C
extension. A drain that finds a lock-free reader in flight puts the tables
it holds back for the next drain to run, and the reader it saw could
already have looked, so nothing was left to free them: they stayed until
the multidict was next used, and if it was dropped instead, the references
their entries held were never released. Builds with the GIL enabled emit
byte-identical code and are unaffected -- by @asvetlov.
Related issues and pull requests on GitHub:
#1565.
Fixed reference cycles through a multidict going uncollected in the
free-threaded build of the C extension. A hash table waiting to be freed
still owns the references of the entries left in it, and those were not
reported to the garbage collector, which then read the values as reachable
from outside the cycle and kept it alive. Builds with the GIL enabled emit
byte-identical code and are unaffected -- by @asvetlov.
Related issues and pull requests on GitHub:
#1566.
Fixed extend(), update() and merge() of the C extension
raising SystemError when given an object whose
__length_hint__() reported close to sys.maxsize together
with keyword arguments, and overflowing the table size computation for
such a hint; a hint too large to reserve for is ignored, as
list.extend() does -- by @asvetlov.
Related issues and pull requests on GitHub:
#1577.
Fixed a use-after-free in the C extension's MultiDict.__eq__
when a value's __eq__ mutated either multidict; comparing
exact str values also became faster -- by @asvetlov.
Related issues and pull requests on GitHub:
#1579.
Fixed a use-after-free in the C extension when a str subclass
key's lower() mutated the source while a
CIMultiDict was built from, extended, updated or
merged with a dict or a multidict of the other case
sensitivity. The extra references this takes are limited to the keys
whose lower() can run Python code, so other updates do not pay for
them -- by @asvetlov.
Related issues and pull requests on GitHub:
#1581, #1598.

