gh-110481: Implement biased reference counting (gh-110764)

This commit is contained in:
Sam Gross 2023-10-30 12:06:09 -04:00 committed by GitHub
parent 05f2f0ac92
commit 6dfb8fe023
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
29 changed files with 511 additions and 52 deletions

View file

@ -106,9 +106,26 @@ check by comparing the reference count field to the immortality reference count.
#define _Py_IMMORTAL_REFCNT (UINT_MAX >> 2)
#endif
// Py_NOGIL builds indicate immortal objects using `ob_ref_local`, which is
// always 32-bits.
#ifdef Py_NOGIL
#define _Py_IMMORTAL_REFCNT_LOCAL UINT32_MAX
#endif
// Make all internal uses of PyObject_HEAD_INIT immortal while preserving the
// C-API expectation that the refcnt will be set to 1.
#ifdef Py_BUILD_CORE
#if defined(Py_NOGIL)
#define PyObject_HEAD_INIT(type) \
{ \
0, \
0, \
0, \
0, \
_Py_IMMORTAL_REFCNT_LOCAL, \
0, \
(type), \
},
#elif defined(Py_BUILD_CORE)
#define PyObject_HEAD_INIT(type) \
{ \
{ _Py_IMMORTAL_REFCNT }, \
@ -142,6 +159,7 @@ check by comparing the reference count field to the immortality reference count.
* by hand. Similarly every pointer to a variable-size Python object can,
* in addition, be cast to PyVarObject*.
*/
#ifndef Py_NOGIL
struct _object {
#if (defined(__GNUC__) || defined(__clang__)) \
&& !(defined __STDC_VERSION__ && __STDC_VERSION__ >= 201112L)
@ -166,6 +184,36 @@ struct _object {
PyTypeObject *ob_type;
};
#else
// Objects that are not owned by any thread use a thread id (tid) of zero.
// This includes both immortal objects and objects whose reference count
// fields have been merged.
#define _Py_UNOWNED_TID 0
// The shared reference count uses the two least-significant bits to store
// flags. The remaining bits are used to store the reference count.
#define _Py_REF_SHARED_SHIFT 2
#define _Py_REF_SHARED_FLAG_MASK 0x3
// The shared flags are initialized to zero.
#define _Py_REF_SHARED_INIT 0x0
#define _Py_REF_MAYBE_WEAKREF 0x1
#define _Py_REF_QUEUED 0x2
#define _Py_REF_MERGED 0x3
// Create a shared field from a refcnt and desired flags
#define _Py_REF_SHARED(refcnt, flags) (((refcnt) << _Py_REF_SHARED_SHIFT) + (flags))
struct _object {
uintptr_t ob_tid; // thread id (or zero)
uint16_t _padding;
uint8_t ob_mutex; // per-object lock
uint8_t ob_gc_bits; // gc-related state
uint32_t ob_ref_local; // local reference count
Py_ssize_t ob_ref_shared; // shared (atomic) reference count
PyTypeObject *ob_type;
};
#endif
/* Cast argument to PyObject* type. */
#define _PyObject_CAST(op) _Py_CAST(PyObject*, (op))
@ -183,9 +231,56 @@ typedef struct {
PyAPI_FUNC(int) Py_Is(PyObject *x, PyObject *y);
#define Py_Is(x, y) ((x) == (y))
#ifndef Py_LIMITED_API
static inline uintptr_t
_Py_ThreadId(void)
{
uintptr_t tid;
#if defined(_MSC_VER) && defined(_M_X64)
tid = __readgsqword(48);
#elif defined(_MSC_VER) && defined(_M_IX86)
