SF patch 576101, by Oren Tirosh: alternative implementation of

interning.  I modified Oren's patch significantly, but the basic idea
and most of the implementation is unchanged.  Interned strings created
with PyString_InternInPlace() are now mortal, and you must keep a
reference to the resulting string around; use the new function
PyString_InternImmortal() to create immortal interned strings.

7 files changed, 171 insertions, 106 deletions

diff --git a/Doc/lib/libfuncs.tex b/Doc/lib/libfuncs.tex
index 1aa59dc..a6428da 100644
--- a/Doc/lib/libfuncs.tex
+++ b/Doc/lib/libfuncs.tex
@@ -518,8 +518,10 @@ def my_import(name):
   be done by a pointer compare instead of a string compare.  Normally,
   the names used in Python programs are automatically interned, and
   the dictionaries used to hold module, class or instance attributes
-  have interned keys.  Interned strings are immortal (never get
-  garbage collected).
+  have interned keys.  \versionchanged[Interned strings are not
+  immortal (like they used to be in Python 2.2 and before);
+  you must keep a reference to the return value of \function{intern()}
+  around to benefit from it]{2.3}
 \end{funcdesc}
 
 \begin{funcdesc}{isinstance}{object, classinfo}
diff --git a/Include/modsupport.h b/Include/modsupport.h
index 2e577ab..7e69921 100644
--- a/Include/modsupport.h
+++ b/Include/modsupport.h
@@ -23,8 +23,8 @@ PyAPI_FUNC(int) PyModule_AddObject(PyObject *, char *, PyObject *);
 PyAPI_FUNC(int) PyModule_AddIntConstant(PyObject *, char *, long);
 PyAPI_FUNC(int) PyModule_AddStringConstant(PyObject *, char *, char *);
 
-#define PYTHON_API_VERSION 1011
-#define PYTHON_API_STRING "1011"
+#define PYTHON_API_VERSION 1012
+#define PYTHON_API_STRING "1012"
 /* The API version is maintained (independently from the Python version)
    so we can detect mismatches between the interpreter and dynamically
    loaded modules.  These are diagnosed by an error message but
@@ -38,6 +38,9 @@ PyAPI_FUNC(int) PyModule_AddStringConstant(PyObject *, char *, char *);
    Please add a line or two to the top of this log for each API
    version change:
 
+   19-Aug-2002  GvR	1012	Changes to string object struct for
+   				interning changes, saving 3 bytes.
+
    17-Jul-2001	GvR	1011	Descr-branch, just to be on the safe side
 
    25-Jan-2001  FLD     1010    Parameters added to PyCode_New() and
diff --git a/Include/stringobject.h b/Include/stringobject.h
index fd0f49a..ff30704 100644
--- a/Include/stringobject.h
+++ b/Include/stringobject.h
@@ -25,7 +25,7 @@ functions should be applied to nil objects.
 */
 
 /* Caching the hash (ob_shash) saves recalculation of a string's hash value.
-   Interning strings (ob_sinterned) tries to ensure that only one string
+   Interning strings (ob_sstate) tries to ensure that only one string
    object with a given value exists, so equality tests can be one pointer
    comparison.  This is generally restricted to strings that "look like"
    Python identifiers, although the intern() builtin can be used to force
@@ -35,10 +35,14 @@ functions should be applied to nil objects.
 typedef struct {
     PyObject_VAR_HEAD
     long ob_shash;
-    PyObject *ob_sinterned;
+    int ob_sstate;
     char ob_sval[1];
 } PyStringObject;
 
+#define SSTATE_NOT_INTERNED 0
+#define SSTATE_INTERNED_MORTAL 1
+#define SSTATE_INTERNED_IMMORTAL 2
+
 PyAPI_DATA(PyTypeObject) PyBaseString_Type;
 PyAPI_DATA(PyTypeObject) PyString_Type;
 
@@ -66,9 +70,13 @@ extern DL_IMPORT(PyObject *) PyString_DecodeEscape(const char *, int,
 						   const char *);
 
 PyAPI_FUNC(void) PyString_InternInPlace(PyObject **);
+PyAPI_FUNC(void) PyString_InternImmortal(PyObject **);
 PyAPI_FUNC(PyObject *) PyString_InternFromString(const char *);
 PyAPI_FUNC(void) _Py_ReleaseInternedStrings(void);
 
