Mirror
/
cpython
mirror of https://github.com/python/cpython
4
1
Fork 0
Code Issues Releases Wiki Activity

GH-115480: Reduce guard strength for binary ops when type of one operand is known already (GH-118050)

This commit is contained in:
Mark Shannon 2024-04-22 13:34:06 +01:00 committed by GitHub
parent ceb6038b05
commit a6647d16ab
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
10 changed files with 316 additions and 103 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
Include/internal/pycore_optimizer.h
View File

@ -98,6 +98,7 @@ extern bool _Py_uop_sym_set_type(_Py_UopsSymbol *sym, PyTypeObject *typ);
extern bool _Py_uop_sym_set_const(_Py_UopsSymbol *sym, PyObject *const_val);
extern bool _Py_uop_sym_is_bottom(_Py_UopsSymbol *sym);
extern int _Py_uop_sym_truthiness(_Py_UopsSymbol *sym);
extern PyTypeObject *_Py_uop_sym_get_type(_Py_UopsSymbol *sym);
extern int _Py_uop_abstractcontext_init(_Py_UOpsContext *ctx);

168
Include/internal/pycore_uop_ids.h generated
View File

@ -118,17 +118,21 @@ extern "C" {
#define _GUARD_IS_NOT_NONE_POP 356
#define _GUARD_IS_TRUE_POP 357
#define _GUARD_KEYS_VERSION 358
#define _GUARD_NOT_EXHAUSTED_LIST 359
#define _GUARD_NOT_EXHAUSTED_RANGE 360
#define _GUARD_NOT_EXHAUSTED_TUPLE 361
#define _GUARD_TYPE_VERSION 362
#define _INIT_CALL_BOUND_METHOD_EXACT_ARGS 363
#define _INIT_CALL_PY_EXACT_ARGS 364
#define _INIT_CALL_PY_EXACT_ARGS_0 365
#define _INIT_CALL_PY_EXACT_ARGS_1 366
#define _INIT_CALL_PY_EXACT_ARGS_2 367
#define _INIT_CALL_PY_EXACT_ARGS_3 368
#define _INIT_CALL_PY_EXACT_ARGS_4 369
#define _GUARD_NOS_FLOAT 359
#define _GUARD_NOS_INT 360
#define _GUARD_NOT_EXHAUSTED_LIST 361
#define _GUARD_NOT_EXHAUSTED_RANGE 362
#define _GUARD_NOT_EXHAUSTED_TUPLE 363
#define _GUARD_TOS_FLOAT 364
#define _GUARD_TOS_INT 365
#define _GUARD_TYPE_VERSION 366
#define _INIT_CALL_BOUND_METHOD_EXACT_ARGS 367
#define _INIT_CALL_PY_EXACT_ARGS 368
#define _INIT_CALL_PY_EXACT_ARGS_0 369
#define _INIT_CALL_PY_EXACT_ARGS_1 370
#define _INIT_CALL_PY_EXACT_ARGS_2 371
#define _INIT_CALL_PY_EXACT_ARGS_3 372
#define _INIT_CALL_PY_EXACT_ARGS_4 373
#define _INSTRUMENTED_CALL INSTRUMENTED_CALL
#define _INSTRUMENTED_CALL_FUNCTION_EX INSTRUMENTED_CALL_FUNCTION_EX
