bpo-40521: Make empty Unicode string per interpreter (GH-21096)

Each interpreter now has its own empty Unicode string singleton.

Include/internal/pycore_interp.h

@@ -71,6 +71,8 @@ struct _Py_bytes_state {
 };
 
 struct _Py_unicode_state {
+    // The empty Unicode object is a singleton to improve performance.
+    PyObject *empty;
     struct _Py_unicode_fs_codec fs_codec;
 };
 

Include/internal/pycore_pylifecycle.h

@@ -31,7 +31,7 @@ PyAPI_FUNC(int) _Py_IsLocaleCoercionTarget(const char *ctype_loc);
 
 /* Various one-time initializers */
 
-extern PyStatus _PyUnicode_Init(void);
+extern PyStatus _PyUnicode_Init(PyThreadState *tstate);
 extern int _PyStructSequence_Init(void);
 extern int _PyLong_Init(PyThreadState *tstate);
 extern PyStatus _PyFaulthandler_Init(int enable);

Misc/NEWS.d/next/Core and Builtins/2020-05-20-01-17-34.bpo-40521.wvAehI.rst

@@ -2,7 +2,7 @@ Each interpreter now its has own free lists, singletons and caches:
 
 * Free lists: float, tuple, list, dict, frame, context,
   asynchronous generator, MemoryError.
-* Singletons: empty tuple, empty bytes string,
+* Singletons: empty tuple, empty bytes string, empty Unicode string,
   single byte character.
 * Slice cache.
 

Objects/stringlib/asciilib.h

@@ -11,7 +11,6 @@
 #define STRINGLIB_CHAR           Py_UCS1
 #define STRINGLIB_TYPE_NAME      "unicode"
 #define STRINGLIB_PARSE_CODE     "U"
-#define STRINGLIB_GET_EMPTY()    unicode_empty
 #define STRINGLIB_ISSPACE        Py_UNICODE_ISSPACE
 #define STRINGLIB_ISLINEBREAK    BLOOM_LINEBREAK
 #define STRINGLIB_ISDECIMAL      Py_UNICODE_ISDECIMAL

Objects/stringlib/partition.h

@@ -4,6 +4,11 @@
 #  error must include "stringlib/fastsearch.h" before including this module
 #endif
 
+#if !STRINGLIB_MUTABLE && !defined(STRINGLIB_GET_EMPTY)
+#  error "STRINGLIB_GET_EMPTY must be defined if STRINGLIB_MUTABLE is zero"
+#endif
+
+
 Py_LOCAL_INLINE(PyObject*)
 STRINGLIB(partition)(PyObject* str_obj,
                     const STRINGLIB_CHAR* str, Py_ssize_t str_len,

Objects/stringlib/stringdefs.h

@@ -1,10 +1,6 @@
 #ifndef STRINGLIB_STRINGDEFS_H
 #define STRINGLIB_STRINGDEFS_H
 
-#ifndef STRINGLIB_GET_EMPTY
-#  error "STRINGLIB_GET_EMPTY macro must be defined"
-#endif
-
 /* this is sort of a hack.  there's at least one place (formatting
    floats) where some stringlib code takes a different path if it's
    compiled as unicode. */

Objects/stringlib/ucs1lib.h

@@ -11,7 +11,6 @@
 #define STRINGLIB_CHAR           Py_UCS1
 #define STRINGLIB_TYPE_NAME      "unicode"
 #define STRINGLIB_PARSE_CODE     "U"
-#define STRINGLIB_GET_EMPTY()    unicode_empty
 #define STRINGLIB_ISSPACE        Py_UNICODE_ISSPACE
 #define STRINGLIB_ISLINEBREAK    BLOOM_LINEBREAK
 #define STRINGLIB_ISDECIMAL      Py_UNICODE_ISDECIMAL

Objects/stringlib/ucs2lib.h

@@ -11,7 +11,6 @@
 #define STRINGLIB_CHAR           Py_UCS2
 #define STRINGLIB_TYPE_NAME      "unicode"
 #define STRINGLIB_PARSE_CODE     "U"
-#define STRINGLIB_GET_EMPTY()    unicode_empty
 #define STRINGLIB_ISSPACE        Py_UNICODE_ISSPACE
 #define STRINGLIB_ISLINEBREAK    BLOOM_LINEBREAK
 #define STRINGLIB_ISDECIMAL      Py_UNICODE_ISDECIMAL

Objects/stringlib/ucs4lib.h

@@ -11,7 +11,6 @@
 #define STRINGLIB_CHAR           Py_UCS4
 #define STRINGLIB_TYPE_NAME      "unicode"
 #define STRINGLIB_PARSE_CODE     "U"
-#define STRINGLIB_GET_EMPTY()    unicode_empty
 #define STRINGLIB_ISSPACE        Py_UNICODE_ISSPACE
 #define STRINGLIB_ISLINEBREAK    BLOOM_LINEBREAK
 #define STRINGLIB_ISDECIMAL      Py_UNICODE_ISDECIMAL

