[3.11] GH-95113: Don't use EXTENDED_ARG_QUICK in unquickened code (GH-95121) (GH-95143)

(cherry picked from commit e402b26b7f)
This commit is contained in:
Brandt Bucher 2022-07-22 11:56:10 -07:00 committed by GitHub
parent e99496e1c2
commit 5a48ab01e9
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
8 changed files with 20 additions and 198 deletions

View File

@ -16,8 +16,6 @@ extern const uint8_t _PyOpcode_Caches[256];
extern const uint8_t _PyOpcode_Deopt[256];
extern const uint8_t _PyOpcode_Original[256];
#ifdef NEED_OPCODE_TABLES
static const uint32_t _PyOpcode_RelativeJump[8] = {
0U,
@ -237,190 +235,6 @@ const uint8_t _PyOpcode_Deopt[256] = {
[WITH_EXCEPT_START] = WITH_EXCEPT_START,
[YIELD_VALUE] = YIELD_VALUE,
};
const uint8_t _PyOpcode_Original[256] = {
[ASYNC_GEN_WRAP] = ASYNC_GEN_WRAP,
[BEFORE_ASYNC_WITH] = BEFORE_ASYNC_WITH,
[BEFORE_WITH] = BEFORE_WITH,
[BINARY_OP] = BINARY_OP,
[BINARY_OP_ADAPTIVE] = BINARY_OP,
[BINARY_OP_ADD_FLOAT] = BINARY_OP,
[BINARY_OP_ADD_INT] = BINARY_OP,
[BINARY_OP_ADD_UNICODE] = BINARY_OP,
[BINARY_OP_INPLACE_ADD_UNICODE] = BINARY_OP,
[BINARY_OP_MULTIPLY_FLOAT] = BINARY_OP,
[BINARY_OP_MULTIPLY_INT] = BINARY_OP,
[BINARY_OP_SUBTRACT_FLOAT] = BINARY_OP,
[BINARY_OP_SUBTRACT_INT] = BINARY_OP,
[BINARY_SUBSCR] = BINARY_SUBSCR,
[BINARY_SUBSCR_ADAPTIVE] = BINARY_SUBSCR,
[BINARY_SUBSCR_DICT] = BINARY_SUBSCR,
[BINARY_SUBSCR_GETITEM] = BINARY_SUBSCR,
[BINARY_SUBSCR_LIST_INT] = BINARY_SUBSCR,
[BINARY_SUBSCR_TUPLE_INT] = BINARY_SUBSCR,
[BUILD_CONST_KEY_MAP] = BUILD_CONST_KEY_MAP,
[BUILD_LIST] = BUILD_LIST,
[BUILD_MAP] = BUILD_MAP,
[BUILD_SET] = BUILD_SET,
[BUILD_SLICE] = BUILD_SLICE,
[BUILD_STRING] = BUILD_STRING,
[BUILD_TUPLE] = BUILD_TUPLE,
[CACHE] = CACHE,
[CALL] = CALL,
[CALL_ADAPTIVE] = CALL,
[CALL_FUNCTION_EX] = CALL_FUNCTION_EX,
[CALL_PY_EXACT_ARGS] = CALL,
[CALL_PY_WITH_DEFAULTS] = CALL,
[CHECK_EG_MATCH] = CHECK_EG_MATCH,
[CHECK_EXC_MATCH] = CHECK_EXC_MATCH,
[COMPARE_OP] = COMPARE_OP,
[COMPARE_OP_ADAPTIVE] = COMPARE_OP,
[COMPARE_OP_FLOAT_JUMP] = COMPARE_OP,
[COMPARE_OP_INT_JUMP] = COMPARE_OP,
[COMPARE_OP_STR_JUMP] = COMPARE_OP,
[CONTAINS_OP] = CONTAINS_OP,
[COPY] = COPY,
[COPY_FREE_VARS] = COPY_FREE_VARS,
[DELETE_ATTR] = DELETE_ATTR,
[DELETE_DEREF] = DELETE_DEREF,
[DELETE_FAST] = DELETE_FAST,
[DELETE_GLOBAL] = DELETE_GLOBAL,
[DELETE_NAME] = DELETE_NAME,
