GH-91095: Specialize calls to normal Python classes. (GH-99331)

8 files changed, 346 insertions, 107 deletions

diff --git a/Python/bytecodes.c b/Python/bytecodes.c
index a7acff6..3ef8ed0 100644
--- a/Python/bytecodes.c
+++ b/Python/bytecodes.c
@@ -2685,6 +2685,7 @@ dummy_func(
             CALL_METHOD_DESCRIPTOR_FAST_WITH_KEYWORDS,
             CALL_NO_KW_METHOD_DESCRIPTOR_NOARGS,
             CALL_NO_KW_METHOD_DESCRIPTOR_FAST,
+            CALL_NO_KW_ALLOC_AND_ENTER_INIT,
         };
 
         // On entry, the stack is either
@@ -2900,6 +2901,68 @@ dummy_func(
             CHECK_EVAL_BREAKER();
         }
 
+        inst(CALL_NO_KW_ALLOC_AND_ENTER_INIT, (unused/1, unused/2, null, callable, args[oparg] -- unused)) {
+            /* This instruction does the following:
+             * 1. Creates the object (by calling ``object.__new__``)
+             * 2. Pushes a shim frame to the frame stack (to cleanup after ``__init__``)
+             * 3. Pushes the frame for ``__init__`` to the frame stack
+             * */
+            assert(kwnames == NULL);
+            _PyCallCache *cache = (_PyCallCache *)next_instr;
+            DEOPT_IF(null != NULL, CALL);
+            DEOPT_IF(!PyType_Check(callable), CALL);
+            PyTypeObject *tp = (PyTypeObject *)callable;
+            DEOPT_IF(tp->tp_version_tag != read_u32(cache->func_version), CALL);
+            PyHeapTypeObject *cls = (PyHeapTypeObject *)callable;
+            PyFunctionObject *init = (PyFunctionObject *)cls->_spec_cache.init;
+            PyCodeObject *code = (PyCodeObject *)init->func_code;
+            DEOPT_IF(code->co_argcount != oparg+1, CALL);
+            DEOPT_IF(!_PyThreadState_HasStackSpace(tstate, code->co_framesize + _Py_InitCleanup.co_framesize), CALL);
+            STAT_INC(CALL, hit);
+            PyObject *self = _PyType_NewManagedObject(tp);
+            if (self == NULL) {
+                goto error;
+            }
+            Py_DECREF(tp);
+            if (_Py_EnterRecursivePy(tstate)) {
+                goto exit_unwind;
+            }
+            _PyInterpreterFrame *shim = _PyFrame_PushTrampolineUnchecked(
+                tstate, (PyCodeObject *)&_Py_InitCleanup, 1, 0);
+            assert(_PyCode_CODE((PyCodeObject *)shim->f_executable)[1].op.code == EXIT_INIT_CHECK);
+            /* Push self onto stack of shim */
+            Py_INCREF(self);
+            shim->localsplus[0] = self;
+            Py_INCREF(init);
+            _PyInterpreterFrame *init_frame = _PyFrame_PushUnchecked(tstate, init, oparg+1);
+            /* Copy self followed by args to __init__ frame */
+            init_frame->localsplus[0] = self;
+            for (int i = 0; i < oparg; i++) {
+                init_frame->localsplus[i+1] = args[i];
+            }
+            SKIP_OVER(INLINE_CACHE_ENTRIES_CALL);
+            frame->prev_instr = next_instr - 1;
+            frame->return_offset = 0;
+            STACK_SHRINK(oparg+2);
+            _PyFrame_SetStackPointer(frame, stack_pointer);
+            /* Link frames */
+            init_frame->previous = shim;
+            shim->previous = frame;
+            frame = cframe.current_frame = init_frame;
+            CALL_STAT_INC(inlined_py_calls);
+            goto start_frame;
+        }
+
+        inst(EXIT_INIT_CHECK, (should_be_none -- )) {
+            assert(STACK_LEVEL() == 2);
+            if (should_be_none != Py_None) {
+                PyErr_Format(PyExc_TypeError,
+                    "__init__() should return None, not '%.200s'",
+                    Py_TYPE(should_be_none)->tp_name);
+                goto error;
+            }
+        }
+
         inst(CALL_BUILTIN_CLASS, (unused/1, unused/2, method, callable, args[oparg] -- res)) {
             int is_meth = method != NULL;
             int total_args = oparg;
diff --git a/Python/ceval.c b/Python/ceval.c
index b1ac4de..5310701 100644
--- a/Python/ceval.c
+++ b/Python/ceval.c
@@ -641,6 +641,8 @@ static const _Py_CODEUNIT _Py_INTERPRETER_TRAMPOLINE_INSTRUCTIONS[] = {
     { .op.code = RESUME, .op.arg = 0 }
 };
 
+extern const struct _PyCode_DEF(8) _Py_InitCleanup;
+
 /* Disable unused label warnings.  They are handy for debugging, even
    if computed gotos aren't used. */
 
@@ -746,7 +748,7 @@ resume_frame:
 
 #ifdef LLTRACE
     {
-        if (frame != &entry_frame) {
+        if (frame != &entry_frame && GLOBALS()) {
             int r = PyDict_Contains(GLOBALS(), &_Py_ID(__lltrace__));
             if (r < 0) {
                 goto exit_unwind;
diff --git a/Python/compile.c b/Python/compile.c
index 69468f7..5a05605 100644
--- a/Python/compile.c
+++ b/Python/compile.c
@@ -825,6 +825,9 @@ stack_effect(int opcode, int oparg, int jump)
         case JUMP_NO_INTERRUPT:
             return 0;
 