Objects/stringlib/unicodedefs.h

@@ -13,7 +13,6 @@
 #define STRINGLIB_CHAR           Py_UNICODE
 #define STRINGLIB_TYPE_NAME      "unicode"
 #define STRINGLIB_PARSE_CODE     "U"
-#define STRINGLIB_GET_EMPTY()    unicode_empty
 #define STRINGLIB_ISSPACE        Py_UNICODE_ISSPACE
 #define STRINGLIB_ISLINEBREAK    BLOOM_LINEBREAK
 #define STRINGLIB_ISDECIMAL      Py_UNICODE_ISDECIMAL

Objects/unicodeobject.c

@@ -222,26 +222,43 @@ extern "C" {
 static PyObject *interned = NULL;
 #endif
 
-/* The empty Unicode object is shared to improve performance. */
+static struct _Py_unicode_state*
-static PyObject *unicode_empty = NULL;
+get_unicode_state(void)
+{
+    PyInterpreterState *interp = _PyInterpreterState_GET();
+    return &interp->unicode;
+}
 
-#define _Py_INCREF_UNICODE_EMPTY()                      \
+
-    do {                                                \
+// Return a borrowed reference to the empty string singleton.
-        if (unicode_empty != NULL)                      \
+// Return NULL if the singleton was not created yet.
-            Py_INCREF(unicode_empty);                   \
+static inline PyObject* unicode_get_empty(void)
-        else {                                          \
+{
-            unicode_empty = PyUnicode_New(0, 0);        \
+    struct _Py_unicode_state *state = get_unicode_state();
-            if (unicode_empty != NULL) {                \
+    return state->empty;
-                Py_INCREF(unicode_empty);               \
+}
-                assert(_PyUnicode_CheckConsistency(unicode_empty, 1)); \
+
-            }                                           \
+static inline PyObject* unicode_new_empty(void)
-        }                                               \
+{
-    } while (0)
+    struct _Py_unicode_state *state = get_unicode_state();
+    PyObject *empty = state->empty;
+    if (empty != NULL) {
+        Py_INCREF(empty);
+    }
+    else {
+        empty = PyUnicode_New(0, 0);
+        if (empty != NULL) {
+            Py_INCREF(empty);
+            assert(_PyUnicode_CheckConsistency(empty, 1));
+            state->empty = empty;
+        }
+    }
+    return empty;
+}
 
 #define _Py_RETURN_UNICODE_EMPTY()   \
     do {                             \
-        _Py_INCREF_UNICODE_EMPTY();                     \
+        return unicode_new_empty();  \
-        return unicode_empty;                           \
     } while (0)
 
 static inline void
@@ -676,11 +693,15 @@ unicode_result_ready(PyObject *unicode)
 
     length = PyUnicode_GET_LENGTH(unicode);
     if (length == 0) {
-        if (unicode != unicode_empty) {
+        PyObject *empty = unicode_get_empty();
+        if (unicode != empty) {
             Py_DECREF(unicode);
-            _Py_RETURN_UNICODE_EMPTY();
+
+            Py_INCREF(empty);
+            return empty;
         }
-        return unicode_empty;
+        // unicode is the empty string singleton
+        return unicode;
     }
 
 #ifdef LATIN1_SINGLETONS
@@ -864,7 +885,7 @@ xmlcharrefreplace(_PyBytesWriter *writer, char *str,
    to keep things simple, we use a single bitmask, using the least 5
    bits from each unicode characters as the bit index. */
 
-/* the linebreak mask is set up by Unicode_Init below */
+/* the linebreak mask is set up by _PyUnicode_Init() below */
 
 #if LONG_BIT >= 128
 #define BLOOM_WIDTH 128
@@ -938,6 +959,8 @@ ensure_unicode(PyObject *obj)
 
 /* Compilation of templated routines */
 
+#define STRINGLIB_GET_EMPTY()    unicode_get_empty()
+
 #include "stringlib/asciilib.h"
 #include "stringlib/fastsearch.h"
 #include "stringlib/partition.h"
@@ -986,6 +1009,8 @@ _Py_COMP_DIAG_IGNORE_DEPR_DECLS
 #include "stringlib/undef.h"
 _Py_COMP_DIAG_POP
 
+#undef STRINGLIB_GET_EMPTY
+
 /* --- Unicode Object ----------------------------------------------------- */
 
 static inline Py_ssize_t
@@ -1234,9 +1259,12 @@ _PyUnicode_New(Py_ssize_t length)
     size_t new_size;
 