Features

Added a public C and Cython API for third-party extensions. Its
MultiDict_ForEach() family hands a visitor each entry's key, value,
identity (the canonical key the mapping looks entries up by, lower-cased
on a CIMultiDict) and the identity's hash
-- by @Vizonex and @asvetlov.
Related issues and pull requests on GitHub:
#1478.
Changed the C extension to track the entries that
getall(), item assignment,
update(), merge()
and to_dict() have visited in a bitmap private
to the call, instead of temporarily marking entry hashes in the table
itself. Concurrent readers on free-threaded builds no longer have to
tolerate marked entries. Item assignment,
to_dict() and lookups of keys with many values
got faster, while update() with many repeated
keys got somewhat slower
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1505.
Replaced the C extension's per-key finder object with a single
callback-driven walk in a new _multilib/walk.h header. The entries a
walk has already reported were tracked by the walk itself rather than
by every caller, and the walk kept its position in registers instead of
reloading it from a struct on each step, so
getall() and the items and keys view set
operations got faster on keys with many values -- by @asvetlov.
Related issues and pull requests on GitHub:
#1522.
Sped up CIMultiDict operations keyed by a plain str in the C
extension. An all-ASCII key was lowered directly instead of through a
str.lower() call, and a key that is already lowercase became its own
identity, which saved both a copy and a rehash. Looking up a lowercase key
got about three times faster and inserting one about twice as fast, keys
that do carry uppercase gained 20% to 35%, and bulk operations on
already-lowercase keys, such as extend() and
construction from a dict, got up to three times faster
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1527.
Sped up building an istr in the C extension. The class gained its own
vectorcall, so the one-argument form stopped packing its argument into a
tuple for type_call() only to unpack it again with
PyArg_ParseTupleAndKeywords(). Building an istr from a
str got about 19% faster, and building one from an istr,
which hands back the original, about three times faster. In exchange
getall() costs nine more instructions per
call, since the added code moves where the compiler draws its inlining
line -- by @asvetlov.
Related issues and pull requests on GitHub:
#1537.
Sped up building and dropping multidicts in the C extension by reusing
freed hash tables instead of returning them to the allocator. Small
tables are kept in a bounded per-size-class pool held in the module
state, the way CPython reuses key objects for dict. Only on
builds with the GIL: a shared pool needs an atomic exchange to take a
table and another to put one back, which measured slower than the
per-thread allocator a free-threaded build already has
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1539.
Sped up copy(), constructing a multidict from
another one, and growing one past its current size in the C extension, by
not zeroing the parts of a hash table that are about to be overwritten. A
copied table is overwritten in full, and a resized one has the entries it
carries over written straight onto it -- by @asvetlov.
Related issues and pull requests on GitHub:
#1540.
Sped up items(),
keys(), values()
and iterating a multidict in the C extension, by reusing the view and
iterator objects they allocate instead of returning them to the
allocator -- by @asvetlov.
Related issues and pull requests on GitHub:
#1541.
Sped up constructing a MultiDict,
CIMultiDict, MultiDictProxy or
CIMultiDictProxy in the C extension, by reusing the
object itself as well as its hash table. Subclasses are unaffected and
keep allocating as before -- by @asvetlov.
Related issues and pull requests on GitHub:
#1542.
Sped up getall() in the C extension, by
shrinking the scratch buffer the walk carries on the stack from 4 KB to
1 KB. The buffer was large enough to stop the compiler inlining the walk
into its callers, which cost more than it saved -- by @asvetlov.
Related issues and pull requests on GitHub:
#1543.
Sped up lock-free reads on the free-threaded build of the C extension.
Every reader that leaves a MultiDict last used to
call into the retired-table drain, whose first act is a sequentially
consistent read-modify-write, even though the retired list is empty
almost every time; a plain load took over that case and the drain
itself moved out of line. key in md got about 10% cheaper,
md[key] about 8% and a missing md.get(key) about 5%, measured as
instruction counts on CPython 3.14. Builds with the GIL enabled emit
byte-identical code and are unaffected -- by @asvetlov.
Related issues and pull requests on GitHub:
#1546.
Added a watchers C API, so another C extension can be notified when a
MultiDict or CIMultiDict changes.
It follows CPython's PyDict_AddWatcher() family, with two
differences: the callback is passed two context pointers, one fixed per
registered callback and one per watched multidict, and events are
delivered once the operation that produced them has finished rather than
mid-mutation. The events are multi-value aware, and each one that concerns
a key carries the identity and its hash alongside it. Up to
MULTIDICT_MAX_WATCHERS (32) watchers can be registered, and a
failing callback is reported through sys.unraisablehook() with the
event and the multidict's type and address, as CPython does for dict
watchers -- by @asvetlov.
Related issues and pull requests on GitHub:
#1551.
Sped up lookups, insertions and deletions in the C extension by always
inlining the hash-table probe's initializer. Left to itself the compiler
emitted it out of line, so every probe opened with a call for five
stores. Free-threaded builds already inlined it and are unchanged
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1554.
Sped up every method call on CIMultiDict and
CIMultiDictProxy in the C extension by binding
their method descriptors to the class itself instead of inheriting
them from MultiDict and
MultiDictProxy; CPython only specializes a
method call when the descriptor belongs to the exact type of the
receiver, so every such call was taking the generic path
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1567.
Made popitem() in the C extension drop the
trailing deleted entries it walked over, so that popping every item
out of a mapping costs linear time rather than quadratic; the
pure-Python implementation already did this -- by @asvetlov.
Related issues and pull requests on GitHub:
#1568.
Sped up iterating over items() in the C
extension by handing out the same (key, value) tuple again when
the caller has already let go of it, as for key, value in d.items() does on every step; a tuple the caller keeps is left
alone and a new one is used from then on -- by @asvetlov.
Related issues and pull requests on GitHub:
#1569.
Sped up lookups on the free-threaded build by comparing the hash before
touching an entry's key and by matching a key that is the very same
object as the one looked up without taking a reference to it, so the
common hit costs two fewer atomic operations -- by @asvetlov.
Related issues and pull requests on GitHub:
#1570.
Sped up mutations on the free-threaded build when several threads
mutated multidicts at the same time: each thread reserved version
numbers in batches instead of bumping one counter shared by every
thread on each mutation, so with six threads a d[key] = value was
about four times faster -- by @asvetlov.
Related issues and pull requests on GitHub:
#1587, #1599.
Sped up the C extension by keeping the table resize on the insert path and
the operations behind Python slots out of line, so they stopped crowding the
hot paths out of the compiler's inlining budget. On GIL builds, building a
multidict from 200 pairs got about 4% faster in instruction count and
inserting a new key about 2%; on free-threaded builds, key lookups got up to
5% faster -- by @asvetlov.
Related issues and pull requests on GitHub:
#1600, #1601.

