bpo-45340: Don't create object dictionaries unless actually needed (GH-28802)

* Never change types' cached keys. It could invalidate inline attribute objects.

* Lazily create object dictionaries.

* Update specialization of LOAD/STORE_ATTR.

* Don't update shared keys version for deletion of value.

* Update gdb support to handle instance values.

* Rename SPLIT_KEYS opcodes to INSTANCE_VALUE.
This commit is contained in:
Mark Shannon 2021-10-13 14:19:34 +01:00 committed by GitHub
parent 97308dfcdc
commit a8b9350964
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
18 changed files with 721 additions and 400 deletions

View file

@ -3618,7 +3618,7 @@ check_eval_breaker:
}
}
TARGET(LOAD_ATTR_SPLIT_KEYS) {
TARGET(LOAD_ATTR_INSTANCE_VALUE) {
assert(cframe.use_tracing == 0);
PyObject *owner = TOP();
PyObject *res;
@ -3629,11 +3629,10 @@ check_eval_breaker:
assert(cache1->tp_version != 0);
DEOPT_IF(tp->tp_version_tag != cache1->tp_version, LOAD_ATTR);
assert(tp->tp_dictoffset > 0);
PyDictObject *dict = *(PyDictObject **)(((char *)owner) + tp->tp_dictoffset);
DEOPT_IF(dict == NULL, LOAD_ATTR);
assert(PyDict_CheckExact((PyObject *)dict));
DEOPT_IF(dict->ma_keys->dk_version != cache1->dk_version_or_hint, LOAD_ATTR);
res = dict->ma_values->values[cache0->index];
assert(tp->tp_inline_values_offset > 0);
PyDictValues *values = *(PyDictValues **)(((char *)owner) + tp->tp_inline_values_offset);
DEOPT_IF(values == NULL, LOAD_ATTR);
res = values->values[cache0->index];
DEOPT_IF(res == NULL, LOAD_ATTR);
STAT_INC(LOAD_ATTR, hit);
record_cache_hit(cache0);
@ -3725,7 +3724,7 @@ check_eval_breaker:
}
}
TARGET(STORE_ATTR_SPLIT_KEYS) {
TARGET(STORE_ATTR_INSTANCE_VALUE) {
assert(cframe.use_tracing == 0);
PyObject *owner = TOP();
PyTypeObject *tp = Py_TYPE(owner);
@ -3735,31 +3734,23 @@ check_eval_breaker:
assert(cache1->tp_version != 0);
DEOPT_IF(tp->tp_version_tag != cache1->tp_version, STORE_ATTR);
assert(tp->tp_dictoffset > 0);
PyDictObject *dict = *(PyDictObject **)(((char *)owner) + tp->tp_dictoffset);
DEOPT_IF(dict == NULL, STORE_ATTR);
assert(PyDict_CheckExact((PyObject *)dict));
DEOPT_IF(dict->ma_keys->dk_version != cache1->dk_version_or_hint, STORE_ATTR);
assert(tp->tp_inline_values_offset > 0);
PyDictValues *values = *(PyDictValues **)(((char *)owner) + tp->tp_inline_values_offset);
DEOPT_IF(values == NULL, STORE_ATTR);
STAT_INC(STORE_ATTR, hit);
record_cache_hit(cache0);
int index = cache0->index;
STACK_SHRINK(1);
PyObject *value = POP();
PyObject *old_value = dict->ma_values->values[index];
dict->ma_values->values[index] = value;
PyObject *old_value = values->values[index];
values->values[index] = value;
if (old_value == NULL) {
assert(index < 16);
dict->ma_values->mv_order = (dict->ma_values->mv_order << 4) | index;
dict->ma_used++;
values->mv_order = (values->mv_order << 4) | index;
}
else {
Py_DECREF(old_value);
}
/* Ensure dict is GC tracked if it needs to be */
if (!_PyObject_GC_IS_TRACKED(dict) && _PyObject_GC_MAY_BE_TRACKED(value)) {
_PyObject_GC_TRACK(dict);
}
/* PEP 509 */
dict->ma_version_tag = DICT_NEXT_VERSION();
Py_DECREF(owner);
DISPATCH();
}
@ -4474,21 +4465,31 @@ check_eval_breaker:
_PyObjectCache *cache2 = &caches[-2].obj;
DEOPT_IF(self_cls->tp_version_tag != cache1->tp_version, LOAD_METHOD);
assert(cache1->dk_version_or_hint != 0);
assert(cache1->tp_version != 0);
assert(self_cls->tp_dictoffset >= 0);
assert(Py_TYPE(self_cls)->tp_dictoffset > 0);
// inline version of _PyObject_GetDictPtr for offset >= 0
PyObject *dict = self_cls->tp_dictoffset != 0 ?
*(PyObject **) ((char *)self + self_cls->tp_dictoffset) : NULL;
// Ensure self.__dict__ didn't modify keys.
// Don't care if self has no dict, it could be builtin or __slots__.
DEOPT_IF(dict != NULL &&
((PyDictObject *)dict)->ma_keys->dk_version !=
cache1->dk_version_or_hint, LOAD_METHOD);
assert(self_cls->tp_dictoffset > 0);
assert(self_cls->tp_inline_values_offset > 0);
PyDictObject *dict = *(PyDictObject **)(((char *)self) + self_cls->tp_dictoffset);
DEOPT_IF(dict != NULL, LOAD_METHOD);
DEOPT_IF(((PyHeapTypeObject *)self_cls)->ht_cached_keys->dk_version != cache1->dk_version_or_hint, LOAD_METHOD);
STAT_INC(LOAD_METHOD, hit);
record_cache_hit(cache0);
PyObject *res = cache2->obj;
assert(res != NULL);
assert(_PyType_HasFeature(Py_TYPE(res), Py_TPFLAGS_METHOD_DESCRIPTOR));
Py_INCREF(res);
SET_TOP(res);
PUSH(self);
DISPATCH();
}
TARGET(LOAD_METHOD_NO_DICT) {
PyObject *self = TOP();
PyTypeObject *self_cls = Py_TYPE(self);
SpecializedCacheEntry *caches = GET_CACHE();
_PyAdaptiveEntry *cache0 = &caches[0].adaptive;
_PyAttrCache *cache1 = &caches[-1].attr;
_PyObjectCache *cache2 = &caches[-2].obj;
DEOPT_IF(self_cls->tp_version_tag != cache1->tp_version, LOAD_METHOD);
assert(self_cls->tp_dictoffset == 0);
STAT_INC(LOAD_METHOD, hit);
record_cache_hit(cache0);
PyObject *res = cache2->obj;
@ -4530,7 +4531,6 @@ check_eval_breaker:
record_cache_hit(cache0);
PyObject *res = cache2->obj;
assert(res != NULL);
assert(_PyType_HasFeature(Py_TYPE(res), Py_TPFLAGS_METHOD_DESCRIPTOR));
Py_INCREF(res);
SET_TOP(NULL);
Py_DECREF(cls);

