gh-148378: Allow multiple consecutive recording ops per macro op (GH-148496)

This commit is contained in:
Hai Zhu 2026-04-14 19:26:53 +08:00 committed by GitHub
parent 21da9d7164
commit 5ce0fe8b6c
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
9 changed files with 318 additions and 82 deletions

View file

@ -6349,7 +6349,10 @@ dummy_func(
ERROR_IF(err < 0);
DISPATCH();
}
Py_CLEAR(tracer->prev_state.recorded_value);
for (int i = 0; i < tracer->prev_state.recorded_count; i++) {
Py_CLEAR(tracer->prev_state.recorded_values[i]);
}
tracer->prev_state.recorded_count = 0;
tracer->prev_state.instr = next_instr;
PyObject *prev_code = PyStackRef_AsPyObjectBorrow(frame->f_executable);
if (tracer->prev_state.instr_code != (PyCodeObject *)prev_code) {
@ -6363,11 +6366,12 @@ dummy_func(
(&next_instr[1])->counter = trigger_backoff_counter();
}
uint8_t record_func_index = _PyOpcode_RecordFunctionIndices[opcode];
if (record_func_index) {
_Py_RecordFuncPtr doesnt_escape = _PyOpcode_RecordFunctions[record_func_index];
doesnt_escape(frame, stack_pointer, oparg, &tracer->prev_state.recorded_value);
const _PyOpcodeRecordEntry *record_entry = &_PyOpcode_RecordEntries[opcode];
for (int i = 0; i < record_entry->count; i++) {
_Py_RecordFuncPtr doesnt_escape = _PyOpcode_RecordFunctions[record_entry->indices[i]];
doesnt_escape(frame, stack_pointer, oparg, &tracer->prev_state.recorded_values[i]);
}
tracer->prev_state.recorded_count = record_entry->count;
DISPATCH_GOTO_NON_TRACING();
#else
(void)prev_instr;

View file

@ -12314,9 +12314,12 @@
}
DISPATCH();
}
_PyFrame_SetStackPointer(frame, stack_pointer);
Py_CLEAR(tracer->prev_state.recorded_value);
stack_pointer = _PyFrame_GetStackPointer(frame);
for (int i = 0; i < tracer->prev_state.recorded_count; i++) {
_PyFrame_SetStackPointer(frame, stack_pointer);
Py_CLEAR(tracer->prev_state.recorded_values[i]);
stack_pointer = _PyFrame_GetStackPointer(frame);
}
tracer->prev_state.recorded_count = 0;
tracer->prev_state.instr = next_instr;
PyObject *prev_code = PyStackRef_AsPyObjectBorrow(frame->f_executable);
if (tracer->prev_state.instr_code != (PyCodeObject *)prev_code) {
@ -12330,11 +12333,12 @@
if (_PyOpcode_Caches[_PyOpcode_Deopt[opcode]]) {
(&next_instr[1])->counter = trigger_backoff_counter();
}
uint8_t record_func_index = _PyOpcode_RecordFunctionIndices[opcode];
if (record_func_index) {
_Py_RecordFuncPtr doesnt_escape = _PyOpcode_RecordFunctions[record_func_index];
doesnt_escape(frame, stack_pointer, oparg, &tracer->prev_state.recorded_value);
const _PyOpcodeRecordEntry *record_entry = &_PyOpcode_RecordEntries[opcode];
for (int i = 0; i < record_entry->count; i++) {
_Py_RecordFuncPtr doesnt_escape = _PyOpcode_RecordFunctions[record_entry->indices[i]];
doesnt_escape(frame, stack_pointer, oparg, &tracer->prev_state.recorded_values[i]);
}
tracer->prev_state.recorded_count = record_entry->count;
DISPATCH_GOTO_NON_TRACING();
#else
(void)prev_instr;

