diff --git a/Include/internal/pycore_frame.h b/Include/internal/pycore_frame.h
index 41d4b6ff5ecf59..3d17421fe48417 100644
--- a/Include/internal/pycore_frame.h
+++ b/Include/internal/pycore_frame.h
@@ -86,7 +86,7 @@ typedef struct _PyInterpreterFrame {
 } _PyInterpreterFrame;
 
 #define NewPyInterpreterFrame_LASTI(IF) \
-    ((int)((IF)->instr_ptr - _PyCode_CODE(_PyFrame_GetCode(IF))))
+    ((int)(((IF)->instr_ptr - 1) - _PyCode_CODE(_PyFrame_GetCode(IF))))
 
 #define _PyInterpreterFrame_LASTI(IF) \
     ((int)((IF)->prev_instr - _PyCode_CODE(_PyFrame_GetCode(IF))))
@@ -160,6 +160,7 @@ _PyFrame_Initialize(
     frame->prev_instr = _PyCode_CODE(code) - 1;
     frame->instr_ptr = _PyCode_CODE(code);
     frame->return_offset = 0;
+    frame->new_return_offset = 0;
     frame->owner = FRAME_OWNED_BY_THREAD;
 
     for (int i = null_locals_from; i < code->co_nlocalsplus; i++) {
@@ -325,6 +326,7 @@ _PyFrame_PushTrampolineUnchecked(PyThreadState *tstate, PyCodeObject *code, int
     frame->instr_ptr = _PyCode_CODE(code) + prev_instr + 1;
     frame->owner = FRAME_OWNED_BY_THREAD;
     frame->return_offset = 0;
+    frame->new_return_offset = 0;
     return frame;
 }
 
diff --git a/Python/bytecodes.c b/Python/bytecodes.c
index 830b6f2aa3514c..018ac883707caa 100644
--- a/Python/bytecodes.c
+++ b/Python/bytecodes.c
@@ -771,7 +771,9 @@ fprintf(stderr, "goto exception_unwind: frame=%p frame->prev_instr=%p frame->ins
             assert(frame == &entry_frame);
             assert(_PyFrame_IsIncomplete(frame));
             /* Restore previous frame and return. */
+fprintf(stderr, "INTERPRETER_EXIT1: frame=%p frame->prev_instr=%p frame->instr_ptr=%p new_return_offset=%d frame->previous = %p\n", frame, frame->prev_instr, frame->instr_ptr, frame->new_return_offset, frame->previous);
             tstate->current_frame = frame->previous;
+if (frame->previous) fprintf(stderr, "INTERPRETER_EXIT2: tstate->current_frame=%p tstate->current_frame->prev_instr=%p tstate->current_frame->instr_ptr=%p new_return_offset=%d \n", tstate->current_frame, tstate->current_frame->prev_instr, tstate->current_frame->instr_ptr, tstate->current_frame->new_return_offset);
             assert(!_PyErr_Occurred(tstate));
             tstate->c_recursion_remaining += PY_EVAL_C_STACK_UNITS;
             return retval;
@@ -792,7 +794,7 @@ fprintf(stderr, "goto exception_unwind: frame=%p frame->prev_instr=%p frame->ins
             _PyInterpreterFrame *dying = frame;
             frame = tstate->current_frame = dying->previous;
             _PyEval_FrameClearAndPop(tstate, dying);
-fprintf(stderr, "_POP_FRAME[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p \n", frame, frame->prev_instr, frame->instr_ptr);
+fprintf(stderr, "_POP_FRAME[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p new_return_offset=%d \n", frame, frame->prev_instr, frame->instr_ptr, frame->new_return_offset);
             frame->prev_instr += frame->return_offset;
             frame->instr_ptr += frame->new_return_offset;
             frame->new_return_offset = 0;
@@ -1080,6 +1082,10 @@ fprintf(stderr, "_POP_FRAME[3]: frame=%p frame->prev_instr=%p frame->instr_ptr=%
             frame = tstate->current_frame = frame->previous;
             gen_frame->previous = NULL;
             _PyFrame_StackPush(frame, retval);
+            frame->instr_ptr += frame->new_return_offset;
+            frame->new_return_offset = 0;
+            frame->prev_instr = frame->instr_ptr - 1;
+fprintf(stderr, "YIELD_VALUE: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
             goto resume_frame;
         }
 