View file

@ -41,7 +41,7 @@ static void *opcode_targets[256] = {
&&TARGET_BINARY_SUBSCR_DICT,
&&TARGET_JUMP_ABSOLUTE_QUICK,
&&TARGET_LOAD_ATTR_ADAPTIVE,
&&TARGET_LOAD_ATTR_SPLIT_KEYS,
&&TARGET_LOAD_ATTR_INSTANCE_VALUE,
&&TARGET_LOAD_ATTR_WITH_HINT,
&&TARGET_LOAD_ATTR_SLOT,
&&TARGET_LOAD_ATTR_MODULE,
@ -87,7 +87,7 @@ static void *opcode_targets[256] = {
&&TARGET_SETUP_ANNOTATIONS,
&&TARGET_YIELD_VALUE,
&&TARGET_LOAD_METHOD_MODULE,
&&TARGET_STORE_ATTR_ADAPTIVE,
&&TARGET_LOAD_METHOD_NO_DICT,
&&TARGET_POP_EXCEPT,
&&TARGET_STORE_NAME,
&&TARGET_DELETE_NAME,
@ -119,36 +119,36 @@ static void *opcode_targets[256] = {
&&TARGET_IS_OP,
&&TARGET_CONTAINS_OP,
&&TARGET_RERAISE,
&&TARGET_STORE_ATTR_SPLIT_KEYS,
&&TARGET_STORE_ATTR_ADAPTIVE,
&&TARGET_JUMP_IF_NOT_EXC_MATCH,
&&TARGET_STORE_ATTR_INSTANCE_VALUE,
&&TARGET_STORE_ATTR_SLOT,
&&TARGET_STORE_ATTR_WITH_HINT,
&&TARGET_LOAD_FAST,
&&TARGET_STORE_FAST,
&&TARGET_DELETE_FAST,
&&TARGET_STORE_ATTR_WITH_HINT,
&&TARGET_LOAD_FAST__LOAD_FAST,
&&TARGET_STORE_FAST__LOAD_FAST,
&&TARGET_GEN_START,
&&TARGET_RAISE_VARARGS,
&&TARGET_CALL_FUNCTION,
&&TARGET_MAKE_FUNCTION,
&&TARGET_BUILD_SLICE,
&&TARGET_LOAD_FAST__LOAD_CONST,
&&TARGET_STORE_FAST__LOAD_FAST,
&&TARGET_MAKE_CELL,
&&TARGET_LOAD_CLOSURE,
&&TARGET_LOAD_DEREF,
&&TARGET_STORE_DEREF,
&&TARGET_DELETE_DEREF,
&&TARGET_LOAD_CONST__LOAD_FAST,
&&TARGET_LOAD_FAST__LOAD_CONST,
&&TARGET_CALL_FUNCTION_KW,
&&TARGET_CALL_FUNCTION_EX,
&&TARGET_STORE_FAST__STORE_FAST,
&&TARGET_LOAD_CONST__LOAD_FAST,
&&TARGET_EXTENDED_ARG,
&&TARGET_LIST_APPEND,
&&TARGET_SET_ADD,
&&TARGET_MAP_ADD,
&&TARGET_LOAD_CLASSDEREF,
&&_unknown_opcode,
&&TARGET_STORE_FAST__STORE_FAST,
&&_unknown_opcode,
&&_unknown_opcode,
&&TARGET_MATCH_CLASS,

