RustPython · Jorge-Polanco-Roque · Aug 30, 2026 · Aug 30, 2026 · Aug 30, 2026 · Aug 30, 2026
diff --git a/Lib/test/test_dict.py b/Lib/test/test_dict.py
@@ -1614,7 +1614,6 @@ class Shenanigans:
         self.assertEqual(holds_reference.ref['data'], 42)
         self.assertEqual(holds_reference.attr, "whatever")
 
-    @unittest.expectedFailure  # TODO: RUSTPYTHON
     def test_unhashable_key(self):
         d = {'a': 1}
         key = [1, 2, 3]

diff --git a/crates/vm/src/builtins/dict.rs b/crates/vm/src/builtins/dict.rs
@@ -153,11 +153,12 @@ impl PyDict {
             Ok(keys_method) => {
                 let keys = keys_method.call((), vm)?.get_iter(vm)?;
                 while let PyIterReturn::Return(key) = keys.next(vm)? {
-                    if !override_existing && dict.contains(vm, &*key)? {
+                    let hash = Self::hash_or_unhashable(&*key, vm)?;
+                    if !override_existing && dict.contains_known_hash(vm, &*key, hash)? {
                         continue;
                     }
                     let val = other.get_item(&*key, vm)?;
-                    dict.insert(vm, &*key, val)?;
+                    dict.insert_known_hash(vm, &*key, hash, val)?;
                 }
                 true
             }
@@ -261,10 +262,11 @@ impl PyDict {
         for (index, element) in iter.iter::<PyObjectRef>(vm)?.enumerate() {
             let (key, value) = Self::update_sequence_pair(element?, index, vm)?;
 
-            if !override_existing && dict.contains(vm, &*key)? {
+            let hash = Self::hash_or_unhashable(&*key, vm)?;
+            if !override_existing && dict.contains_known_hash(vm, &*key, hash)? {
                 continue;
             }
-            dict.insert(vm, &*key, value)?;
+            dict.insert_known_hash(vm, &*key, hash, value)?;
         }
         Ok(())
     }
@@ -293,6 +295,37 @@ impl PyDict {
         self.entries.len() == 0
     }
 
+    /// Hash `key`, turning a hashing failure into the dict-specific
+    /// "cannot use 'X' as a dict key (...)" wording used by CPython.
+    ///
+    /// The returned hash is threaded into the `*_known_hash` operations, so the
+    /// key is hashed exactly once. This is why the message is produced here
+    /// rather than around the operation: a `TypeError` from key *comparison*
+    /// (e.g. a colliding key's `__eq__`) is raised by the operation itself and
+    /// must propagate unchanged, and hashing up front also means a `__hash__`
+    /// that fails only intermittently is still reported (the old approach
+    /// re-hashed on the error path, so a hash that succeeded the second time
+    /// escaped unwrapped).
+    fn hash_or_unhashable<K: DictKey + ?Sized>(key: &K, vm: &VirtualMachine) -> PyResult<PyHash> {
+        match key.key_hash(vm) {
+            Ok(hash) => Ok(hash),
+            // Exact `TypeError` only: a `__hash__` raising a *subclass* of
+            // `TypeError` (e.g. a user `MyTypeError`) must propagate unchanged,
+            // matching CPython's exact-type check.
+            Err(cause) if cause.class().is(vm.ctx.exceptions.type_error) => {
+                let message = cause.as_object().str(vm)?;
+                let key = key.to_pyobject(vm);
+                let err = vm.new_type_error(format!(
+                    "cannot use '{}' as a dict key ({message})",
+                    key.class().fully_qualified_name(vm)
+                ));
+                err.set___cause__(Some(cause));
+                Err(err)
+            }
+            Err(other) => Err(other),
+        }
+    }
+
     /// Set item variant which can be called with multiple
     /// key types, such as str to name a notable one.
     pub fn inner_setitem<K: DictKey + ?Sized>(
@@ -301,15 +334,21 @@ impl PyDict {
         value: PyObjectRef,
         vm: &VirtualMachine,
     ) -> PyResult<()> {
-        self.entries.insert(vm, key, value)
+        let hash = Self::hash_or_unhashable(key, vm)?;
+        self.entries.insert_known_hash(vm, key, hash, value)
     }
 