+        case EXIT_INIT_CHECK:
+            return -1;
+
         /* Exception handling pseudo-instructions */
         case SETUP_FINALLY:
             /* 0 in the normal flow.
diff --git a/Python/generated_cases.c.h b/Python/generated_cases.c.h
index 717b29e..11ca535 100644
--- a/Python/generated_cases.c.h
+++ b/Python/generated_cases.c.h
@@ -3765,7 +3765,7 @@
             PyObject *callable = stack_pointer[-(1 + oparg)];
             PyObject *method = stack_pointer[-(2 + oparg)];
             PyObject *res;
-            #line 2698 "Python/bytecodes.c"
+            #line 2699 "Python/bytecodes.c"
             int is_meth = method != NULL;
             int total_args = oparg;
             if (is_meth) {
@@ -3859,7 +3859,7 @@
         TARGET(CALL_BOUND_METHOD_EXACT_ARGS) {
             PyObject *callable = stack_pointer[-(1 + oparg)];
             PyObject *method = stack_pointer[-(2 + oparg)];
-            #line 2786 "Python/bytecodes.c"
+            #line 2787 "Python/bytecodes.c"
             DEOPT_IF(method != NULL, CALL);
             DEOPT_IF(Py_TYPE(callable) != &PyMethod_Type, CALL);
             STAT_INC(CALL, hit);
@@ -3878,7 +3878,7 @@
             PyObject *callable = stack_pointer[-(1 + oparg)];
             PyObject *method = stack_pointer[-(2 + oparg)];
             uint32_t func_version = read_u32(&next_instr[1].cache);
-            #line 2798 "Python/bytecodes.c"
+            #line 2799 "Python/bytecodes.c"
             assert(kwnames == NULL);
             DEOPT_IF(tstate->interp->eval_frame, CALL);
             int is_meth = method != NULL;
@@ -3912,7 +3912,7 @@
             PyObject *callable = stack_pointer[-(1 + oparg)];
             PyObject *method = stack_pointer[-(2 + oparg)];
             uint32_t func_version = read_u32(&next_instr[1].cache);
-            #line 2826 "Python/bytecodes.c"
+            #line 2827 "Python/bytecodes.c"
             assert(kwnames == NULL);
             DEOPT_IF(tstate->interp->eval_frame, CALL);
             int is_meth = method != NULL;
@@ -3956,7 +3956,7 @@
             PyObject *callable = stack_pointer[-(1 + oparg)];
             PyObject *null = stack_pointer[-(2 + oparg)];
             PyObject *res;
-            #line 2864 "Python/bytecodes.c"
+            #line 2865 "Python/bytecodes.c"
             assert(kwnames == NULL);
             assert(oparg == 1);
             DEOPT_IF(null != NULL, CALL);
@@ -3979,7 +3979,7 @@
             PyObject *callable = stack_pointer[-(1 + oparg)];
             PyObject *null = stack_pointer[-(2 + oparg)];
             PyObject *res;
-            #line 2876 "Python/bytecodes.c"
+            #line 2877 "Python/bytecodes.c"
             assert(kwnames == NULL);
             assert(oparg == 1);
             DEOPT_IF(null != NULL, CALL);
@@ -4004,7 +4004,7 @@
             PyObject *callable = stack_pointer[-(1 + oparg)];
             PyObject *null = stack_pointer[-(2 + oparg)];
             PyObject *res;
-            #line 2890 "Python/bytecodes.c"
+            #line 2891 "Python/bytecodes.c"
             assert(kwnames == NULL);
             assert(oparg == 1);
             DEOPT_IF(null != NULL, CALL);
@@ -4024,12 +4024,84 @@
             DISPATCH();
         }
 
+        TARGET(CALL_NO_KW_ALLOC_AND_ENTER_INIT) {
+            PyObject **args = (stack_pointer - oparg);
+            PyObject *callable = stack_pointer[-(1 + oparg)];
+            PyObject *null = stack_pointer[-(2 + oparg)];
+            #line 2905 "Python/bytecodes.c"
+            /* This instruction does the following:
+             * 1. Creates the object (by calling ``object.__new__``)
+             * 2. Pushes a shim frame to the frame stack (to cleanup after ``__init__``)
+             * 3. Pushes the frame for ``__init__`` to the frame stack
+             * */
+            assert(kwnames == NULL);
+            _PyCallCache *cache = (_PyCallCache *)next_instr;
+            DEOPT_IF(null != NULL, CALL);
+            DEOPT_IF(!PyType_Check(callable), CALL);
+            PyTypeObject *tp = (PyTypeObject *)callable;
+            DEOPT_IF(tp->tp_version_tag != read_u32(cache->func_version), CALL);
+            PyHeapTypeObject *cls = (PyHeapTypeObject *)callable;
+            PyFunctionObject *init = (PyFunctionObject *)cls->_spec_cache.init;
+            PyCodeObject *code = (PyCodeObject *)init->func_code;
+            DEOPT_IF(code->co_argcount != oparg+1, CALL);
+            DEOPT_IF(!_PyThreadState_HasStackSpace(tstate, code->co_framesize + _Py_InitCleanup.co_framesize), CALL);
+            STAT_INC(CALL, hit);
+            PyObject *self = _PyType_NewManagedObject(tp);
+            if (self == NULL) {
+                goto error;
+            }
+            Py_DECREF(tp);
+            if (_Py_EnterRecursivePy(tstate)) {
+                goto exit_unwind;
+            }
+            _PyInterpreterFrame *shim = _PyFrame_PushTrampolineUnchecked(
+                tstate, (PyCodeObject *)&_Py_InitCleanup, 1, 0);
+            assert(_PyCode_CODE((PyCodeObject *)shim->f_executable)[1].op.code == EXIT_INIT_CHECK);
+            /* Push self onto stack of shim */
+            Py_INCREF(self);
+            shim->localsplus[0] = self;
+            Py_INCREF(init);
+            _PyInterpreterFrame *init_frame = _PyFrame_PushUnchecked(tstate, init, oparg+1);
+            /* Copy self followed by args to __init__ frame */
+            init_frame->localsplus[0] = self;
+            for (int i = 0; i < oparg; i++) {
+                init_frame->localsplus[i+1] = args[i];
+            }
+            SKIP_OVER(INLINE_CACHE_ENTRIES_CALL);
+            frame->prev_instr = next_instr - 1;
+            frame->return_offset = 0;
+            STACK_SHRINK(oparg+2);
+            _PyFrame_SetStackPointer(frame, stack_pointer);
+            /* Link frames */
+            init_frame->previous = shim;
+            shim->previous = frame;
+            frame = cframe.current_frame = init_frame;
+            CALL_STAT_INC(inlined_py_calls);
+            goto start_frame;
+            #line 4082 "Python/generated_cases.c.h"
+        }
+
+        TARGET(EXIT_INIT_CHECK) {
+            PyObject *should_be_none = stack_pointer[-1];
+            #line 2957 "Python/bytecodes.c"
+            assert(STACK_LEVEL() == 2);
+            if (should_be_none != Py_None) {
+                PyErr_Format(PyExc_TypeError,
+                    "__init__() should return None, not '%.200s'",
+                    Py_TYPE(should_be_none)->tp_name);
+                goto error;
+            }
+            #line 4095 "Python/generated_cases.c.h"
+            STACK_SHRINK(1);
+            DISPATCH();
+        }
+
         TARGET(CALL_BUILTIN_CLASS) {
             PyObject **args = (stack_pointer - oparg);
             PyObject *callable = stack_pointer[-(1 + oparg)];
             PyObject *method = stack_pointer[-(2 + oparg)];
             PyObject *res;
-            #line 2904 "Python/bytecodes.c"
+            #line 2967 "Python/bytecodes.c"
             int is_meth = method != NULL;
             int total_args = oparg;
             if (is_meth) {
@@ -4051,7 +4123,7 @@
             }
             Py_DECREF(tp);
             if (res == NULL) { STACK_SHRINK(oparg); goto pop_2_error; }
-            #line 4055 "Python/generated_cases.c.h"
+            #line 4127 "Python/generated_cases.c.h"
             STACK_SHRINK(oparg);
             STACK_SHRINK(1);
             stack_pointer[-1] = res;
@@ -4065,7 +4137,7 @@
             PyObject *callable = stack_pointer[-(1 + oparg)];
             PyObject *method = stack_pointer[-(2 + oparg)];
             PyObject *res;
-            #line 2929 "Python/bytecodes.c"
+            #line 2992 "Python/bytecodes.c"
             /* Builtin METH_O functions */
             assert(kwnames == NULL);
             int is_meth = method != NULL;
@@ -4093,7 +4165,7 @@
             Py_DECREF(arg);
             Py_DECREF(callable);
             if (res == NULL) { STACK_SHRINK(oparg); goto pop_2_error; }
-            #line 4097 "Python/generated_cases.c.h"
+            #line 4169 "Python/generated_cases.c.h"
             STACK_SHRINK(oparg);
             STACK_SHRINK(1);
             stack_pointer[-1] = res;
@@ -4107,7 +4179,7 @@
             PyObject *callable = stack_pointer[-(1 + oparg)];
             PyObject *method = stack_pointer[-(2 + oparg)];
             PyObject *res;
-            #line 2960 "Python/bytecodes.c"
+            #line 3023 "Python/bytecodes.c"
             /* Builtin METH_FASTCALL functions, without keywords */
             assert(kwnames == NULL);
             int is_meth = method != NULL;
@@ -4139,7 +4211,7 @@
                    'invalid'). In those cases an exception is set, so we must
                    handle it.
                 */
-            #line 4143 "Python/generated_cases.c.h"
+            #line 4215 "Python/generated_cases.c.h"
             STACK_SHRINK(oparg);
             STACK_SHRINK(1);
             stack_pointer[-1] = res;
@@ -4153,7 +4225,7 @@
             PyObject *callable = stack_pointer[-(1 + oparg)];
             PyObject *method = stack_pointer[-(2 + oparg)];
             PyObject *res;
-            #line 2995 "Python/bytecodes.c"
+            #line 3058 "Python/bytecodes.c"
             /* Builtin METH_FASTCALL | METH_KEYWORDS functions */
             int is_meth = method != NULL;
             int total_args = oparg;
@@ -4185,7 +4257,7 @@
             }
             Py_DECREF(callable);
             if (res == NULL) { STACK_SHRINK(oparg); goto pop_2_error; }
-            #line 4189 "Python/generated_cases.c.h"
+            #line 4261 "Python/generated_cases.c.h"
             STACK_SHRINK(oparg);
             STACK_SHRINK(1);
             stack_pointer[-1] = res;
@@ -4199,7 +4271,7 @@
             PyObject *callable = stack_pointer[-(1 + oparg)];
             PyObject *method = stack_pointer[-(2 + oparg)];
             PyObject *res;
-            #line 3030 "Python/bytecodes.c"
+            #line 3093 "Python/bytecodes.c"
             assert(kwnames == NULL);
             /* len(o) */
             int is_meth = method != NULL;
@@ -4224,7 +4296,7 @@
             Py_DECREF(callable);
             Py_DECREF(arg);
             if (res == NULL) { STACK_SHRINK(oparg); goto pop_2_error; }
-            #line 4228 "Python/generated_cases.c.h"
+            #line 4300 "Python/generated_cases.c.h"
             STACK_SHRINK(oparg);
             STACK_SHRINK(1);
             stack_pointer[-1] = res;
@@ -4237,7 +4309,7 @@
             PyObject *callable = stack_pointer[-(1 + oparg)];
             PyObject *method = stack_pointer[-(2 + oparg)];
             PyObject *res;
-            #line 3057 "Python/bytecodes.c"
+            #line 3120 "Python/bytecodes.c"
             assert(kwnames == NULL);
             /* isinstance(o, o2) */
             int is_meth = method != NULL;
@@ -4264,7 +4336,7 @@
             Py_DECREF(cls);
             Py_DECREF(callable);
             if (res == NULL) { STACK_SHRINK(oparg); goto pop_2_error; }
-            #line 4268 "Python/generated_cases.c.h"
+            #line 4340 "Python/generated_cases.c.h"
             STACK_SHRINK(oparg);
             STACK_SHRINK(1);
             stack_pointer[-1] = res;
@@ -4276,7 +4348,7 @@
             PyObject **args = (stack_pointer - oparg);
             PyObject *self = stack_pointer[-(1 + oparg)];
             PyObject *method = stack_pointer[-(2 + oparg)];
-            #line 3087 "Python/bytecodes.c"
+            #line 3150 "Python/bytecodes.c"
             assert(kwnames == NULL);
             assert(oparg == 1);
             assert(method != NULL);
@@ -4294,14 +4366,14 @@
             SKIP_OVER(INLINE_CACHE_ENTRIES_CALL + 1);
             assert(next_instr[-1].op.code == POP_TOP);
             DISPATCH();
-            #line 4298 "Python/generated_cases.c.h"
+            #line 4370 "Python/generated_cases.c.h"
         }
 