Removals and backward incompatible breaking changes

Made istr a final class to disallow subclassing in both
the C extension and pure-Python implementations -- by @agustin18.
Related issues and pull requests on GitHub:
#1535.

Improved documentation

Documented how MultiDict and
CIMultiDict compare with dict on both the
default and the free-threaded builds, with per-operation measurements,
and how to run the benchmarks in a stable environment
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1532.
Documented what istr keys are worth on a
CIMultiDict, as instruction counts next to the
equivalent plain str keys. The advice to create them once and
reuse them was already there; the numbers behind it were not
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1547.
Documented that a few rows of the benchmarking reference are measured on
an empty or 20-item mapping rather than the 200-item one the section
describes -- by @asvetlov.
Related issues and pull requests on GitHub:
#1558.
Mentioned in the README and docs index that multidict is optimized for
both the GIL and free-threaded builds, and that its performance is typically
within 20-30% of dict for common operations, with a link to the
benchmarks page for the full breakdown -- by @asvetlov.
Related issues and pull requests on GitHub:
#1559.
Fixed the C extension's popone() and pop() docstrings, which
claimed the last occurrence of the key was removed; the first one is,
as on the pure-Python backend. Added a note on rejecting duplicate
values for keys that must carry exactly one, such as security-sensitive
HTTP headers -- by @asvetlov.
Related issues and pull requests on GitHub:
#1578.
Added a C API example of a watcher that invalidates a cache: it unwatches
from its own callback, watches again on the next rebuild, and gives up
on a multidict that keeps changing -- by @asvetlov.
Related issues and pull requests on GitHub:
#1586.
Refreshed the tables in the benchmarking reference for the 7.0.0 release,
and replaced the range quoted for the pure-Python backend with its average
slowdown -- by @asvetlov.
Related issues and pull requests on GitHub:
#1602.

Packaging updates and notes for downstreams

Removed the C extension code path that existed only for free-threaded
Python 3.13, whose support was dropped in v6.8.0. The free-threaded
build started requiring CPython 3.14 or newer and failing with a clear
compiler error on older free-threaded interpreters; builds with the GIL
enabled were unaffected and still go down to Python 3.10
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1524.