#define _INSTRUMENTED_CALL_KW INSTRUMENTED_CALL_KW
@ -145,65 +149,65 @@ extern "C" {
#define _INSTRUMENTED_RETURN_CONST INSTRUMENTED_RETURN_CONST
#define _INSTRUMENTED_RETURN_VALUE INSTRUMENTED_RETURN_VALUE
#define _INSTRUMENTED_YIELD_VALUE INSTRUMENTED_YIELD_VALUE
#define _INTERNAL_INCREMENT_OPT_COUNTER 370
#define _IS_NONE 371
#define _INTERNAL_INCREMENT_OPT_COUNTER 374
#define _IS_NONE 375
#define _IS_OP IS_OP
#define _ITER_CHECK_LIST 372
#define _ITER_CHECK_RANGE 373
#define _ITER_CHECK_TUPLE 374
#define _ITER_JUMP_LIST 375
#define _ITER_JUMP_RANGE 376
#define _ITER_JUMP_TUPLE 377
#define _ITER_NEXT_LIST 378
#define _ITER_NEXT_RANGE 379
#define _ITER_NEXT_TUPLE 380
#define _JUMP_TO_TOP 381
#define _ITER_CHECK_LIST 376
#define _ITER_CHECK_RANGE 377
#define _ITER_CHECK_TUPLE 378
#define _ITER_JUMP_LIST 379
#define _ITER_JUMP_RANGE 380
#define _ITER_JUMP_TUPLE 381
#define _ITER_NEXT_LIST 382
#define _ITER_NEXT_RANGE 383
#define _ITER_NEXT_TUPLE 384
#define _JUMP_TO_TOP 385
#define _LIST_APPEND LIST_APPEND
#define _LIST_EXTEND LIST_EXTEND
#define _LOAD_ASSERTION_ERROR LOAD_ASSERTION_ERROR
#define _LOAD_ATTR 382
#define _LOAD_ATTR_CLASS 383
#define _LOAD_ATTR_CLASS_0 384
#define _LOAD_ATTR_CLASS_1 385
#define _LOAD_ATTR 386
#define _LOAD_ATTR_CLASS 387
#define _LOAD_ATTR_CLASS_0 388
#define _LOAD_ATTR_CLASS_1 389
#define _LOAD_ATTR_GETATTRIBUTE_OVERRIDDEN LOAD_ATTR_GETATTRIBUTE_OVERRIDDEN
#define _LOAD_ATTR_INSTANCE_VALUE 386
#define _LOAD_ATTR_INSTANCE_VALUE_0 387
#define _LOAD_ATTR_INSTANCE_VALUE_1 388
#define _LOAD_ATTR_METHOD_LAZY_DICT 389
#define _LOAD_ATTR_METHOD_NO_DICT 390
#define _LOAD_ATTR_METHOD_WITH_VALUES 391
#define _LOAD_ATTR_MODULE 392
#define _LOAD_ATTR_NONDESCRIPTOR_NO_DICT 393
#define _LOAD_ATTR_NONDESCRIPTOR_WITH_VALUES 394
#define _LOAD_ATTR_INSTANCE_VALUE 390
#define _LOAD_ATTR_INSTANCE_VALUE_0 391
#define _LOAD_ATTR_INSTANCE_VALUE_1 392
#define _LOAD_ATTR_METHOD_LAZY_DICT 393
#define _LOAD_ATTR_METHOD_NO_DICT 394
#define _LOAD_ATTR_METHOD_WITH_VALUES 395
#define _LOAD_ATTR_MODULE 396
#define _LOAD_ATTR_NONDESCRIPTOR_NO_DICT 397
#define _LOAD_ATTR_NONDESCRIPTOR_WITH_VALUES 398
#define _LOAD_ATTR_PROPERTY LOAD_ATTR_PROPERTY