[DELETE_SUBSCR] = DELETE_SUBSCR,
[DICT_MERGE] = DICT_MERGE,
[DICT_UPDATE] = DICT_UPDATE,
[END_ASYNC_FOR] = END_ASYNC_FOR,
[EXTENDED_ARG] = EXTENDED_ARG_QUICK,
[EXTENDED_ARG_QUICK] = EXTENDED_ARG_QUICK,
[FORMAT_VALUE] = FORMAT_VALUE,
[FOR_ITER] = FOR_ITER,
[GET_AITER] = GET_AITER,
[GET_ANEXT] = GET_ANEXT,
[GET_AWAITABLE] = GET_AWAITABLE,
[GET_ITER] = GET_ITER,
[GET_LEN] = GET_LEN,
[GET_YIELD_FROM_ITER] = GET_YIELD_FROM_ITER,
[IMPORT_FROM] = IMPORT_FROM,
[IMPORT_NAME] = IMPORT_NAME,
[IMPORT_STAR] = IMPORT_STAR,
[IS_OP] = IS_OP,
[JUMP_BACKWARD] = JUMP_BACKWARD,
[JUMP_BACKWARD_NO_INTERRUPT] = JUMP_BACKWARD_NO_INTERRUPT,
[JUMP_BACKWARD_QUICK] = JUMP_BACKWARD,
[JUMP_FORWARD] = JUMP_FORWARD,
[JUMP_IF_FALSE_OR_POP] = JUMP_IF_FALSE_OR_POP,
[JUMP_IF_TRUE_OR_POP] = JUMP_IF_TRUE_OR_POP,
[KW_NAMES] = KW_NAMES,
[LIST_APPEND] = LIST_APPEND,
[LIST_EXTEND] = LIST_EXTEND,
[LIST_TO_TUPLE] = LIST_TO_TUPLE,
[LOAD_ASSERTION_ERROR] = LOAD_ASSERTION_ERROR,
[LOAD_ATTR] = LOAD_ATTR,
[LOAD_ATTR_ADAPTIVE] = LOAD_ATTR,
[LOAD_ATTR_INSTANCE_VALUE] = LOAD_ATTR,
[LOAD_ATTR_MODULE] = LOAD_ATTR,
[LOAD_ATTR_SLOT] = LOAD_ATTR,
[LOAD_ATTR_WITH_HINT] = LOAD_ATTR,
[LOAD_BUILD_CLASS] = LOAD_BUILD_CLASS,
[LOAD_CLASSDEREF] = LOAD_CLASSDEREF,
[LOAD_CLOSURE] = LOAD_CLOSURE,
[LOAD_CONST] = LOAD_CONST,
[LOAD_CONST__LOAD_FAST] = LOAD_CONST,
[LOAD_DEREF] = LOAD_DEREF,
[LOAD_FAST] = LOAD_FAST,
[LOAD_FAST__LOAD_CONST] = LOAD_FAST,
[LOAD_FAST__LOAD_FAST] = LOAD_FAST,
[LOAD_GLOBAL] = LOAD_GLOBAL,
[LOAD_GLOBAL_ADAPTIVE] = LOAD_GLOBAL,
[LOAD_GLOBAL_BUILTIN] = LOAD_GLOBAL,
[LOAD_GLOBAL_MODULE] = LOAD_GLOBAL,
[LOAD_METHOD] = LOAD_METHOD,
[LOAD_METHOD_ADAPTIVE] = LOAD_METHOD,
[LOAD_METHOD_CLASS] = LOAD_METHOD,
[LOAD_METHOD_MODULE] = LOAD_METHOD,
[LOAD_METHOD_NO_DICT] = LOAD_METHOD,
[LOAD_METHOD_WITH_DICT] = LOAD_METHOD,
[LOAD_METHOD_WITH_VALUES] = LOAD_METHOD,
[LOAD_NAME] = LOAD_NAME,
[MAKE_CELL] = MAKE_CELL,
[MAKE_FUNCTION] = MAKE_FUNCTION,
[MAP_ADD] = MAP_ADD,
[MATCH_CLASS] = MATCH_CLASS,
[MATCH_KEYS] = MATCH_KEYS,
[MATCH_MAPPING] = MATCH_MAPPING,
[MATCH_SEQUENCE] = MATCH_SEQUENCE,
[NOP] = NOP,
[POP_EXCEPT] = POP_EXCEPT,
[POP_JUMP_BACKWARD_IF_FALSE] = POP_JUMP_BACKWARD_IF_FALSE,
[POP_JUMP_BACKWARD_IF_NONE] = POP_JUMP_BACKWARD_IF_NONE,
[POP_JUMP_BACKWARD_IF_NOT_NONE] = POP_JUMP_BACKWARD_IF_NOT_NONE,
[POP_JUMP_BACKWARD_IF_TRUE] = POP_JUMP_BACKWARD_IF_TRUE,
