diff options
author | Brandt Bucher <brandtbucher@microsoft.com> | 2022-03-07 19:45:00 (GMT) |
---|---|---|
committer | GitHub <noreply@github.com> | 2022-03-07 19:45:00 (GMT) |
commit | f193631387bfee99a812e39b05d5b7e6384b57f5 (patch) | |
tree | 31f161bd1e2f6469f32be8333705c82992486485 /Python | |
parent | 105b9ac00174d7bcc653f9e9dc5052215e197c77 (diff) | |
download | cpython-f193631387bfee99a812e39b05d5b7e6384b57f5.zip cpython-f193631387bfee99a812e39b05d5b7e6384b57f5.tar.gz cpython-f193631387bfee99a812e39b05d5b7e6384b57f5.tar.bz2 |
bpo-46841: Use inline caching for calls (GH-31709)
Diffstat (limited to 'Python')
-rw-r--r-- | Python/ceval.c | 235 | ||||
-rw-r--r-- | Python/opcode_targets.h | 38 | ||||
-rw-r--r-- | Python/pylifecycle.c | 10 | ||||
-rw-r--r-- | Python/specialize.c | 263 |
4 files changed, 191 insertions, 355 deletions
diff --git a/Python/ceval.c b/Python/ceval.c index 7439710..b15c101 100644 --- a/Python/ceval.c +++ b/Python/ceval.c @@ -1321,6 +1321,10 @@ eval_frame_handle_pending(PyThreadState *tstate) #define JUMPTO(x) (next_instr = first_instr + (x)) #define JUMPBY(x) (next_instr += (x)) +// Skip from a PRECALL over a CALL to the next instruction: +#define SKIP_CALL() \ + JUMPBY(INLINE_CACHE_ENTRIES_PRECALL + 1 + INLINE_CACHE_ENTRIES_CALL) + /* Get opcode and oparg from original instructions, not quickened form. */ #define TRACING_NEXTOPARG() do { \ _Py_CODEUNIT word = ((_Py_CODEUNIT *)PyBytes_AS_STRING(frame->f_code->co_code))[INSTR_OFFSET()]; \ @@ -1431,9 +1435,6 @@ eval_frame_handle_pending(PyThreadState *tstate) #define JUMP_TO_INSTRUCTION(op) goto PREDICT_ID(op) -#define GET_CACHE() \ - _GetSpecializedCacheEntryForInstruction(first_instr, INSTR_OFFSET(), oparg) - #define DEOPT_IF(cond, instname) if (cond) { goto instname ## _miss; } @@ -3003,8 +3004,8 @@ handle_eval_breaker: TARGET(LOAD_GLOBAL_ADAPTIVE) { assert(cframe.use_tracing == 0); - uint16_t counter = *next_instr; - if (counter == 0) { + _PyLoadGlobalCache *cache = (_PyLoadGlobalCache *)next_instr; + if (cache->counter == 0) { PyObject *name = GETITEM(names, oparg); next_instr--; if (_Py_Specialize_LoadGlobal(GLOBALS(), BUILTINS(), next_instr, name) < 0) { @@ -3014,7 +3015,7 @@ handle_eval_breaker: } else { STAT_INC(LOAD_GLOBAL, deferred); - *next_instr = counter-1; + cache->counter--; JUMP_TO_INSTRUCTION(LOAD_GLOBAL); } } @@ -4563,20 +4564,12 @@ handle_eval_breaker: We'll be passing `oparg + 1` to call_function, to make it accept the `self` as a first argument. */ - int is_method = (PEEK(oparg + 2) != NULL); - int nargs = oparg + is_method; + int is_meth = is_method(stack_pointer, oparg); + int nargs = oparg + is_meth; /* Move ownership of reference from stack to call_shape * and make sure that NULL is cleared from stack */ PyObject *function = PEEK(nargs + 1); -#ifdef Py_STATS - extern int _PySpecialization_ClassifyCallable(PyObject *); - SpecializationStats *stats = - &_py_stats.opcode_stats[PRECALL].specialization; - stats->failure++; - int kind = _PySpecialization_ClassifyCallable(function); - stats->failure_kinds[kind]++; -#endif - if (!is_method && Py_TYPE(function) == &PyMethod_Type) { + if (!is_meth && Py_TYPE(function) == &PyMethod_Type) { PyObject *meth = ((PyMethodObject *)function)->im_func; PyObject *self = ((PyMethodObject *)function)->im_self; Py_INCREF(meth); @@ -4585,35 +4578,32 @@ handle_eval_breaker: PEEK(oparg+2) = meth; Py_DECREF(function); } + JUMPBY(INLINE_CACHE_ENTRIES_PRECALL); DISPATCH(); } TARGET(PRECALL_BOUND_METHOD) { - SpecializedCacheEntry *cache = GET_CACHE(); - int original_oparg = cache->adaptive.original_oparg; - int is_method = (PEEK(original_oparg + 2) != NULL); - DEOPT_IF(is_method, PRECALL); - PyObject *function = PEEK(original_oparg + 1); + DEOPT_IF(is_method(stack_pointer, oparg), PRECALL); + PyObject *function = PEEK(oparg + 1); DEOPT_IF(Py_TYPE(function) != &PyMethod_Type, PRECALL); STAT_INC(PRECALL, hit); PyObject *meth = ((PyMethodObject *)function)->im_func; PyObject *self = ((PyMethodObject *)function)->im_self; Py_INCREF(meth); Py_INCREF(self); - PEEK(original_oparg+1) = self; - PEEK(original_oparg+2) = meth; + PEEK(oparg + 1) = self; + PEEK(oparg + 2) = meth; Py_DECREF(function); + JUMPBY(INLINE_CACHE_ENTRIES_PRECALL); DISPATCH(); } TARGET(PRECALL_PYFUNC) { - SpecializedCacheEntry *cache = GET_CACHE(); - int original_oparg = cache->adaptive.original_oparg; - int is_method = (PEEK(original_oparg + 2) != NULL); - int nargs = original_oparg + is_method; + int nargs = oparg + is_method(stack_pointer, oparg); PyObject *function = PEEK(nargs + 1); DEOPT_IF(Py_TYPE(function) != &PyFunction_Type, PRECALL); STAT_INC(PRECALL, hit); + JUMPBY(INLINE_CACHE_ENTRIES_PRECALL); DISPATCH(); } @@ -4649,6 +4639,7 @@ handle_eval_breaker: goto error; } _PyFrame_SetStackPointer(frame, stack_pointer); + frame->f_lasti += INLINE_CACHE_ENTRIES_CALL; new_frame->previous = frame; cframe.current_frame = frame = new_frame; CALL_STAT_INC(inlined_py_calls); @@ -4680,21 +4671,20 @@ handle_eval_breaker: if (res == NULL) { goto error; } + JUMPBY(INLINE_CACHE_ENTRIES_CALL); CHECK_EVAL_BREAKER(); DISPATCH(); } TARGET(PRECALL_ADAPTIVE) { - SpecializedCacheEntry *cache = GET_CACHE(); - int original_oparg = cache->adaptive.original_oparg; - if (cache->adaptive.counter == 