#define _LOAD_ATTR_SLOT 395
#define _LOAD_ATTR_SLOT_0 396
#define _LOAD_ATTR_SLOT_1 397
#define _LOAD_ATTR_WITH_HINT 398
#define _LOAD_ATTR_SLOT 399
#define _LOAD_ATTR_SLOT_0 400
#define _LOAD_ATTR_SLOT_1 401
#define _LOAD_ATTR_WITH_HINT 402
#define _LOAD_BUILD_CLASS LOAD_BUILD_CLASS
#define _LOAD_CONST LOAD_CONST
#define _LOAD_CONST_INLINE 399
#define _LOAD_CONST_INLINE_BORROW 400
#define _LOAD_CONST_INLINE_BORROW_WITH_NULL 401
#define _LOAD_CONST_INLINE_WITH_NULL 402
#define _LOAD_CONST_INLINE 403
#define _LOAD_CONST_INLINE_BORROW 404
#define _LOAD_CONST_INLINE_BORROW_WITH_NULL 405
#define _LOAD_CONST_INLINE_WITH_NULL 406
#define _LOAD_DEREF LOAD_DEREF
#define _LOAD_FAST 403
#define _LOAD_FAST_0 404
#define _LOAD_FAST_1 405
#define _LOAD_FAST_2 406
#define _LOAD_FAST_3 407
#define _LOAD_FAST_4 408
#define _LOAD_FAST_5 409
#define _LOAD_FAST_6 410
#define _LOAD_FAST_7 411
#define _LOAD_FAST 407
#define _LOAD_FAST_0 408
#define _LOAD_FAST_1 409
#define _LOAD_FAST_2 410
#define _LOAD_FAST_3 411
#define _LOAD_FAST_4 412
#define _LOAD_FAST_5 413
#define _LOAD_FAST_6 414
#define _LOAD_FAST_7 415
#define _LOAD_FAST_AND_CLEAR LOAD_FAST_AND_CLEAR
#define _LOAD_FAST_CHECK LOAD_FAST_CHECK
#define _LOAD_FAST_LOAD_FAST LOAD_FAST_LOAD_FAST
#define _LOAD_FROM_DICT_OR_DEREF LOAD_FROM_DICT_OR_DEREF
#define _LOAD_FROM_DICT_OR_GLOBALS LOAD_FROM_DICT_OR_GLOBALS
#define _LOAD_GLOBAL 412
#define _LOAD_GLOBAL_BUILTINS 413
#define _LOAD_GLOBAL_MODULE 414
#define _LOAD_GLOBAL 416
#define _LOAD_GLOBAL_BUILTINS 417
#define _LOAD_GLOBAL_MODULE 418
#define _LOAD_LOCALS LOAD_LOCALS
#define _LOAD_NAME LOAD_NAME
#define _LOAD_SUPER_ATTR_ATTR LOAD_SUPER_ATTR_ATTR
@ -217,49 +221,49 @@ extern "C" {
#define _MATCH_SEQUENCE MATCH_SEQUENCE
#define _NOP NOP
#define _POP_EXCEPT POP_EXCEPT
#define _POP_FRAME 415
#define _POP_JUMP_IF_FALSE 416
#define _POP_JUMP_IF_TRUE 417
#define _POP_FRAME 419
#define _POP_JUMP_IF_FALSE 420
#define _POP_JUMP_IF_TRUE 421
#define _POP_TOP POP_TOP
#define _POP_TOP_LOAD_CONST_INLINE_BORROW 418
#define _POP_TOP_LOAD_CONST_INLINE_BORROW 422
#define _PUSH_EXC_INFO PUSH_EXC_INFO
#define _PUSH_FRAME 419
#define _PUSH_FRAME 423
#define _PUSH_NULL PUSH_NULL