+/* Use only if you know it's a string */
+#define PyString_CHECK_INTERNED(op) (((PyStringObject *)(op))->ob_sstate)
+
 /* Macro, trading safety for speed */
 #define PyString_AS_STRING(op) (((PyStringObject *)(op))->ob_sval)
 #define PyString_GET_SIZE(op)  (((PyStringObject *)(op))->ob_size)
diff --git a/Misc/NEWS b/Misc/NEWS
index d60205e..7034d72 100644
--- a/Misc/NEWS
+++ b/Misc/NEWS
@@ -57,6 +57,10 @@ Type/class unification and new-style classes
 
 Core and builtins
 
+- A subtle change to the semantics of the built-in function intern():
+  interned strings are no longer immortal.  You must keep a reference
+  to the return value intern() around to get the benefit.
+
 - Use of 'None' as a variable, argument or attribute name now
   issues a SyntaxWarning.  In the future, None may become a keyword.
 
@@ -514,6 +518,19 @@ Build
 
 C API
 
+- The string object's layout has changed: the pointer member
+  ob_sinterned has been replaced by an int member ob_sstate.  On some
+  platforms (e.g. most 64-bit systems) this may change the offset of
+  the ob_sval member, so as a precaution the API_VERSION has been
+  incremented.  The apparently unused feature of "indirect interned
+  strings", supported by the ob_sinterned member, is gone.  Interned
+  strings are now usually mortal; theres a new API,
+  PyString_InternImmortal() that creates immortal interned strings.
+  (The ob_sstate member can only take three values; however, while
+  making it a char saves a few bytes per string object on average, in
+  it also slowed things down a bit because ob_sval was no longer
+  aligned.)
+
 - The Py_InitModule*() functions now accept NULL for the 'methods'
   argument.  Modules without global functions are becoming more common
   now that factories can be types rather than functions.
diff --git a/Objects/classobject.c b/Objects/classobject.c
index 32e8d2a..cbbb3c4 100644
--- a/Objects/classobject.c
+++ b/Objects/classobject.c
@@ -2300,37 +2300,38 @@ instancemethod_traverse(PyMethodObject *im, visitproc visit, void *arg)
 	return 0;
 }
 
-static char *
-getclassname(PyObject *class)
+static void
+getclassname(PyObject *class, char *buf, int bufsize)
 {
 	PyObject *name;
 
+	assert(bufsize > 1);
+	strcpy(buf, "?"); /* Default outcome */
 	if (class == NULL)
-		name = NULL;
-	else
-		name = PyObject_GetAttrString(class, "__name__");
+		return;
+	name = PyObject_GetAttrString(class, "__name__");
 	if (name == NULL) {
 		/* This function cannot return an exception */
 		PyErr_Clear();
-		return "?";
+		return;
 	}
-	if (!PyString_Check(name)) {
-		Py_DECREF(name);
-		return "?";
+	if (PyString_Check(name)) {
+		strncpy(buf, PyString_AS_STRING(name), bufsize);
+		buf[bufsize-1] = '\0';
 	}
-	PyString_InternInPlace(&name);
 	Py_DECREF(name);
-	return PyString_AS_STRING(name);
 }
 
-static char *
-getinstclassname(PyObject *inst)
+static void
+getinstclassname(PyObject *inst, char *buf, int bufsize)
 {
 	PyObject *class;
-	char *name;
 
-	if (inst == NULL)
-		return "nothing";
+	if (inst == NULL) {
+		assert(bufsize > strlen("nothing"));
+		strcpy(buf, "nothing");
+		return;
+	}
 
 	class = PyObject_GetAttrString(inst, "__class__");
 	if (class == NULL) {
@@ -2339,9 +2340,8 @@ getinstclassname(PyObject *inst)
 		class = (PyObject *)(inst->ob_type);
 		Py_INCREF(class);
 	}
-	name = getclassname(class);
+	getclassname(class, buf, bufsize);
 	Py_XDECREF(class);
-	return name;
 }
 