0) { + _PyPrecallCache *cache = (_PyPrecallCache *)next_instr; + if (cache->counter == 0) { next_instr--; - int is_meth = is_method(stack_pointer, original_oparg); - int nargs = original_oparg + is_meth; + int is_meth = is_method(stack_pointer, oparg); + int nargs = oparg + is_meth; PyObject *callable = PEEK(nargs + 1); - int err = _Py_Specialize_Precall( - callable, next_instr, nargs, - call_shape.kwnames, cache, BUILTINS()); + int err = _Py_Specialize_Precall(callable, next_instr, nargs, + call_shape.kwnames, oparg); if (err < 0) { goto error; } @@ -4702,23 +4692,20 @@ handle_eval_breaker: } else { STAT_INC(PRECALL, deferred); - cache->adaptive.counter--; - oparg = original_oparg; + cache->counter--; JUMP_TO_INSTRUCTION(PRECALL); } } TARGET(CALL_ADAPTIVE) { - SpecializedCacheEntry *cache = GET_CACHE(); - int original_oparg = cache->adaptive.original_oparg; - if (cache->adaptive.counter == 0) { + _PyCallCache *cache = (_PyCallCache *)next_instr; + if (cache->counter == 0) { next_instr--; - int is_meth = is_method(stack_pointer, original_oparg); - int nargs = original_oparg + is_meth; + int is_meth = is_method(stack_pointer, oparg); + int nargs = oparg + is_meth; PyObject *callable = PEEK(nargs + 1); - int err = _Py_Specialize_Call( - callable, next_instr, nargs, - call_shape.kwnames, cache); + int err = _Py_Specialize_Call(callable, next_instr, nargs, + call_shape.kwnames); if (err < 0) { goto error; } @@ -4726,23 +4713,20 @@ handle_eval_breaker: } else { STAT_INC(CALL, deferred); - cache->adaptive.counter--; - oparg = original_oparg; + cache->counter--; goto call_function; } } TARGET(CALL_PY_EXACT_ARGS) { assert(call_shape.kwnames == NULL); - SpecializedCacheEntry *caches = GET_CACHE(); - int original_oparg = caches->adaptive.original_oparg; - int is_meth = is_method(stack_pointer, original_oparg); - int argcount = original_oparg + is_meth; + _PyCallCache *cache = (_PyCallCache *)next_instr; + int is_meth = is_method(stack_pointer, oparg); + int argcount = oparg + is_meth; PyObject *callable = PEEK(argcount + 1); DEOPT_IF(!PyFunction_Check(callable), CALL); - _PyCallCache *cache1 = &caches[-1].call; PyFunctionObject *func = (PyFunctionObject *)callable; - DEOPT_IF(func->func_version != cache1->func_version, CALL); + DEOPT_IF(func->func_version != read_u32(cache->func_version), CALL); PyCodeObject *code = (PyCodeObject *)func->func_code; DEOPT_IF(code->co_argcount != argcount, CALL); STAT_INC(CALL, hit); @@ -4760,6 +4744,7 @@ handle_eval_breaker: } STACK_SHRINK(2-is_meth); _PyFrame_SetStackPointer(frame, stack_pointer); + frame->f_lasti += INLINE_CACHE_ENTRIES_CALL; new_frame->previous = frame; frame = cframe.current_frame = new_frame; goto start_frame; @@ -4767,18 +4752,16 @@ handle_eval_breaker: TARGET(CALL_PY_WITH_DEFAULTS) { assert(call_shape.kwnames == NULL); - SpecializedCacheEntry *caches = GET_CACHE(); - int original_oparg = caches->adaptive.original_oparg; - int is_meth = is_method(stack_pointer, original_oparg); - int argcount = original_oparg + is_meth; + _PyCallCache *cache = (_PyCallCache *)next_instr; + int is_meth = is_method(stack_pointer, oparg); + int argcount = oparg + is_meth; PyObject *callable = PEEK(argcount + 1); DEOPT_IF(!PyFunction_Check(callable), CALL); - _PyCallCache *cache1 = &caches[-1].call; PyFunctionObject *func = (PyFunctionObject *)callable; - DEOPT_IF(func->func_version != cache1->func_version, CALL); + DEOPT_IF(func->func_version != read_u32(cache->func_version), CALL); PyCodeObject *code = (PyCodeObject *)func->func_code; DEOPT_IF(argcount > code->co_argcount, CALL); - int minargs = cache1->min_args; + int minargs = cache->min_args; DEOPT_IF(argcount < minargs, CALL); STAT_INC(CALL, hit); _PyInterpreterFrame *new_frame = _PyFrame_Push(tstate, func); @@ -4790,9 +4773,9 @@ handle_eval_breaker: for (int i = 0; i < argcount; i++) { new_frame->localsplus[i] = stack_pointer[i]; } - int def_offset = cache1->defaults_len - code->co_argcount; for (int i = argcount; i < code->co_argcount; i++) { - PyObject *def = PyTuple_GET_ITEM(func->func_defaults, i + def_offset); + PyObject *def = PyTuple_GET_ITEM(func->func_defaults, + i - minargs); Py_INCREF(def); new_frame->localsplus[i] = def; } @@ -4801,6 +4784,7 @@ handle_eval_breaker: } STACK_SHRINK(2-is_meth); _PyFrame_SetStackPointer(frame, stack_pointer); + frame->f_lasti += INLINE_CACHE_ENTRIES_CALL; new_frame->previous = frame; frame = cframe.current_frame = new_frame; goto start_frame; @@ -4809,13 +4793,13 @@ handle_eval_breaker: TARGET(PRECALL_NO_KW_TYPE_1) { assert(call_shape.kwnames == NULL); assert(cframe.use_tracing == 0); - assert(GET_CACHE()->adaptive.original_oparg == 1); + assert(oparg == 1); DEOPT_IF(is_method(stack_pointer, 1), PRECALL); PyObject *obj = TOP(); PyObject *callable = SECOND(); DEOPT_IF(callable != (PyObject *)&PyType_Type, PRECALL); - next_instr++; // Skip following call STAT_INC(PRECALL, hit); + SKIP_CALL(); PyObject *res = Py_NewRef(Py_TYPE(obj)); Py_DECREF(callable); Py_DECREF(obj); @@ -4827,12 +4811,12 @@ handle_eval_breaker: TARGET(PRECALL_NO_KW_STR_1) { assert(call_shape.kwnames == NULL); assert(cframe.use_tracing == 0); - assert(GET_CACHE()->adaptive.original_oparg == 1); + assert(oparg == 1); DEOPT_IF(is_method(stack_pointer, 1), PRECALL); PyObject *callable = PEEK(2); DEOPT_IF(callable != (PyObject *)&PyUnicode_Type, PRECALL); - next_instr++; // Skip following call STAT_INC(PRECALL, hit); + SKIP_CALL(); PyObject *arg = TOP(); PyObject *res = PyObject_Str(arg); Py_DECREF(arg); @@ -4848,12 +4832,12 @@ handle_eval_breaker: TARGET(PRECALL_NO_KW_TUPLE_1) { assert(call_shape.kwnames == NULL); - assert(GET_CACHE()->adaptive.original_oparg == 1); + assert(oparg == 1); DEOPT_IF(is_method(stack_pointer, 1), PRECALL); PyObject *callable = PEEK(2); DEOPT_IF(callable != (PyObject *)&PyTuple_Type, PRECALL); - next_instr++; // Skip following call STAT_INC(PRECALL, hit); + SKIP_CALL(); PyObject *arg = TOP(); PyObject *res = PySequence_Tuple(arg); Py_DECREF(arg); @@ -4868,16 +4852,15 @@ handle_eval_breaker: } TARGET(PRECALL_BUILTIN_CLASS) { - int original_oparg = GET_CACHE()->adaptive.original_oparg; - int is_meth = is_method(stack_pointer, original_oparg); - int total_args = original_oparg + is_meth; + int is_meth = is_method(stack_pointer, oparg); + int total_args = oparg + is_meth; int kwnames_len = KWNAMES_LEN(); PyObject *callable = PEEK(total_args + 1); DEOPT_IF(!PyType_Check(callable), PRECALL); PyTypeObject *tp = (PyTypeObject *)callable; DEOPT_IF(tp->tp_vectorcall == NULL, PRECALL); - next_instr++; // Skip following call STAT_INC(PRECALL, hit); + SKIP_CALL(); STACK_SHRINK(total_args); PyObject *res = tp->tp_vectorcall((PyObject *)tp, stack_pointer, total_args-kwnames_len, call_shape.kwnames); @@ -4900,16 +4883,14 @@ handle_eval_breaker: assert(cframe.use_tracing == 0); /* Builtin METH_O functions */ assert(call_shape.kwnames == NULL); - SpecializedCacheEntry *caches = GET_CACHE(); - int original_oparg = caches->adaptive.original_oparg; - int is_meth = is_method(stack_pointer, original_oparg); - int total_args = original_oparg + is_meth; + int is_meth = is_method(stack_pointer, oparg); + int total_args = oparg + is_meth; DEOPT_IF(total_args != 1, PRECALL); PyObject *callable = PEEK(total_args + 1); DEOPT_IF(!PyCFunction_CheckExact(callable), PRECALL); DEOPT_IF(PyCFunction_GET_FLAGS(callable) != METH_O, PRECALL); - next_instr++; // Skip following call STAT_INC(PRECALL, hit); + SKIP_CALL(); PyCFunction cfunc = PyCFunction_GET_FUNCTION(callable); // This is slower but CPython promises to check all non-vectorcall // function calls. @@ -4936,16 +4917,14 @@ handle_eval_breaker: assert(cframe.use_tracing == 0); /* Builtin METH_FASTCALL functions, without keywords */ assert(call_shape.kwnames == NULL); - SpecializedCacheEntry *caches = GET_CACHE(); - int original_oparg = caches->adaptive.original_oparg; - int is_meth = is_method(stack_pointer, original_oparg); - int total_args = original_oparg + is_meth; + int is_meth = is_method(stack_pointer, oparg); + int total_args = oparg + is_meth; PyObject *callable = PEEK(total_args + 1); DEOPT_IF(!PyCFunction_CheckExact(callable), PRECALL); DEOPT_IF(PyCFunction_GET_FLAGS(callable) != METH_FASTCALL, PRECALL); - next_instr++; // Skip following call STAT_INC(PRECALL, hit); + SKIP_CALL(); PyCFunction cfunc = PyCFunction_GET_FUNCTION(callable); STACK_SHRINK(total_args); /* res = func(self, args, nargs) */ @@ -4977,16 +4956,14 @@ handle_eval_breaker: TARGET(PRECALL_BUILTIN_FAST_WITH_KEYWORDS) { assert(cframe.use_tracing == 0); /* Builtin METH_FASTCALL | METH_KEYWORDS functions */ - SpecializedCacheEntry *caches = GET_CACHE(); - int original_oparg = caches->adaptive.original_oparg; - int is_meth = is_method(stack_pointer, original_oparg); - int total_args = original_oparg + is_meth; + int is_meth = is_method(stack_pointer, oparg); + int total_args = oparg + is_meth; PyObject *callable = PEEK(total_args + 1); DEOPT_IF(!PyCFunction_CheckExact(callable), PRECALL); DEOPT_IF(PyCFunction_GET_FLAGS(callable) != (METH_FASTCALL | METH_KEYWORDS), PRECALL); - next_instr++; // Skip following call STAT_INC(PRECALL, hit); + SKIP_CALL(); STACK_SHRINK(total_args); /* res = func(self, args, nargs, kwnames) */ _PyCFunctionFastWithKeywords cfunc = @@ -5019,16 +4996,14 @@ handle_eval_breaker: assert(cframe.use_tracing == 0); assert(call_shape.kwnames == NULL); /* len(o) */ - SpecializedCacheEntry *caches = GET_CACHE(); - int original_oparg = caches->adaptive.original_oparg; - int is_meth = is_method(stack_pointer, original_oparg); - int total_args = original_oparg + is_meth; + int is_meth = is_method(stack_pointer, oparg); + int total_args = oparg + is_meth; DEOPT_IF(total_args != 1, PRECALL); - _PyObjectCache *cache1 = &caches[-1].obj; PyObject *callable = PEEK(total_args + 1); - DEOPT_IF(callable != cache1->obj, PRECALL); - next_instr++; // Skip following call + PyInterpreterState *interp = _PyInterpreterState_GET(); + DEOPT_IF(callable != interp->callable_cache.len, PRECALL); STAT_INC(PRECALL, hit); + SKIP_CALL(); PyObject *arg = TOP(); Py_ssize_t len_i = PyObject_Length(arg); if (len_i < 0) { @@ -5051,17 +5026,14 @@ handle_eval_breaker: assert(cframe.use_tracing == 0); assert(call_shape.kwnames == NULL); /* isinstance(o, o2) */ - SpecializedCacheEntry *caches = GET_CACHE(); - int original_oparg = caches->adaptive.original_oparg; - int is_meth = is_method(stack_pointer, original_oparg); - int total_args = original_oparg + is_meth; + int is_meth = is_method(stack_pointer, oparg); + int total_args = oparg + is_meth; PyObject *callable = PEEK(total_args + 1); DEOPT_IF(total_args != 2, PRECALL); - _PyObjectCache *cache1 = &caches[-1].obj; - - DEOPT_IF(callable != cache1->obj, PRECALL); - next_instr++; // Skip following call + PyInterpreterState *interp = _PyInterpreterState_GET(); + DEOPT_IF(callable != interp->callable_cache.isinstance, PRECALL); STAT_INC(PRECALL, hit); + SKIP_CALL(); PyObject *cls = POP(); PyObject *inst = TOP(); int retval = PyObject_IsInstance(inst, cls); @@ -5086,16 +5058,14 @@ handle_eval_breaker: TARGET(PRECALL_NO_KW_LIST_APPEND) { assert(cframe.use_tracing == 0); assert(call_shape.kwnames == NULL); - assert(GET_CACHE()->adaptive.original_oparg == 1); - SpecializedCacheEntry *caches = GET_CACHE(); - _PyObjectCache *cache1 = &caches[-1].obj; - assert(cache1->obj != NULL); + assert(oparg == 1); PyObject *callable = PEEK(3); - DEOPT_IF(callable != cache1->obj, PRECALL); + PyInterpreterState *interp = _PyInterpreterState_GET(); + DEOPT_IF(callable != interp->callable_cache.list_append, PRECALL); PyObject *list = SECOND(); DEOPT_IF(!PyList_Check(list), PRECALL); STAT_INC(PRECALL, hit); - next_instr++; // Skip following call + SKIP_CALL(); PyObject *arg = TOP(); int err = PyList_Append(list, arg); if (err) { @@ -5112,16 +5082,15 @@ handle_eval_breaker: TARGET(PRECALL_NO_KW_METHOD_DESCRIPTOR_O) { assert(call_shape.kwnames == NULL); - int original_oparg = GET_CACHE()->adaptive.original_oparg; - int is_meth = is_method(stack_pointer, original_oparg); - int total_args = original_oparg + is_meth; + int is_meth = is_method(stack_pointer, oparg); + int total_args = oparg + is_meth; PyObject *callable = PEEK(total_args + 1); DEOPT_IF(total_args != 2, PRECALL); DEOPT_IF(!Py_IS_TYPE(callable, &PyMethodDescr_Type), PRECALL); PyMethodDef *meth = ((PyMethodDescrObject *)callable)->d_method; DEOPT_IF(meth->ml_flags != METH_O, PRECALL); - next_instr++; // Skip following call STAT_INC(PRECALL, hit); + SKIP_CALL(); PyCFunction cfunc = meth->ml_meth; // This is slower but CPython promises to check all non-vectorcall // function calls. @@ -5135,7 +5104,7 @@ handle_eval_breaker: assert((res != NULL) ^ (_PyErr_Occurred(tstate) != NULL)); Py_DECREF(self); Py_DECREF(arg); - STACK_SHRINK(original_oparg+1); + STACK_SHRINK(oparg + 1); SET_TOP(res); Py_DECREF(callable); if (res == NULL) { @@ -5147,17 +5116,16 @@ handle_eval_breaker: TARGET(PRECALL_NO_KW_METHOD_DESCRIPTOR_NOARGS) { assert(call_shape.kwnames == NULL); - int original_oparg = GET_CACHE()->adaptive.original_oparg; - assert(original_oparg == 0 || original_oparg == 1); - int is_meth = is_method(stack_pointer, original_oparg); - int total_args = original_oparg + is_meth; + assert(oparg == 0 || oparg == 1); + int is_meth = is_method(stack_pointer, oparg); + int total_args = oparg + is_meth; DEOPT_IF(total_args != 1, PRECALL); PyObject *callable = SECOND(); DEOPT_IF(!Py_IS_TYPE(callable, &PyMethodDescr_Type), PRECALL); PyMethodDef *meth = ((PyMethodDescrObject *)callable)->d_method; DEOPT_IF(meth->ml_flags != METH_NOARGS, PRECALL); - next_instr++; // Skip following call STAT_INC(PRECALL, hit); + SKIP_CALL(); PyCFunction cfunc = meth->ml_meth; // This is slower but CPython promises to check all non-vectorcall // function calls. @@ -5169,7 +5137,7 @@ handle_eval_breaker: _Py_LeaveRecursiveCall(tstate); assert((res != NULL) ^ (_PyErr_Occurred(tstate) != NULL)); Py_DECREF(self); - STACK_SHRINK(original_oparg+1); + STACK_SHRINK(oparg + 1); SET_TOP(res); Py_DECREF(callable); if (res == NULL) { @@ -5181,16 +5149,15 @@ handle_eval_breaker: TARGET(PRECALL_NO_KW_METHOD_DESCRIPTOR_FAST) { assert(call_shape.kwnames == NULL); - int original_oparg = GET_CACHE()->adaptive.original_oparg; - int is_meth = is_method(stack_pointer, original_oparg); - int total_args = original_oparg + is_meth; + int is_meth = is_method(stack_pointer, oparg); + int total_args = oparg + is_meth; PyObject *callable = PEEK(total_args + 1); /* Builtin METH_FASTCALL methods, without keywords */ DEOPT_IF(!Py_IS_TYPE(callable, &PyMethodDescr_Type), PRECALL); PyMethodDef *meth = ((PyMethodDescrObject *)callable)->d_method; DEOPT_IF(meth->ml_flags != METH_FASTCALL, PRECALL); - next_instr++; // Skip following call STAT_INC(PRECALL, hit); + SKIP_CALL(); _PyCFunctionFast cfunc = (_PyCFunctionFast)(void(*)(void))meth->ml_meth; int nargs = total_args-1; STACK_SHRINK(nargs); @@ -5537,22 +5504,6 @@ handle_eval_breaker: /* Specialization misses */ -#define MISS_WITH_CACHE(opname) \ -opname ## _miss: \ - { \ - STAT_INC(opcode, miss); \ - STAT_INC(opname, miss); \ - _PyAdaptiveEntry *cache = &GET_CACHE()->adaptive; \ - cache->counter--; \ - if (cache->counter == 0) { \ - next_instr[-1] = _Py_MAKECODEUNIT(opname ## _ADAPTIVE, _Py_OPARG(next_instr[-1])); \ - STAT_INC(opname, deopt); \ - cache_backoff(cache); \ - } \ - oparg = cache->original_oparg; \ - JUMP_TO_INSTRUCTION(opname); \ - } - #define MISS_WITH_INLINE_CACHE(opname) \ opname ## _miss: \ { \ @@ -5588,8 +5539,8 @@ MISS_WITH_INLINE_CACHE(LOAD_ATTR) MISS_WITH_INLINE_CACHE(STORE_ATTR) MISS_WITH_INLINE_CACHE(LOAD_GLOBAL) MISS_WITH_INLINE_CACHE(LOAD_METHOD) -MISS_WITH_CACHE(PRECALL) -MISS_WITH_CACHE(CALL) +MISS_WITH_INLINE_CACHE(PRECALL) +MISS_WITH_INLINE_CACHE(CALL) MISS_WITH_INLINE_CACHE(BINARY_OP) MISS_WITH_INLINE_CACHE(COMPARE_OP) MISS_WITH_INLINE_CACHE(BINARY_SUBSCR) diff --git a/Python/opcode_targets.h b/Python/opcode_targets.h index 2060793..7c94999 100644 --- a/Python/opcode_targets.h +++ b/Python/opcode_targets.h @@ -1,21 +1,20 @@ static void *opcode_targets[256] = { - &&_unknown_opcode, + &&TARGET_CACHE, &&TARGET_POP_TOP, &&TARGET_PUSH_NULL, - &&TARGET_CACHE, &&TARGET_BINARY_OP_ADAPTIVE, &&TARGET_BINARY_OP_ADD_INT, &&TARGET_BINARY_OP_ADD_FLOAT, &&TARGET_BINARY_OP_ADD_UNICODE, &&TARGET_BINARY_OP_INPLACE_ADD_UNICODE, + &&TARGET_BINARY_OP_MULTIPLY_INT, &&TARGET_NOP, &&TARGET_UNARY_POSITIVE, &&TARGET_UNARY_NEGATIVE, &&TARGET_UNARY_NOT, - &&TARGET_BINARY_OP_MULTIPLY_INT, &&TARGET_BINARY_OP_MULTIPLY_FLOAT, - &&TARGET_UNARY_INVERT, &&TARGET_BINARY_OP_SUBTRACT_INT, + &&TARGET_UNARY_INVERT, &&TARGET_BINARY_OP_SUBTRACT_FLOAT, &&TARGET_COMPARE_OP_ADAPTIVE, &&TARGET_COMPARE_OP_FLOAT_JUMP, @@ -24,18 +23,18 @@ static void *opcode_targets[256] = { &&TARGET_BINARY_SUBSCR_ADAPTIVE, &&TARGET_BINARY_SUBSCR_GETITEM, &&TARGET_BINARY_SUBSCR_LIST_INT, - &&TARGET_BINARY_SUBSCR, &&TARGET_BINARY_SUBSCR_TUPLE_INT, + &&TARGET_BINARY_SUBSCR, &&TARGET_BINARY_SUBSCR_DICT, &&TARGET_STORE_SUBSCR_ADAPTIVE, &&TARGET_STORE_SUBSCR_LIST_INT, + &&TARGET_STORE_SUBSCR_DICT, &&TARGET_GET_LEN, &&TARGET_MATCH_MAPPING, &&TARGET_MATCH_SEQUENCE, &&TARGET_MATCH_KEYS, - &&TARGET_STORE_SUBSCR_DICT, - &&TARGET_PUSH_EXC_INFO, &&TARGET_CALL_ADAPTIVE, + &&TARGET_PUSH_EXC_INFO, &&TARGET_CALL_PY_EXACT_ARGS, &&TARGET_CALL_PY_WITH_DEFAULTS, &&TARGET_JUMP_ABSOLUTE_QUICK, @@ -48,39 +47,40 @@ static void *opcode_targets[256] = { &&TARGET_LOAD_GLOBAL_MODULE, &&TARGET_LOAD_GLOBAL_BUILTIN, &&TARGET_LOAD_METHOD_ADAPTIVE, + &&TARGET_LOAD_METHOD_CLASS, &&TARGET_WITH_EXCEPT_START, &&TARGET_GET_AITER, &&TARGET_GET_ANEXT, &&TARGET_BEFORE_ASYNC_WITH, &&TARGET_BEFORE_WITH, &&TARGET_END_ASYNC_FOR, - &&TARGET_LOAD_METHOD_CLASS, &&TARGET_LOAD_METHOD_MODULE, &&TARGET_LOAD_METHOD_NO_DICT, &&TARGET_LOAD_METHOD_WITH_DICT, &&TARGET_LOAD_METHOD_WITH_VALUES, + &&TARGET_PRECALL_ADAPTIVE, &&TARGET_STORE_SUBSCR, &&TARGET_DELETE_SUBSCR, - &&TARGET_PRECALL_ADAPTIVE, &&TARGET_PRECALL_BUILTIN_CLASS, &&TARGET_PRECALL_NO_KW_BUILTIN_O, &&TARGET_PRECALL_NO_KW_BUILTIN_FAST, &&TARGET_PRECALL_BUILTIN_FAST_WITH_KEYWORDS, &&TARGET_PRECALL_NO_KW_LEN, + &&TARGET_PRECALL_NO_KW_ISINSTANCE, &&TARGET_GET_ITER, &&TARGET_GET_YIELD_FROM_ITER, &&TARGET_PRINT_EXPR, &&TARGET_LOAD_BUILD_CLASS, - &&TARGET_PRECALL_NO_KW_ISINSTANCE, &&TARGET_PRECALL_NO_KW_LIST_APPEND, + &&TARGET_PRECALL_NO_KW_METHOD_DESCRIPTOR_O, &&TARGET_LOAD_ASSERTION_ERROR, &&TARGET_RETURN_GENERATOR, - &&TARGET_PRECALL_NO_KW_METHOD_DESCRIPTOR_O, &&TARGET_PRECALL_NO_KW_METHOD_DESCRIPTOR_NOARGS, &&TARGET_PRECALL_NO_KW_STR_1, &&TARGET_PRECALL_NO_KW_TUPLE_1, &&TARGET_PRECALL_NO_KW_TYPE_1, &&TARGET_PRECALL_NO_KW_METHOD_DESCRIPTOR_FAST, + &&TARGET_PRECALL_BOUND_METHOD, &&TARGET_LIST_TO_TUPLE, &&TARGET_RETURN_VALUE, &&TARGET_IMPORT_STAR, @@ -139,40 +139,39 @@ static void *opcode_targets[256] = { &&TARGET_LOAD_DEREF, &&TARGET_STORE_DEREF, &&TARGET_DELETE_DEREF, - &&TARGET_PRECALL_BOUND_METHOD, &&TARGET_PRECALL_PYFUNC, - &&TARGET_CALL_FUNCTION_EX, &&TARGET_RESUME_QUICK, + &&TARGET_CALL_FUNCTION_EX, + &&TARGET_STORE_ATTR_ADAPTIVE, &&TARGET_EXTENDED_ARG, &&TARGET_LIST_APPEND, &&TARGET_SET_ADD, &&TARGET_MAP_ADD, &&TARGET_LOAD_CLASSDEREF, &&TARGET_COPY_FREE_VARS, - &&TARGET_STORE_ATTR_ADAPTIVE, + &&TARGET_STORE_ATTR_INSTANCE_VALUE, &&TARGET_RESUME, &&TARGET_MATCH_CLASS, - &&TARGET_STORE_ATTR_INSTANCE_VALUE, &&TARGET_STORE_ATTR_SLOT, + &&TARGET_STORE_ATTR_WITH_HINT, &&TARGET_FORMAT_VALUE, &&TARGET_BUILD_CONST_KEY_MAP, &&TARGET_BUILD_STRING, - &&TARGET_STORE_ATTR_WITH_HINT, &&TARGET_UNPACK_SEQUENCE_ADAPTIVE, - &&TARGET_LOAD_METHOD, &&TARGET_UNPACK_SEQUENCE_LIST, + &&TARGET_LOAD_METHOD, + &&TARGET_UNPACK_SEQUENCE_TUPLE, &&TARGET_LIST_EXTEND, &&TARGET_SET_UPDATE, &&TARGET_DICT_MERGE, &&TARGET_DICT_UPDATE, &&TARGET_PRECALL, - &&TARGET_UNPACK_SEQUENCE_TUPLE, &&TARGET_UNPACK_SEQUENCE_TWO_TUPLE, &&TARGET_LOAD_FAST__LOAD_FAST, &&TARGET_STORE_FAST__LOAD_FAST, + &&TARGET_LOAD_FAST__LOAD_CONST, &&TARGET_CALL, &&TARGET_KW_NAMES, - &&TARGET_LOAD_FAST__LOAD_CONST, &&TARGET_LOAD_CONST__LOAD_FAST, &&TARGET_STORE_FAST__STORE_FAST, &&_unknown_opcode, @@ -254,5 +253,6 @@ static void *opcode_targets[256] = { &&_unknown_opcode, &&_unknown_opcode, &&_unknown_opcode, + &&_unknown_opcode, &&TARGET_DO_TRACING }; diff --git a/Python/pylifecycle.c b/Python/pylifecycle.c index 9228778..8abd536 100644 --- a/Python/pylifecycle.c +++ b/Python/pylifecycle.c @@ -774,6 +774,16 @@ pycore_init_builtins(PyThreadState *tstate) Py_INCREF(builtins_dict); interp->builtins = builtins_dict; + PyObject *isinstance = PyDict_GetItem(builtins_dict, &_Py_ID(isinstance)); + assert(isinstance); + interp->callable_cache.isinstance = isinstance; + PyObject *len = PyDict_GetItem(builtins_dict, &_Py_ID(len)); + assert(len); + interp->callable_cache.len = len; + PyObject *list_append = _PyType_Lookup(&PyList_Type, &_Py_ID(append)); + assert(list_append); + interp->callable_cache.list_append = list_append; + if (_PyBuiltins_AddExceptions(bimod) < 0) { return _PyStatus_ERR("failed to add exceptions to builtins"); } diff --git a/Python/specialize.c b/Python/specialize.c index 417eece..dae4e3f 