View file

@ -4,6 +4,7 @@
#include "pycore_dict.h"
#include "pycore_long.h"
#include "pycore_moduleobject.h"
#include "pycore_object.h"
#include "opcode.h"
#include "structmember.h" // struct PyMemberDef, T_OFFSET_EX
@ -462,6 +463,7 @@ specialize_module_load_attr(
PyObject *value = NULL;
PyObject *getattr;
_Py_IDENTIFIER(__getattr__);
assert(owner->ob_type->tp_inline_values_offset == 0);
PyDictObject *dict = (PyDictObject *)m->md_dict;
if (dict == NULL) {
SPECIALIZATION_FAIL(opcode, SPEC_FAIL_NO_DICT);
@ -584,7 +586,7 @@ specialize_dict_access(
PyObject *owner, _Py_CODEUNIT *instr, PyTypeObject *type,
DesciptorClassification kind, PyObject *name,
_PyAdaptiveEntry *cache0, _PyAttrCache *cache1,
int base_op, int split_op, int hint_op)
int base_op, int values_op, int hint_op)
{
assert(kind == NON_OVERRIDING || kind == NON_DESCRIPTOR || kind == ABSENT ||
kind == BUILTIN_CLASSMETHOD || kind == PYTHON_CLASSMETHOD);
@ -595,17 +597,11 @@ specialize_dict_access(
}
if (type->tp_dictoffset > 0) {
PyObject **dictptr = (PyObject **) ((char *)owner + type->tp_dictoffset);
if (*dictptr == NULL || !PyDict_CheckExact(*dictptr)) {
SPECIALIZATION_FAIL(base_op, SPEC_FAIL_NO_DICT);
return 0;
}
// We found an instance with a __dict__.
PyDictObject *dict = (PyDictObject *)*dictptr;
PyDictKeysObject *keys = dict->ma_keys;
if ((type->tp_flags & Py_TPFLAGS_HEAPTYPE)
&& keys == ((PyHeapTypeObject*)type)->ht_cached_keys
) {
// Keys are shared
if (type->tp_inline_values_offset && dict == NULL) {
// Virtual dictionary
PyDictKeysObject *keys = ((PyHeapTypeObject *)type)->ht_cached_keys;
assert(type->tp_inline_values_offset > 0);
assert(PyUnicode_CheckExact(name));
Py_ssize_t index = _PyDictKeys_StringLookup(keys, name);
assert (index != DKIX_ERROR);
@ -613,18 +609,17 @@ specialize_dict_access(
SPECIALIZATION_FAIL(base_op, SPEC_FAIL_OUT_OF_RANGE);
return 0;
}
uint32_t keys_version = _PyDictKeys_GetVersionForCurrentState(keys);
if (keys_version == 0) {
SPECIALIZATION_FAIL(base_op, SPEC_FAIL_OUT_OF_VERSIONS);
return 0;
}
cache1->dk_version_or_hint = keys_version;
cache1->tp_version = type->tp_version_tag;
cache0->index = (uint16_t)index;
*instr = _Py_MAKECODEUNIT(split_op, _Py_OPARG(*instr));
*instr = _Py_MAKECODEUNIT(values_op, _Py_OPARG(*instr));
return 0;
}
else {
if (dict == NULL || !PyDict_CheckExact(dict)) {
SPECIALIZATION_FAIL(base_op, SPEC_FAIL_NO_DICT);
return 0;
}
// We found an instance with a __dict__.
PyObject *value = NULL;
Py_ssize_t hint =
_PyDict_GetItemHint(dict, name, -1, &value);
@ -736,7 +731,7 @@ _Py_Specialize_LoadAttr(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name, Sp
}
int err = specialize_dict_access(
owner, instr, type, kind, name, cache0, cache1,
LOAD_ATTR, LOAD_ATTR_SPLIT_KEYS, LOAD_ATTR_WITH_HINT
LOAD_ATTR, LOAD_ATTR_INSTANCE_VALUE, LOAD_ATTR_WITH_HINT
);
if (err < 0) {
return -1;
@ -818,7 +813,7 @@ _Py_Specialize_StoreAttr(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name, S