         TARGET(CALL_NO_KW_METHOD_DESCRIPTOR_O) {
             PyObject **args = (stack_pointer - oparg);
             PyObject *method = stack_pointer[-(2 + oparg)];
             PyObject *res;
-            #line 3107 "Python/bytecodes.c"
+            #line 3170 "Python/bytecodes.c"
             assert(kwnames == NULL);
             int is_meth = method != NULL;
             int total_args = oparg;
@@ -4332,7 +4404,7 @@
             Py_DECREF(arg);
             Py_DECREF(callable);
             if (res == NULL) { STACK_SHRINK(oparg); goto pop_2_error; }
-            #line 4336 "Python/generated_cases.c.h"
+            #line 4408 "Python/generated_cases.c.h"
             STACK_SHRINK(oparg);
             STACK_SHRINK(1);
             stack_pointer[-1] = res;
@@ -4345,7 +4417,7 @@
             PyObject **args = (stack_pointer - oparg);
             PyObject *method = stack_pointer[-(2 + oparg)];
             PyObject *res;
-            #line 3141 "Python/bytecodes.c"
+            #line 3204 "Python/bytecodes.c"
             int is_meth = method != NULL;
             int total_args = oparg;
             if (is_meth) {
@@ -4374,7 +4446,7 @@
             }
             Py_DECREF(callable);
             if (res == NULL) { STACK_SHRINK(oparg); goto pop_2_error; }
-            #line 4378 "Python/generated_cases.c.h"
+            #line 4450 "Python/generated_cases.c.h"
             STACK_SHRINK(oparg);
             STACK_SHRINK(1);
             stack_pointer[-1] = res;
@@ -4387,7 +4459,7 @@
             PyObject **args = (stack_pointer - oparg);
             PyObject *method = stack_pointer[-(2 + oparg)];
             PyObject *res;
-            #line 3173 "Python/bytecodes.c"
+            #line 3236 "Python/bytecodes.c"
             assert(kwnames == NULL);
             assert(oparg == 0 || oparg == 1);
             int is_meth = method != NULL;
@@ -4416,7 +4488,7 @@
             Py_DECREF(self);
             Py_DECREF(callable);
             if (res == NULL) { STACK_SHRINK(oparg); goto pop_2_error; }
-            #line 4420 "Python/generated_cases.c.h"
+            #line 4492 "Python/generated_cases.c.h"
             STACK_SHRINK(oparg);
             STACK_SHRINK(1);
             stack_pointer[-1] = res;
@@ -4429,7 +4501,7 @@
             PyObject **args = (stack_pointer - oparg);
             PyObject *method = stack_pointer[-(2 + oparg)];
             PyObject *res;
-            #line 3205 "Python/bytecodes.c"
+            #line 3268 "Python/bytecodes.c"
             assert(kwnames == NULL);
             int is_meth = method != NULL;
             int total_args = oparg;
@@ -4457,7 +4529,7 @@
             }
             Py_DECREF(callable);
             if (res == NULL) { STACK_SHRINK(oparg); goto pop_2_error; }
-            #line 4461 "Python/generated_cases.c.h"
+            #line 4533 "Python/generated_cases.c.h"
             STACK_SHRINK(oparg);
             STACK_SHRINK(1);
             stack_pointer[-1] = res;
@@ -4467,9 +4539,9 @@
         }
 
