bpo-44337: Shrink the LOAD_ATTR/STORE_ATTR caches (GH-31517)

diff --git a/Include/internal/pycore_code.h b/Include/internal/pycore_code.h
index d7a514d8dc98841b018013c4000e85f5d67896da..88bbefc2474b4cbce05912c9515551b962024d97 100644 (file)
--- a/Include/internal/pycore_code.h
+++ b/Include/internal/pycore_code.h
@@ -23,7 +23,7 @@ typedef struct {
 
 typedef struct {
     uint32_t tp_version;
-    uint32_t dk_version_or_hint;
+    uint32_t dk_version;
 } _PyAttrCache;
 
 typedef struct {

diff --git a/Misc/NEWS.d/next/Core and Builtins/2022-02-22-17-19-45.bpo-44337.XA-egu.rst b/Misc/NEWS.d/next/Core and Builtins/2022-02-22-17-19-45.bpo-44337.XA-egu.rst
new file mode 100644 (file)
index 0000000..5037aa1
--- /dev/null
+++ b/Misc/NEWS.d/next/Core and Builtins/2022-02-22-17-19-45.bpo-44337.XA-egu.rst
@@ -0,0 +1,2 @@
+Reduce the memory usage of specialized :opcode:`LOAD_ATTR` and
+:opcode:`STORE_ATTR` instructions.