100644 --- a/Python/specialize.c +++ b/Python/specialize.c @@ -56,13 +56,6 @@ static uint8_t adaptive_opcodes[256] = { [UNPACK_SEQUENCE] = UNPACK_SEQUENCE_ADAPTIVE, }; -/* The number of cache entries required for a "family" of instructions. */ -static uint8_t cache_requirements[256] = { - [STORE_SUBSCR] = 0, - [CALL] = 2, /* _PyAdaptiveEntry and _PyObjectCache/_PyCallCache */ - [PRECALL] = 2, /* _PyAdaptiveEntry and _PyObjectCache/_PyCallCache */ -}; - Py_ssize_t _Py_QuickenedCount = 0; #ifdef Py_STATS PyStats _py_stats = { 0 }; @@ -282,137 +275,41 @@ _Py_PrintSpecializationStats(int to_file) #define SPECIALIZATION_FAIL(opcode, kind) ((void)0) #endif -static SpecializedCacheOrInstruction * -allocate(int cache_count, int instruction_count) +static _Py_CODEUNIT * +allocate(int instruction_count) { - assert(sizeof(SpecializedCacheOrInstruction) == 2*sizeof(int32_t)); - assert(sizeof(SpecializedCacheEntry) == 2*sizeof(int32_t)); - assert(cache_count > 0); assert(instruction_count > 0); - int count = cache_count + (instruction_count + INSTRUCTIONS_PER_ENTRY -1)/INSTRUCTIONS_PER_ENTRY; - SpecializedCacheOrInstruction *array = (SpecializedCacheOrInstruction *) - PyMem_Malloc(sizeof(SpecializedCacheOrInstruction) * count); + void *array = PyMem_Malloc(sizeof(_Py_CODEUNIT) * instruction_count); if (array == NULL) { PyErr_NoMemory(); return NULL; } _Py_QuickenedCount++; - array[0].entry.zero.cache_count = cache_count; - return array; + return (_Py_CODEUNIT *)array; } -static int -get_cache_count(SpecializedCacheOrInstruction *quickened) { - return quickened[0].entry.zero.cache_count; -} -/* Return the oparg for the cache_offset and instruction index. - * - * If no cache is needed then return the original oparg. - * If a cache is needed, but cannot be accessed because - * oparg would be too large, then return -1. - * - * Also updates the cache_offset, as it may need to be incremented by - * more than the cache requirements, if many instructions do not need caches. - * - * See pycore_code.h for details of how the cache offset, - * instruction index and oparg are related */ -static int -oparg_from_instruction_and_update_offset(int index, int opcode, int original_oparg, int *cache_offset) { - /* The instruction pointer in the interpreter points to the next - * instruction, so we compute the offset using nexti (index + 1) */ - int nexti = index + 1; - uint8_t need = cache_requirements[opcode]; - if (need == 0) { - return original_oparg; - } - assert(adaptive_opcodes[opcode] != 0); - int oparg = oparg_from_offset_and_nexti(*cache_offset, nexti); - assert(*cache_offset == offset_from_oparg_and_nexti(oparg, nexti)); - /* Some cache space is wasted here as the minimum possible offset is (nexti>>1) */ - if (oparg < 0) { - oparg = 0; - *cache_offset = offset_from_oparg_and_nexti(oparg, nexti); - } - else if (oparg > 255) { - return -1; - } - *cache_offset += need; - return oparg; -} - -static int -entries_needed(const _Py_CODEUNIT *code, int len) -{ - int cache_offset = 0; - int previous_opcode = -1; - for (int i = 0; i < len; i++) { - uint8_t opcode = _Py_OPCODE(code[i]); - if (previous_opcode != EXTENDED_ARG) { - oparg_from_instruction_and_update_offset(i, opcode, 0, &cache_offset); - } - previous_opcode = opcode; - } - return cache_offset + 1; // One extra for the count entry -} - -static inline _Py_CODEUNIT * -first_instruction(SpecializedCacheOrInstruction *quickened) -{ - return &quickened[get_cache_count(quickened)].code[0]; -} - -/** Insert adaptive instructions and superinstructions. - * - * Skip instruction preceded by EXTENDED_ARG for adaptive - * instructions as those are both very rare and tricky - * to handle. - */ +// Insert adaptive instructions and superinstructions. static void -optimize(SpecializedCacheOrInstruction *quickened, int len) +optimize(_Py_CODEUNIT *instructions, int len) { - _Py_CODEUNIT *instructions = first_instruction(quickened); - int cache_offset = 0; int previous_opcode = -1; - int previous_oparg = 0; + int previous_oparg = -1; for(int i = 0; i < len; i++) { int opcode = _Py_OPCODE(instructions[i]); int oparg = _Py_OPARG(instructions[i]); uint8_t adaptive_opcode = adaptive_opcodes[opcode]; if (adaptive_opcode) { - if (_PyOpcode_InlineCacheEntries[opcode]) { - instructions[i] = _Py_MAKECODEUNIT(adaptive_opcode, oparg); - previous_opcode = -1; - i += _PyOpcode_InlineCacheEntries[opcode]; - } - else if (previous_opcode != EXTENDED_ARG) { - int new_oparg = oparg_from_instruction_and_update_offset( - i, opcode, oparg, &cache_offset - ); - if (new_oparg < 0) { - /* Not possible to allocate a cache for this instruction */ - previous_opcode = opcode; - continue; - } - previous_opcode = adaptive_opcode; - int entries_needed = cache_requirements[opcode]; - if (entries_needed) { - /* Initialize the adpative cache entry */ - int cache0_offset = cache_offset-entries_needed; - SpecializedCacheEntry *cache = - _GetSpecializedCacheEntry(instructions, cache0_offset); - cache->adaptive.original_oparg = oparg; - cache->adaptive.counter = 0; - } else { - // oparg is the adaptive cache counter - new_oparg = 0; - } - instructions[i] = _Py_MAKECODEUNIT(adaptive_opcode, new_oparg); - } + instructions[i] = _Py_MAKECODEUNIT(adaptive_opcode, oparg); + int caches = _PyOpcode_InlineCacheEntries[opcode]; + // Make sure the adaptive counter is zero: + assert((caches ? instructions[i + 1] : oparg) == 0); + previous_opcode = -1; + previous_oparg = -1; + i += caches; } else { - /* Super instructions don't use the cache, - * so no need to update the offset. */ + assert(!