gh-144005: Eliminate redundant refcounting in the JIT for BINARY_OP_EXTEND (#144006)

This commit is contained in:
AN Long 2026-01-24 02:19:01 +09:00 committed by GitHub
parent 70e67f579e
commit 2f42f83344
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
11 changed files with 109 additions and 53 deletions

View File

@ -1094,7 +1094,7 @@ const struct opcode_metadata _PyOpcode_opcode_metadata[267] = {
[BINARY_OP_ADD_FLOAT] = { true, INSTR_FMT_IXC0000, HAS_EXIT_FLAG | HAS_ERROR_FLAG | HAS_ERROR_NO_POP_FLAG },
[BINARY_OP_ADD_INT] = { true, INSTR_FMT_IXC0000, HAS_EXIT_FLAG },
[BINARY_OP_ADD_UNICODE] = { true, INSTR_FMT_IXC0000, HAS_EXIT_FLAG | HAS_ERROR_FLAG | HAS_ERROR_NO_POP_FLAG },
[BINARY_OP_EXTEND] = { true, INSTR_FMT_IXC0000, HAS_DEOPT_FLAG | HAS_ERROR_FLAG | HAS_ESCAPES_FLAG },
[BINARY_OP_EXTEND] = { true, INSTR_FMT_IXC0000, HAS_DEOPT_FLAG | HAS_ERROR_FLAG | HAS_ERROR_NO_POP_FLAG | HAS_ESCAPES_FLAG },
[BINARY_OP_INPLACE_ADD_UNICODE] = { true, INSTR_FMT_IXC0000, HAS_LOCAL_FLAG | HAS_DEOPT_FLAG | HAS_EXIT_FLAG | HAS_ERROR_FLAG | HAS_ESCAPES_FLAG },
[BINARY_OP_MULTIPLY_FLOAT] = { true, INSTR_FMT_IXC0000, HAS_EXIT_FLAG | HAS_ERROR_FLAG | HAS_ERROR_NO_POP_FLAG },
[BINARY_OP_MULTIPLY_INT] = { true, INSTR_FMT_IXC0000, HAS_EXIT_FLAG },
@ -1347,7 +1347,7 @@ _PyOpcode_macro_expansion[256] = {
[BINARY_OP_ADD_FLOAT] = { .nuops = 5, .uops = { { _GUARD_TOS_FLOAT, OPARG_SIMPLE, 0 }, { _GUARD_NOS_FLOAT, OPARG_SIMPLE, 0 }, { _BINARY_OP_ADD_FLOAT, OPARG_SIMPLE, 5 }, { _POP_TOP_FLOAT, OPARG_SIMPLE, 5 }, { _POP_TOP_FLOAT, OPARG_SIMPLE, 5 } } },
[BINARY_OP_ADD_INT] = { .nuops = 5, .uops = { { _GUARD_TOS_INT, OPARG_SIMPLE, 0 }, { _GUARD_NOS_INT, OPARG_SIMPLE, 0 }, { _BINARY_OP_ADD_INT, OPARG_SIMPLE, 5 }, { _POP_TOP_INT, OPARG_SIMPLE, 5 }, { _POP_TOP_INT, OPARG_SIMPLE, 5 } } },
[BINARY_OP_ADD_UNICODE] = { .nuops = 5, .uops = { { _GUARD_TOS_UNICODE, OPARG_SIMPLE, 0 }, { _GUARD_NOS_UNICODE, OPARG_SIMPLE, 0 }, { _BINARY_OP_ADD_UNICODE, OPARG_SIMPLE, 5 }, { _POP_TOP_UNICODE, OPARG_SIMPLE, 5 }, { _POP_TOP_UNICODE, OPARG_SIMPLE, 5 } } },
[BINARY_OP_EXTEND] = { .nuops = 2, .uops = { { _GUARD_BINARY_OP_EXTEND, 4, 1 }, { _BINARY_OP_EXTEND, 4, 1 } } },
[BINARY_OP_EXTEND] = { .nuops = 4, .uops = { { _GUARD_BINARY_OP_EXTEND, 4, 1 }, { _BINARY_OP_EXTEND, 4, 1 }, { _POP_TOP, OPARG_SIMPLE, 5 }, { _POP_TOP, OPARG_SIMPLE, 5 } } },
[BINARY_OP_INPLACE_ADD_UNICODE] = { .nuops = 3, .uops = { { _GUARD_TOS_UNICODE, OPARG_SIMPLE, 0 }, { _GUARD_NOS_UNICODE, OPARG_SIMPLE, 0 }, { _BINARY_OP_INPLACE_ADD_UNICODE, OPARG_SIMPLE, 5 } } },
[BINARY_OP_MULTIPLY_FLOAT] = { .nuops = 5, .uops = { { _GUARD_TOS_FLOAT, OPARG_SIMPLE, 0 }, { _GUARD_NOS_FLOAT, OPARG_SIMPLE, 0 }, { _BINARY_OP_MULTIPLY_FLOAT, OPARG_SIMPLE, 5 }, { _POP_TOP_FLOAT, OPARG_SIMPLE, 5 }, { _POP_TOP_FLOAT, OPARG_SIMPLE, 5 } } },
[BINARY_OP_MULTIPLY_INT] = { .nuops = 5, .uops = { { _GUARD_TOS_INT, OPARG_SIMPLE, 0 }, { _GUARD_NOS_INT, OPARG_SIMPLE, 0 }, { _BINARY_OP_MULTIPLY_INT, OPARG_SIMPLE, 5 }, { _POP_TOP_INT, OPARG_SIMPLE, 5 }, { _POP_TOP_INT, OPARG_SIMPLE, 5 } } },