tid = __readfsdword(24);
#elif defined(_MSC_VER) && defined(_M_ARM64)
tid = __getReg(18);
#elif defined(__i386__)
__asm__("movl %%gs:0, %0" : "=r" (tid)); // 32-bit always uses GS
#elif defined(__MACH__) && defined(__x86_64__)
__asm__("movq %%gs:0, %0" : "=r" (tid)); // x86_64 macOSX uses GS
#elif defined(__x86_64__)
__asm__("movq %%fs:0, %0" : "=r" (tid)); // x86_64 Linux, BSD uses FS
#elif defined(__arm__)
__asm__ ("mrc p15, 0, %0, c13, c0, 3\nbic %0, %0, #3" : "=r" (tid));
#elif defined(__aarch64__) && defined(__APPLE__)
__asm__ ("mrs %0, tpidrro_el0" : "=r" (tid));
#elif defined(__aarch64__)
__asm__ ("mrs %0, tpidr_el0" : "=r" (tid));
#else
# error "define _Py_ThreadId for this platform"
#endif
return tid;
}
#endif
#if defined(Py_NOGIL) && !defined(Py_LIMITED_API)
static inline Py_ALWAYS_INLINE int
_Py_IsOwnedByCurrentThread(PyObject *ob)
{
return ob->ob_tid == _Py_ThreadId();
}
#endif
static inline Py_ssize_t Py_REFCNT(PyObject *ob) {
#if !defined(Py_NOGIL)
return ob->ob_refcnt;
#else
uint32_t local = _Py_atomic_load_uint32_relaxed(&ob->ob_ref_local);
if (local == _Py_IMMORTAL_REFCNT_LOCAL) {
return _Py_IMMORTAL_REFCNT;
}
Py_ssize_t shared = _Py_atomic_load_ssize_relaxed(&ob->ob_ref_shared);
return _Py_STATIC_CAST(Py_ssize_t, local) +
Py_ARITHMETIC_RIGHT_SHIFT(Py_ssize_t, shared, _Py_REF_SHARED_SHIFT);
#endif
}
#if !defined(Py_LIMITED_API) || Py_LIMITED_API+0 < 0x030b0000
# define Py_REFCNT(ob) Py_REFCNT(_PyObject_CAST(ob))
@ -216,7 +311,9 @@ static inline Py_ssize_t Py_SIZE(PyObject *ob) {
static inline Py_ALWAYS_INLINE int _Py_IsImmortal(PyObject *op)
{
#if SIZEOF_VOID_P > 4
#if defined(Py_NOGIL)
return op->ob_ref_local == _Py_IMMORTAL_REFCNT_LOCAL;
#elif SIZEOF_VOID_P > 4
return _Py_CAST(PY_INT32_T, op->ob_refcnt) < 0;
#else
return op->ob_refcnt == _Py_IMMORTAL_REFCNT;
@ -240,7 +337,24 @@ static inline void Py_SET_REFCNT(PyObject *ob, Py_ssize_t refcnt) {
if (_Py_IsImmortal(ob)) {
return;
}
#if !defined(Py_NOGIL)
ob->ob_refcnt = refcnt;
#else
if (_Py_IsOwnedByCurrentThread(ob)) {
// Set local refcount to desired refcount and shared refcount to zero,
// but preserve the shared refcount flags.
assert(refcnt < UINT32_MAX);
ob->ob_ref_local = _Py_STATIC_CAST(uint32_t, refcnt);
ob->ob_ref_shared &= _Py_REF_SHARED_FLAG_MASK;
}
else {
// Set local refcount to zero and shared refcount to desired refcount.
// Mark the object as merged.
ob->ob_tid = _Py_UNOWNED_TID;
ob->ob_ref_local = 0;
ob->ob_ref_shared = _Py_REF_SHARED(refcnt, _Py_REF_MERGED);
}
#endif
}
#if !defined(Py_LIMITED_API) || Py_LIMITED_API+0 < 0x030b0000
# define Py_SET_REFCNT(ob, refcnt) Py_SET_REFCNT(_PyObject_CAST(ob), (refcnt))
@ -618,7 +732,19 @@ static inline Py_ALWAYS_INLINE void Py_INCREF(PyObject *op)
#else
// Non-limited C API and limited C API for Python 3.9 and older access
// directly PyObject.ob_refcnt.