#define _REPLACE_WITH_TRUE 420
#define _REPLACE_WITH_TRUE 424
#define _RESUME_CHECK RESUME_CHECK
#define _SAVE_RETURN_OFFSET 421
#define _SEND 422
#define _SAVE_RETURN_OFFSET 425
#define _SEND 426
#define _SEND_GEN SEND_GEN
#define _SETUP_ANNOTATIONS SETUP_ANNOTATIONS
#define _SET_ADD SET_ADD
#define _SET_FUNCTION_ATTRIBUTE SET_FUNCTION_ATTRIBUTE
#define _SET_UPDATE SET_UPDATE
#define _SIDE_EXIT 423
#define _START_EXECUTOR 424
#define _STORE_ATTR 425
#define _STORE_ATTR_INSTANCE_VALUE 426
#define _STORE_ATTR_SLOT 427
#define _SIDE_EXIT 427
#define _START_EXECUTOR 428
#define _STORE_ATTR 429
#define _STORE_ATTR_INSTANCE_VALUE 430
#define _STORE_ATTR_SLOT 431
#define _STORE_ATTR_WITH_HINT STORE_ATTR_WITH_HINT
#define _STORE_DEREF STORE_DEREF
#define _STORE_FAST 428
#define _STORE_FAST_0 429
#define _STORE_FAST_1 430
#define _STORE_FAST_2 431
#define _STORE_FAST_3 432
#define _STORE_FAST_4 433
#define _STORE_FAST_5 434
#define _STORE_FAST_6 435
#define _STORE_FAST_7 436
#define _STORE_FAST 432
#define _STORE_FAST_0 433
#define _STORE_FAST_1 434
#define _STORE_FAST_2 435
#define _STORE_FAST_3 436
#define _STORE_FAST_4 437
#define _STORE_FAST_5 438
#define _STORE_FAST_6 439
#define _STORE_FAST_7 440
#define _STORE_FAST_LOAD_FAST STORE_FAST_LOAD_FAST
#define _STORE_FAST_STORE_FAST STORE_FAST_STORE_FAST
#define _STORE_GLOBAL STORE_GLOBAL
#define _STORE_NAME STORE_NAME
#define _STORE_SLICE STORE_SLICE
#define _STORE_SUBSCR 437
#define _STORE_SUBSCR 441
#define _STORE_SUBSCR_DICT STORE_SUBSCR_DICT
#define _STORE_SUBSCR_LIST_INT STORE_SUBSCR_LIST_INT
#define _SWAP SWAP
#define _TO_BOOL 438
#define _TO_BOOL 442
#define _TO_BOOL_BOOL TO_BOOL_BOOL
#define _TO_BOOL_INT TO_BOOL_INT
#define _TO_BOOL_LIST TO_BOOL_LIST
@ -269,12 +273,12 @@ extern "C" {
#define _UNARY_NEGATIVE UNARY_NEGATIVE
#define _UNARY_NOT UNARY_NOT
#define _UNPACK_EX UNPACK_EX
#define _UNPACK_SEQUENCE 439
#define _UNPACK_SEQUENCE 443
#define _UNPACK_SEQUENCE_LIST UNPACK_SEQUENCE_LIST
#define _UNPACK_SEQUENCE_TUPLE UNPACK_SEQUENCE_TUPLE
#define _UNPACK_SEQUENCE_TWO_TUPLE UNPACK_SEQUENCE_TWO_TUPLE
#define _WITH_EXCEPT_START WITH_EXCEPT_START
#define MAX_UOP_ID 439
#define MAX_UOP_ID 443
#ifdef __cplusplus
}