Contributor-facing changes

Added CodSpeed benchmarks for getall() on a large table, update()
with duplicate keys, merge(), to_dict() and items view containment
with duplicate keys -- by @asvetlov.
Related issues and pull requests on GitHub:
#1506.
Added CodSpeed benchmarks for getall() on keys with many values, both
in a small table and in a table large enough to need a heap allocated
bitmap -- by @asvetlov.
Related issues and pull requests on GitHub:
#1509.
Moved the compiler hint macros used by the C extension out of
htkeys.h into compiler.h and dropped the HT_ prefix from
their names -- by @asvetlov.
Related issues and pull requests on GitHub:
#1510.
Moved the debug-only ASSERT_CONSISTENT macro and the
_md_check_consistency() and _md_dump() helpers used by the C
extension out of hashtable.h into their own debug.h header
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1513.
Added CodSpeed benchmarks for CIMultiDict keyed by plain str instead
of istr, covering already-lowercase, mixed-case and non-ASCII keys
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1521.
Documented that the pure-Python test leg needs --no-c-extensions
alongside MULTIDICT_NO_EXTENSIONS=1; the environment variable picks
the backend that gets imported, but only the flag deselects the
C-extension half of the test matrix -- by @asvetlov.
Related issues and pull requests on GitHub:
#1523.
Documented how to run the Hypothesis property tests, which a default
pytest run excludes via the -m "not hypothesis" marker
expression in pytest.ini and whose pin make install-dev does
not install -- by @asvetlov.
Related issues and pull requests on GitHub:
#1525.
Added tools/codegen_diff.py, which reports the functions GCC emits
differently before and after a change to the C extension. Addresses,
RIP-relative displacements and branch offsets inside a symbol are
normalized away first, so a single instruction-length change no longer
makes every later function look modified, while constants and structure
offsets are left alone so that a real change is not folded away. A tree
that fails to build is reported by name rather than compared
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1529.
Fixed benchmarks/istr.py to import pyperf under its current
name; it had been failing to start since the package was renamed from
perf -- by @asvetlov.
Related issues and pull requests on GitHub:
#1531.
Rewrote the benchmark scripts around a single registry of benchmarked
operations, and added a Callgrind driver that records deterministic
instruction counts per operation. The measurements of destructive
operations such as pop() and clear() stopped counting the cost
of rebuilding the mapping -- by @asvetlov.
Related issues and pull requests on GitHub:
#1532.
Moved the bulk extend(), update() and merge() paths of the C
extension out of hashtable.h into their own bulk_update.h
header, leaving hashtable.h to the table itself and the single-key
operations -- by @asvetlov.
Related issues and pull requests on GitHub:
#1533.
Moved the two list-reading helpers behind itemsview.__contains__() and
the extend()/update()/merge() sequence parser out of
hashtable.h and into a new multidict/_multilib/unpack.h, together with
a shared unpack_pair() that both call to read a (key, value)
pair out of an exact two-element tuple or list. C extension only; the pure
Python implementation has no header layout to mirror
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1534.
Added a MULTIDICT_NO_FREELIST=1 build option that stops the C
extension reusing freed blocks. A reused block never reaches free(),
so AddressSanitizer can neither poison it nor report a use-after-free on
it; an instrumented run wants a second pass with this set, as
AGENTS.md describes.
Also added a private _freelist_clear() to the C extension, for tests
that inject an allocation failure: a reused block lets an operation run
without calling the allocator at all, which would put the recovery path
out of reach -- by @asvetlov.
Related issues and pull requests on GitHub:
#1539.
Added a visitor to the public C API test harnesses that mutates the multidict
it is walking, covering the reentrancy guard in MultiDict_ForEach() from
both multidict._testcapi and multidict._testcyapi
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1544.
Added benchmarks for the shapes where an allocation is most of the work:
an empty mapping, a 20-item one built or copied, a view or an iterator
taken on its own, and a proxy. The allocation an operation makes is a
fixed cost, so at the existing 200-item size it was divided across 200
entries and all but vanished; d.items() measured 1.7% faster where
it is 46% faster -- by @asvetlov.
Related issues and pull requests on GitHub:
#1556.
Sorted docs/spelling_wordlist.txt case-insensitively, and documented
in AGENTS.md that it is to be re-sorted rather than appended to. The
list had grown in append order, which made a word hard to find and put
every PR adding one on the same last line -- by @asvetlov.
Related issues and pull requests on GitHub:
#1557.
Stopped asking every performance pull request to regenerate the tables in
the benchmarking reference, since each run shifts every row slightly and the
churn hides the rows that moved, and moved the refresh to once per release.
Added RELEASE.md, which documents the release procedure, including
that refresh -- by @asvetlov.
Related issues and pull requests on GitHub:
#1564.
Renamed the C extension's internal helpers in the _multilib headers
so that a leading underscore marks exactly the functions used only
within their own header; type slots and method callbacks keep their
names -- by @asvetlov.
Related issues and pull requests on GitHub:
#1571.
Spelled out in AGENTS.md that every verb in a news fragment is
past tense, and added a bad and good example pair plus a matching
"Things not to do" entry -- by @asvetlov.
Related issues and pull requests on GitHub:
#1588.
Changed the release workflow to rewrite the Sphinx roles of the changelog
section as Markdown before publishing it as the GitHub Release body, so
that class, method and exception references, contributor handles and
documentation links render there instead of showing as raw markup
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1596.