 static PyObject *
@@ -2366,14 +2366,18 @@ instancemethod_call(PyObject *func, PyObject *arg, PyObject *kw)
 				return NULL;
 		}
 		if (!ok) {
+			char clsbuf[256];
+			char instbuf[256];
+			getclassname(class, clsbuf, sizeof(clsbuf));
+			getinstclassname(self, instbuf, sizeof(instbuf));
 			PyErr_Format(PyExc_TypeError,
 				     "unbound method %s%s must be called with "
 				     "%s instance as first argument "
 				     "(got %s%s instead)",
 				     PyEval_GetFuncName(func),
 				     PyEval_GetFuncDesc(func),
-				     getclassname(class),
-				     getinstclassname(self),
+				     clsbuf,
+				     instbuf,
 				     self == NULL ? "" : " instance");
 			return NULL;
 		}
diff --git a/Objects/dictobject.c b/Objects/dictobject.c
index 86c74ba..b7b499e 100644
--- a/Objects/dictobject.c
+++ b/Objects/dictobject.c
@@ -511,15 +511,9 @@ PyDict_SetItem(register PyObject *op, PyObject *key, PyObject *value)
 	}
 	mp = (dictobject *)op;
 	if (PyString_CheckExact(key)) {
-		if (((PyStringObject *)key)->ob_sinterned != NULL) {
-			key = ((PyStringObject *)key)->ob_sinterned;
-			hash = ((PyStringObject *)key)->ob_shash;
-		}
-		else {
-			hash = ((PyStringObject *)key)->ob_shash;
-			if (hash == -1)
-				hash = PyObject_Hash(key);
-		}
+		hash = ((PyStringObject *)key)->ob_shash;
+		if (hash == -1)
+			hash = PyObject_Hash(key);
 	}
 	else {
 		hash = PyObject_Hash(key);
diff --git a/Objects/stringobject.c b/Objects/stringobject.c
index 7d3881e..7a48627 100644
--- a/Objects/stringobject.c
+++ b/Objects/stringobject.c
@@ -15,6 +15,17 @@ int null_strings, one_strings;
 static PyStringObject *characters[UCHAR_MAX + 1];
 static PyStringObject *nullstring;
 
+/* This dictionary holds all interned strings.  Note that references to
+   strings in this dictionary are *not* counted in the string's ob_refcnt.
+   When the interned string reaches a refcnt of 0 the string deallocation
+   function will delete the reference from this dictionary.
+
+   Another way to look at this is that to say that the actual reference 
+   count of a string is:  s->ob_refcnt + (s->ob_sstate?2:0)
+*/
+static PyObject *interned;
+
+
 /*
    For both PyString_FromString() and PyString_FromStringAndSize(), the
    parameter `size' denotes number of characters to allocate, not counting any
@@ -69,7 +80,7 @@ PyString_FromStringAndSize(const char *str, int size)
 		return PyErr_NoMemory();
 	PyObject_INIT_VAR(op, &PyString_Type, size);
 	op->ob_shash = -1;
-	op->ob_sinterned = NULL;
+	op->ob_sstate = SSTATE_NOT_INTERNED;
 	if (str != NULL)
 		memcpy(op->ob_sval, str, size);
 	op->ob_sval[size] = '\0';
@@ -125,7 +136,7 @@ PyString_FromString(const char *str)
 		return PyErr_NoMemory();
 	PyObject_INIT_VAR(op, &PyString_Type, size);
 	op->ob_shash = -1;
-	op->ob_sinterned = NULL;
+	op->ob_sstate = SSTATE_NOT_INTERNED;
 	memcpy(op->ob_sval, str, size+1);
 	/* share short strings */
 	if (size == 0) {
@@ -486,6 +497,24 @@ PyObject *PyString_AsEncodedString(PyObject *str,
 static void
 string_dealloc(PyObject *op)
 {
+	switch (PyString_CHECK_INTERNED(op)) {
+		case SSTATE_NOT_INTERNED:
+			break;
+
+		case SSTATE_INTERNED_MORTAL:
+			/* revive dead object temporarily for DelItem */
+			op->ob_refcnt = 3;
+			if (PyDict_DelItem(interned, op) != 0)
+				Py_FatalError(
+					"deletion of interned string failed");
+			break;
+
+		case SSTATE_INTERNED_IMMORTAL:
+			Py_FatalError("Immortal interned string died.");
+
+		default:
+			Py_FatalError("Inconsistent interned string state.");
+	}
 	op->ob_type->tp_free(op);
 }
 
