perf: look up instance __dict__ by the interned name (cached hash) in getattr

sharifhsn · sharifhsn · commit dd4ad34b3037 · 2026-06-17T20:57:55.000-04:00
generic_getattr_opt looked up the instance dict by name_str.as_wtf8() -- the &amp;Wtf8 content, whose DictKey::key_hash recomputes the SipHash from the raw bytes on every attribute access. Look it up by name_str (the &amp;Py&lt;PyStr&gt;) instead: its DictKey returns the hash PyStr already caches, and adds a pointer-equality fast path for the key compare. (generic_setattr already passed the PyStr, so writes were never affected.)

Free -- no memory added; the hash cache already exists. Benchmarks (interleaved A/B vs main, median): attribute-read loop -17.3%, pystone -3.0%, OOP method loop -7.9%; neutral on non-attribute code. Verified: 12 attribute/descriptor/class/dataclass test modules pass; clippy clean.

Co-Authored-By: Claude Opus 4.8 (1M context) &lt;noreply@anthropic.com&gt;

Assisted-by: Claude:claude-opus-4
diff --git a/crates/vm/src/protocol/object.rs b/crates/vm/src/protocol/object.rs
@@ -230,7 +230,6 @@ impl PyObject {
         dict: Option<PyDictRef>,
         vm: &VirtualMachine,
     ) -> PyResult<Option<PyObjectRef>> {
-        let name = name_str.as_wtf8();
         let obj_cls = self.class();
         let cls_attr_name = vm.ctx.interned_str(name_str);
         let cls_attr = match cls_attr_name.and_then(|name| obj_cls.get_attr(name)) {
@@ -251,7 +250,9 @@ impl PyObject {
         let dict = dict.or_else(|| self.dict());
 
         let attr = if let Some(dict) = dict {
-            dict.get_item_opt(name, vm)?
+            // Look up by the `PyStr` object (cached hash + pointer-equality fast
+            // path) rather than its `&Wtf8` content, which re-hashes every call.
+            dict.get_item_opt(name_str, vm)?
         } else {
             None
         };