@@ -2665,7 +2671,7 @@ fprintf(stderr, "_POP_FRAME[3]: frame=%p frame->prev_instr=%p frame->instr_ptr=%
             assert(next_instr[oparg].op.code == END_FOR ||
                    next_instr[oparg].op.code == INSTRUMENTED_END_FOR);
             frame->return_offset = oparg;
-            frame->new_return_offset = oparg;
+            frame->new_return_offset = oparg + next_instr - frame->instr_ptr;
             DISPATCH_INLINED(gen_frame);
         }
 
@@ -2913,7 +2919,7 @@ fprintf(stderr, "_POP_FRAME[3]: frame=%p frame->prev_instr=%p frame->instr_ptr=%
         // When calling Python, inline the call using DISPATCH_INLINED().
         inst(CALL, (unused/1, unused/2, callable, self_or_null, args[oparg] -- res)) {
             // oparg counts all of the args, but *not* self:
-fprintf(stderr, "CALL1: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+fprintf(stderr, "CALL1: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
             int total_args = oparg;
             if (self_or_null != NULL) {
                 args--;
@@ -2929,7 +2935,7 @@ fprintf(stderr, "CALL1: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_i
             STAT_INC(CALL, deferred);
             DECREMENT_ADAPTIVE_COUNTER(cache->counter);
             #endif  /* ENABLE_SPECIALIZATION */
-fprintf(stderr, "CALL2: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+fprintf(stderr, "CALL2: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
             if (self_or_null == NULL && Py_TYPE(callable) == &PyMethod_Type) {
                 args--;
                 total_args++;
@@ -2940,7 +2946,8 @@ fprintf(stderr, "CALL2: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_i
                 Py_DECREF(callable);
                 callable = method;
             }
-fprintf(stderr, "CALL3: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+fprintf(stderr, "CALL3: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
+
             // Check if the call can be inlined or not
             if (Py_TYPE(callable) == &PyFunction_Type &&
                 tstate->interp->eval_frame == NULL &&
@@ -2969,10 +2976,12 @@ fprintf(stderr, "CALL-DISPATCH_INLINED: frame=%p frame->prev_instr=%p frame->ins
                 DISPATCH_INLINED(new_frame);
             }
             /* Callable is not a normal Python function */
+fprintf(stderr, "CALL-not normal python func1: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
             res = PyObject_Vectorcall(
                 callable, args,
                 total_args | PY_VECTORCALL_ARGUMENTS_OFFSET,
                 NULL);
+fprintf(stderr, "CALL-not normal python func2: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
             if (opcode == INSTRUMENTED_CALL) {
                 PyObject *arg = total_args == 0 ?
                     &_PyInstrumentation_MISSING : args[0];
@@ -2998,6 +3007,8 @@ fprintf(stderr, "CALL-DISPATCH_INLINED: frame=%p frame->prev_instr=%p frame->ins
 
             ERROR_IF(res == NULL, error);
 fprintf(stderr, "CALL-END: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+            frame->return_offset = 0;
+            frame->new_return_offset = 0;
             CHECK_EVAL_BREAKER();
         }
 
@@ -3675,8 +3686,7 @@ fprintf(stderr, "_PUSH_FRAME: frame=%p frame->prev_instr=%p frame->instr_ptr=%p\
                         goto error;
                     }
                     frame->return_offset = 0;
-                    frame->instr_ptr = next_instr - frame->new_return_offset;
-                    frame->new_return_offset = 0;
+                    frame->new_return_offset = next_instr - frame->instr_ptr;
                     DISPATCH_INLINED(new_frame);
                 }
                 result = PyObject_Call(func, callargs, kwargs);
@@ -3946,11 +3956,11 @@ fprintf(stderr, "_PUSH_FRAME: frame=%p frame->prev_instr=%p frame->instr_ptr=%p\
 