diff --git a/Python/ceval.c b/Python/ceval.c
index 308271b8953de027495554738300bd8565db7303..807757019020ab7a146c3e1c74742936c965f348 100644 (file)
--- a/Python/ceval.c
+++ b/Python/ceval.c
@@ -1450,11 +1450,10 @@ eval_frame_handle_pending(PyThreadState *tstate)
 #define LOAD_MODULE_ATTR_OR_METHOD(attr_or_method) \
     SpecializedCacheEntry *caches = GET_CACHE(); \
     _PyAdaptiveEntry *cache0 = &caches[0].adaptive; \
-    _PyAttrCache *cache1 = &caches[-1].attr; \
     DEOPT_IF(!PyModule_CheckExact(owner), LOAD_##attr_or_method); \
     PyDictObject *dict = (PyDictObject *)((PyModuleObject *)owner)->md_dict; \
     assert(dict != NULL); \
-    DEOPT_IF(dict->ma_keys->dk_version != cache1->dk_version_or_hint, \
+    DEOPT_IF(dict->ma_keys->dk_version != cache0->version, \
         LOAD_##attr_or_method); \
     assert(dict->ma_keys->dk_kind == DICT_KEYS_UNICODE); \
     assert(cache0->index < dict->ma_keys->dk_nentries); \
@@ -3452,9 +3451,8 @@ handle_eval_breaker:
             PyTypeObject *tp = Py_TYPE(owner);
             SpecializedCacheEntry *caches = GET_CACHE();
             _PyAdaptiveEntry *cache0 = &caches[0].adaptive;
-            _PyAttrCache *cache1 = &caches[-1].attr;
-            assert(cache1->tp_version != 0);
-            DEOPT_IF(tp->tp_version_tag != cache1->tp_version, LOAD_ATTR);
+            assert(cache0->version != 0);
+            DEOPT_IF(tp->tp_version_tag != cache0->version, LOAD_ATTR);
             assert(tp->tp_dictoffset < 0);
             assert(tp->tp_flags & Py_TPFLAGS_MANAGED_DICT);
             PyDictValues *values = *_PyObject_ValuesPointer(owner);
@@ -3486,15 +3484,14 @@ handle_eval_breaker:
             PyTypeObject *tp = Py_TYPE(owner);
             SpecializedCacheEntry *caches = GET_CACHE();
             _PyAdaptiveEntry *cache0 = &caches[0].adaptive;
-            _PyAttrCache *cache1 = &caches[-1].attr;
-            assert(cache1->tp_version != 0);
-            DEOPT_IF(tp->tp_version_tag != cache1->tp_version, LOAD_ATTR);
+            assert(cache0->version != 0);
+            DEOPT_IF(tp->tp_version_tag != cache0->version, LOAD_ATTR);
             assert(tp->tp_flags & Py_TPFLAGS_MANAGED_DICT);
             PyDictObject *dict = *(PyDictObject **)_PyObject_ManagedDictPointer(owner);
             DEOPT_IF(dict == NULL, LOAD_ATTR);
             assert(PyDict_CheckExact((PyObject *)dict));
             PyObject *name = GETITEM(names, cache0->original_oparg);
-            uint32_t hint = cache1->dk_version_or_hint;
+            uint16_t hint = cache0->index;
             DEOPT_IF(hint >= (size_t)dict->ma_keys->dk_nentries, LOAD_ATTR);
             PyDictKeyEntry *ep = DK_ENTRIES(dict->ma_keys) + hint;
             DEOPT_IF(ep->me_key != name, LOAD_ATTR);
@@ -3514,9 +3511,8 @@ handle_eval_breaker:
             PyTypeObject *tp = Py_TYPE(owner);
             SpecializedCacheEntry *caches = GET_CACHE();
             _PyAdaptiveEntry *cache0 = &caches[0].adaptive;
-            _PyAttrCache *cache1 = &caches[-1].attr;
-            assert(cache1->tp_version != 0);
-            DEOPT_IF(tp->tp_version_tag != cache1->tp_version, LOAD_ATTR);
+            assert(cache0->version != 0);
+            DEOPT_IF(tp->tp_version_tag != cache0->version, LOAD_ATTR);
             char *addr = (char *)owner + cache0->index;
             res = *(PyObject **)addr;
             DEOPT_IF(res == NULL, LOAD_ATTR);
@@ -3553,9 +3549,8 @@ handle_eval_breaker:
             PyTypeObject *tp = Py_TYPE(owner);
             SpecializedCacheEntry *caches = GET_CACHE();
             _PyAdaptiveEntry *cache0 = &caches[0].adaptive;
-            _PyAttrCache *cache1 = &caches[-1].attr;
-            assert(cache1->tp_version != 0);
-            DEOPT_IF(tp->tp_version_tag != cache1->tp_version, STORE_ATTR);
+            assert(cache0->version != 0);
+            DEOPT_IF(tp->tp_version_tag != cache0->version, STORE_ATTR);
             assert(tp->tp_flags & Py_TPFLAGS_MANAGED_DICT);
             PyDictValues *values = *_PyObject_ValuesPointer(owner);
             DEOPT_IF(values == NULL, STORE_ATTR);
@@ -3581,15 +3576,14 @@ handle_eval_breaker:
             PyTypeObject *tp = Py_TYPE(owner);
             SpecializedCacheEntry *caches = GET_CACHE();
             _PyAdaptiveEntry *cache0 = &caches[0].adaptive;
-            _PyAttrCache *cache1 = &caches[-1].attr;
-            assert(cache1->tp_version != 0);
-            DEOPT_IF(tp->tp_version_tag != cache1->tp_version, STORE_ATTR);
+            assert(cache0->version != 0);
+            DEOPT_IF(tp->tp_version_tag != cache0->version, STORE_ATTR);
             assert(tp->tp_flags & Py_TPFLAGS_MANAGED_DICT);
             PyDictObject *dict = *(PyDictObject **)_PyObject_ManagedDictPointer(owner);
             DEOPT_IF(dict == NULL, STORE_ATTR);
             assert(PyDict_CheckExact((PyObject *)dict));
             PyObject *name = GETITEM(names, cache0->original_oparg);
-            uint32_t hint = cache1->dk_version_or_hint;
+            uint16_t hint = cache0->index;
             DEOPT_IF(hint >= (size_t)dict->ma_keys->dk_nentries, STORE_ATTR);
             PyDictKeyEntry *ep = DK_ENTRIES(dict->ma_keys) + hint;
             DEOPT_IF(ep->me_key != name, STORE_ATTR);
@@ -3616,9 +3610,8 @@ handle_eval_breaker:
             PyTypeObject *tp = Py_TYPE(owner);
             SpecializedCacheEntry *caches = GET_CACHE();
             _PyAdaptiveEntry *cache0 = &caches[0].adaptive;
-            _PyAttrCache *cache1 = &caches[-1].attr;
-            assert(cache1->tp_version != 0);
-            DEOPT_IF(tp->tp_version_tag != cache1->tp_version, STORE_ATTR);
+            assert(cache0->version != 0);
+            DEOPT_IF(tp->tp_version_tag != cache0->version, STORE_ATTR);
             char *addr = (char *)owner + cache0->index;
             STAT_INC(STORE_ATTR, hit);
             STACK_SHRINK(1);
@@ -4416,7 +4409,7 @@ handle_eval_breaker:
             assert(self_cls->tp_flags & Py_TPFLAGS_MANAGED_DICT);
             PyDictObject *dict = *(PyDictObject**)_PyObject_ManagedDictPointer(self);
             DEOPT_IF(dict != NULL, LOAD_METHOD);
-            DEOPT_IF(((PyHeapTypeObject *)self_cls)->ht_cached_keys->dk_version != cache1->dk_version_or_hint, LOAD_METHOD);
+            DEOPT_IF(((PyHeapTypeObject *)self_cls)->ht_cached_keys->dk_version != cache1->dk_version, LOAD_METHOD);
             STAT_INC(LOAD_METHOD, hit);
             PyObject *res = cache2->obj;
             assert(res != NULL);