         TARGET(INSTRUMENTED_CALL_FUNCTION_EX) {
-            #line 3236 "Python/bytecodes.c"
+            #line 3299 "Python/bytecodes.c"
             GO_TO_INSTRUCTION(CALL_FUNCTION_EX);
-            #line 4473 "Python/generated_cases.c.h"
+            #line 4545 "Python/generated_cases.c.h"
         }
 
         TARGET(CALL_FUNCTION_EX) {
@@ -4478,7 +4550,7 @@
             PyObject *callargs = stack_pointer[-(1 + ((oparg & 1) ? 1 : 0))];
             PyObject *func = stack_pointer[-(2 + ((oparg & 1) ? 1 : 0))];
             PyObject *result;
-            #line 3240 "Python/bytecodes.c"
+            #line 3303 "Python/bytecodes.c"
             // DICT_MERGE is called before this opcode if there are kwargs.
             // It converts all dict subtypes in kwargs into regular dicts.
             assert(kwargs == NULL || PyDict_CheckExact(kwargs));
@@ -4540,14 +4612,14 @@
                 }
                 result = PyObject_Call(func, callargs, kwargs);
             }
-            #line 4544 "Python/generated_cases.c.h"
+            #line 4616 "Python/generated_cases.c.h"
             Py_DECREF(func);
             Py_DECREF(callargs);
             Py_XDECREF(kwargs);
-            #line 3302 "Python/bytecodes.c"
+            #line 3365 "Python/bytecodes.c"
             assert(PEEK(3 + (oparg & 1)) == NULL);
             if (result == NULL) { STACK_SHRINK(((oparg & 1) ? 1 : 0)); goto pop_3_error; }
-            #line 4551 "Python/generated_cases.c.h"
+            #line 4623 "Python/generated_cases.c.h"
             STACK_SHRINK(((oparg & 1) ? 1 : 0));
             STACK_SHRINK(2);
             stack_pointer[-1] = result;
@@ -4558,7 +4630,7 @@
         TARGET(MAKE_FUNCTION) {
             PyObject *codeobj = stack_pointer[-1];
             PyObject *func;
-            #line 3308 "Python/bytecodes.c"
+            #line 3371 "Python/bytecodes.c"
 
             PyFunctionObject *func_obj = (PyFunctionObject *)
                 PyFunction_New(codeobj, GLOBALS());
@@ -4570,7 +4642,7 @@
 
             func_obj->func_version = ((PyCodeObject *)codeobj)->co_version;
             func = (PyObject *)func_obj;
-            #line 4574 "Python/generated_cases.c.h"
+            #line 4646 "Python/generated_cases.c.h"
             stack_pointer[-1] = func;
             DISPATCH();
         }
@@ -4578,7 +4650,7 @@
         TARGET(SET_FUNCTION_ATTRIBUTE) {
             PyObject *func = stack_pointer[-1];
             PyObject *attr = stack_pointer[-2];
-            #line 3322 "Python/bytecodes.c"
+            #line 3385 "Python/bytecodes.c"
             assert(PyFunction_Check(func));
             PyFunctionObject *func_obj = (PyFunctionObject *)func;
             switch(oparg) {
@@ -4603,14 +4675,14 @@
                 default:
                     Py_UNREACHABLE();
             }
-            #line 4607 "Python/generated_cases.c.h"
+            #line 4679 "Python/generated_cases.c.h"
             STACK_SHRINK(1);
             stack_pointer[-1] = func;
             DISPATCH();
         }
 
         TARGET(RETURN_GENERATOR) {
-            #line 3349 "Python/bytecodes.c"
+            #line 3412 "Python/bytecodes.c"
             assert(PyFunction_Check(frame->f_funcobj));
             PyFunctionObject *func = (PyFunctionObject *)frame->f_funcobj;
             PyGenObject *gen = (PyGenObject *)_Py_MakeCoro(func);
@@ -4631,7 +4703,7 @@
             frame = cframe.current_frame = prev;
             _PyFrame_StackPush(frame, (PyObject *)gen);
             goto resume_frame;
-            #line 4635 "Python/generated_cases.c.h"
+            #line 4707 "Python/generated_cases.c.h"
         }
 
