René's URL Explorer Experiment

Title: Use dict keys-version stamps and entry-index hints in LOAD_ATTR/STORE_ATTR specializations by youknowone · Pull Request #8350 · RustPython/RustPython · GitHub

Open Graph Title: Use dict keys-version stamps and entry-index hints in LOAD_ATTR/STORE_ATTR specializations by youknowone · Pull Request #8350 · RustPython/RustPython

X Title: Use dict keys-version stamps and entry-index hints in LOAD_ATTR/STORE_ATTR specializations by youknowone · Pull Request #8350 · RustPython/RustPython

Description: Even after LOAD_ATTR specializes, the specialized handlers still paid for a full hash+probe dict lookup on every execution — LoadAttrWithHint ignored its hint, LoadAttrMethodWithValues probed the instance dict as its shadow check, and StoreAttr* re-probed on every store. This PR replaces those probes with two cheap guards. Changes Dict keys-version stamp (dict_inner.rs) Dict gets a keys_version: AtomicU32. Stamps are allocated from a global counter and never reused, so equal nonzero stamps guarantee an unchanged key set. Any key-set change (new key, deletion, pop, clear, drain_entries) resets the stamp while holding the write lock, before the mutation; value-only updates keep it. New get_with_hint (lookup returning the entry index) and insert_with_hint (identity-checked value replace by entry index, falling back to a full insert). LOAD_ATTR_METHOD_WITH_VALUES / LOAD_ATTR_NONDESCRIPTOR_WITH_VALUES — the shadow check caches the instance dict's stamp in the instruction's pointer-cache slot. While the stamp matches, the per-call dict probe is skipped entirely (an integer compare, like CPython's keys-version check). The stamp is taken before the absence probe so a concurrent key-set change can never leave a stale skip. Dict subclasses never get a stamp and always probe. LOAD_ATTR_WITH_HINT — the specializer now stores the entry index in the cache. A hit is a single identity check on the hinted entry's key — no hashing, no probe sequence. On miss the handler re-probes and refreshes the hint in place. STORE_ATTR_WITH_HINT / STORE_ATTR_INSTANCE_VALUE — same treatment via set_item_with_hint: value replace through the cached entry index, self-refreshing on miss. Results (Apple M5 Max, release, vs CPython 3.14.5) benchmark before after CPython c.x read overhead vs local read 25.4 ns ~16 ns (−37%) 2.2 ns self.value r/w + c.add(i) loop 183 ns/iter ~163 ns/iter (−11%) 24.4 ns Profiles confirm the Dict::get hash-probe frames (~7–10% of samples) are gone from both loops; the remaining gap is dominated by call overhead, int boxing, and dispatch (tracked separately). Testing cargo test workspace (the pre-existing rustpython-capi pyo3 build failure is unrelated) -m test for test_dict, test_descr, test_class, test_weakref, test_gc, test_property, test_super, test_builtin, test_types, test_copy, test_pickle, test_ordered_dict, test_userdict, test_collections — all SUCCESS New regression tests in extra_tests/snippets/vm_specialization.py (shadowing after warmup, value-only updates, hint invalidation across key churn / __dict__ replacement / clear()), also validated against CPython 3.14 Multithreaded stress test of concurrent shadowing/deletion and key-set churn against the stamp and hint paths Fixes youknowone#42 🤖 Generated with Claude Code Summary by CodeRabbit Performance Improved instance attribute access and dict lookups using refreshed hint-based fast paths that account for per-dictionary keyset/layout changes. Enhanced hinted attribute store/update so cached attribute resolution remains valid with fewer slow-path fallbacks. Bug Fixes Strengthened cache invalidation on keyset mutations (including clears/drains and related updates) to avoid stale results. Improved correctness when attributes/methods are shadowed and when dicts have “holes” after deletions. Tests Expanded VM specialization regression coverage for method/value shadowing, hint survival across attribute/dict churn, shared layout behavior, and hole handling.

Open Graph Description: Even after LOAD_ATTR specializes, the specialized handlers still paid for a full hash+probe dict lookup on every execution — LoadAttrWithHint ignored its hint, LoadAttrMethodWithValues probed the i...

X Description: Even after LOAD_ATTR specializes, the specialized handlers still paid for a full hash+probe dict lookup on every execution — LoadAttrWithHint ignored its hint, LoadAttrMethodWithValues probed the i...

Opengraph URL: https://github.com/RustPython/RustPython/pull/8350

X: @github

direct link

Domain: github.com

Links:

Viewport: width=device-width