#if SIZEOF_VOID_P > 4
#if defined(Py_NOGIL)
uint32_t local = _Py_atomic_load_uint32_relaxed(&op->ob_ref_local);
uint32_t new_local = local + 1;
if (new_local == 0) {
return;
}
if (_Py_IsOwnedByCurrentThread(op)) {
_Py_atomic_store_uint32_relaxed(&op->ob_ref_local, new_local);
}
else {
_Py_atomic_add_ssize(&op->ob_ref_shared, (1 << _Py_REF_SHARED_SHIFT));
}
#elif SIZEOF_VOID_P > 4
// Portable saturated add, branching on the carry flag and set low bits
PY_UINT32_T cur_refcnt = op->ob_refcnt_split[PY_BIG_ENDIAN];
PY_UINT32_T new_refcnt = cur_refcnt + 1;
@ -643,6 +769,19 @@ static inline Py_ALWAYS_INLINE void Py_INCREF(PyObject *op)
# define Py_INCREF(op) Py_INCREF(_PyObject_CAST(op))
#endif
#if !defined(Py_LIMITED_API) && defined(Py_NOGIL)
// Implements Py_DECREF on objects not owned by the current thread.
PyAPI_FUNC(void) _Py_DecRefShared(PyObject *);
PyAPI_FUNC(void) _Py_DecRefSharedDebug(PyObject *, const char *, int);
// Called from Py_DECREF by the owning thread when the local refcount reaches
// zero. The call will deallocate the object if the shared refcount is also
// zero. Otherwise, the thread gives up ownership and merges the reference
// count fields.
PyAPI_FUNC(void) _Py_MergeZeroLocalRefcount(PyObject *);
#endif
#if defined(Py_LIMITED_API) && (Py_LIMITED_API+0 >= 0x030c0000 || defined(Py_REF_DEBUG))
// Stable ABI implements Py_DECREF() as a function call on limited C API
// version 3.12 and newer, and on Python built in debug mode. _Py_DecRef() was
@ -657,6 +796,52 @@ static inline void Py_DECREF(PyObject *op) {
}
#define Py_DECREF(op) Py_DECREF(_PyObject_CAST(op))
#elif defined(Py_NOGIL) && defined(Py_REF_DEBUG)
static inline void Py_DECREF(const char *filename, int lineno, PyObject *op)
{
uint32_t local = _Py_atomic_load_uint32_relaxed(&op->ob_ref_local);
if (local == _Py_IMMORTAL_REFCNT_LOCAL) {
return;
}
_Py_DECREF_STAT_INC();
_Py_DECREF_DecRefTotal();
if (_Py_IsOwnedByCurrentThread(op)) {
if (local == 0) {
_Py_NegativeRefcount(filename, lineno, op);
}
local--;
_Py_atomic_store_uint32_relaxed(&op->ob_ref_local, local);
if (local == 0) {
_Py_MergeZeroLocalRefcount(op);
}
}
else {
_Py_DecRefSharedDebug(op, filename, lineno);
}
}
#define Py_DECREF(op) Py_DECREF(__FILE__, __LINE__, _PyObject_CAST(op))
#elif defined(Py_NOGIL)
static inline void Py_DECREF(PyObject *op)
{
uint32_t local = _Py_atomic_load_uint32_relaxed(&op->ob_ref_local);
if (local == _Py_IMMORTAL_REFCNT_LOCAL) {
return;
}
_Py_DECREF_STAT_INC();
if (_Py_IsOwnedByCurrentThread(op)) {
local--;
_Py_atomic_store_uint32_relaxed(&op->ob_ref_local, local);
if (local == 0) {
_Py_MergeZeroLocalRefcount(op);
}
}
else {
_Py_DecRefShared(op);
}
}
#define Py_DECREF(op) Py_DECREF(_PyObject_CAST(op))
#elif defined(Py_REF_DEBUG)
static inline void Py_DECREF(const char *filename, int lineno, PyObject *op)
{