         TARGET(BUILD_SLICE) {
@@ -4639,15 +4711,15 @@
             PyObject *stop = stack_pointer[-(1 + ((oparg == 3) ? 1 : 0))];
             PyObject *start = stack_pointer[-(2 + ((oparg == 3) ? 1 : 0))];
             PyObject *slice;
-            #line 3372 "Python/bytecodes.c"
+            #line 3435 "Python/bytecodes.c"
             slice = PySlice_New(start, stop, step);
-            #line 4645 "Python/generated_cases.c.h"
+            #line 4717 "Python/generated_cases.c.h"
             Py_DECREF(start);
             Py_DECREF(stop);
             Py_XDECREF(step);
-            #line 3374 "Python/bytecodes.c"
+            #line 3437 "Python/bytecodes.c"
             if (slice == NULL) { STACK_SHRINK(((oparg == 3) ? 1 : 0)); goto pop_2_error; }
-            #line 4651 "Python/generated_cases.c.h"
+            #line 4723 "Python/generated_cases.c.h"
             STACK_SHRINK(((oparg == 3) ? 1 : 0));
             STACK_SHRINK(1);
             stack_pointer[-1] = slice;
@@ -4657,14 +4729,14 @@
         TARGET(CONVERT_VALUE) {
             PyObject *value = stack_pointer[-1];
             PyObject *result;
-            #line 3378 "Python/bytecodes.c"
+            #line 3441 "Python/bytecodes.c"
             convertion_func_ptr  conv_fn;
             assert(oparg >= FVC_STR && oparg <= FVC_ASCII);
             conv_fn = CONVERSION_FUNCTIONS[oparg];
             result = conv_fn(value);
             Py_DECREF(value);
             if (result == NULL) goto pop_1_error;
-            #line 4668 "Python/generated_cases.c.h"
+            #line 4740 "Python/generated_cases.c.h"
             stack_pointer[-1] = result;
             DISPATCH();
         }
@@ -4672,7 +4744,7 @@
         TARGET(FORMAT_SIMPLE) {
             PyObject *value = stack_pointer[-1];
             PyObject *res;
-            #line 3387 "Python/bytecodes.c"
+            #line 3450 "Python/bytecodes.c"
             /* If value is a unicode object, then we know the result
              * of format(value) is value itself. */
             if (!PyUnicode_CheckExact(value)) {
@@ -4683,7 +4755,7 @@
             else {
                 res = value;
             }
-            #line 4687 "Python/generated_cases.c.h"
+            #line 4759 "Python/generated_cases.c.h"
             stack_pointer[-1] = res;
             DISPATCH();
         }
@@ -4692,12 +4764,12 @@
             PyObject *fmt_spec = stack_pointer[-1];
             PyObject *value = stack_pointer[-2];
             PyObject *res;
-            #line 3400 "Python/bytecodes.c"
+            #line 3463 "Python/bytecodes.c"
             res = PyObject_Format(value, fmt_spec);
             Py_DECREF(value);
             Py_DECREF(fmt_spec);
             if (res == NULL) goto pop_2_error;
-            #line 4701 "Python/generated_cases.c.h"
+            #line 4773 "Python/generated_cases.c.h"
             STACK_SHRINK(1);
             stack_pointer[-1] = res;
             DISPATCH();
@@ -4706,10 +4778,10 @@
         TARGET(COPY) {
             PyObject *bottom = stack_pointer[-(1 + (oparg-1))];
             PyObject *top;
-            #line 3407 "Python/bytecodes.c"
+            #line 3470 "Python/bytecodes.c"
             assert(oparg > 0);
             top = Py_NewRef(bottom);
-            #line 4713 "Python/generated_cases.c.h"
+            #line 4785 "Python/generated_cases.c.h"
             STACK_GROW(1);
             stack_pointer[-1] = top;
             DISPATCH();
@@ -4721,7 +4793,7 @@
             PyObject *rhs = stack_pointer[-1];
             PyObject *lhs = stack_pointer[-2];
             PyObject *res;
-            #line 3412 "Python/bytecodes.c"
+            #line 3475 "Python/bytecodes.c"
             #if ENABLE_SPECIALIZATION
             _PyBinaryOpCache *cache = (_PyBinaryOpCache *)next_instr;
             if (ADAPTIVE_COUNTER_IS_ZERO(cache->counter)) {
@@ -4736,12 +4808,12 @@
             assert((unsigned)oparg < Py_ARRAY_LENGTH(binary_ops));
             assert(binary_ops[oparg]);
             res = binary_ops[oparg](lhs, rhs);
-            #line 4740 "Python/generated_cases.c.h"
+            #line 4812 "Python/generated_cases.c.h"
             Py_DECREF(lhs);
             Py_DECREF(rhs);
-            #line 3427 "Python/bytecodes.c"
+            #line 3490 "Python/bytecodes.c"
             if (res == NULL) goto pop_2_error;
-            #line 4745 "Python/generated_cases.c.h"
+            #line 4817 "Python/generated_cases.c.h"
             STACK_SHRINK(1);
             stack_pointer[-1] = res;
             next_instr += 1;
@@ -4751,16 +4823,16 @@
         TARGET(SWAP) {
             PyObject *top = stack_pointer[-1];
             PyObject *bottom = stack_pointer[-(2 + (oparg-2))];
-            #line 3432 "Python/bytecodes.c"
+            #line 3495 "Python/bytecodes.c"
             assert(oparg >= 2);
-            #line 4757 "Python/generated_cases.c.h"
+            #line 4829 "Python/generated_cases.c.h"
             stack_pointer[-1] = bottom;
             stack_pointer[-(2 + (oparg-2))] = top;
             DISPATCH();
         }
 
         TARGET(INSTRUMENTED_INSTRUCTION) {
-            #line 3436 "Python/bytecodes.c"
+            #line 3499 "Python/bytecodes.c"
             int next_opcode = _Py_call_instrumentation_instruction(
                 tstate, frame, next_instr-1);
             if (next_opcode < 0) goto error;
@@ -4772,26 +4844,26 @@
             assert(next_opcode > 0 && next_opcode < 256);
             opcode = next_opcode;
             DISPATCH_GOTO();
-            #line 4776 "Python/generated_cases.c.h"
+            #line 4848 "Python/generated_cases.c.h"
         }
 
         TARGET(INSTRUMENTED_JUMP_FORWARD) {
-            #line 3450 "Python/bytecodes.c"
+            #line 3513 "Python/bytecodes.c"
             INSTRUMENTED_JUMP(next_instr-1, next_instr+oparg, PY_MONITORING_EVENT_JUMP);
-            #line 4782 "Python/generated_cases.c.h"
+            #line 4854 "Python/generated_cases.c.h"
             DISPATCH();
         }
 
         TARGET(INSTRUMENTED_JUMP_BACKWARD) {
-            #line 3454 "Python/bytecodes.c"
+            #line 3517 "Python/bytecodes.c"
             INSTRUMENTED_JUMP(next_instr-1, next_instr+1-oparg, PY_MONITORING_EVENT_JUMP);
-            #line 4789 "Python/generated_cases.c.h"
+            #line 4861 "Python/generated_cases.c.h"
             CHECK_EVAL_BREAKER();
             DISPATCH();
         }
 
         TARGET(INSTRUMENTED_POP_JUMP_IF_TRUE) {
-            #line 3459 "Python/bytecodes.c"
+            #line 3522 "Python/bytecodes.c"
             PyObject *cond = POP();
             int err = PyObject_IsTrue(cond);
             Py_DECREF(cond);
@@ -4800,12 +4872,12 @@
             assert(err == 0 || err == 1);
             int offset = err*oparg;
             INSTRUMENTED_JUMP(here, next_instr + offset, PY_MONITORING_EVENT_BRANCH);
-            #line 4804 "Python/generated_cases.c.h"
+            #line 4876 "Python/generated_cases.c.h"
             DISPATCH();
         }
 
         TARGET(INSTRUMENTED_POP_JUMP_IF_FALSE) {
-            #line 3470 "Python/bytecodes.c"
+            #line 3533 "Python/bytecodes.c"
             PyObject *cond = POP();
             int err = PyObject_IsTrue(cond);
             Py_DECREF(cond);
@@ -4814,12 +4886,12 @@
             assert(err == 0 || err == 1);
             int offset = (1-err)*oparg;
             INSTRUMENTED_JUMP(here, next_instr + offset, PY_MONITORING_EVENT_BRANCH);
-            #line 4818 "Python/generated_cases.c.h"
+            #line 4890 "Python/generated_cases.c.h"
             DISPATCH();
         }
 
         TARGET(INSTRUMENTED_POP_JUMP_IF_NONE) {
-            #line 3481 "Python/bytecodes.c"
+            #line 3544 "Python/bytecodes.c"
             PyObject *value = POP();
             _Py_CODEUNIT *here = next_instr-1;
             int offset;
@@ -4831,12 +4903,12 @@
                 offset = 0;
             }
             INSTRUMENTED_JUMP(here, next_instr + offset, PY_MONITORING_EVENT_BRANCH);
-            #line 4835 "Python/generated_cases.c.h"
+            #line 4907 "Python/generated_cases.c.h"
             DISPATCH();
         }
 