     /* Optimization for empty strings */
-    if (length == 0 && unicode_empty != NULL) {
+    if (length == 0) {
-        Py_INCREF(unicode_empty);
+        PyObject *empty = unicode_get_empty();
-        return (PyUnicodeObject*)unicode_empty;
+        if (empty != NULL) {
+            Py_INCREF(empty);
+            return (PyUnicodeObject *)empty;
+        }
     }
 
     /* Ensure we won't overflow the size. */
@@ -1386,6 +1414,15 @@ _PyUnicode_Dump(PyObject *op)
 PyObject *
 PyUnicode_New(Py_ssize_t size, Py_UCS4 maxchar)
 {
+    /* Optimization for empty strings */
+    if (size == 0) {
+        PyObject *empty = unicode_get_empty();
+        if (empty != NULL) {
+            Py_INCREF(empty);
+            return empty;
+        }
+    }
+
     PyObject *obj;
     PyCompactUnicodeObject *unicode;
     void *data;
@@ -1394,12 +1431,6 @@ PyUnicode_New(Py_ssize_t size, Py_UCS4 maxchar)
     Py_ssize_t char_size;
     Py_ssize_t struct_size;
 
-    /* Optimization for empty strings */
-    if (size == 0 && unicode_empty != NULL) {
-        Py_INCREF(unicode_empty);
-        return unicode_empty;
-    }
-
     is_ascii = 0;
     is_sharing = 0;
     struct_size = sizeof(PyCompactUnicodeObject);
@@ -1970,7 +2001,8 @@ unicode_dealloc(PyObject *unicode)
 static int
 unicode_is_singleton(PyObject *unicode)
 {
-    if (unicode == unicode_empty) {
+    struct _Py_unicode_state *state = get_unicode_state();
+    if (unicode == state->empty) {
         return 1;
     }
 #ifdef LATIN1_SINGLETONS
@@ -2026,10 +2058,10 @@ unicode_resize(PyObject **p_unicode, Py_ssize_t length)
         return 0;
 
     if (length == 0) {
-        _Py_INCREF_UNICODE_EMPTY();
+        PyObject *empty = unicode_new_empty();
-        if (!unicode_empty)
+        if (!empty)
             return -1;
-        Py_SETREF(*p_unicode, unicode_empty);
+        Py_SETREF(*p_unicode, empty);
         return 0;
     }
 
@@ -10836,10 +10868,10 @@ replace(PyObject *self, PyObject *str1,
         }
         new_size = slen + n * (len2 - len1);
         if (new_size == 0) {
-            _Py_INCREF_UNICODE_EMPTY();
+            PyObject *empty = unicode_new_empty();
-            if (!unicode_empty)
+            if (!empty)
                 goto error;
-            u = unicode_empty;
+            u = empty;
             goto done;
         }
         if (new_size > (PY_SSIZE_T_MAX / rkind)) {
@@ -11497,10 +11529,13 @@ PyUnicode_Concat(PyObject *left, PyObject *right)
         return NULL;
 
     /* Shortcuts */
-    if (left == unicode_empty)
+    PyObject *empty = unicode_get_empty();  // Borrowed reference
+    if (left == empty) {
         return PyUnicode_FromObject(right);
-    if (right == unicode_empty)
+    }
+    if (right == empty) {
         return PyUnicode_FromObject(left);
+    }
 
     left_len = PyUnicode_GET_LENGTH(left);
     right_len = PyUnicode_GET_LENGTH(right);
@@ -11551,14 +11586,16 @@ PyUnicode_Append(PyObject **p_left, PyObject *right)
         goto error;
 
     /* Shortcuts */
-    if (left == unicode_empty) {
+    PyObject *empty = unicode_get_empty();  // Borrowed reference
+    if (left == empty) {
         Py_DECREF(left);
         Py_INCREF(right);
         *p_left = right;
         return;
     }
-    if (right == unicode_empty)
+    if (right == empty) {
         return;
+    }
 