[POP_JUMP_FORWARD_IF_FALSE] = POP_JUMP_FORWARD_IF_FALSE,
[POP_JUMP_FORWARD_IF_NONE] = POP_JUMP_FORWARD_IF_NONE,
[POP_JUMP_FORWARD_IF_NOT_NONE] = POP_JUMP_FORWARD_IF_NOT_NONE,
[POP_JUMP_FORWARD_IF_TRUE] = POP_JUMP_FORWARD_IF_TRUE,
[POP_TOP] = POP_TOP,
[PRECALL] = PRECALL,
[PRECALL_ADAPTIVE] = PRECALL,
[PRECALL_BOUND_METHOD] = PRECALL,
[PRECALL_BUILTIN_CLASS] = PRECALL,
[PRECALL_BUILTIN_FAST_WITH_KEYWORDS] = PRECALL,
[PRECALL_METHOD_DESCRIPTOR_FAST_WITH_KEYWORDS] = PRECALL,
[PRECALL_NO_KW_BUILTIN_FAST] = PRECALL,
[PRECALL_NO_KW_BUILTIN_O] = PRECALL,
[PRECALL_NO_KW_ISINSTANCE] = PRECALL,
[PRECALL_NO_KW_LEN] = PRECALL,
[PRECALL_NO_KW_LIST_APPEND] = PRECALL,
[PRECALL_NO_KW_METHOD_DESCRIPTOR_FAST] = PRECALL,
[PRECALL_NO_KW_METHOD_DESCRIPTOR_NOARGS] = PRECALL,
[PRECALL_NO_KW_METHOD_DESCRIPTOR_O] = PRECALL,
[PRECALL_NO_KW_STR_1] = PRECALL,
[PRECALL_NO_KW_TUPLE_1] = PRECALL,
[PRECALL_NO_KW_TYPE_1] = PRECALL,
[PRECALL_PYFUNC] = PRECALL,
[PREP_RERAISE_STAR] = PREP_RERAISE_STAR,
[PRINT_EXPR] = PRINT_EXPR,
[PUSH_EXC_INFO] = PUSH_EXC_INFO,
[PUSH_NULL] = PUSH_NULL,
[RAISE_VARARGS] = RAISE_VARARGS,
[RERAISE] = RERAISE,
[RESUME] = RESUME,
[RESUME_QUICK] = RESUME,
[RETURN_GENERATOR] = RETURN_GENERATOR,
[RETURN_VALUE] = RETURN_VALUE,
[SEND] = SEND,
[SETUP_ANNOTATIONS] = SETUP_ANNOTATIONS,
[SET_ADD] = SET_ADD,
[SET_UPDATE] = SET_UPDATE,
[STORE_ATTR] = STORE_ATTR,
[STORE_ATTR_ADAPTIVE] = STORE_ATTR,
[STORE_ATTR_INSTANCE_VALUE] = STORE_ATTR,
[STORE_ATTR_SLOT] = STORE_ATTR,
[STORE_ATTR_WITH_HINT] = STORE_ATTR,
[STORE_DEREF] = STORE_DEREF,
[STORE_FAST] = STORE_FAST,
[STORE_FAST__LOAD_FAST] = STORE_FAST,
[STORE_FAST__STORE_FAST] = STORE_FAST,
[STORE_GLOBAL] = STORE_GLOBAL,
[STORE_NAME] = STORE_NAME,
[STORE_SUBSCR] = STORE_SUBSCR,
[STORE_SUBSCR_ADAPTIVE] = STORE_SUBSCR,
[STORE_SUBSCR_DICT] = STORE_SUBSCR,
[STORE_SUBSCR_LIST_INT] = STORE_SUBSCR,
[SWAP] = SWAP,
[UNARY_INVERT] = UNARY_INVERT,
[UNARY_NEGATIVE] = UNARY_NEGATIVE,
[UNARY_NOT] = UNARY_NOT,
[UNARY_POSITIVE] = UNARY_POSITIVE,
[UNPACK_EX] = UNPACK_EX,
[UNPACK_SEQUENCE] = UNPACK_SEQUENCE,
[UNPACK_SEQUENCE_ADAPTIVE] = UNPACK_SEQUENCE,
[UNPACK_SEQUENCE_LIST] = UNPACK_SEQUENCE,
[UNPACK_SEQUENCE_TUPLE] = UNPACK_SEQUENCE,
[UNPACK_SEQUENCE_TWO_TUPLE] = UNPACK_SEQUENCE,
[WITH_EXCEPT_START] = WITH_EXCEPT_START,
[YIELD_VALUE] = YIELD_VALUE,
};
#endif // NEED_OPCODE_TABLES
#ifdef Py_DEBUG