     pub(crate) fn inner_delitem<K: DictKey + ?Sized>(
         &self,
         key: &K,
         vm: &VirtualMachine,
     ) -> PyResult<()> {
-        self.entries.delete(vm, key)
+        let hash = Self::hash_or_unhashable(key, vm)?;
+        if self.entries.delete_if_exists_known_hash(vm, key, hash)? {
+            Ok(())
+        } else {
+            Err(vm.new_key_error(key.to_pyobject(vm)))
+        }
     }
 
     pub fn get_or_insert(
@@ -318,7 +357,8 @@ impl PyDict {
         key: PyObjectRef,
         default: impl FnOnce() -> PyObjectRef,
     ) -> PyResult {
-        self.entries.setdefault(vm, &*key, default)
+        let hash = Self::hash_or_unhashable(&*key, vm)?;
+        self.entries.setdefault(vm, &*key, hash, default)
     }
 
     pub fn from_attributes(attrs: PyAttributes, vm: &VirtualMachine) -> PyResult<Self> {
@@ -426,7 +466,8 @@ impl PyDict {
     }
 
     fn __contains__(&self, key: PyObjectRef, vm: &VirtualMachine) -> PyResult<bool> {
-        self.entries.contains(vm, &*key)
+        let hash = Self::hash_or_unhashable(&*key, vm)?;
+        self.entries.contains_known_hash(vm, &*key, hash)
     }
 
     fn __delitem__(&self, key: PyObjectRef, vm: &VirtualMachine) -> PyResult<()> {
@@ -454,10 +495,9 @@ impl PyDict {
         default: OptionalArg<PyObjectRef>,
         vm: &VirtualMachine,
     ) -> PyResult {
-        Ok(self
-            .entries
-            .get(vm, &*key)?
-            .unwrap_or_else(|| default.unwrap_or_none(vm)))
+        let hash = Self::hash_or_unhashable(&*key, vm)?;
+        let found = self.entries.get_known_hash(vm, &*key, hash)?;
+        Ok(found.unwrap_or_else(|| default.unwrap_or_none(vm)))
     }
 
     #[pymethod]
@@ -467,8 +507,9 @@ impl PyDict {
         default: OptionalArg<PyObjectRef>,
         vm: &VirtualMachine,
     ) -> PyResult {
+        let hash = Self::hash_or_unhashable(&*key, vm)?;
         self.entries
-            .setdefault(vm, &*key, || default.unwrap_or_none(vm))
+            .setdefault(vm, &*key, hash, || default.unwrap_or_none(vm))
     }
 
     #[pymethod]
@@ -512,7 +553,8 @@ impl PyDict {
         default: OptionalArg<PyObjectRef>,
         vm: &VirtualMachine,
     ) -> PyResult {
-        match self.entries.pop(vm, &*key)? {
+        let hash = Self::hash_or_unhashable(&*key, vm)?;
+        match self.entries.pop(vm, &*key, hash)? {
             Some(value) => Ok(value),
             None => default.ok_or_else(|| vm.new_key_error(key)),
         }
@@ -660,9 +702,12 @@ impl AsMapping for PyDict {
 impl AsSequence for PyDict {
     fn as_sequence() -> &'static PySequenceMethods {
         static AS_SEQUENCE: LazyLock<PySequenceMethods> = LazyLock::new(|| PySequenceMethods {
-            contains: atomic_func!(|seq, target, vm| PyDict::sequence_downcast(seq)
-                .entries
-                .contains(vm, target)),
+            contains: atomic_func!(|seq, target, vm| {
+                let hash = PyDict::hash_or_unhashable(target, vm)?;
+                PyDict::sequence_downcast(seq)
+                    .entries
+                    .contains_known_hash(vm, target, hash)
+            }),
             ..PySequenceMethods::NOT_IMPLEMENTED
         });
         &AS_SEQUENCE
@@ -765,7 +810,9 @@ impl Py<PyDict> {
         key: &K,
         vm: &VirtualMachine,
     ) -> PyResult<PyObjectRef> {
-        if let Some(value) = self.entries.get(vm, key)? {
+        let hash = PyDict::hash_or_unhashable(key, vm)?;
+        let found = self.entries.get_known_hash(vm, key, hash)?;
+        if let Some(value) = found {
             Ok(value)
         } else if let Some(value) = self.missing_opt(key, vm)? {
             Ok(value)