     left_len = PyUnicode_GET_LENGTH(left);
     right_len = PyUnicode_GET_LENGTH(right);
@@ -13255,12 +13292,12 @@ PyUnicode_Partition(PyObject *str_obj, PyObject *sep_obj)
     len1 = PyUnicode_GET_LENGTH(str_obj);
     len2 = PyUnicode_GET_LENGTH(sep_obj);
     if (kind1 < kind2 || len1 < len2) {
-        _Py_INCREF_UNICODE_EMPTY();
+        PyObject *empty = unicode_get_empty();  // Borrowed reference
-        if (!unicode_empty)
+        if (!empty) {
             out = NULL;
+        }
         else {
-            out = PyTuple_Pack(3, str_obj, unicode_empty, unicode_empty);
+            out = PyTuple_Pack(3, str_obj, empty, empty);
-            Py_DECREF(unicode_empty);
         }
         return out;
     }
@@ -13313,12 +13350,12 @@ PyUnicode_RPartition(PyObject *str_obj, PyObject *sep_obj)
     len1 = PyUnicode_GET_LENGTH(str_obj);
     len2 = PyUnicode_GET_LENGTH(sep_obj);
     if (kind1 < kind2 || len1 < len2) {
-        _Py_INCREF_UNICODE_EMPTY();
+        PyObject *empty = unicode_get_empty();  // Borrowed reference
-        if (!unicode_empty)
+        if (!empty) {
             out = NULL;
+        }
         else {
-            out = PyTuple_Pack(3, unicode_empty, unicode_empty, str_obj);
+            out = PyTuple_Pack(3, empty, empty, str_obj);
-            Py_DECREF(unicode_empty);
         }
         return out;
     }
@@ -15538,10 +15575,10 @@ PyTypeObject PyUnicode_Type = {
 /* Initialize the Unicode implementation */
 
 PyStatus
-_PyUnicode_Init(void)
+_PyUnicode_Init(PyThreadState *tstate)
 {
     /* XXX - move this array to unicodectype.c ? */
-    Py_UCS2 linebreak[] = {
+    const Py_UCS2 linebreak[] = {
         0x000A, /* LINE FEED */
         0x000D, /* CARRIAGE RETURN */
         0x001C, /* FILE SEPARATOR */
@@ -15553,21 +15590,22 @@ _PyUnicode_Init(void)
     };
 
     /* Init the implementation */
-    _Py_INCREF_UNICODE_EMPTY();
+    PyObject *empty = unicode_new_empty();
-    if (!unicode_empty) {
+    if (!empty) {
-        return _PyStatus_ERR("Can't create empty string");
+        return _PyStatus_NO_MEMORY();
-    }
-    Py_DECREF(unicode_empty);
-
-    if (PyType_Ready(&PyUnicode_Type) < 0) {
-        return _PyStatus_ERR("Can't initialize unicode type");
     }
+    Py_DECREF(empty);
 
+    if (_Py_IsMainInterpreter(tstate)) {
         /* initialize the linebreak bloom filter */
         bloom_linebreak = make_bloom_mask(
             PyUnicode_2BYTE_KIND, linebreak,
             Py_ARRAY_LENGTH(linebreak));
 
+        if (PyType_Ready(&PyUnicode_Type) < 0) {
+            return _PyStatus_ERR("Can't initialize unicode type");
+        }
+
         if (PyType_Ready(&EncodingMapType) < 0) {
              return _PyStatus_ERR("Can't initialize encoding map type");
         }
@@ -15577,6 +15615,7 @@ _PyUnicode_Init(void)
         if (PyType_Ready(&PyFormatterIter_Type) < 0) {
             return _PyStatus_ERR("Can't initialize formatter iter type");
         }
+    }
     return _PyStatus_OK();
 }
 
@@ -16205,7 +16244,10 @@ _PyUnicode_EnableLegacyWindowsFSEncoding(void)
 void
 _PyUnicode_Fini(PyThreadState *tstate)
 {
-    if (_Py_IsMainInterpreter(tstate)) {
+    struct _Py_unicode_state *state = &tstate->interp->unicode;
+
+    int is_main_interp = _Py_IsMainInterpreter(tstate);
+    if (is_main_interp) {
 #if defined(WITH_VALGRIND) || defined(__INSURE__)
         /* Insure++ is a memory analysis tool that aids in discovering
          * memory leaks and other memory problems.  On Python exit, the
@@ -16218,9 +16260,11 @@ _PyUnicode_Fini(PyThreadState *tstate)
          */
         unicode_release_interned();
 #endif /* __INSURE__ */
+    }
 
-        Py_CLEAR(unicode_empty);
+    Py_CLEAR(state->empty);
 
+    if (is_main_interp) {
 #ifdef LATIN1_SINGLETONS
         for (Py_ssize_t i = 0; i < 256; i++) {
             Py_CLEAR(unicode_latin1[i]);

Python/pylifecycle.c

@@ -595,12 +595,10 @@ pycore_init_types(PyThreadState *tstate)
         return _PyStatus_ERR("can't init longs");
     }
 
-    if (is_main_interp) {
+    status = _PyUnicode_Init(tstate);
-        status = _PyUnicode_Init();
     if (_PyStatus_EXCEPTION(status)) {
         return status;
     }
-    }
 
     status = _PyExc_Init(tstate);
     if (_PyStatus_EXCEPTION(status)) {