int err = specialize_dict_access(
owner, instr, type, kind, name, cache0, cache1,
STORE_ATTR, STORE_ATTR_SPLIT_KEYS, STORE_ATTR_WITH_HINT
STORE_ATTR, STORE_ATTR_INSTANCE_VALUE, STORE_ATTR_WITH_HINT
);
if (err < 0) {
return -1;
@ -875,6 +870,27 @@ load_method_fail_kind(DesciptorClassification kind)
}
#endif
static int
specialize_class_load_method(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name,
_PyAttrCache *cache1, _PyObjectCache *cache2)
{
PyObject *descr = NULL;
DesciptorClassification kind = 0;
kind = analyze_descriptor((PyTypeObject *)owner, name, &descr, 0);
switch (kind) {
case METHOD:
case NON_DESCRIPTOR:
cache1->tp_version = ((PyTypeObject *)owner)->tp_version_tag;
cache2->obj = descr;
*instr = _Py_MAKECODEUNIT(LOAD_METHOD_CLASS, _Py_OPARG(*instr));
return 0;
default:
SPECIALIZATION_FAIL(LOAD_METHOD, load_method_fail_kind(kind));
return -1;
}
}
// Please collect stats carefully before and after modifying. A subtle change
// can cause a significant drop in cache hits. A possible test is
// python.exe -m test_typing test_re test_dis test_zlib.
@ -886,7 +902,6 @@ _Py_Specialize_LoadMethod(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name,
_PyObjectCache *cache2 = &cache[-2].obj;
PyTypeObject *owner_cls = Py_TYPE(owner);
PyDictObject *owner_dict = NULL;
if (PyModule_CheckExact(owner)) {
int err = specialize_module_load_attr(owner, instr, name, cache0, cache1,
LOAD_METHOD, LOAD_METHOD_MODULE);
@ -900,9 +915,12 @@ _Py_Specialize_LoadMethod(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name,
return -1;
}
}
if (Py_TYPE(owner_cls)->tp_dictoffset < 0) {
SPECIALIZATION_FAIL(LOAD_METHOD, SPEC_FAIL_OUT_OF_RANGE);
goto fail;
if (PyType_Check(owner)) {
int err = specialize_class_load_method(owner, instr, name, cache1, cache2);
if (err) {
goto fail;
}
goto success;
}
// Technically this is fine for bound method calls, but it's uncommon and
// slightly slower at runtime to get dict.
@ -910,66 +928,45 @@ _Py_Specialize_LoadMethod(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name,
SPECIALIZATION_FAIL(LOAD_METHOD, SPEC_FAIL_OUT_OF_RANGE);
goto fail;
}
PyObject **owner_dictptr = _PyObject_GetDictPtr(owner);
int owner_has_dict = (owner_dictptr != NULL && *owner_dictptr != NULL);
owner_dict = owner_has_dict ? (PyDictObject *)*owner_dictptr : NULL;
// Make sure dict doesn't get GC-ed halfway.
Py_XINCREF(owner_dict);
// Check for classmethods.
int owner_is_class = PyType_Check(owner);
owner_cls = owner_is_class ? (PyTypeObject *)owner : owner_cls;
if ((owner_cls->tp_flags & Py_TPFLAGS_VALID_VERSION_TAG) == 0 ||
owner_cls->tp_version_tag == 0) {
SPECIALIZATION_FAIL(LOAD_METHOD, SPEC_FAIL_OUT_OF_VERSIONS);
goto fail;
}
PyObject *descr = NULL;
DesciptorClassification kind = 0;
kind = analyze_descriptor(owner_cls, name, &descr, 0);
// Store the version right away, in case it's modified halfway through.
cache1->tp_version = owner_cls->tp_version_tag;