         op(_SAVE_CURRENT_IP, (--)) {
             #if TIER_ONE
-fprintf(stderr, "_SAVE_CURRENT_IP[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+fprintf(stderr, "_SAVE_CURRENT_IP[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
             frame->prev_instr = next_instr - 1;
-            frame->instr_ptr = next_instr - frame->new_return_offset;
-            frame->new_return_offset = 0;
-fprintf(stderr, "_SAVE_CURRENT_IP[2]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+            assert(frame->new_return_offset == 0);
+            frame->new_return_offset = next_instr - frame->instr_ptr;
+fprintf(stderr, "_SAVE_CURRENT_IP[2]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
             #endif
             #if TIER_TWO
             // Relies on a preceding _SET_IP
diff --git a/Python/ceval_macros.h b/Python/ceval_macros.h
index 785e39a553edbb..f718ae1604f811 100644
--- a/Python/ceval_macros.h
+++ b/Python/ceval_macros.h
@@ -71,10 +71,10 @@
 #else
 #define INSTRUCTION_START(op) \
     do { \
-fprintf(stderr, "-- %s: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", _PyOpcode_OpName[op], frame, frame->prev_instr, frame->instr_ptr, next_instr); \
+fprintf(stderr, "--- %s: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", _PyOpcode_OpName[op], frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset); \
         frame->instr_ptr = next_instr; \
         frame->prev_instr = next_instr++; \
-fprintf(stderr, "== %s: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", _PyOpcode_OpName[op], frame, frame->prev_instr, frame->instr_ptr, next_instr); \
+fprintf(stderr, "=== %s: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", _PyOpcode_OpName[op], frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset); \
     } while(0)
 #endif
 
@@ -387,6 +387,9 @@ static inline void _Py_LeaveRecursiveCallPy(PyThreadState *tstate)  {
 
 #define LOAD_IP() do { \
         assert(frame->prev_instr + 1 == frame->instr_ptr); \
+        frame->instr_ptr += frame->new_return_offset; \
+        frame->new_return_offset = 0; \
+fprintf(stderr, "LOAD_IP: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset); \
         next_instr = frame->instr_ptr; \
     } while (0)
 
diff --git a/Python/executor_cases.c.h b/Python/executor_cases.c.h
index a1288ca1017c24..7729cccd1fcdf4 100644
--- a/Python/executor_cases.c.h
+++ b/Python/executor_cases.c.h
@@ -692,7 +692,7 @@
             _PyInterpreterFrame *dying = frame;
             frame = tstate->current_frame = dying->previous;
             _PyEval_FrameClearAndPop(tstate, dying);
-fprintf(stderr, "_POP_FRAME[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p \n", frame, frame->prev_instr, frame->instr_ptr);
+fprintf(stderr, "_POP_FRAME[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p new_return_offset=%d \n", frame, frame->prev_instr, frame->instr_ptr, frame->new_return_offset);
             frame->prev_instr += frame->return_offset;
             frame->instr_ptr += frame->new_return_offset;
             frame->new_return_offset = 0;
@@ -2978,11 +2978,11 @@ fprintf(stderr, "_POP_FRAME[3]: frame=%p frame->prev_instr=%p frame->instr_ptr=%
 
