gh-111968: Use per-thread freelists for dict in free-threading

Include/internal/pycore_dict.h

@@ -9,6 +9,7 @@ extern "C" {
 #  error "this header requires Py_BUILD_CORE define"
 #endif
 
+#include "pycore_freelist.h"      // _PyFreeListState
 #include "pycore_identifier.h"    // _Py_Identifier
 #include "pycore_object.h"        // PyDictOrValues
 
@@ -69,7 +70,7 @@ extern PyObject* _PyDictView_Intersect(PyObject* self, PyObject *other);
 
 /* runtime lifecycle */
 
-extern void _PyDict_Fini(PyInterpreterState *interp);
+extern void _PyDict_Fini(_PyFreeListState *state);
 
 
 /* other API */
@@ -210,7 +211,7 @@ static inline PyDictUnicodeEntry* DK_UNICODE_ENTRIES(PyDictKeysObject *dk) {
 #define DICT_VERSION_MASK (DICT_VERSION_INCREMENT - 1)
 
 #define DICT_NEXT_VERSION(INTERP) \
-    ((INTERP)->dict_state.global_version += DICT_VERSION_INCREMENT)
+    ((INTERP)->global_dict_state.global_version += DICT_VERSION_INCREMENT)
 
 void
 _PyDict_SendEvent(int watcher_bits,

Include/internal/pycore_freelist.h

@@ -17,13 +17,15 @@ extern "C" {
 #  define PyTuple_NFREELISTS PyTuple_MAXSAVESIZE
 #  define PyTuple_MAXFREELIST 2000
 #  define PyList_MAXFREELIST 80
+#  define PyDict_MAXFREELIST 80
 #  define PyFloat_MAXFREELIST 100
 #  define PyContext_MAXFREELIST 255
 # define _PyAsyncGen_MAXFREELIST 80
 #else
 #  define PyTuple_NFREELISTS 0
 #  define PyTuple_MAXFREELIST 0
 #  define PyList_MAXFREELIST 0
+#  define PyDict_MAXFREELIST 0
 #  define PyFloat_MAXFREELIST 0
 #  define PyContext_MAXFREELIST 0
 #  define _PyAsyncGen_MAXFREELIST 0
@@ -63,6 +65,16 @@ struct _Py_float_state {
 #endif
 };
 
+struct _Py_dict_state {
+#ifdef WITH_FREELISTS
+    /* Dictionary reuse scheme to save calls to malloc and free */
+    PyDictObject *free_list[PyDict_MAXFREELIST];
+    PyDictKeysObject *keys_free_list[PyDict_MAXFREELIST];
+    int numfree;
+    int keys_numfree;
+#endif
+};
+
 struct _Py_slice_state {
 #ifdef WITH_FREELISTS
     /* Using a cache is very effective since typically only a single slice is
@@ -97,6 +109,7 @@ typedef struct _Py_freelist_state {
     struct _Py_float_state float_state;
     struct _Py_tuple_state tuple_state;
     struct _Py_list_state list_state;
+    struct _Py_dict_state dict_state;
     struct _Py_slice_state slice_state;
     struct _Py_context_state context_state;
     struct _Py_async_gen_state async_gen_state;

Include/internal/pycore_gc.h

@@ -250,7 +250,7 @@ extern void _PyTuple_ClearFreeList(_PyFreeListState *state, int is_finalization)
 extern void _PyFloat_ClearFreeList(_PyFreeListState *state, int is_finalization);
 extern void _PyList_ClearFreeList(_PyFreeListState *state, int is_finalization);
 extern void _PySlice_ClearCache(_PyFreeListState *state);
-extern void _PyDict_ClearFreeList(PyInterpreterState *interp);
+extern void _PyDict_ClearFreeList(_PyFreeListState *state, int is_finalization);
 extern void _PyAsyncGen_ClearFreeLists(_PyFreeListState *state, int is_finalization);
 extern void _PyContext_ClearFreeList(_PyFreeListState *state, int is_finalization);
 extern void _Py_ScheduleGC(PyInterpreterState *interp);

