improve hash function

CHANGELOG.md

@@ -12,6 +12,7 @@
 - fix implementation of Hamming.normalized_similarity
 - fix default score_cutoff of Hamming.similarity
 - fix implementation of LCSseq.distance when used in the process module
+- treat hash for -1 and -2 as different
 
 ### [2.0.15] - 2022-06-24
 #### Fixed

rapidfuzz/cpp_common.pxd

@@ -116,6 +116,11 @@ cdef extern from "cpp_common.hpp":
 
     vector[T] vector_slice[T](const vector[T]& vec, int start, int stop, int step) except +
 
+cdef inline uint64_t rf_hash(val) except *:
+    if val == -1:
+        return <uint64_t>-1
+    return <uint64_t>hash(val)
+
 cdef inline RF_String hash_array(arr) except *:
     # TODO on Cpython this does not require any copies
     cdef RF_String s_proc
@@ -156,7 +161,7 @@ cdef inline RF_String hash_array(arr) except *:
         else: # float/double are hashed
             s_proc.kind = RF_StringType.RF_UINT64
             for i in range(s_proc.length):
-                (<uint64_t*>s_proc.data)[i] = <uint64_t>hash(arr[i])
+                (<uint64_t*>s_proc.data)[i] = rf_hash(arr[i])
     except Exception as e:
         free(s_proc.data)
         s_proc.data = NULL
@@ -183,7 +188,7 @@ cdef inline RF_String hash_sequence(seq) except *:
             if isinstance(elem, str) and len(elem) == 1:
                 (<uint64_t*>s_proc.data)[i] = <uint64_t><Py_UCS4>elem
             else:
-                (<uint64_t*>s_proc.data)[i] = <uint64_t>hash(elem)
+                (<uint64_t*>s_proc.data)[i] = rf_hash(elem)
     except Exception as e:
         free(s_proc.data)
         s_proc.data = NULL

tests/distance/test_Levenshtein.py

@@ -60,6 +60,7 @@ def test_cross_type_matching():
     assert Levenshtein.distance("aaaa", ["a", "a", "a", "a"]) == 0
     # todo add support in pure python
     assert Levenshtein_cpp.distance("aaaa", [ord("a"), ord("a"), "a", "a"]) == 0
+    assert Levenshtein_cpp.distance([0, -1], [0, -2]) == 1
 
 
 def test_word_error_rate():