         case _SAVE_CURRENT_IP: {
             #if TIER_ONE
-fprintf(stderr, "_SAVE_CURRENT_IP[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+fprintf(stderr, "_SAVE_CURRENT_IP[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
             frame->prev_instr = next_instr - 1;
-            frame->instr_ptr = next_instr - frame->new_return_offset;
-            frame->new_return_offset = 0;
-fprintf(stderr, "_SAVE_CURRENT_IP[2]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+            assert(frame->new_return_offset == 0);
+            frame->new_return_offset = next_instr - frame->instr_ptr;
+fprintf(stderr, "_SAVE_CURRENT_IP[2]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
             #endif
             #if TIER_TWO
             // Relies on a preceding _SET_IP
diff --git a/Python/generated_cases.c.h b/Python/generated_cases.c.h
index 786d66abb55012..be6d5d5f55547a 100644
--- a/Python/generated_cases.c.h
+++ b/Python/generated_cases.c.h
@@ -979,7 +979,9 @@ fprintf(stderr, "goto exception_unwind: frame=%p frame->prev_instr=%p frame->ins
             assert(frame == &entry_frame);
             assert(_PyFrame_IsIncomplete(frame));
             /* Restore previous frame and return. */
+fprintf(stderr, "INTERPRETER_EXIT1: frame=%p frame->prev_instr=%p frame->instr_ptr=%p new_return_offset=%d frame->previous = %p\n", frame, frame->prev_instr, frame->instr_ptr, frame->new_return_offset, frame->previous);
             tstate->current_frame = frame->previous;
+if (frame->previous) fprintf(stderr, "INTERPRETER_EXIT2: tstate->current_frame=%p tstate->current_frame->prev_instr=%p tstate->current_frame->instr_ptr=%p new_return_offset=%d \n", tstate->current_frame, tstate->current_frame->prev_instr, tstate->current_frame->instr_ptr, tstate->current_frame->new_return_offset);
             assert(!_PyErr_Occurred(tstate));
             tstate->c_recursion_remaining += PY_EVAL_C_STACK_UNITS;
             return retval;
@@ -990,11 +992,11 @@ fprintf(stderr, "goto exception_unwind: frame=%p frame->prev_instr=%p frame->ins
             // _SAVE_CURRENT_IP
             {
                 #if TIER_ONE
-    fprintf(stderr, "_SAVE_CURRENT_IP[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+    fprintf(stderr, "_SAVE_CURRENT_IP[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
                 frame->prev_instr = next_instr - 1;
-                frame->instr_ptr = next_instr - frame->new_return_offset;
-                frame->new_return_offset = 0;
-    fprintf(stderr, "_SAVE_CURRENT_IP[2]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+                assert(frame->new_return_offset == 0);
+                frame->new_return_offset = next_instr - frame->instr_ptr;
+    fprintf(stderr, "_SAVE_CURRENT_IP[2]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
                 #endif
                 #if TIER_TWO
                 // Relies on a preceding _SET_IP
@@ -1015,7 +1017,7 @@ fprintf(stderr, "goto exception_unwind: frame=%p frame->prev_instr=%p frame->ins
                 _PyInterpreterFrame *dying = frame;
                 frame = tstate->current_frame = dying->previous;
                 _PyEval_FrameClearAndPop(tstate, dying);
-    fprintf(stderr, "_POP_FRAME[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p \n", frame, frame->prev_instr, frame->instr_ptr);
+    fprintf(stderr, "_POP_FRAME[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p new_return_offset=%d \n", frame, frame->prev_instr, frame->instr_ptr, frame->new_return_offset);
                 frame->prev_instr += frame->return_offset;
                 frame->instr_ptr += frame->new_return_offset;
                 frame->new_return_offset = 0;
@@ -1069,11 +1071,11 @@ fprintf(stderr, "goto exception_unwind: frame=%p frame->prev_instr=%p frame->ins
             // _SAVE_CURRENT_IP
             {
                 #if TIER_ONE
-    fprintf(stderr, "_SAVE_CURRENT_IP[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+    fprintf(stderr, "_SAVE_CURRENT_IP[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
                 frame->prev_instr = next_instr - 1;
-                frame->instr_ptr = next_instr - frame->new_return_offset;
-                frame->new_return_offset = 0;
-    fprintf(stderr, "_SAVE_CURRENT_IP[2]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+                assert(frame->new_return_offset == 0);
+                frame->new_return_offset = next_instr - frame->instr_ptr;
+    fprintf(stderr, "_SAVE_CURRENT_IP[2]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
                 #endif
                 #if TIER_TWO
                 // Relies on a preceding _SET_IP
@@ -1093,7 +1095,7 @@ fprintf(stderr, "goto exception_unwind: frame=%p frame->prev_instr=%p frame->ins
                 _PyInterpreterFrame *dying = frame;
                 frame = tstate->current_frame = dying->previous;
                 _PyEval_FrameClearAndPop(tstate, dying);
-    fprintf(stderr, "_POP_FRAME[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p \n", frame, frame->prev_instr, frame->instr_ptr);
+    fprintf(stderr, "_POP_FRAME[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p new_return_offset=%d \n", frame, frame->prev_instr, frame->instr_ptr, frame->new_return_offset);
                 frame->prev_instr += frame->return_offset;
                 frame->instr_ptr += frame->new_return_offset;
                 frame->new_return_offset = 0;
@@ -1381,6 +1383,10 @@ fprintf(stderr, "goto exception_unwind: frame=%p frame->prev_instr=%p frame->ins
             frame = tstate->current_frame = frame->previous;
             gen_frame->previous = NULL;
             _PyFrame_StackPush(frame, retval);
+            frame->instr_ptr += frame->new_return_offset;
+            frame->new_return_offset = 0;
+            frame->prev_instr = frame->instr_ptr - 1;
+fprintf(stderr, "YIELD_VALUE: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
             goto resume_frame;
         }
 