assert(descr != NULL || kind == ABSENT || kind == GETSET_OVERRIDDEN);
if (kind != METHOD) {
SPECIALIZATION_FAIL(LOAD_METHOD, load_method_fail_kind(kind));
goto fail;
}
// If o.__dict__ changes, the method might be found in o.__dict__
// instead of old type lookup. So record o.__dict__'s keys.
uint32_t keys_version = UINT32_MAX;
if (owner_has_dict) {
// _PyDictKeys_GetVersionForCurrentState isn't accurate for
// custom dict subclasses at the moment.
if (!PyDict_CheckExact(owner_dict)) {
SPECIALIZATION_FAIL(LOAD_METHOD, SPEC_FAIL_DICT_SUBCLASS);
if (owner_cls->tp_inline_values_offset) {
PyObject **owner_dictptr = _PyObject_DictPointer(owner);
assert(owner_dictptr);
if (*owner_dictptr) {
SPECIALIZATION_FAIL(LOAD_METHOD, SPEC_FAIL_IS_ATTR);
goto fail;
}
assert(PyUnicode_CheckExact(name));
Py_hash_t hash = PyObject_Hash(name);
if (hash == -1) {
return -1;
PyDictKeysObject *keys = ((PyHeapTypeObject *)owner_cls)->ht_cached_keys;
Py_ssize_t index = _PyDictKeys_StringLookup(keys, name);
if (index != DKIX_EMPTY) {
SPECIALIZATION_FAIL(LOAD_METHOD, SPEC_FAIL_IS_ATTR);
goto fail;
}
PyObject *value = NULL;
if (!owner_is_class) {
// Instance methods shouldn't be in o.__dict__. That makes
// it an attribute.
Py_ssize_t ix = _Py_dict_lookup(owner_dict, name, hash, &value);
assert(ix != DKIX_ERROR);
if (ix != DKIX_EMPTY) {
SPECIALIZATION_FAIL(LOAD_METHOD, SPEC_FAIL_IS_ATTR);
goto fail;
}
}
keys_version = _PyDictKeys_GetVersionForCurrentState(owner_dict->ma_keys);
uint32_t keys_version = _PyDictKeys_GetVersionForCurrentState(keys);
if (keys_version == 0) {
SPECIALIZATION_FAIL(LOAD_ATTR, SPEC_FAIL_OUT_OF_VERSIONS);
SPECIALIZATION_FAIL(LOAD_METHOD, SPEC_FAIL_OUT_OF_VERSIONS);
goto fail;
}
// Fall through.
} // Else owner is maybe a builtin with no dict, or __slots__. Doesn't matter.
cache1->dk_version_or_hint = keys_version;
*instr = _Py_MAKECODEUNIT(LOAD_METHOD_CACHED, _Py_OPARG(*instr));
}
else {
if (owner_cls->tp_dictoffset == 0) {
*instr = _Py_MAKECODEUNIT(LOAD_METHOD_NO_DICT, _Py_OPARG(*instr));
}
else {
SPECIALIZATION_FAIL(LOAD_METHOD, SPEC_FAIL_IS_ATTR);
goto fail;
}
}
/* `descr` is borrowed. This is safe for methods (even inherited ones from
* super classes!) as long as tp_version_tag is validated for two main reasons:
*
@ -984,19 +981,15 @@ _Py_Specialize_LoadMethod(PyObject *owner, _Py_CODEUNIT *instr, PyObject *name,
* PyType_Modified usages in typeobject.c). The MCACHE has been
* working since Python 2.6 and it's battle-tested.
*/
cache1->tp_version = owner_cls->tp_version_tag;
cache2->obj = descr;
cache1->dk_version_or_hint = keys_version;
*instr = _Py_MAKECODEUNIT(owner_is_class ? LOAD_METHOD_CLASS :
LOAD_METHOD_CACHED, _Py_OPARG(*instr));
// Fall through.
success:
Py_XDECREF(owner_dict);
STAT_INC(LOAD_METHOD, specialization_success);
assert(!PyErr_Occurred());
cache0->counter = saturating_start();
return 0;
fail:
Py_XDECREF(owner_dict);
STAT_INC(LOAD_METHOD, specialization_failure);
assert(!PyErr_Occurred());
cache_backoff(cache0);