perf: run pyperformance in CI and close the gaps it finds - #8663

The dispatch loop wrote prev_line (a locations lookup plus a Cell store)
on every instruction only so f_lineno could read it. f_lineno now derives
the line from lasti and code.locations, prev_line is only maintained on
the cold tracing path for 'line' event de-duplication, the two use_tracing
reads per instruction are merged into one, and the pre-dispatch lasti
reload is replaced by the value just stored.

When a trace function is installed on an already-running frame (bdb
stepping back into an untraced caller, frame.f_trace assignment, a local
trace returned from the 'call' event on a resumed generator) prev_line is
synced from lasti so no spurious 'line' event fires.

Assisted-by: Claude Code:claude-fable-5-1

set_current_frame ran on every frame enter/exit and touched three separate
thread_local! statics (each a _tlv_get_addr call on macOS). Merge them into
a single thread-local struct so the hot path does one TLS access.

Assisted-by: Claude Code:claude-fable-5-1

eval_breaker_tripped ran three checks per bytecode: the finalizing flag,
a thread-local stop_requested lookup, and the EVAL_BREAKER load. Add
STOP_BIT and FINALIZING_BIT so the per-instruction check is a single
relaxed atomic load; the slow path (check_signals) still re-derives the
exact per-thread answer. The bits are set alongside the per-thread flags
and cleared under the stop-the-world exclusion.

Assisted-by: Claude Code:claude-fable-5-1

gc_bits is zero for a fresh allocation (both the freelist reuse path and
PyInner::new zero it), so track_new_object can initialise the bit with a
relaxed store instead of a fetch_or. Re-tracking resurrected objects keeps
the read-modify-write variant.

Assisted-by: Claude Code:claude-fable-5-1

Generator/coroutine resume pushes and pops an exc_info slot on every
send. When that slot is None the value returned by topmost_exception
cannot change (it skips None entries), so the thread-local exception
publish in push_exception/pop_exception is unnecessary.

Assisted-by: Claude Code:claude-fable-5-1

Keyword calls to plain functions where every parameter is filled exactly
once are reordered in place into the fast-locals layout (no IndexMap, no
keyword name clones); any mismatch falls back to the existing slow path
so error messages are unchanged. super().attr now walks the type's MRO
by reference instead of cloning every class into two Vecs per lookup.

Assisted-by: Claude Code:claude-fable-5-1

enumerate kept its counter as a BigInt behind a write lock and did bignum
arithmetic on every item; use a machine integer that promotes to BigInt
only for huge/negative starts or on overflow. zip pre-sizes its per-item
Vec so the common two-iterable case allocates once.

Assisted-by: Claude Code:claude-fable-5-1

Task registration re-resolved _asyncio module attributes and re-imported
contextvars on every task; the eager path additionally went through generic
call_method for set add/discard. Cache the module objects per thread
(validated by identity), split the contextvars cache from the task
containers, and use PySet/PyDict methods directly.

Assisted-by: Claude Code:claude-fable-5-1

json.encoder fell back to the pure-Python _make_iterencode because _json
had no make_encoder, making json.dumps ~25x slower than CPython. Implement
the C encoder semantics natively: str/None/bool/int/float/list/tuple/dict
ordering, key coercion and skipkeys, sort_keys via the VM's fallible sort
(comparison errors propagate instead of panicking), circular reference
detection, indent handling, allow_nan, default() fallback with recursion
guard, and CPython's add_note annotations. Removes the now-passing
expectedFailure markers in test_json.

Assisted-by: Claude Code:claude-fable-5-1

Adds mimalloc (0.1.52, default-features = false) as the #[global_allocator]
for the rustpython binary, gated behind a new `mimalloc` feature enabled
by default on non-wasm targets (dep declared under
target.'cfg(not(target_arch = "wasm32"))'.dependencies, so wasm builds
never pull it in regardless of default-features).