View File

@ -389,7 +389,7 @@ extern "C" {
#define _BINARY_OP_ADD_UNICODE_r03 586
#define _BINARY_OP_ADD_UNICODE_r13 587
#define _BINARY_OP_ADD_UNICODE_r23 588
#define _BINARY_OP_EXTEND_r21 589
#define _BINARY_OP_EXTEND_r23 589
#define _BINARY_OP_INPLACE_ADD_UNICODE_r21 590
#define _BINARY_OP_MULTIPLY_FLOAT_r03 591
#define _BINARY_OP_MULTIPLY_FLOAT_r13 592

View File

@ -115,7 +115,7 @@ const uint32_t _PyUop_Flags[MAX_UOP_ID+1] = {
[_BINARY_OP_ADD_UNICODE] = HAS_ERROR_FLAG | HAS_ERROR_NO_POP_FLAG | HAS_PURE_FLAG,
[_BINARY_OP_INPLACE_ADD_UNICODE] = HAS_LOCAL_FLAG | HAS_DEOPT_FLAG | HAS_ERROR_FLAG | HAS_ESCAPES_FLAG,
[_GUARD_BINARY_OP_EXTEND] = HAS_DEOPT_FLAG | HAS_ESCAPES_FLAG,
[_BINARY_OP_EXTEND] = HAS_ERROR_FLAG | HAS_ESCAPES_FLAG,
[_BINARY_OP_EXTEND] = HAS_ERROR_FLAG | HAS_ERROR_NO_POP_FLAG | HAS_ESCAPES_FLAG,
[_BINARY_SLICE] = HAS_ERROR_FLAG | HAS_ESCAPES_FLAG,
[_STORE_SLICE] = HAS_ERROR_FLAG | HAS_ESCAPES_FLAG,
[_BINARY_OP_SUBSCR_LIST_INT] = HAS_DEOPT_FLAG | HAS_ESCAPES_FLAG,
@ -1113,7 +1113,7 @@ const _PyUopCachingInfo _PyUop_Caching[MAX_UOP_ID+1] = {
.entries = {
{ -1, -1, -1 },
{ -1, -1, -1 },
{ 1, 2, _BINARY_OP_EXTEND_r21 },
{ 3, 2, _BINARY_OP_EXTEND_r23 },
{ -1, -1, -1 },
},
},
@ -3670,7 +3670,7 @@ const uint16_t _PyUop_Uncached[MAX_UOP_REGS_ID+1] = {
[_BINARY_OP_ADD_UNICODE_r23] = _BINARY_OP_ADD_UNICODE,
[_BINARY_OP_INPLACE_ADD_UNICODE_r21] = _BINARY_OP_INPLACE_ADD_UNICODE,
[_GUARD_BINARY_OP_EXTEND_r22] = _GUARD_BINARY_OP_EXTEND,
[_BINARY_OP_EXTEND_r21] = _BINARY_OP_EXTEND,
[_BINARY_OP_EXTEND_r23] = _BINARY_OP_EXTEND,
[_BINARY_SLICE_r31] = _BINARY_SLICE,
[_STORE_SLICE_r30] = _STORE_SLICE,
[_BINARY_OP_SUBSCR_LIST_INT_r23] = _BINARY_OP_SUBSCR_LIST_INT,
@ -4239,7 +4239,7 @@ const char *const _PyOpcode_uop_name[MAX_UOP_REGS_ID+1] = {
[_BINARY_OP_ADD_UNICODE_r13] = "_BINARY_OP_ADD_UNICODE_r13",
[_BINARY_OP_ADD_UNICODE_r23] = "_BINARY_OP_ADD_UNICODE_r23",
[_BINARY_OP_EXTEND] = "_BINARY_OP_EXTEND",
[_BINARY_OP_EXTEND_r21] = "_BINARY_OP_EXTEND_r21",
[_BINARY_OP_EXTEND_r23] = "_BINARY_OP_EXTEND_r23",
[_BINARY_OP_INPLACE_ADD_UNICODE] = "_BINARY_OP_INPLACE_ADD_UNICODE",
[_BINARY_OP_INPLACE_ADD_UNICODE_r21] = "_BINARY_OP_INPLACE_ADD_UNICODE_r21",
[_BINARY_OP_MULTIPLY_FLOAT] = "_BINARY_OP_MULTIPLY_FLOAT",