_PyOpcode_InlineCacheEntries[opcode]); switch (opcode) { case JUMP_ABSOLUTE: instructions[i] = _Py_MAKECODEUNIT(JUMP_ABSOLUTE_QUICK, oparg); @@ -423,23 +320,28 @@ optimize(SpecializedCacheOrInstruction *quickened, int len) case LOAD_FAST: switch(previous_opcode) { case LOAD_FAST: + assert(0 <= previous_oparg); instructions[i-1] = _Py_MAKECODEUNIT(LOAD_FAST__LOAD_FAST, previous_oparg); break; case STORE_FAST: + assert(0 <= previous_oparg); instructions[i-1] = _Py_MAKECODEUNIT(STORE_FAST__LOAD_FAST, previous_oparg); break; case LOAD_CONST: + assert(0 <= previous_oparg); instructions[i-1] = _Py_MAKECODEUNIT(LOAD_CONST__LOAD_FAST, previous_oparg); break; } break; case STORE_FAST: if (previous_opcode == STORE_FAST) { + assert(0 <= previous_oparg); instructions[i-1] = _Py_MAKECODEUNIT(STORE_FAST__STORE_FAST, previous_oparg); } break; case LOAD_CONST: if (previous_opcode == LOAD_FAST) { + assert(0 <= previous_oparg); instructions[i-1] = _Py_MAKECODEUNIT(LOAD_FAST__LOAD_CONST, previous_oparg); } break; @@ -448,7 +350,6 @@ optimize(SpecializedCacheOrInstruction *quickened, int len) previous_oparg = oparg; } } - assert(cache_offset+1 == get_cache_count(quickened)); } int @@ -462,16 +363,14 @@ _Py_Quicken(PyCodeObject *code) { code->co_warmup = QUICKENING_WARMUP_COLDEST; return 0; } - int entry_count = entries_needed(code->co_firstinstr, instr_count); - SpecializedCacheOrInstruction *quickened = allocate(entry_count, instr_count); + _Py_CODEUNIT *quickened = allocate(instr_count); if (quickened == NULL) { return -1; } - _Py_CODEUNIT *new_instructions = first_instruction(quickened); - memcpy(new_instructions, code->co_firstinstr, size); + memcpy(quickened, code->co_firstinstr, size); optimize(quickened, instr_count); code->co_quickened = quickened; - code->co_firstinstr = new_instructions; + code->co_firstinstr = quickened; return 0; } @@ -1516,9 +1415,8 @@ success: } static int -specialize_class_call( - PyObject *callable, _Py_CODEUNIT *instr, - int nargs, PyObject *kwnames, SpecializedCacheEntry *cache) +specialize_class_call(PyObject *callable, _Py_CODEUNIT *instr, int nargs, + PyObject *kwnames, int oparg) { assert(_Py_OPCODE(*instr) == PRECALL_ADAPTIVE); PyTypeObject *tp = _PyType_CAST(callable); @@ -1527,7 +1425,7 @@ specialize_class_call( return -1; } if (tp->tp_flags & Py_TPFLAGS_IMMUTABLETYPE) { - if (nargs == 1 && kwnames == NULL && cache->adaptive.original_oparg == 1) { + if (nargs == 1 && kwnames == NULL && oparg == 1) { if (tp == &PyUnicode_Type) { *instr = _Py_MAKECODEUNIT(PRECALL_NO_KW_STR_1, _Py_OPARG(*instr)); return 0; @@ -1577,28 +1475,15 @@ builtin_call_fail_kind(int ml_flags) } #endif -static PyMethodDescrObject *_list_append = NULL; - static int -specialize_method_descriptor( - PyMethodDescrObject *descr, _Py_CODEUNIT *instr, - int nargs, PyObject *kwnames, SpecializedCacheEntry *cache) +specialize_method_descriptor(PyMethodDescrObject *descr, _Py_CODEUNIT *instr, + int nargs, PyObject *kwnames, int oparg) { assert(_Py_OPCODE(*instr) == PRECALL_ADAPTIVE); if (kwnames) { SPECIALIZATION_FAIL(PRECALL, SPEC_FAIL_CALL_KWNAMES); return -1; } - if (_list_append == NULL) { - _list_append = (PyMethodDescrObject *)_PyType_Lookup(&PyList_Type, - &_Py_ID(append)); - } - assert(_list_append != NULL); - if (nargs == 2 && descr == _list_append && cache->adaptive.original_oparg == 1) { - cache[-1].obj.obj = (PyObject *)_list_append; - *instr = _Py_MAKECODEUNIT(PRECALL_NO_KW_LIST_APPEND, _Py_OPARG(*instr)); - return 0; - } switch (descr->d_method->ml_flags & (METH_VARARGS | METH_FASTCALL | METH_NOARGS | METH_O | @@ -1614,9 +1499,16 @@ specialize_method_descriptor( } case METH_O: { if (nargs != 2) { - SPECIALIZATION_FAIL(PRECALL, SPEC_FAIL_OUT_OF_RANGE); + SPECIALIZATION_FAIL(PRECALL, SPEC_FAIL_WRONG_NUMBER_ARGUMENTS); return -1; } + PyInterpreterState *interp = _PyInterpreterState_GET(); + PyObject *list_append = interp->callable_cache.list_append; + if ((PyObject *)descr == list_append && oparg == 1) { + *instr = _Py_MAKECODEUNIT(PRECALL_NO_KW_LIST_APPEND, + _Py_OPARG(*instr)); + return 0; + } *instr = _Py_MAKECODEUNIT(PRECALL_NO_KW_METHOD_DESCRIPTOR_O, _Py_OPARG(*instr)); return 0; @@ -1632,12 +1524,11 @@ specialize_method_descriptor( } static int -specialize_py_call( - PyFunctionObject *func, _Py_CODEUNIT *instr, - int nargs, PyObject *kwnames, SpecializedCacheEntry *cache) +specialize_py_call(PyFunctionObject *func, _Py_CODEUNIT *instr, int nargs, + PyObject *kwnames) { + _PyCallCache *cache = (_PyCallCache *)(instr + 1); assert(_Py_OPCODE(*instr) == CALL_ADAPTIVE); - _PyCallCache *cache1 = &cache[-1].call; PyCodeObject *code = (PyCodeObject *)func->func_code; int kind = function_kind(code); if (kwnames) { @@ -1649,10 +1540,6 @@ specialize_py_call( return -1; } int argcount = code->co_argcount; - if (argcount > 0xffff) { - SPECIALIZATION_FAIL(CALL, SPEC_FAIL_OUT_OF_RANGE); - return -1; - } int defcount = func->func_defaults == NULL ? 