@@ -3446,7 +3452,7 @@ fprintf(stderr, "goto exception_unwind: frame=%p frame->prev_instr=%p frame->ins
             assert(next_instr[oparg].op.code == END_FOR ||
                    next_instr[oparg].op.code == INSTRUMENTED_END_FOR);
             frame->return_offset = oparg;
-            frame->new_return_offset = oparg;
+            frame->new_return_offset = oparg + next_instr - frame->instr_ptr;
             DISPATCH_INLINED(gen_frame);
         }
 
@@ -3747,7 +3753,7 @@ fprintf(stderr, "goto exception_unwind: frame=%p frame->prev_instr=%p frame->ins
             self_or_null = stack_pointer[-1 - oparg];
             callable = stack_pointer[-2 - oparg];
             // oparg counts all of the args, but *not* self:
-fprintf(stderr, "CALL1: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+fprintf(stderr, "CALL1: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
             int total_args = oparg;
             if (self_or_null != NULL) {
                 args--;
@@ -3763,7 +3769,7 @@ fprintf(stderr, "CALL1: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_i
             STAT_INC(CALL, deferred);
             DECREMENT_ADAPTIVE_COUNTER(cache->counter);
             #endif  /* ENABLE_SPECIALIZATION */
-fprintf(stderr, "CALL2: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+fprintf(stderr, "CALL2: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
             if (self_or_null == NULL && Py_TYPE(callable) == &PyMethod_Type) {
                 args--;
                 total_args++;
@@ -3774,7 +3780,8 @@ fprintf(stderr, "CALL2: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_i
                 Py_DECREF(callable);
                 callable = method;
             }
-fprintf(stderr, "CALL3: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+fprintf(stderr, "CALL3: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
+
             // Check if the call can be inlined or not
             if (Py_TYPE(callable) == &PyFunction_Type &&
                 tstate->interp->eval_frame == NULL &&
@@ -3803,10 +3810,12 @@ fprintf(stderr, "CALL-DISPATCH_INLINED: frame=%p frame->prev_instr=%p frame->ins
                 DISPATCH_INLINED(new_frame);
             }
             /* Callable is not a normal Python function */
+fprintf(stderr, "CALL-not normal python func1: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
             res = PyObject_Vectorcall(
                 callable, args,
                 total_args | PY_VECTORCALL_ARGUMENTS_OFFSET,
                 NULL);
+fprintf(stderr, "CALL-not normal python func2: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
             if (opcode == INSTRUMENTED_CALL) {
                 PyObject *arg = total_args == 0 ?
                     &_PyInstrumentation_MISSING : args[0];
@@ -3832,6 +3841,8 @@ fprintf(stderr, "CALL-DISPATCH_INLINED: frame=%p frame->prev_instr=%p frame->ins
 