Profiling a `t = (a, b, c)` loop showed malloc/free at ~40% of samples on
macOS with the system allocator. Switching to mimalloc gives consistent
wins across small-object-heavy workloads, medians over 5-9 interleaved
A/B rounds (release build, this machine):

  tuple (5M `(a,b,c)`)         1.020s -> 0.863s  (1.18x)
  4-thread tuple (5M total)    1.643s -> 1.382s  (1.19x)
  class instances (2M)         0.874s -> 0.797s  (1.10x)
  dict-heavy (200k x 20)       2.196s -> 1.992s  (1.10x)
  string-heavy (join x 20)     0.662s -> 0.592s  (1.12x)
  nqueens (pyperformance)      0.327s -> 0.286s  (1.14x)
  chaos (pyperformance)        0.204s -> 0.191s  (1.06x)
  richards (pyperformance)     0.154s -> 0.151s  (1.02x)
  go (pyperformance)           0.411s -> 0.405s  (1.01x)
  listcomp / fib               ~neutral (<=1%)

Peak RSS (/usr/bin/time -l): instances workload roughly unchanged
(29.0MB -> 29.1MB); dict-heavy workload actually drops (155MB -> 106MB),
likely because mimalloc reclaims freed pages more aggressively than the
system allocator across repeated large-dict churn.

`cargo build` and `cargo build --release` are warning-free with the
default feature set, and `cargo build --release --no-default-features
--features stdlib,threading,importlib,stdio,host_env` (mimalloc
excluded) also builds cleanly. CI's wasm32-wasip1 and other
--no-default-features build matrices in .github/workflows already omit
`mimalloc` from their feature lists and target wasm via the
cfg(not(wasm32)) dependency gate, so no workflow changes needed.

`rustpython -m test -j4 test_gc test_weakref test_threading test_dict
test_list test_tuple test_str test_bytes test_memoryview test_ctypes
test_mmap test_array` passes identically (2541 tests, 241 skipped) on
both the baseline and mimalloc binaries.

Assisted-by: Claude Code:claude-fable-5-1

asyncio's eager task path (and its per-task bookkeeping sets in general)
repeatedly adds and discards distinct task objects from a handful of
long-lived sets (_eager_tasks, _scheduled_tasks). Two dict_inner issues
turned that steady-state churn into the dominant cost of async_tree_eager:

1. resize() picked a target size of `used * 2` with no floor, so once a
   churny set settled around 1-2 live entries it shrank to a table that
   the very next insert would already overflow, forcing a resize on
   almost every single add. Clamping to the original minimum table size
   (8) restores the headroom dict/set start with.

2. `entries` only ever grows (holes from deletions are never reclaimed),
   so resize()'s rebuild pass scanned every entry ever inserted into the
   dict's lifetime, not just the live ones - a set with a stable working
   set size but heavy churn grew this scan, and the dict's memory, without
   bound. resize() now compacts holes out of `entries` first, matching
   CPython's compacting hash table and keeping the shape-stamp fast path
   (which already requires a hole-free layout) reachable more often.

Measured with a standalone async_tree micro-benchmark (6 levels x 6
branches, pyperformance's bm_async_tree "none"/"eager" shape) run inside
one process for 10 asyncio.run() iterations, comparing medians of
interleaved binary runs on an otherwise idle machine:

- eager: ~2.91s -> ~0.86s median (about 3.4x)
- none:  ~1.98s -> ~1.23s median (about 1.6x)

A single fresh-process run (no in-process accumulation) still improves
by roughly 15-20% for eager, since the resize-storm from point 1 hits
even a single run's worth of task churn.

test_asyncio (34 files, all passing), test_dict, test_set, test_weakref,
test_weakset, test_ordered_dict, test_gc, test_descr, test_class, and
`cargo test -p rustpython-vm -p rustpython-stdlib` all pass.

Assisted-by: Claude Code:claude-fable-5-1

test_plistlib's four invalid-document tests and test_xmlrpc's
test_loads_unsupported pass now that pyexpat no longer swallows
handler exceptions and keeps parser state across chunked feeds.

Assisted-by: Claude Code:claude-fable-5-1

…arse()

8802451f3 fixed multi-chunk Parse()/ParseFile() feeds by accumulating every
fed byte and re-parsing the whole buffer from scratch on each call, which is
O(total bytes) per chunk -- quadratic in document size. xml-rs's EventReader
has no API to pause/resume across separate Read sources, so instead one
EventReader is now driven to completion on a dedicated background thread fed
through a condvar-backed queue (FeedBuffer/FeedReader); Parse()/ParseFile()
push their chunk, wake the thread, and drain (on the calling thread only)
whatever events that unblocks, stopping once the thread reports it is
blocked waiting for more input. Handlers always run on the calling thread.
A handler exception stops dispatch without tearing down the stream, so the
next Parse() resumes where it left off, mirroring libexpat. CurrentByteIndex
is now computed incrementally per line (with an amortized per-query cursor,
needed since real documents are often a single very long line) instead of
scanning the whole buffer. ParserCreate() never spawns a thread; one is
spawned lazily by the first Parse()/ParseFile() call, and dropping the
parser always stops and joins it.