0 : (int)PyTuple_GET_SIZE(func->func_defaults); assert(defcount <= argcount); int min_args = argcount-defcount; @@ -1663,7 +1550,7 @@ specialize_py_call( assert(nargs <= argcount && nargs >= min_args); assert(min_args >= 0 && defcount >= 0); assert(defcount == 0 || func->func_defaults != NULL); - if (min_args > 0xffff || defcount > 0xffff) { + if (min_args > 0xffff) { SPECIALIZATION_FAIL(CALL, SPEC_FAIL_OUT_OF_RANGE); return -1; } @@ -1672,10 +1559,8 @@ specialize_py_call( SPECIALIZATION_FAIL(CALL, SPEC_FAIL_OUT_OF_VERSIONS); return -1; } - cache[0].adaptive.index = nargs; - cache1->func_version = version; - cache1->min_args = min_args; - cache1->defaults_len = defcount; + write_u32(cache->func_version, version); + cache->min_args = min_args; if (argcount == nargs) { *instr = _Py_MAKECODEUNIT(CALL_PY_EXACT_ARGS, _Py_OPARG(*instr)); } @@ -1687,10 +1572,9 @@ specialize_py_call( static int specialize_c_call(PyObject *callable, _Py_CODEUNIT *instr, int nargs, - PyObject *kwnames, SpecializedCacheEntry *cache, PyObject *builtins) + PyObject *kwnames) { assert(_Py_OPCODE(*instr) == PRECALL_ADAPTIVE); - _PyObjectCache *cache1 = &cache[-1].obj; if (PyCFunction_GET_FUNCTION(callable) == NULL) { return 1; } @@ -1707,9 +1591,8 @@ specialize_c_call(PyObject *callable, _Py_CODEUNIT *instr, int nargs, return 1; } /* len(o) */ - PyObject *builtin_len = PyDict_GetItemString(builtins, "len"); - if (callable == builtin_len) { - cache1->obj = builtin_len; // borrowed + PyInterpreterState *interp = _PyInterpreterState_GET(); + if (callable == interp->callable_cache.len) { *instr = _Py_MAKECODEUNIT(PRECALL_NO_KW_LEN, _Py_OPARG(*instr)); return 0; @@ -1725,10 +1608,8 @@ specialize_c_call(PyObject *callable, _Py_CODEUNIT *instr, int nargs, } if (nargs == 2) { /* isinstance(o1, o2) */ - PyObject *builtin_isinstance = PyDict_GetItemString( - builtins, "isinstance"); - if (callable == builtin_isinstance) { - cache1->obj = builtin_isinstance; // borrowed + PyInterpreterState *interp = _PyInterpreterState_GET(); + if (callable == interp->callable_cache.isinstance) { *instr = _Py_MAKECODEUNIT(PRECALL_NO_KW_ISINSTANCE, _Py_OPARG(*instr)); return 0; @@ -1793,44 +1674,44 @@ call_fail_kind(PyObject *callable) int -_Py_Specialize_Precall( - PyObject *callable, _Py_CODEUNIT *instr, - int nargs, PyObject *kwnames, - SpecializedCacheEntry *cache, PyObject *builtins) +_Py_Specialize_Precall(PyObject *callable, _Py_CODEUNIT *instr, int nargs, + PyObject *kwnames, int oparg) { - _PyAdaptiveEntry *cache0 = &cache->adaptive; + assert(_PyOpcode_InlineCacheEntries[PRECALL] == + INLINE_CACHE_ENTRIES_PRECALL); + _PyPrecallCache *cache = (_PyPrecallCache *)(instr + 1); int fail; if (PyCFunction_CheckExact(callable)) { - fail = specialize_c_call(callable, instr, nargs, kwnames, cache, builtins); + fail = specialize_c_call(callable, instr, nargs, kwnames); } else if (PyFunction_Check(callable)) { *instr = _Py_MAKECODEUNIT(PRECALL_PYFUNC, _Py_OPARG(*instr)); fail = 0; } else if (PyType_Check(callable)) { - fail = specialize_class_call(callable, instr, nargs, kwnames, cache); + fail = specialize_class_call(callable, instr, nargs, kwnames, oparg); } else if (Py_IS_TYPE(callable, &PyMethodDescr_Type)) { - fail = specialize_method_descriptor( - (PyMethodDescrObject *)callable, instr, nargs, kwnames, cache); + fail = specialize_method_descriptor((PyMethodDescrObject *)callable, + instr, nargs, kwnames, oparg); } else if (Py_TYPE(callable) == &PyMethod_Type) { *instr = _Py_MAKECODEUNIT(PRECALL_BOUND_METHOD, _Py_OPARG(*instr)); fail = 0; } else { - SPECIALIZATION_FAIL(CALL, call_fail_kind(callable)); + SPECIALIZATION_FAIL(PRECALL, call_fail_kind(callable)); fail = -1; } if (fail) { - STAT_INC(CALL, failure); + STAT_INC(PRECALL, failure); assert(!PyErr_Occurred()); - cache_backoff(cache0); + cache->counter = ADAPTIVE_CACHE_BACKOFF; } else { - STAT_INC(CALL, success); + STAT_INC(PRECALL, success); assert(!PyErr_Occurred()); - cache0->counter = initial_counter_value(); + cache->counter = initial_counter_value(); } return 0; } @@ -1840,15 +1721,15 @@ _Py_Specialize_Precall( - Specialize calling classes. */ int -_Py_Specialize_Call( - PyObject *callable, _Py_CODEUNIT *instr, - int nargs, PyObject *kwnames, - SpecializedCacheEntry *cache) +_Py_Specialize_Call(PyObject *callable, _Py_CODEUNIT *instr, int nargs, + PyObject *kwnames) { - _PyAdaptiveEntry *cache0 = &cache->adaptive; + assert(_PyOpcode_InlineCacheEntries[CALL] == INLINE_CACHE_ENTRIES_CALL); + _PyCallCache *cache = (_PyCallCache *)(instr + 1); int fail; if (PyFunction_Check(callable)) { - fail = specialize_py_call((PyFunctionObject *)callable, instr, nargs, kwnames, cache); + fail = specialize_py_call((PyFunctionObject *)callable, instr, nargs, + kwnames); } else { SPECIALIZATION_FAIL(CALL, call_fail_kind(callable)); @@ -1857,12 +1738,12 @@ _Py_Specialize_Call( if (fail) { STAT_INC(CALL, failure); assert(!PyErr_Occurred()); - cache_backoff(cache0); + cache->counter = ADAPTIVE_CACHE_BACKOFF; } else { STAT_INC(CALL, success); assert(!PyErr_Occurred()); - cache0->counter = initial_counter_value(); + cache->counter = initial_counter_value(); } return 0; } @@ -2238,10 +2119,4 @@ int return SPEC_FAIL_OTHER; } -int -_PySpecialization_ClassifyCallable(PyObject *callable) -{ - return call_fail_kind(callable); -} - #endif |