         TARGET(INSTRUMENTED_POP_JUMP_IF_NOT_NONE) {
-            #line 3495 "Python/bytecodes.c"
+            #line 3558 "Python/bytecodes.c"
             PyObject *value = POP();
             _Py_CODEUNIT *here = next_instr-1;
             int offset;
@@ -4848,30 +4920,30 @@
                  offset = oparg;
             }
             INSTRUMENTED_JUMP(here, next_instr + offset, PY_MONITORING_EVENT_BRANCH);
-            #line 4852 "Python/generated_cases.c.h"
+            #line 4924 "Python/generated_cases.c.h"
             DISPATCH();
         }
 
         TARGET(EXTENDED_ARG) {
-            #line 3509 "Python/bytecodes.c"
+            #line 3572 "Python/bytecodes.c"
             assert(oparg);
             opcode = next_instr->op.code;
             oparg = oparg << 8 | next_instr->op.arg;
             PRE_DISPATCH_GOTO();
             DISPATCH_GOTO();
-            #line 4863 "Python/generated_cases.c.h"
+            #line 4935 "Python/generated_cases.c.h"
         }
 
         TARGET(CACHE) {
-            #line 3517 "Python/bytecodes.c"
+            #line 3580 "Python/bytecodes.c"
             assert(0 && "Executing a cache.");
             Py_UNREACHABLE();
-            #line 4870 "Python/generated_cases.c.h"
+            #line 4942 "Python/generated_cases.c.h"
         }
 