Fixes test_xml_etree{,_c}'s test_events_comment, previously xfailed because
the old whole-buffer replay lost top-level comment events across feeds.

Measured (M4 Pro, release build):
- Scaling (ElementTree.parse, newline-separated doc): new backend 1/4/16 MB
  = 0.15/0.61/2.35s (linear); CPython 3.14 same file = 0.015/0.055/0.23s.
- pyperformance's bm_xml_etree doc (156KB, effectively one line): old
  backend single etree.parse() = 1.99s, single iterparse() = 3.53s; new
  backend = 0.025s / 0.031s (~80-110x). Full bench_parse/iterparse/generate/
  process (30/10/10/1 reps) now finish in ~1.5s/0.6s/0.4s/0.3s instead of
  timing out past pyperformance's 400s limit.
- 10k ParserCreate() (never fed / partially fed / fully fed and dropped):
  0.32s, no residual native threads, no RSS growth.

test -j4 test_pyexpat test_xml_etree test_xml_etree_c test_sax test_minidom
test_xml_dom_minicompat test_pulldom test_plistlib test_xmlrpc test_html
test_htmlparser test_threading: all pass (1350 tests, 89 skipped).

Assisted-by: Claude Code:claude-fable-5-1

2f58e9531 made pyexpat always drive its EventReader on a background
thread (ParserStream::spawn), panicking via .expect() if the thread
couldn't be spawned. pyexpat is registered unconditionally, including
on wasm32, where std::thread::spawn always fails at runtime, so the
first Parse() call panicked there.

Introduce a Backend enum: Backend::Threaded (the existing ParserStream)
on targets that can spawn threads, and a new Backend::Sync, the
accumulate-and-reparse implementation from 8802451f3, used on wasm32
and as a runtime fallback if spawning a thread ever fails. Both share
LineTracker for identical CurrentLineNumber/CurrentColumnNumber/
CurrentByteIndex and handler/error semantics.

Fixed two correctness bugs while porting the Sync backend forward:
marking the parser permanently "finished" as soon as a from-scratch
reparse happened to reach EndDocument (which can be a false positive
on a not-yet-complete buffer, e.g. right before epilog content
arrives) rather than only on the caller's real isfinal signal; and
counting that same EndDocument towards the dispatched-event index,
which could make a real event land on an index a later, longer reparse
had already skipped past. Verified by force-selecting Backend::Sync
under a test-only env var (removed) and diffing its output against
Backend::Threaded for the xml/sax/etree/plistlib/xmlrpc test suites
and a 2MB document fed in 1KB chunks.

Assisted-by: Claude Code:claude-fable-5-1

Adds `crates/stdlib/src/pickle.rs`, a native `_pickle` implementing
Pickler, Unpickler, PickleBuffer, PicklerMemoProxy/UnpicklerMemoProxy,
dump/dumps/load/loads and the PickleError family, so `Lib/pickle.py`
stops falling back to `_Pickler`/`_Unpickler`.

Protocols 0-5 are supported and produce byte-identical output to the
pure-Python pickler, including framing for protocol >= 4, memoization,
`fast` mode, dispatch_table, reducer_override, persistent_id /
persistent_load, find_class with `_compat_pickle` fix_imports, buffer
callbacks and out-of-band protocol 5 buffers. This also turns on
`pickle.PickleBuffer`, which RustPython did not have before.

Speedups over the pure-Python implementation (best of 5, us/op):
pickle 835 -> 21 (39x), pickle_list 525 -> 4.4 (119x),
pickle_dict 4018 -> 32 (125x), unpickle 630 -> 24 (26x),
unpickle_list 317 -> 5.0 (63x). CPython 3.14's C accelerator runs the
same workloads 1.5-2.9x faster than this.

test_pickle now runs 1008 tests (was 470) since the C-implementation
test classes are enabled; test_pickle, test_pickletools,
test_picklebuffer are green and 22 obsolete expectedFailure markers are
removed.

