GH-117581: Specialize binary operators by refcount as well as type.

Include/cpython/pystats.h

@@ -154,9 +154,9 @@ typedef struct _stats {
     OptimizationStats optimization_stats;
     RareEventStats rare_event_stats;
     GCStats *gc_stats;
+    uint64_t binary_specialization_failure[1<<15];
 } PyStats;
 
-
 // Export for shared extensions like 'math'
 PyAPI_DATA(PyStats*) _Py_stats;
 

Include/internal/pycore_code.h

@@ -51,6 +51,7 @@ typedef struct {
 
 typedef struct {
     _Py_BackoffCounter counter;
+    uint16_t type_versions;
 } _PyBinaryOpCache;
 
 #define INLINE_CACHE_ENTRIES_BINARY_OP CACHE_ENTRIES(_PyBinaryOpCache)
@@ -298,6 +299,8 @@ extern void _Py_Specialize_Send(PyObject *receiver, _Py_CODEUNIT *instr);
 extern void _Py_Specialize_ToBool(PyObject *value, _Py_CODEUNIT *instr);
 extern void _Py_Specialize_ContainsOp(PyObject *value, _Py_CODEUNIT *instr);
 
+
+PyAPI_DATA(const binaryfunc) _Py_BinaryFunctionTable[];
 #ifdef Py_STATS
 
 #include "pycore_bitutils.h"  // _Py_bit_length

Include/internal/pycore_list.h

@@ -60,6 +60,8 @@ typedef struct {
 
 PyAPI_FUNC(PyObject *)_PyList_FromArraySteal(PyObject *const *src, Py_ssize_t n);
 
+extern PyObject *_PyList_Concat(PyListObject *a, PyListObject *b);
+
 #ifdef __cplusplus
 }
 #endif

Include/internal/pycore_long.h

@@ -121,9 +121,17 @@ PyAPI_DATA(PyObject*) _PyLong_Rshift(PyObject *, size_t);
 // Export for 'math' shared extension
 PyAPI_DATA(PyObject*) _PyLong_Lshift(PyObject *, size_t);
 
+PyAPI_FUNC(PyObject*) _PyLong_Add_1X(PyLongObject *left, PyLongObject *right);
+PyAPI_FUNC(PyObject*) _PyLong_Add_X1(PyLongObject *left, PyLongObject *right);
 PyAPI_FUNC(PyObject*) _PyLong_Add(PyLongObject *left, PyLongObject *right);
 PyAPI_FUNC(PyObject*) _PyLong_Multiply(PyLongObject *left, PyLongObject *right);
 PyAPI_FUNC(PyObject*) _PyLong_Subtract(PyLongObject *left, PyLongObject *right);
+PyObject* _PyLong_And(PyLongObject *left, PyLongObject *right);
+PyObject* _PyLong_Or(PyLongObject *left, PyLongObject *right);
+PyObject* _PyLong_Xor(PyLongObject *left, PyLongObject *right);
+PyObject* _PyLong_FloorDiv(PyLongObject *left, PyLongObject *right);
+PyObject* _PyLong_LShiftObject(PyLongObject *left, PyLongObject *right);
+PyObject* _PyLong_RShiftObject(PyLongObject *left, PyLongObject *right);
 
 // Export for 'binascii' shared extension.
 PyAPI_DATA(unsigned char) _PyLong_DigitValue[256];
@@ -311,6 +319,8 @@ _PyLong_FlipSign(PyLongObject *op) {
 #define _PyLong_FALSE_TAG TAG_FROM_SIGN_AND_SIZE(0, 0)
 #define _PyLong_TRUE_TAG TAG_FROM_SIGN_AND_SIZE(1, 1)
 
+extern double _PyLong_AsDouble(PyLongObject *l);
+
 #ifdef __cplusplus
 }
 #endif

Include/internal/pycore_optimizer.h

@@ -122,6 +122,7 @@ extern _Py_UopsSymbol *_Py_uop_sym_new_type(
 extern _Py_UopsSymbol *_Py_uop_sym_new_const(_Py_UOpsContext *ctx, PyObject *const_val);
 extern _Py_UopsSymbol *_Py_uop_sym_new_null(_Py_UOpsContext *ctx);
 extern bool _Py_uop_sym_has_type(_Py_UopsSymbol *sym);
+extern PyTypeObject *_Py_uop_sym_get_type(_Py_UopsSymbol *sym);
 extern bool _Py_uop_sym_matches_type(_Py_UopsSymbol *sym, PyTypeObject *typ);
 extern void _Py_uop_sym_set_null(_Py_UOpsContext *ctx, _Py_UopsSymbol *sym);
 extern void _Py_uop_sym_set_non_null(_Py_UOpsContext *ctx, _Py_UopsSymbol *sym);
@@ -131,7 +132,6 @@ extern bool _Py_uop_sym_is_bottom(_Py_UopsSymbol *sym);
 extern int _Py_uop_sym_truthiness(_Py_UopsSymbol *sym);
 extern PyTypeObject *_Py_uop_sym_get_type(_Py_UopsSymbol *sym);
 
-
 extern void _Py_uop_abstractcontext_init(_Py_UOpsContext *ctx);
 extern void _Py_uop_abstractcontext_fini(_Py_UOpsContext *ctx);
 

Include/internal/pycore_runtime_init.h

@@ -140,7 +140,7 @@ extern PyTypeObject _PyExc_MemoryError;
             .double_format = _py_float_format_unknown, \
         }, \
         .types = { \
-            .next_version_tag = 1, \
+            .next_version_tag = _Py_TYPE_VERSIONS_PREALLOCATED, \
         }, \
         .static_objects = { \
             .singletons = { \

Include/internal/pycore_typeobject.h

@@ -14,6 +14,24 @@ extern "C" {
 
 /* state */
 
+#define _Py_TYPE_VERSION_INT 1
+#define _Py_TYPE_VERSION_FLOAT 2
+#define _Py_TYPE_VERSION_LIST 3
+#define _Py_TYPE_VERSION_TUPLE 4
+#define _Py_TYPE_VERSION_STR 5
+#define _Py_TYPE_VERSION_SET 6
+#define _Py_TYPE_VERSION_FROZEN_SET 7
+#define _Py_TYPE_VERSION_ARRAY 8
+#define _Py_TYPE_VERSION_DICT 9
+#define _Py_TYPE_VERSION_BYTES 10
+#define _Py_TYPE_VERSION_COMPLEX 11
+#define _Py_TYPE_VERSION_DICTITEMS 12
+#define _Py_TYPE_VERSION_BYTEARRAY 13
+
+#define _Py_TYPE_VERSIONS_PREALLOCATED 16
+
+extern PyTypeObject *const _Py_PreAllocatedTypes[_Py_TYPE_VERSIONS_PREALLOCATED];
+
 #define _Py_TYPE_BASE_VERSION_TAG (2<<16)
 #define _Py_MAX_GLOBAL_TYPE_VERSION_TAG (_Py_TYPE_BASE_VERSION_TAG - 1)