diff --git a/Python/specialize.c b/Python/specialize.c
index 0fc992d30bdc8489d14748cd2649c89ad2556bc3..816cca1074e83883ba5878667f05c2c90295a06e 100644 (file)
--- a/Python/specialize.c
+++ b/Python/specialize.c
@@ -57,14 +57,14 @@ static uint8_t adaptive_opcodes[256] = {
 
 /* The number of cache entries required for a "family" of instructions. */
 static uint8_t cache_requirements[256] = {
-    [LOAD_ATTR] = 2, /* _PyAdaptiveEntry and _PyAttrCache */
+    [LOAD_ATTR] = 1,  // _PyAdaptiveEntry
     [LOAD_GLOBAL] = 2, /* _PyAdaptiveEntry and _PyLoadGlobalCache */
     [LOAD_METHOD] = 3, /* _PyAdaptiveEntry, _PyAttrCache and _PyObjectCache */
     [BINARY_SUBSCR] = 2, /* _PyAdaptiveEntry, _PyObjectCache */
     [STORE_SUBSCR] = 0,
     [CALL] = 2, /* _PyAdaptiveEntry and _PyObjectCache/_PyCallCache */
     [PRECALL] = 2, /* _PyAdaptiveEntry and _PyObjectCache/_PyCallCache */
-    [STORE_ATTR] = 2, /* _PyAdaptiveEntry and _PyAttrCache */
+    [STORE_ATTR] = 1,  // _PyAdaptiveEntry
     [BINARY_OP] = 1,  // _PyAdaptiveEntry
     [COMPARE_OP] = 1, /* _PyAdaptiveEntry */
     [UNPACK_SEQUENCE] = 1,  // _PyAdaptiveEntry
@@ -638,7 +638,7 @@ initial_counter_value(void) {
 static int
 specialize_module_load_attr(
     PyObject *owner, _Py_CODEUNIT *instr, PyObject *name,
-    _PyAdaptiveEntry *cache0, _PyAttrCache *cache1, int opcode,
+    _PyAdaptiveEntry *cache0, int opcode,
     int opcode_module)
 {
     PyModuleObject *m = (PyModuleObject *)owner;
@@ -671,7 +671,7 @@ specialize_module_load_attr(
         SPECIALIZATION_FAIL(opcode, SPEC_FAIL_OUT_OF_VERSIONS);
         return -1;
     }
-    cache1->dk_version_or_hint = keys_version;
+    cache0->version = keys_version;
     cache0->index = (uint16_t)index;
     *instr = _Py_MAKECODEUNIT(opcode_module, _Py_OPARG(*instr));
     return 0;
@@ -760,7 +760,7 @@ static int
 specialize_dict_access(
     PyObject *owner, _Py_CODEUNIT *instr, PyTypeObject *type,
     DescriptorClassification kind, PyObject *name,
-    _PyAdaptiveEntry *cache0, _PyAttrCache *cache1,
+    _PyAdaptiveEntry *cache0,
     int base_op, int values_op, int hint_op)
 {
     assert(kind == NON_OVERRIDING || kind == NON_DESCRIPTOR || kind == ABSENT ||
@@ -782,7 +782,7 @@ specialize_dict_access(
             SPECIALIZATION_FAIL(base_op, SPEC_FAIL_OUT_OF_RANGE);
             return 0;
         }
-        cache1->tp_version = type->tp_version_tag;
+        cache0->version = type->tp_version_tag;
         cache0->index = (uint16_t)index;
         *instr = _Py_MAKECODEUNIT(values_op, _Py_OPARG(*instr));
     }
@@ -795,12 +795,12 @@ specialize_dict_access(
         PyObject *value = NULL;
         Py_ssize_t hint =
             _PyDict_GetItemHint(dict, name, -1, &value);
-        if (hint != (uint32_t)hint) {
+        if (hint != (uint16_t)hint) {
             SPECIALIZATION_FAIL(base_op, SPEC_FAIL_OUT_OF_RANGE);
             return 0;
         }
-        cache1->dk_version_or_hint = (uint32_t)hint;
-        cache1->tp_version = type->tp_version_tag;
+        cache0->index = (uint16_t)hint;
+        cache0->version = type->tp_version_tag;
         *instr = _Py_MAKECODEUNIT(hint_op, _Py_OPARG(*instr));
     }
     return 1;
@@ -810,9 +810,8 @@ int
 _Py_Specialize_LoadAttr(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name, SpecializedCacheEntry *cache)
 {
     _PyAdaptiveEntry *cache0 = &cache->adaptive;
-    _PyAttrCache *cache1 = &cache[-1].attr;
     if (PyModule_CheckExact(owner)) {
-        int err = specialize_module_load_attr(owner, instr, name, cache0, cache1,