             if (res == NULL) { STACK_SHRINK(oparg); goto pop_2_error; }
 fprintf(stderr, "CALL-END: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+            frame->return_offset = 0;
+            frame->new_return_offset = 0;
             STACK_SHRINK(oparg);
             STACK_SHRINK(1);
             stack_pointer[-1] = res;
@@ -3905,11 +3916,11 @@ fprintf(stderr, "CALL-END: frame=%p frame->prev_instr=%p frame->instr_ptr=%p nex
             next_instr += 3;
             {
                 #if TIER_ONE
-    fprintf(stderr, "_SAVE_CURRENT_IP[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+    fprintf(stderr, "_SAVE_CURRENT_IP[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
                 frame->prev_instr = next_instr - 1;
-                frame->instr_ptr = next_instr - frame->new_return_offset;
-                frame->new_return_offset = 0;
-    fprintf(stderr, "_SAVE_CURRENT_IP[2]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+                assert(frame->new_return_offset == 0);
+                frame->new_return_offset = next_instr - frame->instr_ptr;
+    fprintf(stderr, "_SAVE_CURRENT_IP[2]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
                 #endif
                 #if TIER_TWO
                 // Relies on a preceding _SET_IP
@@ -3989,11 +4000,11 @@ fprintf(stderr, "CALL-END: frame=%p frame->prev_instr=%p frame->instr_ptr=%p nex
             next_instr += 3;
             {
                 #if TIER_ONE
-    fprintf(stderr, "_SAVE_CURRENT_IP[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+    fprintf(stderr, "_SAVE_CURRENT_IP[1]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
                 frame->prev_instr = next_instr - 1;
-                frame->instr_ptr = next_instr - frame->new_return_offset;
-                frame->new_return_offset = 0;
-    fprintf(stderr, "_SAVE_CURRENT_IP[2]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p\n", frame, frame->prev_instr, frame->instr_ptr, next_instr);
+                assert(frame->new_return_offset == 0);
+                frame->new_return_offset = next_instr - frame->instr_ptr;
+    fprintf(stderr, "_SAVE_CURRENT_IP[2]: frame=%p frame->prev_instr=%p frame->instr_ptr=%p next_instr=%p new_return_offset=%d\n", frame, frame->prev_instr, frame->instr_ptr, next_instr, frame->new_return_offset);
                 #endif
                 #if TIER_TWO
                 // Relies on a preceding _SET_IP
@@ -4807,8 +4818,7 @@ fprintf(stderr, "CALL-END: frame=%p frame->prev_instr=%p frame->instr_ptr=%p nex
                         goto error;
                     }
                     frame->return_offset = 0;
-                    frame->instr_ptr = next_instr - frame->new_return_offset;
-                    frame->new_return_offset = 0;
+                    frame->new_return_offset = next_instr - frame->instr_ptr;
                     DISPATCH_INLINED(new_frame);
                 }
                 result = PyObject_Call(func, callargs, kwargs);
diff --git a/Python/traceback.c b/Python/traceback.c
index 16f5d34e46286d..a7c4c7c6ed820d 100644
--- a/Python/traceback.c
+++ b/Python/traceback.c
@@ -239,8 +239,8 @@ _PyTraceBack_FromFrame(PyObject *tb_next, PyFrameObject *frame)
     int newline = PyCode_Addr2Line(_PyFrame_GetCode(frame->f_frame), new_addr);
     int line = PyCode_Addr2Line(_PyFrame_GetCode(frame->f_frame), addr);
 if (true || line != newline) {
-    fprintf(stderr, "frame=%p frame->prev_instr=%p frame->instr_ptr=%p line=%d newline=%d %s\n",
-            frame->f_frame, frame->f_frame->prev_instr, frame->f_frame->instr_ptr, line, newline, newline - line ? "XXX" : "");
+    fprintf(stderr, "frame=%p frame->prev_instr=%p frame->instr_ptr=%p frame->new_return_offset=%d  line=%d newline=%d %s\n",
+            frame->f_frame, frame->f_frame->prev_instr, frame->f_frame->instr_ptr, frame->f_frame->new_return_offset, line, newline, newline - line ? "XXX" : "");
 }
     assert(line == newline);
     return tb_create_raw((PyTracebackObject *)tb_next, frame, addr,