cpython/Python/index_pool.c
mpage 2e95c5ba3b
gh-115999: Implement thread-local bytecode and enable specialization for BINARY_OP (#123926)
Each thread specializes a thread-local copy of the bytecode, created on the first RESUME, in free-threaded builds. All copies of the bytecode for a code object are stored in the co_tlbc array on the code object. Threads reserve a globally unique index identifying its copy of the bytecode in all co_tlbc arrays at thread creation and release the index at thread destruction. The first entry in every co_tlbc array always points to the "main" copy of the bytecode that is stored at the end of the code object. This ensures that no bytecode is copied for programs that do not use threads.

Thread-local bytecode can be disabled at runtime by providing either -X tlbc=0 or PYTHON_TLBC=0. Disabling thread-local bytecode also disables specialization.

Concurrent modifications to the bytecode made by the specializing interpreter and instrumentation use atomics, with specialization taking care not to overwrite an instruction that was instrumented concurrently.
2024-11-04 11:13:32 -08:00

194 lines
4.5 KiB
C

#include <stdbool.h>
#include "Python.h"
#include "pycore_index_pool.h"
#include "pycore_lock.h"
#ifdef Py_GIL_DISABLED
static inline void
swap(int32_t *values, Py_ssize_t i, Py_ssize_t j)
{
int32_t tmp = values[i];
values[i] = values[j];
values[j] = tmp;
}
static bool
heap_try_swap(_PyIndexHeap *heap, Py_ssize_t i, Py_ssize_t j)
{
if (i < 0 || i >= heap->size) {
return 0;
}
if (j < 0 || j >= heap->size) {
return 0;
}
if (i <= j) {
if (heap->values[i] <= heap->values[j]) {
return 0;
}
}
else if (heap->values[j] <= heap->values[i]) {
return 0;
}
swap(heap->values, i, j);
return 1;
}
static inline Py_ssize_t
parent(Py_ssize_t i)
{
return (i - 1) / 2;
}
static inline Py_ssize_t
left_child(Py_ssize_t i)
{
return 2 * i + 1;
}
static inline Py_ssize_t
right_child(Py_ssize_t i)
{
return 2 * i + 2;
}
static void
heap_add(_PyIndexHeap *heap, int32_t val)
{
assert(heap->size < heap->capacity);
// Add val to end
heap->values[heap->size] = val;
heap->size++;
// Sift up
for (Py_ssize_t cur = heap->size - 1; cur > 0; cur = parent(cur)) {
if (!heap_try_swap(heap, cur, parent(cur))) {
break;
}
}
}
static Py_ssize_t
heap_min_child(_PyIndexHeap *heap, Py_ssize_t i)
{
if (left_child(i) < heap->size) {
if (right_child(i) < heap->size) {
Py_ssize_t lval = heap->values[left_child(i)];
Py_ssize_t rval = heap->values[right_child(i)];
return lval < rval ? left_child(i) : right_child(i);
}
return left_child(i);
}
else if (right_child(i) < heap->size) {
return right_child(i);
}
return -1;
}
static int32_t
heap_pop(_PyIndexHeap *heap)
{
assert(heap->size > 0);
// Pop smallest and replace with the last element
int32_t result = heap->values[0];
heap->values[0] = heap->values[heap->size - 1];
heap->size--;
// Sift down
for (Py_ssize_t cur = 0; cur < heap->size;) {
Py_ssize_t min_child = heap_min_child(heap, cur);
if (min_child > -1 && heap_try_swap(heap, cur, min_child)) {
cur = min_child;
}
else {
break;
}
}
return result;
}
static int
heap_ensure_capacity(_PyIndexHeap *heap, Py_ssize_t limit)
{
assert(limit > 0);
if (heap->capacity > limit) {
return 0;
}
Py_ssize_t new_capacity = heap->capacity ? heap->capacity : 1024;
while (new_capacity && new_capacity < limit) {
new_capacity <<= 1;
}
if (!new_capacity) {
return -1;
}
int32_t *new_values = PyMem_RawCalloc(new_capacity, sizeof(int32_t));
if (new_values == NULL) {
return -1;
}
if (heap->values != NULL) {
memcpy(new_values, heap->values, heap->capacity);
PyMem_RawFree(heap->values);
}
heap->values = new_values;
heap->capacity = new_capacity;
return 0;
}
static void
heap_fini(_PyIndexHeap *heap)
{
if (heap->values != NULL) {
PyMem_RawFree(heap->values);
heap->values = NULL;
}
heap->size = -1;
heap->capacity = -1;
}
#define LOCK_POOL(pool) PyMutex_LockFlags(&pool->mutex, _Py_LOCK_DONT_DETACH)
#define UNLOCK_POOL(pool) PyMutex_Unlock(&pool->mutex)
int32_t
_PyIndexPool_AllocIndex(_PyIndexPool *pool)
{
LOCK_POOL(pool);
int32_t index;
_PyIndexHeap *free_indices = &pool->free_indices;
if (free_indices->size == 0) {
// No free indices. Make sure the heap can always store all of the
// indices that have been allocated to avoid having to allocate memory
// (which can fail) when freeing an index. Freeing indices happens when
// threads are being destroyed, which makes error handling awkward /
// impossible. This arrangement shifts handling of allocation failures
// to when indices are allocated, which happens at thread creation,
// where we are better equipped to deal with failure.
if (heap_ensure_capacity(free_indices, pool->next_index + 1) < 0) {
UNLOCK_POOL(pool);
PyErr_NoMemory();
return -1;
}
index = pool->next_index++;
}
else {
index = heap_pop(free_indices);
}
UNLOCK_POOL(pool);
return index;
}
void
_PyIndexPool_FreeIndex(_PyIndexPool *pool, int32_t index)
{
LOCK_POOL(pool);
heap_add(&pool->free_indices, index);
UNLOCK_POOL(pool);
}
void
_PyIndexPool_Fini(_PyIndexPool *pool)
{
heap_fini(&pool->free_indices);
}
#endif // Py_GIL_DISABLED