16
Include/internal/pycore_uop_metadata.h generated
View File

@ -59,10 +59,14 @@ const uint16_t _PyUop_Flags[MAX_UOP_ID+1] = {
[_REPLACE_WITH_TRUE] = 0,
[_UNARY_INVERT] = HAS_ERROR_FLAG | HAS_ESCAPES_FLAG,
[_GUARD_BOTH_INT] = HAS_EXIT_FLAG,
[_GUARD_NOS_INT] = HAS_EXIT_FLAG,
[_GUARD_TOS_INT] = HAS_EXIT_FLAG,
[_BINARY_OP_MULTIPLY_INT] = HAS_ERROR_FLAG | HAS_PURE_FLAG,
[_BINARY_OP_ADD_INT] = HAS_ERROR_FLAG | HAS_PURE_FLAG,
[_BINARY_OP_SUBTRACT_INT] = HAS_ERROR_FLAG | HAS_PURE_FLAG,
[_GUARD_BOTH_FLOAT] = HAS_EXIT_FLAG,
[_GUARD_NOS_FLOAT] = HAS_EXIT_FLAG,
[_GUARD_TOS_FLOAT] = HAS_EXIT_FLAG,
[_BINARY_OP_MULTIPLY_FLOAT] = HAS_PURE_FLAG,
[_BINARY_OP_ADD_FLOAT] = HAS_PURE_FLAG,
[_BINARY_OP_SUBTRACT_FLOAT] = HAS_PURE_FLAG,
@ -352,9 +356,13 @@ const char *const _PyOpcode_uop_name[MAX_UOP_ID+1] = {
[_GUARD_IS_NOT_NONE_POP] = "_GUARD_IS_NOT_NONE_POP",
[_GUARD_IS_TRUE_POP] = "_GUARD_IS_TRUE_POP",
[_GUARD_KEYS_VERSION] = "_GUARD_KEYS_VERSION",
[_GUARD_NOS_FLOAT] = "_GUARD_NOS_FLOAT",
[_GUARD_NOS_INT] = "_GUARD_NOS_INT",
[_GUARD_NOT_EXHAUSTED_LIST] = "_GUARD_NOT_EXHAUSTED_LIST",
[_GUARD_NOT_EXHAUSTED_RANGE] = "_GUARD_NOT_EXHAUSTED_RANGE",
[_GUARD_NOT_EXHAUSTED_TUPLE] = "_GUARD_NOT_EXHAUSTED_TUPLE",
[_GUARD_TOS_FLOAT] = "_GUARD_TOS_FLOAT",
[_GUARD_TOS_INT] = "_GUARD_TOS_INT",
[_GUARD_TYPE_VERSION] = "_GUARD_TYPE_VERSION",
[_INIT_CALL_BOUND_METHOD_EXACT_ARGS] = "_INIT_CALL_BOUND_METHOD_EXACT_ARGS",
[_INIT_CALL_PY_EXACT_ARGS] = "_INIT_CALL_PY_EXACT_ARGS",
@ -566,6 +574,10 @@ int _PyUop_num_popped(int opcode, int oparg)
return 1;
case _GUARD_BOTH_INT:
return 2;
case _GUARD_NOS_INT:
return 2;
case _GUARD_TOS_INT:
return 1;
case _BINARY_OP_MULTIPLY_INT:
return 2;
case _BINARY_OP_ADD_INT:
@ -574,6 +586,10 @@ int _PyUop_num_popped(int opcode, int oparg)
return 2;
case _GUARD_BOTH_FLOAT:
return 2;
case _GUARD_NOS_FLOAT:
return 2;
case _GUARD_TOS_FLOAT:
return 1;
case _BINARY_OP_MULTIPLY_FLOAT:
return 2;
case _BINARY_OP_ADD_FLOAT:

44
Lib/test/test_capi/test_opt.py
View File

@ -903,10 +903,50 @@ class TestUopsOptimization(unittest.TestCase):
self.assertTrue(res)
self.assertIsNotNone(ex)
uops = get_opnames(ex)
guard_both_float_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_BOTH_INT"]
self.assertLessEqual(len(guard_both_float_count), 1)
guard_both_int_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_BOTH_INT"]
self.assertLessEqual(len(guard_both_int_count), 1)
self.assertIn("_COMPARE_OP_INT", uops)
def test_compare_op_type_propagation_int_partial(self):
def testfunc(n):
a = 1
for _ in range(n):
if a > 2:
x = 0
if a < 2:
x = 1
return x
res, ex = self._run_with_optimizer(testfunc, 32)
self.assertEqual(res, 1)
self.assertIsNotNone(ex)
uops = get_opnames(ex)
guard_left_int_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_NOS_INT"]
guard_both_int_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_BOTH_INT"]
self.assertLessEqual(len(guard_left_int_count), 1)
self.assertEqual(len(guard_both_int_count), 0)
self.assertIn("_COMPARE_OP_INT", uops)
def test_compare_op_type_propagation_float_partial(self):
def testfunc(n):
a = 1.0
for _ in range(n):
if a > 2.0:
x = 0
if a < 2.0:
x = 1
return x
res, ex = self._run_with_optimizer(testfunc, 32)
self.assertEqual(res, 1)
self.assertIsNotNone(ex)
uops = get_opnames(ex)
guard_left_float_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_NOS_FLOAT"]
guard_both_float_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_BOTH_FLOAT"]
self.assertLessEqual(len(guard_left_float_count), 1)
self.assertEqual(len(guard_both_float_count), 0)
self.assertIn("_COMPARE_OP_FLOAT", uops)
def test_compare_op_type_propagation_unicode(self):
def testfunc(n):
a = ""