View file

@ -866,6 +866,7 @@ _PyJit_translate_single_bytecode_to_trace(
assert(nuops > 0);
uint32_t orig_oparg = oparg; // For OPARG_TOP/BOTTOM
uint32_t orig_target = target;
int record_idx = 0;
for (int i = 0; i < nuops; i++) {
oparg = orig_oparg;
target = orig_target;
@ -946,8 +947,9 @@ _PyJit_translate_single_bytecode_to_trace(
operand = next->op.arg;
}
else if (_PyUop_Flags[uop] & HAS_RECORDS_VALUE_FLAG) {
PyObject *recorded_value = tracer->prev_state.recorded_value;
tracer->prev_state.recorded_value = NULL;
PyObject *recorded_value = tracer->prev_state.recorded_values[record_idx];
tracer->prev_state.recorded_values[record_idx] = NULL;
record_idx++;
operand = (uintptr_t)recorded_value;
}
// All other instructions
@ -1060,12 +1062,16 @@ _PyJit_TryInitializeTracing(
tracer->prev_state.instr_frame = frame;
tracer->prev_state.instr_oparg = oparg;
tracer->prev_state.instr_stacklevel = tracer->initial_state.stack_depth;
tracer->prev_state.recorded_value = NULL;
uint8_t record_func_index = _PyOpcode_RecordFunctionIndices[curr_instr->op.code];
if (record_func_index) {
_Py_RecordFuncPtr record_func = _PyOpcode_RecordFunctions[record_func_index];
record_func(frame, stack_pointer, oparg, &tracer->prev_state.recorded_value);
tracer->prev_state.recorded_count = 0;
for (int i = 0; i < MAX_RECORDED_VALUES; i++) {
tracer->prev_state.recorded_values[i] = NULL;
}
const _PyOpcodeRecordEntry *record_entry = &_PyOpcode_RecordEntries[curr_instr->op.code];
for (int i = 0; i < record_entry->count; i++) {
_Py_RecordFuncPtr record_func = _PyOpcode_RecordFunctions[record_entry->indices[i]];
record_func(frame, stack_pointer, oparg, &tracer->prev_state.recorded_values[i]);
}
tracer->prev_state.recorded_count = record_entry->count;
assert(curr_instr->op.code == JUMP_BACKWARD_JIT || curr_instr->op.code == RESUME_CHECK_JIT || (exit != NULL));
tracer->initial_state.jump_backward_instr = curr_instr;
@ -1117,7 +1123,10 @@ _PyJit_FinalizeTracing(PyThreadState *tstate, int err)
Py_CLEAR(tracer->initial_state.func);
Py_CLEAR(tracer->initial_state.executor);
Py_CLEAR(tracer->prev_state.instr_code);
Py_CLEAR(tracer->prev_state.recorded_value);
for (int i = 0; i < MAX_RECORDED_VALUES; i++) {
Py_CLEAR(tracer->prev_state.recorded_values[i]);
}
tracer->prev_state.recorded_count = 0;
uop_buffer_init(buffer, &tracer->uop_array[0], UOP_MAX_TRACE_LENGTH);
tracer->is_tracing = false;
}