Assisted-by: Claude Code:claude-fable-5-1

Unpickler.find_class now raises `sys.audit('pickle.find_class', module,
name)` like CPython, so audit hooks can veto global lookups
(test_audit.test_pickle). Also drops four expectedFailure markers that
the native module makes pass: test_pydoc's builtin-method description
tests, test_memoryview.test_picklebuffer_reference_loop and
test_pyclbr.test_others.

Full suite: 456 test files / 42,648 tests green.

Assisted-by: Claude Code:claude-fable-5-1

Opcode plus payload is now appended straight to the output buffer
(`write_pair`) instead of building a `Vec` per int, str, bytes and memo
opcode, `save_tuple` no longer copies the element slice, and the
`when serializing ...` note text is only formatted when an error is
actually raised.

bm_pickle "pickle" workload: 21.4 -> 17.0 us/op.

Assisted-by: Claude Code:claude-fable-5-1

fork() only keeps the calling thread. test_platform's
test_mac_ver_with_fork crashed the child with SIGTRAP: plistlib parses
a plist through pyexpat (spawning a background parser thread) before
os.fork(), then parses again in the child. Bisecting away every other
explanation (a plain threading.Thread survives the same fork fine,
since it goes through this VM's own fork-safe thread bookkeeping)
showed the crash is not from touching the *inherited* ParserStream
(already joined by refcounting before the fork in this repro) but from
spawning a *brand new* parser thread from the child, which reliably
crashed during the new thread's own startup.

Register a pthread_atfork child hook (lazily, the first time a parser
thread is spawned) that sets a permanent, process-wide flag the
instant this process is ever the child of a fork(). Once set:
- ParserStream::try_spawn refuses to spawn any further thread and
  Backend::new falls back to Backend::Sync instead, exactly as on
  targets that cannot spawn threads at all.
- Feeding an inherited Backend::Threaded raises a clear ExpatError
  instead of touching its FeedBuffer/channel (whose mutex/condvar may
  look held by a thread that no longer exists), and dropping it leaks
  its JoinHandle/buffer/channel via ManuallyDrop rather than joining or
  destroying them, since even destroying a POSIX mutex that looks held
  is undefined behavior.

The parent is unaffected: the pthread_atfork hook only runs in the
child. Verified with a 200-iteration fork stress test (fresh parsers
and parsers forked mid-parse) and the full test_platform/xml/fork/
subprocess/multiprocessing_fork test modules.

Assisted-by: Claude Code:claude-fable-5-1

self.intern (an always-present dict by default) was stored but never
consulted; every StartElement/EndElement/attribute name allocated a
fresh PyStr even when the same tag or attribute name repeated across
many sibling events. Wire it up the way libexpat does
(PyDict_SetDefault(self->intern, name, name)): element and attribute
*names* are now memoized so a repeated name reuses the same PyStr
object (and its cached hash) instead of reallocating and rehashing.
Attribute values are left out of the cache, same as libexpat.

This fixes test_pyexpat's previously-xfailed InterningTest.test, which
asserts object identity across repeated StartElementHandler/
EndElementHandler calls.

Measured on the pyperformance xml_etree parse/iterparse/generate/
process variants: no measurable wall-clock change (the dominant cost
there is Python-level TreeBuilder/Element bytecode execution and
thread hand-off, not name allocation), so this is a correctness/
feature-completion fix rather than a throughput win for that
benchmark; kept because it is a real, previously-inert compatibility
gap and is otherwise free.

Assisted-by: Claude Code:claude-fable-5-1

Adds crates/stdlib/src/elementtree.rs providing the accelerator names
xml/etree/ElementTree.py picks up at import time. Element carries the
full API (tag/text/tail/attrib, indexing and slicing, find*/iter/
itertext, copy/deepcopy/pickle state) with CPython's lazily created
attrib dict and deferred text join, is subclassable, weak-referenceable
and GC-traversable, and drops every displaced object after releasing its
lock so a __del__ can re-enter safely.

xml_etree, best of 3 (s): parse 1.64 -> 1.14, iterparse 0.68 -> 0.54,
generate 0.44 -> 0.25, process 0.27 -> 0.16.