View File

@ -638,7 +638,7 @@ def extended_arg_quick():
%3d 0 RESUME 0
%3d 2 LOAD_CONST 1 (Ellipsis)
4 EXTENDED_ARG_QUICK 1
4 EXTENDED_ARG 1
6 UNPACK_EX 256
8 STORE_FAST 0 (_)
10 STORE_FAST 0 (_)

View File

@ -0,0 +1,4 @@
Replace all ``EXTENDED_ARG_QUICK`` instructions with basic
:opcode:`EXTENDED_ARG` instructions in unquickened code. Consumers of
non-adaptive bytecode should be able to handle extended arguments the same
way they were handled in CPython 3.10 and older.

View File

@ -1416,7 +1416,7 @@ deopt_code(_Py_CODEUNIT *instructions, Py_ssize_t len)
{
for (int i = 0; i < len; i++) {
_Py_CODEUNIT instruction = instructions[i];
int opcode = _PyOpcode_Original[_Py_OPCODE(instruction)];
int opcode = _PyOpcode_Deopt[_Py_OPCODE(instruction)];
int caches = _PyOpcode_Caches[opcode];
instructions[i] = _Py_MAKECODEUNIT(opcode, _Py_OPARG(instruction));
while (caches--) {

View File

@ -5590,8 +5590,16 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
assert(oparg);
oparg <<= 8;
oparg |= _Py_OPARG(*next_instr);
// We might be tracing. To avoid breaking tracing guarantees in
// quickened instructions, always deoptimize the next opcode:
opcode = _PyOpcode_Deopt[_Py_OPCODE(*next_instr)];
PRE_DISPATCH_GOTO();
// CPython hasn't traced the following instruction historically
// (DO_TRACING would clobber our extended oparg anyways), so just
// skip our usual cframe.use_tracing check before dispatch. Also,
// make sure the next instruction isn't a RESUME, since that needs
// to trace properly (and shouldn't have an extended arg anyways):
assert(opcode != RESUME);
DISPATCH_GOTO();
}

View File

@ -211,13 +211,13 @@ write_instr(_Py_CODEUNIT *codestr, struct instr *instruction, int ilen)
int caches = _PyOpcode_Caches[opcode];
switch (ilen - caches) {
case 4:
*codestr++ = _Py_MAKECODEUNIT(EXTENDED_ARG_QUICK, (oparg >> 24) & 0xFF);
*codestr++ = _Py_MAKECODEUNIT(EXTENDED_ARG, (oparg >> 24) & 0xFF);
/* fall through */
case 3:
*codestr++ = _Py_MAKECODEUNIT(EXTENDED_ARG_QUICK, (oparg >> 16) & 0xFF);
*codestr++ = _Py_MAKECODEUNIT(EXTENDED_ARG, (oparg >> 16) & 0xFF);
/* fall through */
case 2:
*codestr++ = _Py_MAKECODEUNIT(EXTENDED_ARG_QUICK, (oparg >> 8) & 0xFF);
*codestr++ = _Py_MAKECODEUNIT(EXTENDED_ARG, (oparg >> 8) & 0xFF);
/* fall through */
case 1:
*codestr++ = _Py_MAKECODEUNIT(opcode, oparg & 0xFF);

View File

@ -275,6 +275,9 @@ _PyCode_Quicken(PyCodeObject *code)
else {
assert(!_PyOpcode_Caches[opcode]);
switch (opcode) {
case EXTENDED_ARG:
_Py_SET_OPCODE(instructions[i], EXTENDED_ARG_QUICK);
break;
case JUMP_BACKWARD:
_Py_SET_OPCODE(instructions[i], JUMP_BACKWARD_QUICK);
break;

View File

@ -117,7 +117,6 @@ def main(opcode_py, outfile='Include/opcode.h', internaloutfile='Include/interna
iobj.write("\nextern const uint8_t _PyOpcode_Caches[256];\n")
iobj.write("\nextern const uint8_t _PyOpcode_Deopt[256];\n")
iobj.write("\nextern const uint8_t _PyOpcode_Original[256];\n")
iobj.write("\n#ifdef NEED_OPCODE_TABLES\n")
write_int_array_from_ops("_PyOpcode_RelativeJump", opcode['hasjrel'], iobj)
write_int_array_from_ops("_PyOpcode_Jump", opcode['hasjrel'] + opcode['hasjabs'], iobj)
@ -138,12 +137,6 @@ def main(opcode_py, outfile='Include/opcode.h', internaloutfile='Include/interna
for opt, deopt in sorted(deoptcodes.items()):
iobj.write(f" [{opt}] = {deopt},\n")
iobj.write("};\n")
iobj.write("\nconst uint8_t _PyOpcode_Original[256] = {\n")
for opt, deopt in sorted(deoptcodes.items()):
if opt.startswith("EXTENDED_ARG"):
deopt = "EXTENDED_ARG_QUICK"
iobj.write(f" [{opt}] = {deopt},\n")
iobj.write("};\n")
iobj.write("#endif // NEED_OPCODE_TABLES\n")
fobj.write("\n")