@@ -885,7 +914,7 @@ string_concat(register PyStringObject *a, register PyObject *bb)
 		return PyErr_NoMemory();
 	PyObject_INIT_VAR(op, &PyString_Type, size);
 	op->ob_shash = -1;
-	op->ob_sinterned = NULL;
+	op->ob_sstate = SSTATE_NOT_INTERNED;
 	memcpy(op->ob_sval, a->ob_sval, (int) a->ob_size);
 	memcpy(op->ob_sval + a->ob_size, b->ob_sval, (int) b->ob_size);
 	op->ob_sval[size] = '\0';
@@ -928,7 +957,7 @@ string_repeat(register PyStringObject *a, register int n)
 		return PyErr_NoMemory();
 	PyObject_INIT_VAR(op, &PyString_Type, size);
 	op->ob_shash = -1;
-	op->ob_sinterned = NULL;
+	op->ob_sstate = SSTATE_NOT_INTERNED;
 	for (i = 0; i < size; i += a->ob_size)
 		memcpy(op->ob_sval+i, a->ob_sval, (int) a->ob_size);
 	op->ob_sval[size] = '\0';
@@ -1093,9 +1122,6 @@ string_hash(PyStringObject *a)
 
 	if (a->ob_shash != -1)
 		return a->ob_shash;
-	if (a->ob_sinterned != NULL)
-		return (a->ob_shash =
-			((PyStringObject *)(a->ob_sinterned))->ob_shash);
 	len = a->ob_size;
 	p = (unsigned char *) a->ob_sval;
 	x = *p << 7;
@@ -3067,8 +3093,7 @@ str_subtype_new(PyTypeObject *type, PyObject *args, PyObject *kwds)
 		memcpy(PyString_AS_STRING(pnew), PyString_AS_STRING(tmp), n+1);
 		((PyStringObject *)pnew)->ob_shash =
 			((PyStringObject *)tmp)->ob_shash;
-		((PyStringObject *)pnew)->ob_sinterned =
-			((PyStringObject *)tmp)->ob_sinterned;
+		((PyStringObject *)pnew)->ob_sstate = SSTATE_NOT_INTERNED;
 	}
 	Py_DECREF(tmp);
 	return pnew;
@@ -3983,22 +4008,6 @@ PyString_Format(PyObject *format, PyObject *args)
 	return NULL;
 }
 
-
-
-/* This dictionary will leak at PyString_Fini() time.  That's acceptable
- * because PyString_Fini() specifically frees interned strings that are
- * only referenced by this dictionary.  The CVS log entry for revision 2.45
- * says:
- *
- *    Change the Fini function to only remove otherwise unreferenced
- *    strings from the interned table.  There are references in
- *    hard-to-find static variables all over the interpreter, and it's not
- *    worth trying to get rid of all those; but "uninterning" isn't fair
- *    either and may cause subtle failures later -- so we have to keep them
- *    in the interned table.
- */
-static PyObject *interned;
-
 void
 PyString_InternInPlace(PyObject **p)
 {
@@ -4006,49 +4015,57 @@ PyString_InternInPlace(PyObject **p)
 	PyObject *t;
 	if (s == NULL || !PyString_Check(s))
 		Py_FatalError("PyString_InternInPlace: strings only please!");
-	if ((t = s->ob_sinterned) != NULL) {
-		if (t == (PyObject *)s)
-			return;
-		Py_INCREF(t);
-		*p = t;
-		Py_DECREF(s);
+	if (PyString_CHECK_INTERNED(s))
 		return;
-	}
 	if (interned == NULL) {
 		interned = PyDict_New();
-		if (interned == NULL)
+		if (interned == NULL) {
+			PyErr_Clear(); /* Don't leave an exception */
 			return;
+		}
 	}
 	if ((t = PyDict_GetItem(interned, (PyObject *)s)) != NULL) {
 		Py_INCREF(t);
-		*p = s->ob_sinterned = t;
-		Py_DECREF(s);
+		Py_DECREF(*p);
+		*p = t;
 		return;
 	}
-	/* Ensure that only true string objects appear in the intern dict,
-	   and as the value of ob_sinterned. */
-	if (PyString_CheckExact(s)) {
-		t = (PyObject *)s;
-		if (PyDict_SetItem(interned, t, t) == 0) {
-			s->ob_sinterned = t;
-			return;
-		}
-	}
-	else {
+	/* Ensure that only true string objects appear in the intern dict */
+	if (!PyString_CheckExact(s)) {
 		t = PyString_FromStringAndSize(PyString_AS_STRING(s),
 						PyString_GET_SIZE(s));
-		if (t != NULL) {
-			if (PyDict_SetItem(interned, t, t) == 0) {
-				*p = s->ob_sinterned = t;
-				Py_DECREF(s);
-				return;
-			}
-			Py_DECREF(t);
+		if (t == NULL) {
+			PyErr_Clear();
+			return;
 		}
+	} else {
+		t = (PyObject*) s;
+		Py_INCREF(t);
+	}
+
+	if (PyDict_SetItem(interned, t, t) == 0) {
+		/* The two references in interned are not counted by
+		refcnt.  The string deallocator will take care of this */
+		((PyObject *)t)->ob_refcnt-=2;
+		PyString_CHECK_INTERNED(t) = SSTATE_INTERNED_MORTAL;
+		Py_DECREF(*p);
+		*p = t;
+		return;
 	}
+	Py_DECREF(t);
 	PyErr_Clear();
 }
 