View file

@ -99,41 +99,42 @@ void _PyOpcode_RecordFunction_CODE(_PyInterpreterFrame *frame, _PyStackRef *stac
#define _RECORD_BOUND_METHOD_INDEX 6
#define _RECORD_CALLABLE_KW_INDEX 7
#define _RECORD_4OS_INDEX 8
const uint8_t _PyOpcode_RecordFunctionIndices[256] = {
[TO_BOOL_ALWAYS_TRUE] = _RECORD_TOS_TYPE_INDEX,
[BINARY_OP_SUBSCR_GETITEM] = _RECORD_NOS_INDEX,
[SEND_GEN] = _RECORD_3OS_GEN_FUNC_INDEX,
[LOAD_SUPER_ATTR_METHOD] = _RECORD_NOS_INDEX,
[LOAD_ATTR_INSTANCE_VALUE] = _RECORD_TOS_TYPE_INDEX,
[LOAD_ATTR_WITH_HINT] = _RECORD_TOS_TYPE_INDEX,
[LOAD_ATTR_SLOT] = _RECORD_TOS_TYPE_INDEX,
[LOAD_ATTR_CLASS_WITH_METACLASS_CHECK] = _RECORD_TOS_TYPE_INDEX,
[LOAD_ATTR_PROPERTY] = _RECORD_TOS_TYPE_INDEX,
[STORE_ATTR_INSTANCE_VALUE] = _RECORD_TOS_TYPE_INDEX,
[STORE_ATTR_WITH_HINT] = _RECORD_TOS_TYPE_INDEX,
[STORE_ATTR_SLOT] = _RECORD_TOS_TYPE_INDEX,
[FOR_ITER_GEN] = _RECORD_NOS_GEN_FUNC_INDEX,
[LOAD_ATTR_METHOD_WITH_VALUES] = _RECORD_TOS_TYPE_INDEX,
[LOAD_ATTR_METHOD_NO_DICT] = _RECORD_TOS_TYPE_INDEX,
[LOAD_ATTR_NONDESCRIPTOR_WITH_VALUES] = _RECORD_TOS_TYPE_INDEX,
[LOAD_ATTR_NONDESCRIPTOR_NO_DICT] = _RECORD_TOS_TYPE_INDEX,
[LOAD_ATTR_METHOD_LAZY_DICT] = _RECORD_TOS_TYPE_INDEX,
[CALL_PY_GENERAL] = _RECORD_CALLABLE_INDEX,
[CALL_BOUND_METHOD_GENERAL] = _RECORD_BOUND_METHOD_INDEX,
[CALL_NON_PY_GENERAL] = _RECORD_CALLABLE_INDEX,
[CALL_BOUND_METHOD_EXACT_ARGS] = _RECORD_BOUND_METHOD_INDEX,
[CALL_PY_EXACT_ARGS] = _RECORD_CALLABLE_INDEX,
[CALL_ALLOC_AND_ENTER_INIT] = _RECORD_CALLABLE_INDEX,
[CALL_BUILTIN_CLASS] = _RECORD_CALLABLE_INDEX,
[CALL_BUILTIN_O] = _RECORD_CALLABLE_INDEX,
[CALL_BUILTIN_FAST] = _RECORD_CALLABLE_INDEX,
[CALL_BUILTIN_FAST_WITH_KEYWORDS] = _RECORD_CALLABLE_INDEX,
[CALL_METHOD_DESCRIPTOR_O] = _RECORD_CALLABLE_INDEX,
[CALL_METHOD_DESCRIPTOR_FAST_WITH_KEYWORDS] = _RECORD_CALLABLE_INDEX,
[CALL_METHOD_DESCRIPTOR_NOARGS] = _RECORD_CALLABLE_INDEX,
[CALL_KW_PY] = _RECORD_CALLABLE_KW_INDEX,
[CALL_KW_BOUND_METHOD] = _RECORD_CALLABLE_KW_INDEX,
[CALL_EX_PY] = _RECORD_4OS_INDEX,
const _PyOpcodeRecordEntry _PyOpcode_RecordEntries[256] = {
[TO_BOOL_ALWAYS_TRUE] = {1, {_RECORD_TOS_TYPE_INDEX}},
[BINARY_OP_SUBSCR_GETITEM] = {1, {_RECORD_NOS_INDEX}},
[SEND_GEN] = {1, {_RECORD_3OS_GEN_FUNC_INDEX}},
[LOAD_SUPER_ATTR_METHOD] = {1, {_RECORD_NOS_INDEX}},
[LOAD_ATTR_INSTANCE_VALUE] = {1, {_RECORD_TOS_TYPE_INDEX}},
[LOAD_ATTR_WITH_HINT] = {1, {_RECORD_TOS_TYPE_INDEX}},
[LOAD_ATTR_SLOT] = {1, {_RECORD_TOS_TYPE_INDEX}},
[LOAD_ATTR_CLASS_WITH_METACLASS_CHECK] = {1, {_RECORD_TOS_TYPE_INDEX}},
[LOAD_ATTR_PROPERTY] = {1, {_RECORD_TOS_TYPE_INDEX}},
[STORE_ATTR_INSTANCE_VALUE] = {1, {_RECORD_TOS_TYPE_INDEX}},
[STORE_ATTR_WITH_HINT] = {1, {_RECORD_TOS_TYPE_INDEX}},
[STORE_ATTR_SLOT] = {1, {_RECORD_TOS_TYPE_INDEX}},
[FOR_ITER_GEN] = {1, {_RECORD_NOS_GEN_FUNC_INDEX}},
[LOAD_ATTR_METHOD_WITH_VALUES] = {1, {_RECORD_TOS_TYPE_INDEX}},
[LOAD_ATTR_METHOD_NO_DICT] = {1, {_RECORD_TOS_TYPE_INDEX}},
[LOAD_ATTR_NONDESCRIPTOR_WITH_VALUES] = {1, {_RECORD_TOS_TYPE_INDEX}},
[LOAD_ATTR_NONDESCRIPTOR_NO_DICT] = {1, {_RECORD_TOS_TYPE_INDEX}},
[LOAD_ATTR_METHOD_LAZY_DICT] = {1, {_RECORD_TOS_TYPE_INDEX}},
[CALL_PY_GENERAL] = {1, {_RECORD_CALLABLE_INDEX}},
[CALL_BOUND_METHOD_GENERAL] = {1, {_RECORD_BOUND_METHOD_INDEX}},
[CALL_NON_PY_GENERAL] = {1, {_RECORD_CALLABLE_INDEX}},
[CALL_BOUND_METHOD_EXACT_ARGS] = {1, {_RECORD_BOUND_METHOD_INDEX}},
[CALL_PY_EXACT_ARGS] = {1, {_RECORD_CALLABLE_INDEX}},
[CALL_ALLOC_AND_ENTER_INIT] = {1, {_RECORD_CALLABLE_INDEX}},
[CALL_BUILTIN_CLASS] = {1, {_RECORD_CALLABLE_INDEX}},
[CALL_BUILTIN_O] = {1, {_RECORD_CALLABLE_INDEX}},
[CALL_BUILTIN_FAST] = {1, {_RECORD_CALLABLE_INDEX}},
[CALL_BUILTIN_FAST_WITH_KEYWORDS] = {1, {_RECORD_CALLABLE_INDEX}},
[CALL_METHOD_DESCRIPTOR_O] = {1, {_RECORD_CALLABLE_INDEX}},
[CALL_METHOD_DESCRIPTOR_FAST_WITH_KEYWORDS] = {1, {_RECORD_CALLABLE_INDEX}},
[CALL_METHOD_DESCRIPTOR_NOARGS] = {1, {_RECORD_CALLABLE_INDEX}},
[CALL_KW_PY] = {1, {_RECORD_CALLABLE_KW_INDEX}},
[CALL_KW_BOUND_METHOD] = {1, {_RECORD_CALLABLE_KW_INDEX}},
[CALL_EX_PY] = {1, {_RECORD_4OS_INDEX}},
};
const _Py_RecordFuncPtr _PyOpcode_RecordFunctions[9] = {