diff --git a/crates/vm/src/dict_inner.rs b/crates/vm/src/dict_inner.rs
@@ -573,6 +573,17 @@ impl<T: Clone> Dict<T> {
         self._get_inner(vm, key, hash)
     }
 
+    /// [`Self::get`] with a known hash. Same contract as
+    /// [`Self::insert_known_hash`].
+    pub(crate) fn get_known_hash<K: DictKey + ?Sized>(
+        &self,
+        vm: &VirtualMachine,
+        key: &K,
+        hash: HashValue,
+    ) -> PyResult<Option<T>> {
+        self._get_inner(vm, key, hash)
+    }
+
     /// Return a stable entry hint for `key` if present.
     ///
     /// The hint is the internal entry index and can be used with
@@ -872,12 +883,19 @@ impl<T: Clone> Dict<T> {
         Ok(())
     }
 
-    pub(crate) fn setdefault<K, F>(&self, vm: &VirtualMachine, key: &K, default: F) -> PyResult<T>
+    /// Get the value for `key`, inserting `default()` if it is absent, given a
+    /// known hash. Same contract as [`Self::insert_known_hash`].
+    pub(crate) fn setdefault<K, F>(
+        &self,
+        vm: &VirtualMachine,
+        key: &K,
+        hash: HashValue,
+        default: F,
+    ) -> PyResult<T>
     where
         K: DictKey + ?Sized,
         F: FnOnce() -> T,
     {
-        let hash = key.key_hash(vm)?;
         let mut default = Some(default);
         loop {
             let (index_entry, index_index) = self.lookup(vm, key, hash, None)?;
@@ -907,43 +925,6 @@ impl<T: Clone> Dict<T> {
         }
     }
 
-    #[allow(dead_code)]
-    pub(crate) fn setdefault_entry<K, F>(
-        &self,
-        vm: &VirtualMachine,
-        key: &K,
-        default: F,
-    ) -> PyResult<(PyObjectRef, T)>
-    where
-        K: DictKey + ?Sized,
-        F: FnOnce() -> T,
-    {
-        let hash = key.key_hash(vm)?;
-        let mut default = Some(default);
-        loop {
-            let (index_entry, index_index) = self.lookup(vm, key, hash, None)?;
-            if let Some(index) = index_entry.index() {
-                let inner = self.read();
-                if let Some(entry) = inner.get_entry_checked(index, index_index) {
-                    return Ok((entry.key.clone(), entry.value.clone()));
-                }
-                continue;
-            }
-            let mut inner = self.write();
-            if inner.indices.get(index_index) != Some(&index_entry) {
-                continue;
-            }
-            let value = default
-                .take()
-                .expect("default must only be computed on insertion")();
-            let key_obj = key.to_pyobject(vm);
-            let ret = (key_obj.clone(), value.clone());
-            self.invalidate_keys_version();
-            inner.unchecked_push(index_index, hash, key_obj, value, index_entry);
-            return Ok(ret);
-        }
-    }
-
     pub(crate) fn len(&self) -> usize {
         self.read().used
     }
@@ -1228,13 +1209,14 @@ impl<T: Clone> Dict<T> {
         Ok(ControlFlow::Break(removed))
     }
 