Include/internal/pycore_dict_state.h → Include/internal/pycore_global_dict_state.h

@@ -1,5 +1,5 @@
-#ifndef Py_INTERNAL_DICT_STATE_H
-#define Py_INTERNAL_DICT_STATE_H
+#ifndef Py_INTERNAL_GLOBAL_DICT_STATE_H
+#define Py_INTERNAL_GLOBAL_DICT_STATE_H
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -8,33 +8,14 @@ extern "C" {
 #  error "this header requires Py_BUILD_CORE define"
 #endif
 
-
-#ifndef WITH_FREELISTS
-// without freelists
-#  define PyDict_MAXFREELIST 0
-#endif
-
-#ifndef PyDict_MAXFREELIST
-#  define PyDict_MAXFREELIST 80
-#endif
-
 #define DICT_MAX_WATCHERS 8
 
-struct _Py_dict_state {
+struct _Py_global_dict_state {
     /*Global counter used to set ma_version_tag field of dictionary.
      * It is incremented each time that a dictionary is created and each
      * time that a dictionary is modified. */
     uint64_t global_version;
     uint32_t next_keys_version;
-
-#if PyDict_MAXFREELIST > 0
-    /* Dictionary reuse scheme to save calls to malloc and free */
-    PyDictObject *free_list[PyDict_MAXFREELIST];
-    PyDictKeysObject *keys_free_list[PyDict_MAXFREELIST];
-    int numfree;
-    int keys_numfree;
-#endif
-
     PyDict_WatchCallback watchers[DICT_MAX_WATCHERS];
 };
 
@@ -47,4 +28,4 @@ struct _Py_dict_state {
 #ifdef __cplusplus
 }
 #endif
-#endif   /* !Py_INTERNAL_DICT_STATE_H */
+#endif   /* !Py_INTERNAL_GLOBAL_DICT_STATE_H */

Include/internal/pycore_interp.h

@@ -8,33 +8,34 @@ extern "C" {
 #  error "this header requires Py_BUILD_CORE define"
 #endif
 