16
Python/bytecodes.c
View File

@ -426,6 +426,14 @@ dummy_func(
EXIT_IF(!PyLong_CheckExact(right));
}
op(_GUARD_NOS_INT, (left, unused -- left, unused)) {
EXIT_IF(!PyLong_CheckExact(left));
}
op(_GUARD_TOS_INT, (value -- value)) {
EXIT_IF(!PyLong_CheckExact(value));
}
pure op(_BINARY_OP_MULTIPLY_INT, (left, right -- res)) {
STAT_INC(BINARY_OP, hit);
res = _PyLong_Multiply((PyLongObject *)left, (PyLongObject *)right);
@ -462,6 +470,14 @@ dummy_func(
EXIT_IF(!PyFloat_CheckExact(right));
}
op(_GUARD_NOS_FLOAT, (left, unused -- left, unused)) {
EXIT_IF(!PyFloat_CheckExact(left));
}
op(_GUARD_TOS_FLOAT, (value -- value)) {
EXIT_IF(!PyFloat_CheckExact(value));
}
pure op(_BINARY_OP_MULTIPLY_FLOAT, (left, right -- res)) {
STAT_INC(BINARY_OP, hit);
double dres =

40
Python/executor_cases.c.h generated
View File

@ -447,6 +447,26 @@
break;
}
case _GUARD_NOS_INT: {
PyObject *left;
left = stack_pointer[-2];
if (!PyLong_CheckExact(left)) {
UOP_STAT_INC(uopcode, miss);
JUMP_TO_JUMP_TARGET();
}
break;
}
case _GUARD_TOS_INT: {
PyObject *value;
value = stack_pointer[-1];
if (!PyLong_CheckExact(value)) {
UOP_STAT_INC(uopcode, miss);
JUMP_TO_JUMP_TARGET();
}
break;
}
case _BINARY_OP_MULTIPLY_INT: {
PyObject *right;
PyObject *left;
@ -511,6 +531,26 @@
break;
}
case _GUARD_NOS_FLOAT: {
PyObject *left;
left = stack_pointer[-2];
if (!PyFloat_CheckExact(left)) {
UOP_STAT_INC(uopcode, miss);
JUMP_TO_JUMP_TARGET();
}
break;
}
case _GUARD_TOS_FLOAT: {
PyObject *value;
value = stack_pointer[-1];
if (!PyFloat_CheckExact(value)) {
UOP_STAT_INC(uopcode, miss);
JUMP_TO_JUMP_TARGET();
}
break;
}
case _BINARY_OP_MULTIPLY_FLOAT: {
PyObject *right;
PyObject *left;

1
Python/optimizer_analysis.c
View File

@ -320,6 +320,7 @@ remove_globals(_PyInterpreterFrame *frame, _PyUOpInstruction *buffer,
#define sym_new_const _Py_uop_sym_new_const
#define sym_new_null _Py_uop_sym_new_null
#define sym_has_type _Py_uop_sym_has_type
#define sym_get_type _Py_uop_sym_get_type
#define sym_matches_type _Py_uop_sym_matches_type
#define sym_set_null _Py_uop_sym_set_null
#define sym_set_non_null _Py_uop_sym_set_non_null