View File

@ -2897,6 +2897,23 @@ class TestUopsOptimization(unittest.TestCase):
self.assertIn("_POP_TOP_NOP", uops)
self.assertLessEqual(count_ops(ex, "_POP_TOP"), 2)
def test_binary_op_extend_float_long_add_refcount_elimination(self):
def testfunc(n):
a = 1.5
b = 2
res = 0.0
for _ in range(n):
res = a + b
return res
res, ex = self._run_with_optimizer(testfunc, TIER2_THRESHOLD)
self.assertEqual(res, 3.5)
self.assertIsNotNone(ex)
uops = get_opnames(ex)
self.assertIn("_BINARY_OP_EXTEND", uops)
self.assertIn("_POP_TOP_NOP", uops)
self.assertLessEqual(count_ops(ex, "_POP_TOP"), 2)
def test_remove_guard_for_slice_list(self):
def f(n):
for i in range(n):

View File

@ -0,0 +1 @@
Eliminate redundant refcounting from ``BINARY_OP_EXTEND``.

View File

@ -317,6 +317,9 @@
_PyStackRef left;
_PyStackRef right;
_PyStackRef res;
_PyStackRef l;
_PyStackRef r;
_PyStackRef value;
/* Skip 1 cache entry */
// _GUARD_BINARY_OP_EXTEND
{
@ -348,25 +351,32 @@
STAT_INC(BINARY_OP, hit);
_PyFrame_SetStackPointer(frame, stack_pointer);
PyObject *res_o = d->action(left_o, right_o);
_PyStackRef tmp = right;
right = PyStackRef_NULL;
stack_pointer[-1] = right;
PyStackRef_CLOSE(tmp);
tmp = left;
left = PyStackRef_NULL;
stack_pointer[-2] = left;
PyStackRef_CLOSE(tmp);
stack_pointer = _PyFrame_GetStackPointer(frame);
stack_pointer += -2;
ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
if (res_o == NULL) {
JUMP_TO_LABEL(error);
}
res = PyStackRef_FromPyObjectSteal(res_o);
l = left;
r = right;
}
// _POP_TOP
{
value = r;
stack_pointer[-2] = res;
stack_pointer[-1] = l;
_PyFrame_SetStackPointer(frame, stack_pointer);
PyStackRef_XCLOSE(value);
stack_pointer = _PyFrame_GetStackPointer(frame);
}
// _POP_TOP
{
value = l;
stack_pointer += -1;
ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
_PyFrame_SetStackPointer(frame, stack_pointer);
PyStackRef_XCLOSE(value);
stack_pointer = _PyFrame_GetStackPointer(frame);
}
stack_pointer[0] = res;
stack_pointer += 1;
ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
DISPATCH();
}

View File

@ -829,7 +829,7 @@ dummy_func(
DEOPT_IF(!res);
}
op(_BINARY_OP_EXTEND, (descr/4, left, right -- res)) {
op(_BINARY_OP_EXTEND, (descr/4, left, right -- res, l, r)) {
PyObject *left_o = PyStackRef_AsPyObjectBorrow(left);
PyObject *right_o = PyStackRef_AsPyObjectBorrow(right);
assert(INLINE_CACHE_ENTRIES_BINARY_OP == 5);
@ -838,13 +838,18 @@ dummy_func(
STAT_INC(BINARY_OP, hit);
PyObject *res_o = d->action(left_o, right_o);
DECREF_INPUTS();
ERROR_IF(res_o == NULL);
if (res_o == NULL) {
ERROR_NO_POP();
}
res = PyStackRef_FromPyObjectSteal(res_o);
l = left;
r = right;
DEAD(left);
DEAD(right);
}
macro(BINARY_OP_EXTEND) =
unused/1 + _GUARD_BINARY_OP_EXTEND + rewind/-4 + _BINARY_OP_EXTEND;
unused/1 + _GUARD_BINARY_OP_EXTEND + rewind/-4 + _BINARY_OP_EXTEND + POP_TOP + POP_TOP;
macro(BINARY_OP_INPLACE_ADD_UNICODE) =
_GUARD_TOS_UNICODE + _GUARD_NOS_UNICODE + unused/5 + _BINARY_OP_INPLACE_ADD_UNICODE;