Assisted-by: Claude Code:claude-fable-5-1

TreeBuilder keeps CPython's element/data/tail bookkeeping in Rust,
including the pull-parser event hooks _setevents installs. XMLParser
drives a pyexpat parser whose handlers are this type's own native
methods, so a start or end tag reaches the builder without a Python
frame: the universal-name cache, attribute dict and TreeBuilder call all
happen in Rust when the target is our own builder, while foreign targets
still go through the documented start/end/data/comment/pi protocol.
ElementTree.parse() also picks up the _parse_whole fast path.

xml_etree, best of 3 (s): parse 1.14 -> 0.53, iterparse 0.54 -> 0.28,
generate 0.25 -> 0.24, process 0.16 -> 0.16. Against the pure-Python
baseline that is parse 3.1x, iterparse 2.4x, generate 1.8x, process
1.7x.

Assisted-by: Claude Code:claude-fable-5-1

LineTracker converted a TextPosition column to a byte offset by
validating and decoding its entire buffered line tail on every query,
and appended fed bytes to that tail one at a time. On a document that is
one long line -- what xml.etree.tostring() writes, and what the
xml_etree benchmark parses -- the tail is the whole read-ahead, so a
per-event lookup became quadratic. Walk only the characters the query
actually asks for, over the deque's two halves in place, and record fed
bytes a line at a time. Byte indices are unchanged.

xml_etree, best of 3 (s): parse 0.52 -> 0.38, iterparse 0.28 -> 0.25.

Assisted-by: Claude Code:claude-fable-5-1

A chain of elements is copied by native recursion, so a deep enough tree
overflowed the real stack instead of raising. Wrap the copy in
with_recursion, as CPython does, and drop the skip on
test_deeply_nested_deepcopy for the accelerated run.

Assisted-by: Claude Code:claude-fable-5-1

The parser thread reports NeedMore when it finds the feed buffer empty
and parks. If it reached that point before the chunk Parse() was about
to push became visible -- which it does whenever the thread happens to
be scheduled first, so more often under load -- pump() consumed that
report as if it answered the new chunk and handed control back to Python
having dispatched nothing, silently losing the whole document.

Stamp each NeedMore with the push count observed when the buffer was
found empty and ignore reports older than the push being pumped. The
push notifies the condvar under the same mutex the report was sent
from, so the parser thread is guaranteed to wake and report again.

Fixes an intermittent failure of test_pyexpat's InterningTest under
parallel load (about 5 losses per 20000 parses before, none after).

Assisted-by: Claude Code:claude-fable-5-1

Mirrors CPython's lookdict: two exact `int` keys now compare by BigInt
value directly instead of going through the generic rich-compare
dispatch (identical_or_equal -> bool_eq -> rich_compare_bool -> __eq__
slot lookup). Falls back to bool_eq whenever either side isn't exactly
`int` (subclasses may override __eq__), so behavior is unchanged there.

Measured (min-of-6 rounds, 2M-iteration micros, this tree's build has
~4% layout noise so min is used as the noise floor):
- dict_int_lookup (`d[i]`, int keys): 70.9ns -> 62.0ns (-12.5%)
- dict_str_lookup / dict_tuple_lookup / set_str_contains: unchanged
  (within noise), confirming no regression on non-int keys.

pyperformance sanity checks (single round each, base vs candidate):
bm_mdp median ~5027ms/loop vs ~5003ms/loop (no regression, hashing
isn't the dominant cost there); bm_nqueens, bm_go, bm_deltablue,
bm_hexiom, bm_chaos all within ~1% (noise).

Hash/equality values cross-checked bit-identical against CPython 3.14
for 0, 1, -1, True, False, 2**61-1, 2**61, -(2**63), 2**100, -(2**100),
and int/float/bool/Decimal/Fraction(1) hash-equality invariants.

Also tried an exact-type fast dispatch in PyObject::hash (skipping the
slots.hash vtable call + with_recursion for exact int/str) - it only
moved hash_small_int/hash_large_int by ~2%, under the 3% bar, so it was
not committed.

Tests: `rustpython -m test -j4 test_hash test_int test_long test_tuple
test_dict test_set test_dictviews test_ordered_dict test_collections
test_fractions test_numeric_tower test_float test_bool test_str
test_userdict test_weakref` all pass (1796 run, 72 skipped). `cargo
test -p rustpython-vm --release -p rustpython-common` passes. clippy
clean on the touched lines.