51
Python/optimizer_bytecodes.c
View File

@ -21,6 +21,7 @@ typedef struct _Py_UOpsAbstractFrame _Py_UOpsAbstractFrame;
#define sym_new_const _Py_uop_sym_new_const
#define sym_new_null _Py_uop_sym_new_null
#define sym_matches_type _Py_uop_sym_matches_type
#define sym_get_type _Py_uop_sym_get_type
#define sym_has_type _Py_uop_sym_has_type
#define sym_set_null _Py_uop_sym_set_null
#define sym_set_non_null _Py_uop_sym_set_non_null
@ -99,9 +100,18 @@ dummy_func(void) {
}
op(_GUARD_BOTH_INT, (left, right -- left, right)) {
if (sym_matches_type(left, &PyLong_Type) &&
sym_matches_type(right, &PyLong_Type)) {
REPLACE_OP(this_instr, _NOP, 0, 0);
if (sym_matches_type(left, &PyLong_Type)) {
if (sym_matches_type(right, &PyLong_Type)) {
REPLACE_OP(this_instr, _NOP, 0, 0);
}
else {
REPLACE_OP(this_instr, _GUARD_TOS_INT, 0, 0);
}
}
else {
if (sym_matches_type(right, &PyLong_Type)) {
REPLACE_OP(this_instr, _GUARD_NOS_INT, 0, 0);
}
}
if (!sym_set_type(left, &PyLong_Type)) {
goto hit_bottom;
@ -112,9 +122,18 @@ dummy_func(void) {
}
op(_GUARD_BOTH_FLOAT, (left, right -- left, right)) {
if (sym_matches_type(left, &PyFloat_Type) &&
sym_matches_type(right, &PyFloat_Type)) {
REPLACE_OP(this_instr, _NOP, 0 ,0);
if (sym_matches_type(left, &PyFloat_Type)) {
if (sym_matches_type(right, &PyFloat_Type)) {
REPLACE_OP(this_instr, _NOP, 0, 0);
}
else {
REPLACE_OP(this_instr, _GUARD_TOS_FLOAT, 0, 0);
}
}
else {
if (sym_matches_type(right, &PyFloat_Type)) {
REPLACE_OP(this_instr, _GUARD_NOS_FLOAT, 0, 0);
}
}
if (!sym_set_type(left, &PyFloat_Type)) {
goto hit_bottom;
@ -137,6 +156,25 @@ dummy_func(void) {
}
}
op(_BINARY_OP, (left, right -- res)) {
PyTypeObject *ltype = sym_get_type(left);
PyTypeObject *rtype = sym_get_type(right);
if (ltype != NULL && (ltype == &PyLong_Type || ltype == &PyFloat_Type) &&
rtype != NULL && (rtype == &PyLong_Type || rtype == &PyFloat_Type))
{
if (oparg != NB_TRUE_DIVIDE && oparg != NB_INPLACE_TRUE_DIVIDE &&
ltype == &PyLong_Type && rtype == &PyLong_Type) {
/* If both inputs are ints and the op is not division the result is an int */
OUT_OF_SPACE_IF_NULL(res = sym_new_type(ctx, &PyLong_Type));
}
else {
/* For any other op combining ints/floats the result is a float */
OUT_OF_SPACE_IF_NULL(res = sym_new_type(ctx, &PyFloat_Type));
}
}
OUT_OF_SPACE_IF_NULL(res = sym_new_unknown(ctx));
}
op(_BINARY_OP_ADD_INT, (left, right -- res)) {
if (sym_is_const(left) && sym_is_const(right) &&
sym_matches_type(left, &PyLong_Type) && sym_matches_type(right, &PyLong_Type))
@ -424,7 +462,6 @@ dummy_func(void) {
OUT_OF_SPACE_IF_NULL(null = sym_new_null(ctx));
}
op(_COPY, (bottom, unused[oparg-1] -- bottom, unused[oparg-1], top)) {
assert(oparg > 0);
top = bottom;