View File

@ -5143,12 +5143,14 @@
break;
}
case _BINARY_OP_EXTEND_r21: {
case _BINARY_OP_EXTEND_r23: {
CHECK_CURRENT_CACHED_VALUES(2);
assert(WITHIN_STACK_BOUNDS_IGNORING_CACHE());
_PyStackRef right;
_PyStackRef left;
_PyStackRef res;
_PyStackRef l;
_PyStackRef r;
_PyStackRef _stack_item_0 = _tos_cache0;
_PyStackRef _stack_item_1 = _tos_cache1;
right = _stack_item_1;
@ -5165,26 +5167,20 @@
ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
_PyFrame_SetStackPointer(frame, stack_pointer);
PyObject *res_o = d->action(left_o, right_o);
_PyStackRef tmp = right;
right = PyStackRef_NULL;
stack_pointer[-1] = right;
PyStackRef_CLOSE(tmp);
tmp = left;
left = PyStackRef_NULL;
stack_pointer[-2] = left;
PyStackRef_CLOSE(tmp);
stack_pointer = _PyFrame_GetStackPointer(frame);
stack_pointer += -2;
ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
if (res_o == NULL) {
SET_CURRENT_CACHED_VALUES(0);
JUMP_TO_ERROR();
}
res = PyStackRef_FromPyObjectSteal(res_o);
l = left;
r = right;
_tos_cache2 = r;
_tos_cache1 = l;
_tos_cache0 = res;
_tos_cache1 = PyStackRef_ZERO_BITS;
_tos_cache2 = PyStackRef_ZERO_BITS;
SET_CURRENT_CACHED_VALUES(1);
SET_CURRENT_CACHED_VALUES(3);
stack_pointer += -2;
ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
assert(WITHIN_STACK_BOUNDS_IGNORING_CACHE());
break;
}

View File

@ -317,6 +317,9 @@
_PyStackRef left;
_PyStackRef right;
_PyStackRef res;
_PyStackRef l;
_PyStackRef r;
_PyStackRef value;
/* Skip 1 cache entry */
// _GUARD_BINARY_OP_EXTEND
{
@ -348,25 +351,32 @@
STAT_INC(BINARY_OP, hit);
_PyFrame_SetStackPointer(frame, stack_pointer);
PyObject *res_o = d->action(left_o, right_o);
_PyStackRef tmp = right;
right = PyStackRef_NULL;
stack_pointer[-1] = right;
PyStackRef_CLOSE(tmp);
tmp = left;
left = PyStackRef_NULL;
stack_pointer[-2] = left;
PyStackRef_CLOSE(tmp);
stack_pointer = _PyFrame_GetStackPointer(frame);
stack_pointer += -2;
ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
if (res_o == NULL) {
JUMP_TO_LABEL(error);
}
res = PyStackRef_FromPyObjectSteal(res_o);
l = left;
r = right;
}
// _POP_TOP
{
value = r;
stack_pointer[-2] = res;
stack_pointer[-1] = l;
_PyFrame_SetStackPointer(frame, stack_pointer);
PyStackRef_XCLOSE(value);
stack_pointer = _PyFrame_GetStackPointer(frame);
}
// _POP_TOP
{
value = l;
stack_pointer += -1;
ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
_PyFrame_SetStackPointer(frame, stack_pointer);
PyStackRef_XCLOSE(value);
stack_pointer = _PyFrame_GetStackPointer(frame);
}
stack_pointer[0] = res;
stack_pointer += 1;
ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
DISPATCH();
}

View File

@ -311,6 +311,12 @@ dummy_func(void) {
r = right;
}
op(_BINARY_OP_EXTEND, (left, right -- res, l, r)) {
res = sym_new_not_null(ctx);
l = left;
r = right;
}
op(_BINARY_OP_INPLACE_ADD_UNICODE, (left, right -- res)) {
if (sym_is_const(ctx, left) && sym_is_const(ctx, right)) {
assert(PyUnicode_CheckExact(sym_get_const(ctx, left)));

View File

@ -882,11 +882,22 @@
}
case _BINARY_OP_EXTEND: {
JitOptRef right;
JitOptRef left;
JitOptRef res;
JitOptRef l;
JitOptRef r;
right = stack_pointer[-1];
left = stack_pointer[-2];
PyObject *descr = (PyObject *)this_instr->operand0;
res = sym_new_not_null(ctx);
CHECK_STACK_BOUNDS(-1);
l = left;
r = right;
CHECK_STACK_BOUNDS(1);
stack_pointer[-2] = res;
stack_pointer += -1;
stack_pointer[-1] = l;
stack_pointer[0] = r;
stack_pointer += 1;
ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
break;
}