Assisted-by: Claude Code:claude-fable-5-1

CPython's ternary dispatch offers the operation to the third operand's type
whenever its slot differs from the two already tried, including the case where
neither operand had one. The port required both of those slots to exist, so
`pow(10, 2, Decimal(7))` never reached `Decimal` and raised TypeError.

Assisted-by: Claude Code:claude-fable-5-1

A new `rustpython-decimal` crate implements the General Decimal Arithmetic
Specification over `malachite-bigint`: construction and the numeric-string
grammar, the full arithmetic set, comparison and total ordering, the digit-wise
and exponent operations, correctly rounded sqrt/exp/ln/log10/power, and the
format-specification mini-language.

The algorithms are ported from CPython's `Lib/_pydecimal.py`, which is the
executable specification for the parts the standard leaves open; the coefficient
is a `BigUint` with a cached digit count instead of a digit string, so what the
Python does with slicing is done here with divisions by powers of ten.

Conditions follow libmpdec rather than `_pydecimal`: an operation never raises,
it accumulates the conditions it hit into a status word that the caller merges
into a context. That is what lets a trapped signal report every condition an
operation produced rather than only the first.

Assisted-by: Claude Code:claude-fable-5-1

`Lib/decimal.py` starts with `from _decimal import *` and falls back to
`Lib/_pydecimal.py` when that fails. There was no `_decimal`, so every program
using decimals ran the pure-Python implementation.

This wires the `rustpython-decimal` engine up as that module: the `Decimal` and
`Context` types with their whole method surface, the signal hierarchy with
libmpdec's condition ordering, the live `flags`/`traps` views, the
context-variable holding the current context, and `localcontext`/`IEEEContext`.
Signatures carry CPython's text signatures, so `inspect.signature` and the
suite's C/Python signature comparison agree.

test_decimal now runs its C classes as well as its Python ones -- 721 tests
including all 143 IBM `decimaltestdata` files -- and passes. A differential
check of 16000 random operations against CPython's `_decimal` agrees on every
result and every status flag.

pyperformance telco: 348 ms/iteration -> 15.4 ms (23x; CPython 3.14 is 4.5 ms).
Decimal('1.23')*Decimal('4.56') 5275 ns -> 267 ns, quantize 6814 ns -> 423 ns,
str() 1658 ns -> 497 ns.

Assisted-by: Claude Code:claude-fable-5-1

Counts `(prev_op, op)` in the dispatch loop and dumps the nonzero pairs at
exit, so superinstruction and specialization candidates can be ranked from
what pyperformance-shaped code actually dispatches rather than guessed at.

Entirely behind the non-default `opcode-histogram` feature: the module is
`#[cfg]`-ed out and the loop's two-line hook with it, so a default build is
byte-identical to one without the patch.

Assisted-by: Claude Code:claude-fable-5-1

`TO_BOOL*`, `COMPARE_OP*`, `CONTAINS_OP*` and `IS_OP` materialised a bool
only for the `POP_JUMP_IF_*` one dispatch later to pop and test it, and the
compiler's `NOT_TAKEN` marker cost a third dispatch on the fall-through edge.
Generalise the `COMPARE_OP_INT`-only jump fusion so every boolean producer
resolves the jump itself, and step over `NOT_TAKEN` from there and from the
plain `POP_JUMP_IF_*` handlers. Both back off to ordinary dispatch when the
successor is an instrumented opcode or while tracing, so `sys.monitoring`
branch events and `sys.settrace` opcode events are unaffected.

Dispatches over 13 standalone pyperformance benchmarks: -5.0% total
(richards -11.7%, unpickle_pure_python -8.9%, sympy -8.4%, hexiom -8.1%,
django_template -7.1%, go -6.1%; nbody unchanged). Times, best of 7
interleaved rounds: deltablue -4.0%, hexiom -3.5%, go -3.4%,
unpickle_pure_python -3.3%, richards -3.2%, raytrace -3.0%, nqueens -1.9%,
geomean -1.0%; a branch-heavy microbenchmark -4.7% against a straight-line
control at 0%.

Assisted-by: Claude Code:claude-fable-5-1