-    /// Retrieve and delete a key
+    /// Retrieve and delete a key, given a known hash. Same contract as
+    /// [`Self::insert_known_hash`].
     pub(crate) fn pop<K: DictKey + ?Sized>(
         &self,
         vm: &VirtualMachine,
         key: &K,
+        hash_value: HashValue,
     ) -> PyResult<Option<T>> {
-        let hash_value = key.key_hash(vm)?;
         let removed = loop {
             let lookup = self.lookup(vm, key, hash_value, None)?;
             match self.pop_inner(lookup) {

diff --git a/extra_tests/snippets/builtin_dict.py b/extra_tests/snippets/builtin_dict.py
@@ -408,3 +408,112 @@ def test_func(**kwargs):
 assert list(result.keys()) == expected_keys, (
     f"Expected {expected_keys}, got {list(result.keys())}"
 )
+
+
+# A TypeError raised while *comparing* keys (e.g. from a colliding key's
+# __eq__) must propagate unchanged, not be rewritten as an unhashable-key
+# error. Only genuine hashing failures get the dict-specific "unhashable"
+# wording.
+class BadEq:
+    def __hash__(self):
+        return 42  # fixed hash forces a collision
+
+    def __eq__(self, other):
+        raise TypeError("nope")
+
+
+bad = {}
+bad[BadEq()] = 1
+with assert_raises(TypeError) as cm:
+    bad[BadEq()]  # hashes fine (42), then compares against the colliding key
+assert "nope" in str(cm.exception), str(cm.exception)
+assert "dict key" not in str(cm.exception), (
+    f"comparison error mislabeled as unhashable: {cm.exception}"
+)
+
+# A genuinely unhashable key still reports the dict-specific message.
+with assert_raises(TypeError) as cm:
+    {}[[]]
+assert "as a dict key" in str(cm.exception), str(cm.exception)
+
+# The message reaches every insertion path, not just __setitem__: the
+# constructor, update() and |= all go through the same wrapping.
+for make in (
+    lambda: dict([([], 1)]),
+    lambda: {}.update([([], 1)]),
+    lambda: {}.__ior__([([], 1)]),
+):
+    with assert_raises(TypeError) as cm:
+        make()
+    assert "as a dict key" in str(cm.exception), str(cm.exception)
+
+
+# The key is hashed once up front, so a __hash__ that fails only on its first
+# call is still reported (a re-hash on the error path would let it escape).
+class FlakyHash:
+    _calls = 0
+
+    def __hash__(self):
+        FlakyHash._calls += 1
+        if FlakyHash._calls == 1:
+            raise TypeError("first call fails")
+        return 0
+
+
+with assert_raises(TypeError) as cm:
+    {}[FlakyHash()]
+assert "as a dict key" in str(cm.exception), str(cm.exception)
+
+
+# The type name is the fully qualified one, like CPython's %T.
+def _make_nested():
+    class Nested:
+        __hash__ = None
+
+    return Nested
+
+
+with assert_raises(TypeError) as cm:
+    {}[_make_nested()()]
+assert "_make_nested.<locals>.Nested" in str(cm.exception), str(cm.exception)
+
+
+# A __hash__ raising a *subclass* of TypeError is left unchanged (CPython
+# checks the exact type), so `except MySubclass` still catches it.
+class MyTypeError(TypeError):
+    pass
+
+
+class SubclassHash:
+    def __hash__(self):
+        raise MyTypeError("custom")
+
+
+with assert_raises(MyTypeError) as cm:
+    {}[SubclassHash()]
+assert "as a dict key" not in str(cm.exception), str(cm.exception)
+
+
+# Every operation hashes the key exactly once (the hash is threaded into the
+# inner map), so a __hash__ that would fail on a second call is never called
+# twice. setdefault() and pop() went through their own lookup before.
+class CountingHash:
+    calls = 0
+
+    def __hash__(self):
+        CountingHash.calls += 1
+        if CountingHash.calls >= 2:
+            raise TypeError("must not hash twice")
+        return 7
+
+
+CountingHash.calls = 0
+{}.setdefault(CountingHash(), 1)
+assert CountingHash.calls == 1, CountingHash.calls
+
+CountingHash.calls = 0
+with assert_raises(KeyError):
+    # A non-empty dict so the lookup must hash the key (CPython skips hashing
+    # entirely when popping from an empty dict).
+    {1: 1}.pop(CountingHash())
+assert CountingHash.calls == 1, CountingHash.calls