+        int err = specialize_module_load_attr(owner, instr, name, cache0,
             LOAD_ATTR, LOAD_ATTR_MODULE);
         if (err) {
             goto fail;
@@ -853,7 +852,7 @@ _Py_Specialize_LoadAttr(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name, Sp
             assert(dmem->type == T_OBJECT_EX);
             assert(offset > 0);
             cache0->index = (uint16_t)offset;
-            cache1->tp_version = type->tp_version_tag;
+            cache0->version = type->tp_version_tag;
             *instr = _Py_MAKECODEUNIT(LOAD_ATTR_SLOT, _Py_OPARG(*instr));
             goto success;
         }
@@ -862,7 +861,7 @@ _Py_Specialize_LoadAttr(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name, Sp
             Py_ssize_t offset = offsetof(PyObject, ob_type);
             assert(offset == (uint16_t)offset);
             cache0->index = (uint16_t)offset;
-            cache1->tp_version = type->tp_version_tag;
+            cache0->version = type->tp_version_tag;
             *instr = _Py_MAKECODEUNIT(LOAD_ATTR_SLOT, _Py_OPARG(*instr));
             goto success;
         }
@@ -883,7 +882,7 @@ _Py_Specialize_LoadAttr(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name, Sp
             break;
     }
     int err = specialize_dict_access(
-        owner, instr, type, kind, name, cache0, cache1,
+        owner, instr, type, kind, name, cache0,
         LOAD_ATTR, LOAD_ATTR_INSTANCE_VALUE, LOAD_ATTR_WITH_HINT
     );
     if (err < 0) {
@@ -908,7 +907,6 @@ int
 _Py_Specialize_StoreAttr(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name, SpecializedCacheEntry *cache)
 {
     _PyAdaptiveEntry *cache0 = &cache->adaptive;
-    _PyAttrCache *cache1 = &cache[-1].attr;
     PyTypeObject *type = Py_TYPE(owner);
     if (PyModule_CheckExact(owner)) {
         SPECIALIZATION_FAIL(STORE_ATTR, SPEC_FAIL_OVERRIDDEN);
@@ -942,7 +940,7 @@ _Py_Specialize_StoreAttr(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name, S
             assert(dmem->type == T_OBJECT_EX);
             assert(offset > 0);
             cache0->index = (uint16_t)offset;
-            cache1->tp_version = type->tp_version_tag;
+            cache0->version = type->tp_version_tag;
             *instr = _Py_MAKECODEUNIT(STORE_ATTR_SLOT, _Py_OPARG(*instr));
             goto success;
         }
@@ -965,7 +963,7 @@ _Py_Specialize_StoreAttr(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name, S
     }
 
     int err = specialize_dict_access(
-        owner, instr, type, kind, name, cache0, cache1,
+        owner, instr, type, kind, name, cache0,
         STORE_ATTR, STORE_ATTR_INSTANCE_VALUE, STORE_ATTR_WITH_HINT
     );
     if (err < 0) {
@@ -1066,7 +1064,7 @@ _Py_Specialize_LoadMethod(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name,
 
     PyTypeObject *owner_cls = Py_TYPE(owner);
     if (PyModule_CheckExact(owner)) {
-        int err = specialize_module_load_attr(owner, instr, name, cache0, cache1,
+        int err = specialize_module_load_attr(owner, instr, name, cache0,
             LOAD_METHOD, LOAD_METHOD_MODULE);
         if (err) {
             goto fail;
@@ -1111,7 +1109,7 @@ _Py_Specialize_LoadMethod(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name,
             SPECIALIZATION_FAIL(LOAD_METHOD, SPEC_FAIL_OUT_OF_VERSIONS);
             goto fail;
         }
-        cache1->dk_version_or_hint = keys_version;
+        cache1->dk_version = keys_version;
         *instr = _Py_MAKECODEUNIT(LOAD_METHOD_CACHED, _Py_OPARG(*instr));
     }
     else {