         TARGET(RESERVED) {
-            #line 3522 "Python/bytecodes.c"
+            #line 3585 "Python/bytecodes.c"
             assert(0 && "Executing RESERVED instruction.");
             Py_UNREACHABLE();
-            #line 4877 "Python/generated_cases.c.h"
+            #line 4949 "Python/generated_cases.c.h"
         }
diff --git a/Python/instrumentation.c b/Python/instrumentation.c
index fcaccc0..524a82c 100644
--- a/Python/instrumentation.c
+++ b/Python/instrumentation.c
@@ -1526,7 +1526,7 @@ _Py_Instrument(PyCodeObject *code, PyInterpreterState *interp)
         return 0;
     }
     /* Insert instrumentation */
-    for (int i = 0; i < code_len; i+= _PyInstruction_GetLength(code, i)) {
+    for (int i = code->_co_firsttraceable; i < code_len; i+= _PyInstruction_GetLength(code, i)) {
         _Py_CODEUNIT *instr = &_PyCode_CODE(code)[i];
         CHECK(instr->op.code != 0);
         int base_opcode = _Py_GetBaseOpcode(code, i);
diff --git a/Python/opcode_metadata.h b/Python/opcode_metadata.h
index aa1db72..421d979 100644
--- a/Python/opcode_metadata.h
+++ b/Python/opcode_metadata.h
@@ -368,6 +368,10 @@ _PyOpcode_num_popped(int opcode, int oparg, bool jump) {
             return oparg + 2;
         case CALL_NO_KW_TUPLE_1:
             return oparg + 2;
+        case CALL_NO_KW_ALLOC_AND_ENTER_INIT:
+            return oparg + 2;
+        case EXIT_INIT_CHECK:
+            return 1;
         case CALL_BUILTIN_CLASS:
             return oparg + 2;
         case CALL_NO_KW_BUILTIN_O:
@@ -790,6 +794,10 @@ _PyOpcode_num_pushed(int opcode, int oparg, bool jump) {
             return 1;
         case CALL_NO_KW_TUPLE_1:
             return 1;
+        case CALL_NO_KW_ALLOC_AND_ENTER_INIT:
+            return 1;
+        case EXIT_INIT_CHECK:
+            return 0;
         case CALL_BUILTIN_CLASS:
             return 1;
         case CALL_NO_KW_BUILTIN_O:
@@ -1057,6 +1065,8 @@ const struct opcode_metadata _PyOpcode_opcode_metadata[512] = {
     [CALL_NO_KW_TYPE_1] = { true, INSTR_FMT_IBC00, HAS_ARG_FLAG },
     [CALL_NO_KW_STR_1] = { true, INSTR_FMT_IBC00, HAS_ARG_FLAG },
     [CALL_NO_KW_TUPLE_1] = { true, INSTR_FMT_IBC00, HAS_ARG_FLAG },
+    [CALL_NO_KW_ALLOC_AND_ENTER_INIT] = { true, INSTR_FMT_IBC00, HAS_ARG_FLAG },
+    [EXIT_INIT_CHECK] = { true, INSTR_FMT_IX, 0 },
     [CALL_BUILTIN_CLASS] = { true, INSTR_FMT_IBC00, HAS_ARG_FLAG },
     [CALL_NO_KW_BUILTIN_O] = { true, INSTR_FMT_IBC00, HAS_ARG_FLAG },
     [CALL_NO_KW_BUILTIN_FAST] = { true, INSTR_FMT_IBC00, HAS_ARG_FLAG },
diff --git a/Python/opcode_targets.h b/Python/opcode_targets.h
index 4de4bf0..781d72f 100644
--- a/Python/opcode_targets.h
+++ b/Python/opcode_targets.h
@@ -15,39 +15,39 @@ static void *opcode_targets[256] = {
     &&TARGET_BINARY_OP_ADD_FLOAT,
     &&TARGET_BINARY_OP_SUBTRACT_FLOAT,
     &&TARGET_UNARY_INVERT,
-    &&TARGET_BINARY_OP_ADD_UNICODE,
+    &&TARGET_EXIT_INIT_CHECK,
     &&TARGET_RESERVED,
+    &&TARGET_BINARY_OP_ADD_UNICODE,
     &&TARGET_BINARY_OP_INPLACE_ADD_UNICODE,
     &&TARGET_BINARY_SUBSCR_DICT,
     &&TARGET_BINARY_SUBSCR_GETITEM,
     &&TARGET_BINARY_SUBSCR_LIST_INT,
     &&TARGET_BINARY_SUBSCR_TUPLE_INT,
-    &&TARGET_STORE_SUBSCR_DICT,
     &&TARGET_MAKE_FUNCTION,
     &&TARGET_BINARY_SUBSCR,
     &&TARGET_BINARY_SLICE,
     &&TARGET_STORE_SLICE,
+    &&TARGET_STORE_SUBSCR_DICT,
     &&TARGET_STORE_SUBSCR_LIST_INT,
-    &&TARGET_SEND_GEN,
     &&TARGET_GET_LEN,
     &&TARGET_MATCH_MAPPING,
     &&TARGET_MATCH_SEQUENCE,
     &&TARGET_MATCH_KEYS,
-    &&TARGET_UNPACK_SEQUENCE_TWO_TUPLE,
+    &&TARGET_SEND_GEN,
     &&TARGET_PUSH_EXC_INFO,
     &&TARGET_CHECK_EXC_MATCH,
     &&TARGET_CHECK_EG_MATCH,
+    &&TARGET_UNPACK_SEQUENCE_TWO_TUPLE,
     &&TARGET_UNPACK_SEQUENCE_TUPLE,
-    &&TARGET_UNPACK_SEQUENCE_LIST,
     &&TARGET_FORMAT_SIMPLE,
     &&TARGET_FORMAT_WITH_SPEC,
+    &&TARGET_UNPACK_SEQUENCE_LIST,
     &&TARGET_STORE_ATTR_INSTANCE_VALUE,
     &&TARGET_STORE_ATTR_SLOT,
     &&TARGET_STORE_ATTR_WITH_HINT,
     &&TARGET_LOAD_GLOBAL_MODULE,
     &&TARGET_LOAD_GLOBAL_BUILTIN,
     &&TARGET_LOAD_SUPER_ATTR_ATTR,
-    &&TARGET_LOAD_SUPER_ATTR_METHOD,
     &&TARGET_WITH_EXCEPT_START,
     &&TARGET_GET_AITER,
     &&TARGET_GET_ANEXT,
@@ -55,39 +55,39 @@ static void *opcode_targets[256] = {
     &&TARGET_BEFORE_WITH,
     &&TARGET_END_ASYNC_FOR,
     &&TARGET_CLEANUP_THROW,
+    &&TARGET_LOAD_SUPER_ATTR_METHOD,
     &&TARGET_LOAD_ATTR_INSTANCE_VALUE,
     &&TARGET_LOAD_ATTR_MODULE,
     &&TARGET_LOAD_ATTR_WITH_HINT,
-    &&TARGET_LOAD_ATTR_SLOT,
     &&TARGET_STORE_SUBSCR,
     &&TARGET_DELETE_SUBSCR,
+    &&TARGET_LOAD_ATTR_SLOT,
     &&TARGET_LOAD_ATTR_CLASS,
     &&TARGET_LOAD_ATTR_PROPERTY,
     &&TARGET_LOAD_ATTR_GETATTRIBUTE_OVERRIDDEN,
     &&TARGET_LOAD_ATTR_METHOD_WITH_VALUES,
     &&TARGET_LOAD_ATTR_METHOD_NO_DICT,
-    &&TARGET_LOAD_ATTR_METHOD_LAZY_DICT,
     &&TARGET_GET_ITER,
     &&TARGET_GET_YIELD_FROM_ITER,
-    &&TARGET_COMPARE_OP_FLOAT,
+    &&TARGET_LOAD_ATTR_METHOD_LAZY_DICT,
     &&TARGET_LOAD_BUILD_CLASS,
+    &&TARGET_COMPARE_OP_FLOAT,
     &&TARGET_COMPARE_OP_INT,
-    &&TARGET_COMPARE_OP_STR,
     &&TARGET_LOAD_ASSERTION_ERROR,
     &&TARGET_RETURN_GENERATOR,
+    &&TARGET_COMPARE_OP_STR,
     &&TARGET_FOR_ITER_LIST,
     &&TARGET_FOR_ITER_TUPLE,
     &&TARGET_FOR_ITER_RANGE,
     &&TARGET_FOR_ITER_GEN,
     &&TARGET_CALL_BOUND_METHOD_EXACT_ARGS,
     &&TARGET_CALL_PY_EXACT_ARGS,
-    &&TARGET_CALL_PY_WITH_DEFAULTS,
     &&TARGET_RETURN_VALUE,
-    &&TARGET_CALL_NO_KW_TYPE_1,
+    &&TARGET_CALL_PY_WITH_DEFAULTS,
     &&TARGET_SETUP_ANNOTATIONS,
-    &&TARGET_CALL_NO_KW_STR_1,
+    &&TARGET_CALL_NO_KW_TYPE_1,
     &&TARGET_LOAD_LOCALS,
-    &&TARGET_CALL_NO_KW_TUPLE_1,
+    &&TARGET_CALL_NO_KW_STR_1,
     &&TARGET_POP_EXCEPT,
     &&TARGET_STORE_NAME,
     &&TARGET_DELETE_NAME,
@@ -110,9 +110,9 @@ static void *opcode_targets[256] = {
     &&TARGET_IMPORT_NAME,
     &&TARGET_IMPORT_FROM,
     &&TARGET_JUMP_FORWARD,
+    &&TARGET_CALL_NO_KW_TUPLE_1,
     &&TARGET_CALL_BUILTIN_CLASS,
     &&TARGET_CALL_NO_KW_BUILTIN_O,
-    &&TARGET_CALL_NO_KW_BUILTIN_FAST,
     &&TARGET_POP_JUMP_IF_FALSE,
     &&TARGET_POP_JUMP_IF_TRUE,
     &&TARGET_LOAD_GLOBAL,
@@ -131,7 +131,7 @@ static void *opcode_targets[256] = {
     &&TARGET_POP_JUMP_IF_NONE,
     &&TARGET_RAISE_VARARGS,
     &&TARGET_GET_AWAITABLE,
-    &&TARGET_CALL_BUILTIN_FAST_WITH_KEYWORDS,
+    &&TARGET_CALL_NO_KW_BUILTIN_FAST,
     &&TARGET_BUILD_SLICE,
     &&TARGET_JUMP_BACKWARD_NO_INTERRUPT,
     &&TARGET_MAKE_CELL,
@@ -147,26 +147,26 @@ static void *opcode_targets[256] = {
     &&TARGET_LIST_APPEND,
     &&TARGET_SET_ADD,
     &&TARGET_MAP_ADD,
-    &&TARGET_CALL_NO_KW_LEN,
+    &&TARGET_CALL_BUILTIN_FAST_WITH_KEYWORDS,
     &&TARGET_COPY_FREE_VARS,
     &&TARGET_YIELD_VALUE,
     &&TARGET_RESUME,
     &&TARGET_MATCH_CLASS,
+    &&TARGET_CALL_NO_KW_LEN,
     &&TARGET_CALL_NO_KW_ISINSTANCE,
     &&TARGET_CALL_NO_KW_LIST_APPEND,
-    &&TARGET_CALL_NO_KW_METHOD_DESCRIPTOR_O,
     &&TARGET_BUILD_CONST_KEY_MAP,
     &&TARGET_BUILD_STRING,
     &&TARGET_CONVERT_VALUE,
+    &&TARGET_CALL_NO_KW_METHOD_DESCRIPTOR_O,
     &&TARGET_CALL_METHOD_DESCRIPTOR_FAST_WITH_KEYWORDS,
     &&TARGET_CALL_NO_KW_METHOD_DESCRIPTOR_NOARGS,
-    &&TARGET_CALL_NO_KW_METHOD_DESCRIPTOR_FAST,
     &&TARGET_LIST_EXTEND,
     &&TARGET_SET_UPDATE,
     &&TARGET_DICT_MERGE,
     &&TARGET_DICT_UPDATE,
-    &&_unknown_opcode,
-    &&_unknown_opcode,
+    &&TARGET_CALL_NO_KW_METHOD_DESCRIPTOR_FAST,
+    &&TARGET_CALL_NO_KW_ALLOC_AND_ENTER_INIT,
     &&TARGET_LOAD_FAST_LOAD_FAST,
     &&TARGET_STORE_FAST_LOAD_FAST,
     &&TARGET_STORE_FAST_STORE_FAST,
diff --git a/Python/specialize.c b/Python/specialize.c
index 44b14c5..0006aa7 100644
--- a/Python/specialize.c
+++ b/Python/specialize.c
@@ -391,7 +391,7 @@ _PyCode_Quicken(PyCodeObject *code)
 #define SPEC_FAIL_CALL_METH_DESCR_VARARGS_KEYWORDS 18
 #define SPEC_FAIL_CALL_METH_DESCR_METHOD_FASTCALL_KEYWORDS 19
 #define SPEC_FAIL_CALL_BAD_CALL_FLAGS 20
-#define SPEC_FAIL_CALL_PYTHON_CLASS 21
+#define SPEC_FAIL_CALL_INIT_NOT_PYTHON 21
 #define SPEC_FAIL_CALL_PEP_523 22
 #define SPEC_FAIL_CALL_BOUND_METHOD 23
 #define SPEC_FAIL_CALL_STR 24
@@ -400,6 +400,7 @@ _PyCode_Quicken(PyCodeObject *code)
 #define SPEC_FAIL_CALL_KWNAMES 27
 #define SPEC_FAIL_CALL_METHOD_WRAPPER 28
 #define SPEC_FAIL_CALL_OPERATOR_WRAPPER 29
+#define SPEC_FAIL_CALL_INIT_NOT_SIMPLE 30
 