-#include <stdbool.h>              // bool
-
-#include "pycore_ast_state.h"     // struct ast_state
-#include "pycore_atexit.h"        // struct atexit_state
-#include "pycore_ceval_state.h"   // struct _ceval_state
-#include "pycore_code.h"          // struct callable_cache
-#include "pycore_context.h"       // struct _Py_context_state
-#include "pycore_crossinterp.h"   // struct _xidregistry
-#include "pycore_dict_state.h"    // struct _Py_dict_state
-#include "pycore_dtoa.h"          // struct _dtoa_state
-#include "pycore_exceptions.h"    // struct _Py_exc_state
-#include "pycore_floatobject.h"   // struct _Py_float_state
-#include "pycore_function.h"      // FUNC_MAX_WATCHERS
-#include "pycore_gc.h"            // struct _gc_runtime_state
-#include "pycore_genobject.h"     // struct _Py_async_gen_state
-#include "pycore_global_objects.h"// struct _Py_interp_cached_objects
-#include "pycore_import.h"        // struct _import_state
-#include "pycore_instruments.h"   // _PY_MONITORING_EVENTS
-#include "pycore_list.h"          // struct _Py_list_state
-#include "pycore_mimalloc.h"      // struct _mimalloc_interp_state
-#include "pycore_object_state.h"  // struct _py_object_state
-#include "pycore_obmalloc.h"      // struct _obmalloc_state
-#include "pycore_tstate.h"        // _PyThreadStateImpl
-#include "pycore_tuple.h"         // struct _Py_tuple_state
-#include "pycore_typeobject.h"    // struct types_state
-#include "pycore_unicodeobject.h" // struct _Py_unicode_state
-#include "pycore_warnings.h"      // struct _warnings_runtime_state
+#include <stdbool.h>                  // bool
+
+#include "pycore_ast_state.h"         // struct ast_state
+#include "pycore_atexit.h"            // struct atexit_state
+#include "pycore_ceval_state.h"       // struct _ceval_state
+#include "pycore_code.h"              // struct callable_cache
+#include "pycore_context.h"           // struct _Py_context_state
+#include "pycore_crossinterp.h"       // struct _xidregistry
+#include "pycore_dtoa.h"              // struct _dtoa_state
+#include "pycore_exceptions.h"        // struct _Py_exc_state
+#include "pycore_floatobject.h"       // struct _Py_float_state
+#include "pycore_freelist.h"          // struct _Py_freelist_state
+#include "pycore_function.h"          // FUNC_MAX_WATCHERS
+#include "pycore_gc.h"                // struct _gc_runtime_state
+#include "pycore_genobject.h"         // struct _Py_async_gen_state
+#include "pycore_global_dict_state.h" // struct _Py_global_dict_state
+#include "pycore_global_objects.h"    // struct _Py_interp_cached_objects
+#include "pycore_import.h"            // struct _import_state
+#include "pycore_instruments.h"       // _PY_MONITORING_EVENTS
+#include "pycore_list.h"              // struct _Py_list_state
+#include "pycore_mimalloc.h"          // struct _mimalloc_interp_state
+#include "pycore_object_state.h"      // struct _py_object_state
+#include "pycore_obmalloc.h"          // struct _obmalloc_state
+#include "pycore_tstate.h"            // _PyThreadStateImpl
+#include "pycore_tuple.h"             // struct _Py_tuple_state
+#include "pycore_typeobject.h"        // struct types_state
+#include "pycore_unicodeobject.h"     // struct _Py_unicode_state
+#include "pycore_warnings.h"          // struct _warnings_runtime_state
 
 
 struct _Py_long_state {
@@ -188,8 +189,7 @@ struct _is {
     struct _dtoa_state dtoa;
     struct _py_func_state func_state;
 
-    struct _Py_tuple_state tuple;
-    struct _Py_dict_state dict_state;
+    struct _Py_global_dict_state global_dict_state;
     struct _Py_exc_state exc_state;
 
     struct ast_state ast;

Include/internal/pycore_runtime_init.h

@@ -167,7 +167,7 @@ extern PyTypeObject _PyExc_MemoryError;
         }, \
         .object_state = _py_object_state_INIT(INTERP), \
         .dtoa = _dtoa_state_INIT(&(INTERP)), \
-        .dict_state = _dict_state_INIT, \
+        .global_dict_state = _dict_state_INIT, \
         .func_state = { \
             .next_version = 1, \
         }, \

Makefile.pre.in

@@ -1821,7 +1821,6 @@ PYTHON_HEADERS= \
 		$(srcdir)/Include/internal/pycore_critical_section.h \
 		$(srcdir)/Include/internal/pycore_crossinterp.h \
 		$(srcdir)/Include/internal/pycore_dict.h \
-		$(srcdir)/Include/internal/pycore_dict_state.h \
 		$(srcdir)/Include/internal/pycore_descrobject.h \
 		$(srcdir)/Include/internal/pycore_dtoa.h \
 		$(srcdir)/Include/internal/pycore_exceptions.h \
@@ -1835,6 +1834,7 @@ PYTHON_HEADERS= \
 		$(srcdir)/Include/internal/pycore_genobject.h \
 		$(srcdir)/Include/internal/pycore_getopt.h \
 		$(srcdir)/Include/internal/pycore_gil.h \
+		$(srcdir)/Include/internal/pycore_global_dict_state.h \
 		$(srcdir)/Include/internal/pycore_global_objects.h \
 		$(srcdir)/Include/internal/pycore_global_objects_fini_generated.h \
 		$(srcdir)/Include/internal/pycore_hamt.h \