Miscellaneous internal changes

Changed the internal htkeys_build_indices() helper in the C
extension to return void; it cannot fail, so both callers were
checking for an error that never occurred -- by @asvetlov.
Related issues and pull requests on GitHub:
#1516.
Routed every field of the C extension's hash table that more than one
thread can reach through a named accessor in the new
_multilib/freethreading.h header, instead of each call site choosing
between an atomic and a plain access behind #ifdef Py_GIL_DISABLED.
Each accessor kept the operation and memory ordering its call sites
already used, so neither build changed behavior -- by @asvetlov.
Related issues and pull requests on GitHub:
#1518.
Removed get_mod_state_by_cls(), an unused helper in the C extension
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1520.
Documented why md_clear() keeps clearing entry fields with
Py_CLEAR() on the GIL build rather than routing them through
_multilib/freethreading.h: Py_CLEAR() skips the store when a
field is already NULL, so no accessor there is a substitute for it
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1526.
Collapsed the two #ifdef Py_GIL_DISABLED arms that fill in a freshly
inserted hash table entry in the C extension into one. Both builds started
writing the entry's fields in the order the free-threaded build needs, and
publishing the value without first loading and releasing the one already
there, since a fresh entry holds none -- by @asvetlov.
Related issues and pull requests on GitHub:
#1528.
Merged the C extension's deferred-decref accumulator into reflist_t,
so one collector of strong references, in the _multilib/reflist.h
header, came to serve both the callers that build a result list and the
ones that postpone a decref until a mutation finishes. Its storage became an
inline array that spills into a chain of heap blocks, which also made
getall() a little faster on keys with many
values, since the collected values stopped being copied to a bigger
buffer as they accumulate -- by @asvetlov.
Related issues and pull requests on GitHub:
#1530.
Dropped two module state lookups in the C extension that had already been
answered elsewhere. MultiDict.__init__() and CIMultiDict.__init__()
switched to reading the state pointer that tp_new() stored on the object instead
of resolving it from the type a second time, and an istr stopped
carrying a state pointer of its own, which was written on every creation
and never read. Constructing a subclass of either container got about 1%
cheaper, and every istr is eight bytes smaller
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1536.
Narrowed the critical section of the single-key methods (add(),
__setitem__(), __delitem__(), setdefault(), pop(),
popone(), popall() and getall()): a key is converted to its
identity and hashed before the lock is taken, rather than while a
free-threaded build holds the multidict locked -- by @asvetlov.
Related issues and pull requests on GitHub:
#1538.
Changed MultiDict_ForEach() to walk every entry with a linear scan
instead of the resumable iterator cursor, which its caller does not need while
it holds the critical section for the whole walk -- by @asvetlov.
Related issues and pull requests on GitHub:
#1544.
Sped up the argument parsing of the C extension. The all-positional form
of every method taking key started binding its arguments inline instead of
calling into the keyword machinery, and keyword names were matched against
interned parameter names by identity before falling back to a string
comparison -- by @asvetlov.
Related issues and pull requests on GitHub:
#1549.
Stopped building the implicit None default of MultiDict.get()
on every call in the C extension, and produced it only when the key is
actually missing -- by @asvetlov.
Related issues and pull requests on GitHub:
#1549.
Laid out the common branch in line in the single-key lookup probe loop and in
the key to identity conversion: the deleted-slot check and the str
subclass check were marked as the rare cases, so the compiler stopped putting
the frequent path behind a taken branch -- by @asvetlov.
Related issues and pull requests on GitHub:
#1550.
Stopped building the implicit None default of
MultiDict.setdefault() on every call in the C extension -- by
@asvetlov.
Related issues and pull requests on GitHub:
#1553.
Reordered the fields of the C extension's multidict object so the ones a
lookup reads share a cache line, and dropped its separate module
reference in favor of the one the module state already holds, which
made every multidict 16 bytes smaller -- by @asvetlov.
Related issues and pull requests on GitHub:
#1584.
Dropped the C extension's multidict_tp_traverse() and
multidict_tp_clear() wrappers; md_traverse() and md_clear()
now serve as the type's GC slots directly -- by @asvetlov.
Related issues and pull requests on GitHub:
#1589.
Removed a redundant NULL check and a stray PyErr_Format()
argument from the C constructors, and an always true condition from the
pure Python getall() -- by @asvetlov.
Related issues and pull requests on GitHub:
#1590.
Merged the duplicated __init__() bodies of the C MultiDict and
CIMultiDict into one helper -- by @asvetlov.
Related issues and pull requests on GitHub:
#1591.
Replaced the allocate, set state and take the module reference sequence
repeated across the C constructors with one helper -- by @asvetlov.
Related issues and pull requests on GitHub:
#1592.
Renamed internal helpers whose names no longer matched what they did,
and fixed misspelled docstring identifiers in the C extension
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1593.
Renamed the C extension's functions that are used only as type slots
after the slot they fill, as <type>_tp_<slot>, <type>_nb_<slot>,
<type>_sq_<slot> or <type>_mp_<slot> -- by @asvetlov.
Related issues and pull requests on GitHub:
#1594.
Replaced the custom tp_new slots that blocked direct instantiation of
the C view and iterator types with Py_TPFLAGS_DISALLOW_INSTANTIATION
-- by @asvetlov.
Related issues and pull requests on GitHub:
#1595.