68
Python/optimizer_cases.c.h generated
View File

@ -225,9 +225,18 @@
_Py_UopsSymbol *left;
right = stack_pointer[-1];
left = stack_pointer[-2];
if (sym_matches_type(left, &PyLong_Type) &&
sym_matches_type(right, &PyLong_Type)) {
REPLACE_OP(this_instr, _NOP, 0, 0);
if (sym_matches_type(left, &PyLong_Type)) {
if (sym_matches_type(right, &PyLong_Type)) {
REPLACE_OP(this_instr, _NOP, 0, 0);
}
else {
REPLACE_OP(this_instr, _GUARD_TOS_INT, 0, 0);
}
}
else {
if (sym_matches_type(right, &PyLong_Type)) {
REPLACE_OP(this_instr, _GUARD_NOS_INT, 0, 0);
}
}
if (!sym_set_type(left, &PyLong_Type)) {
goto hit_bottom;
@ -238,6 +247,14 @@
break;
}
case _GUARD_NOS_INT: {
break;
}
case _GUARD_TOS_INT: {
break;
}
case _BINARY_OP_MULTIPLY_INT: {
_Py_UopsSymbol *right;
_Py_UopsSymbol *left;
@ -333,9 +350,18 @@
_Py_UopsSymbol *left;
right = stack_pointer[-1];
left = stack_pointer[-2];
if (sym_matches_type(left, &PyFloat_Type) &&
sym_matches_type(right, &PyFloat_Type)) {
REPLACE_OP(this_instr, _NOP, 0 ,0);
if (sym_matches_type(left, &PyFloat_Type)) {
if (sym_matches_type(right, &PyFloat_Type)) {
REPLACE_OP(this_instr, _NOP, 0, 0);
}
else {
REPLACE_OP(this_instr, _GUARD_TOS_FLOAT, 0, 0);
}
}
else {
if (sym_matches_type(right, &PyFloat_Type)) {
REPLACE_OP(this_instr, _GUARD_NOS_FLOAT, 0, 0);
}
}
if (!sym_set_type(left, &PyFloat_Type)) {
goto hit_bottom;
@ -346,6 +372,14 @@
break;
}
case _GUARD_NOS_FLOAT: {
break;
}
case _GUARD_TOS_FLOAT: {
break;
}
case _BINARY_OP_MULTIPLY_FLOAT: {
_Py_UopsSymbol *right;
_Py_UopsSymbol *left;
@ -1852,9 +1886,27 @@
}
case _BINARY_OP: {
_Py_UopsSymbol *right;
_Py_UopsSymbol *left;
_Py_UopsSymbol *res;
res = sym_new_not_null(ctx);
if (res == NULL) goto out_of_space;
right = stack_pointer[-1];
left = stack_pointer[-2];
PyTypeObject *ltype = sym_get_type(left);
PyTypeObject *rtype = sym_get_type(right);
if (ltype != NULL && (ltype == &PyLong_Type || ltype == &PyFloat_Type) &&
rtype != NULL && (rtype == &PyLong_Type || rtype == &PyFloat_Type))
{
if (oparg != NB_TRUE_DIVIDE && oparg != NB_INPLACE_TRUE_DIVIDE &&
ltype == &PyLong_Type && rtype == &PyLong_Type) {
/* If both inputs are ints and the op is not division the result is an int */
OUT_OF_SPACE_IF_NULL(res = sym_new_type(ctx, &PyLong_Type));
}
else {
/* For any other op combining ints/floats the result is a float */
OUT_OF_SPACE_IF_NULL(res = sym_new_type(ctx, &PyFloat_Type));
}
}
OUT_OF_SPACE_IF_NULL(res = sym_new_unknown(ctx));
stack_pointer[-2] = res;
stack_pointer += -1;
break;

14
Python/optimizer_symbols.c
View File

@ -231,6 +231,15 @@ _Py_uop_sym_new_null(_Py_UOpsContext *ctx)
return null_sym;
}
PyTypeObject *
_Py_uop_sym_get_type(_Py_UopsSymbol *sym)
{
if (_Py_uop_sym_is_bottom(sym)) {
return NULL;
}
return sym->typ;
}
bool
_Py_uop_sym_has_type(_Py_UopsSymbol *sym)
{
@ -244,10 +253,7 @@ bool
_Py_uop_sym_matches_type(_Py_UopsSymbol *sym, PyTypeObject *typ)
{
assert(typ != NULL && PyType_Check(typ));
if (_Py_uop_sym_is_bottom(sym)) {
return false;
}
return sym->typ == typ;
return _Py_uop_sym_get_type(sym) == typ;
}
int