 /* COMPARE_OP */
 #define SPEC_FAIL_COMPARE_OP_DIFFERENT_TYPES 12
@@ -1491,15 +1492,46 @@ success:
     cache->counter = adaptive_counter_cooldown();
 }
 
+/* Returns a borrowed reference.
+ * The reference is only valid if guarded by a type version check.
+ */
+static PyFunctionObject *
+get_init_for_simple_managed_python_class(PyTypeObject *tp)
+{
+    assert(tp->tp_new == PyBaseObject_Type.tp_new);
+    if (tp->tp_alloc != PyType_GenericAlloc) {
+        SPECIALIZATION_FAIL(CALL, SPEC_FAIL_OVERRIDDEN);
+        return NULL;
+    }
+    if ((tp->tp_flags & Py_TPFLAGS_MANAGED_DICT) == 0) {
+        SPECIALIZATION_FAIL(CALL, SPEC_FAIL_NO_DICT);
+        return NULL;
+    }
+    if (!(tp->tp_flags & Py_TPFLAGS_HEAPTYPE)) {
+        /* Is this possible? */
+        SPECIALIZATION_FAIL(CALL, SPEC_FAIL_EXPECTED_ERROR);
+        return NULL;
+    }
+    PyObject *init = _PyType_Lookup(tp, &_Py_ID(__init__));
+    if (init == NULL || !PyFunction_Check(init)) {
+        SPECIALIZATION_FAIL(CALL, SPEC_FAIL_CALL_INIT_NOT_PYTHON);
+        return NULL;
+    }
+    int kind = function_kind((PyCodeObject *)PyFunction_GET_CODE(init));
+    if (kind != SIMPLE_FUNCTION) {
+        SPECIALIZATION_FAIL(CALL, SPEC_FAIL_CALL_INIT_NOT_SIMPLE);
+        return NULL;
+    }
+    ((PyHeapTypeObject *)tp)->_spec_cache.init = init;
+    return (PyFunctionObject *)init;
+}
+
 static int
 specialize_class_call(PyObject *callable, _Py_CODEUNIT *instr, int nargs,
                       PyObject *kwnames)
 {
+    assert(PyType_Check(callable));
     PyTypeObject *tp = _PyType_CAST(callable);
-    if (tp->tp_new == PyBaseObject_Type.tp_new) {
-        SPECIALIZATION_FAIL(CALL, SPEC_FAIL_CALL_PYTHON_CLASS);
-        return -1;
-    }
     if (tp->tp_flags & Py_TPFLAGS_IMMUTABLETYPE) {
         int oparg = instr->op.arg;
         if (nargs == 1 && kwnames == NULL && oparg == 1) {
@@ -1524,6 +1556,24 @@ specialize_class_call(PyObject *callable, _Py_CODEUNIT *instr, int nargs,
             SPEC_FAIL_CALL_STR : SPEC_FAIL_CALL_CLASS_NO_VECTORCALL);
         return -1;
     }
+    if (tp->tp_new == PyBaseObject_Type.tp_new) {
+        PyFunctionObject *init = get_init_for_simple_managed_python_class(tp);
+        if (init != NULL) {
+            if (((PyCodeObject *)init->func_code)->co_argcount != nargs+1) {
+                SPECIALIZATION_FAIL(CALL, SPEC_FAIL_WRONG_NUMBER_ARGUMENTS);
+                return -1;
+            }
+            if (kwnames) {
+                SPECIALIZATION_FAIL(CALL, SPEC_FAIL_CALL_KWNAMES);
+                return -1;
+            }
+            _PyCallCache *cache = (_PyCallCache *)(instr + 1);
+            write_u32(cache->func_version, tp->tp_version_tag);
+            _Py_SET_OPCODE(*instr, CALL_NO_KW_ALLOC_AND_ENTER_INIT);
+            return 0;
+        }
+        return -1;
+    }
     SPECIALIZATION_FAIL(CALL, SPEC_FAIL_CALL_CLASS_MUTABLE);
     return -1;
 }
@@ -2229,3 +2279,42 @@ success:
     STAT_INC(SEND, success);
     cache->counter = adaptive_counter_cooldown();
 }
+
+/* Code init cleanup.
+ * CALL_NO_KW_ALLOC_AND_ENTER_INIT will set up
+ * the frame to execute the EXIT_INIT_CHECK
+ * instruction.
+ * Ends with a RESUME so that it is not traced.
+ * This is used as a plain code object, not a function,
+ * so must not access globals or builtins.
+ */
+
+#define NO_LOC_4 (128 | (PY_CODE_LOCATION_INFO_NONE << 3) | 3)
+
+static const PyBytesObject no_location = {
+    PyVarObject_HEAD_INIT(&PyBytes_Type, 1)
+    .ob_sval = { NO_LOC_4 }
+};
+
+const struct _PyCode_DEF(8) _Py_InitCleanup = {
+    _PyVarObject_HEAD_INIT(&PyCode_Type, 4)
+    .co_consts = (PyObject *)&_Py_SINGLETON(tuple_empty),
+    .co_names = (PyObject *)&_Py_SINGLETON(tuple_empty),
+    .co_exceptiontable = (PyObject *)&_Py_SINGLETON(bytes_empty),
+    .co_flags = CO_OPTIMIZED,
+    .co_localsplusnames = (PyObject *)&_Py_SINGLETON(tuple_empty),
+    .co_localspluskinds = (PyObject *)&_Py_SINGLETON(bytes_empty),
+    .co_filename = &_Py_ID(__init__),
+    .co_name = &_Py_ID(__init__),
+    .co_qualname = &_Py_ID(__init__),
+    .co_linetable = (PyObject *)&no_location,
+    ._co_firsttraceable = 4,
+    .co_stacksize = 2,
+    .co_framesize = 2 + FRAME_SPECIALS_SIZE,
+    .co_code_adaptive = {
+        NOP, 0,
+        EXIT_INIT_CHECK, 0,
+        RETURN_VALUE, 0,
+        RESUME, 0,
+    }
+};