+void
+PyString_InternImmortal(PyObject **p)
+{
+	PyString_InternInPlace(p);
+	if (PyString_CHECK_INTERNED(*p) != SSTATE_INTERNED_IMMORTAL) {
+		PyString_CHECK_INTERNED(*p) = SSTATE_INTERNED_IMMORTAL;
+		Py_INCREF(*p);
+	}
+}
+
 
 PyObject *
 PyString_InternFromString(const char *cp)
@@ -4070,28 +4087,48 @@ PyString_Fini(void)
 	}
 	Py_XDECREF(nullstring);
 	nullstring = NULL;
-	if (interned) {
-		int pos, changed;
-		PyObject *key, *value;
-		do {
-			changed = 0;
-			pos = 0;
-			while (PyDict_Next(interned, &pos, &key, &value)) {
-				if (key->ob_refcnt == 2 && key == value) {
-					PyDict_DelItem(interned, key);
-					changed = 1;
-				}
-			}
-		} while (changed);
-	}
 }
 
 void _Py_ReleaseInternedStrings(void)
 {
-	if (interned) {
-		fprintf(stderr, "releasing interned strings\n");
-		PyDict_Clear(interned);
-		Py_DECREF(interned);
-		interned = NULL;
+	PyObject *keys;
+	PyStringObject *s;
+	int i, n;
+
+	if (interned == NULL || !PyDict_Check(interned))
+		return;
+	keys = PyDict_Keys(interned);
+	if (keys == NULL || !PyList_Check(keys)) {
+		PyErr_Clear();
+		return;
+	}
+
+	/* Since _Py_ReleaseInternedStrings() is intended to help a leak
+	   detector, interned strings are not forcibly deallocated; rather, we
+	   give them their stolen references back, and then clear and DECREF
+	   the interned dict. */
+	   
+	fprintf(stderr, "releasing interned strings\n");
+	n = PyList_GET_SIZE(keys);
+	for (i = 0; i < n; i++) {
+		s = (PyStringObject *) PyList_GET_ITEM(keys, i);
+		switch (s->ob_sstate) {
+		case SSTATE_NOT_INTERNED:
+			/* XXX Shouldn't happen */
+			break;
+		case SSTATE_INTERNED_IMMORTAL:
+			s->ob_refcnt += 1;
+			break;
+		case SSTATE_INTERNED_MORTAL:
+			s->ob_refcnt += 2;
+			break;
+		default:
+			Py_FatalError("Inconsistent interned string state.");
+		}
+		s->ob_sstate = SSTATE_NOT_INTERNED;
 	}
+	Py_DECREF(keys);
+	PyDict_Clear(interned);
+	Py_DECREF(interned);
+	interned = NULL;
 }