Backport PR pandas-dev#41988: BUG: nan-objects lookup fails with Python3.10

Author: realead

pandas/_libs/src/klib/khash_python.h

@@ -251,12 +251,105 @@ int PANDAS_INLINE pyobject_cmp(PyObject* a, PyObject* b) {
 }
 
 
-khint32_t PANDAS_INLINE kh_python_hash_func(PyObject* key){
+Py_hash_t PANDAS_INLINE _Pandas_HashDouble(double val) {
+    //Since Python3.10, nan is no longer has hash 0
+    if (Py_IS_NAN(val)) {
+        return 0;
+    }
+#if PY_VERSION_HEX < 0x030A0000
+    return _Py_HashDouble(val);
+#else
+    return _Py_HashDouble(NULL, val);
+#endif
+}
+
+
+Py_hash_t PANDAS_INLINE floatobject_hash(PyFloatObject* key) {
+    return _Pandas_HashDouble(PyFloat_AS_DOUBLE(key));
+}
+
+
+// replaces _Py_HashDouble with _Pandas_HashDouble
+Py_hash_t PANDAS_INLINE complexobject_hash(PyComplexObject* key) {
+    Py_uhash_t realhash = (Py_uhash_t)_Pandas_HashDouble(key->cval.real);
+    Py_uhash_t imaghash = (Py_uhash_t)_Pandas_HashDouble(key->cval.imag);
+    if (realhash == (Py_uhash_t)-1 || imaghash == (Py_uhash_t)-1) {
+        return -1;
+    }
+    Py_uhash_t combined = realhash + _PyHASH_IMAG * imaghash;
+    if (combined == (Py_uhash_t)-1) {
+        return -2;
+    }
+    return (Py_hash_t)combined;
+}
+
+
+khint32_t PANDAS_INLINE kh_python_hash_func(PyObject* key);
+
+//we could use any hashing algorithm, this is the original CPython's for tuples
+
+#if SIZEOF_PY_UHASH_T > 4
+#define _PandasHASH_XXPRIME_1 ((Py_uhash_t)11400714785074694791ULL)
+#define _PandasHASH_XXPRIME_2 ((Py_uhash_t)14029467366897019727ULL)
+#define _PandasHASH_XXPRIME_5 ((Py_uhash_t)2870177450012600261ULL)
+#define _PandasHASH_XXROTATE(x) ((x << 31) | (x >> 33))  /* Rotate left 31 bits */
+#else
+#define _PandasHASH_XXPRIME_1 ((Py_uhash_t)2654435761UL)
+#define _PandasHASH_XXPRIME_2 ((Py_uhash_t)2246822519UL)
+#define _PandasHASH_XXPRIME_5 ((Py_uhash_t)374761393UL)
+#define _PandasHASH_XXROTATE(x) ((x << 13) | (x >> 19))  /* Rotate left 13 bits */
+#endif
+
+Py_hash_t PANDAS_INLINE tupleobject_hash(PyTupleObject* key) {
+    Py_ssize_t i, len = Py_SIZE(key);
+    PyObject **item = key->ob_item;
+
+    Py_uhash_t acc = _PandasHASH_XXPRIME_5;
+    for (i = 0; i < len; i++) {
+        Py_uhash_t lane = kh_python_hash_func(item[i]);
+        if (lane == (Py_uhash_t)-1) {
+            return -1;
+        }
+        acc += lane * _PandasHASH_XXPRIME_2;
+        acc = _PandasHASH_XXROTATE(acc);
+        acc *= _PandasHASH_XXPRIME_1;
+    }
+
+    /* Add input length, mangled to keep the historical value of hash(()). */
+    acc += len ^ (_PandasHASH_XXPRIME_5 ^ 3527539UL);
+
+    if (acc == (Py_uhash_t)-1) {
+        return 1546275796;
+    }
+    return acc;
+}
+
+
+khint32_t PANDAS_INLINE kh_python_hash_func(PyObject* key) {
+    Py_hash_t hash;
     // For PyObject_Hash holds:
     //    hash(0.0) == 0 == hash(-0.0)
-    //    hash(X) == 0 if X is a NaN-value
-    // so it is OK to use it directly for doubles
-    Py_hash_t hash = PyObject_Hash(key);
+    //    yet for different nan-objects different hash-values
+    //    are possible
+    if (PyFloat_CheckExact(key)) {
+        // we cannot use kh_float64_hash_func
+        // becase float(k) == k holds for any int-object k
+        // and kh_float64_hash_func doesn't respect it
+        hash = floatobject_hash((PyFloatObject*)key);
+    }
+    else if (PyComplex_CheckExact(key)) {
+        // we cannot use kh_complex128_hash_func
+        // becase complex(k,0) == k holds for any int-object k
+        // and kh_complex128_hash_func doesn't respect it
+        hash = complexobject_hash((PyComplexObject*)key);
+    }
+    else if (PyTuple_CheckExact(key)) {
+        hash = tupleobject_hash((PyTupleObject*)key);
+    }
+    else {
+        hash = PyObject_Hash(key);
+    }
+
 	if (hash == -1) {
 		PyErr_Clear();
 		return 0;

pandas/tests/libs/test_hashtable.py

@@ -339,6 +339,29 @@ def test_unique(self, table_type, dtype):
         assert np.all(np.isnan(unique)) and len(unique) == 1
 
 
+def test_unique_for_nan_objects_floats():
+    table = ht.PyObjectHashTable()
+    keys = np.array([float("nan") for i in range(50)], dtype=np.object_)
+    unique = table.unique(keys)
+    assert len(unique) == 1
+
+
+def test_unique_for_nan_objects_complex():
+    table = ht.PyObjectHashTable()
+    keys = np.array([complex(float("nan"), 1.0) for i in range(50)], dtype=np.object_)
+    unique = table.unique(keys)
+    assert len(unique) == 1
+
+
+def test_unique_for_nan_objects_tuple():
+    table = ht.PyObjectHashTable()
+    keys = np.array(
+        [1] + [(1.0, (float("nan"), 1.0)) for i in range(50)], dtype=np.object_
+    )
+    unique = table.unique(keys)
+    assert len(unique) == 2
+
+
 def get_ht_function(fun_name, type_suffix):
     return getattr(ht, fun_name)
 
@@ -497,3 +520,11 @@ def test_ismember_tuple_with_nans():
     result = isin(values, comps)
     expected = np.array([True, False], dtype=np.bool_)
     tm.assert_numpy_array_equal(result, expected)
+
+
+def test_float_complex_int_are_equal_as_objects():
+    values = ["a", 5, 5.0, 5.0 + 0j]
+    comps = list(range(129))
+    result = isin(values, comps)
+    expected = np.array([False, True, True, True], dtype=np.bool_)
+    tm.assert_numpy_array_equal(result, expected)