diff --git a/Include/internal/pycore_uop_ids.h b/Include/internal/pycore_uop_ids.h
index 1e6ef8e54a221a..702231e7cd7e5f 100644
--- a/Include/internal/pycore_uop_ids.h
+++ b/Include/internal/pycore_uop_ids.h
@@ -238,39 +238,40 @@ extern "C" {
 #define _REPLACE_WITH_TRUE 432
 #define _RESUME_CHECK RESUME_CHECK
 #define _RETURN_GENERATOR RETURN_GENERATOR
-#define _SAVE_RETURN_OFFSET 433
-#define _SEND 434
+#define _RETURN_OFFSET 433
+#define _SAVE_RETURN_OFFSET 434
+#define _SEND 435
 #define _SEND_GEN SEND_GEN
 #define _SETUP_ANNOTATIONS SETUP_ANNOTATIONS
 #define _SET_ADD SET_ADD
 #define _SET_FUNCTION_ATTRIBUTE SET_FUNCTION_ATTRIBUTE
 #define _SET_UPDATE SET_UPDATE
-#define _START_EXECUTOR 435
-#define _STORE_ATTR 436
-#define _STORE_ATTR_INSTANCE_VALUE 437
-#define _STORE_ATTR_SLOT 438
+#define _START_EXECUTOR 436
+#define _STORE_ATTR 437
+#define _STORE_ATTR_INSTANCE_VALUE 438
+#define _STORE_ATTR_SLOT 439
 #define _STORE_ATTR_WITH_HINT STORE_ATTR_WITH_HINT
 #define _STORE_DEREF STORE_DEREF
-#define _STORE_FAST 439
-#define _STORE_FAST_0 440
-#define _STORE_FAST_1 441
-#define _STORE_FAST_2 442
-#define _STORE_FAST_3 443
-#define _STORE_FAST_4 444
-#define _STORE_FAST_5 445
-#define _STORE_FAST_6 446
-#define _STORE_FAST_7 447
+#define _STORE_FAST 440
+#define _STORE_FAST_0 441
+#define _STORE_FAST_1 442
+#define _STORE_FAST_2 443
+#define _STORE_FAST_3 444
+#define _STORE_FAST_4 445
+#define _STORE_FAST_5 446
+#define _STORE_FAST_6 447
+#define _STORE_FAST_7 448
 #define _STORE_FAST_LOAD_FAST STORE_FAST_LOAD_FAST
 #define _STORE_FAST_STORE_FAST STORE_FAST_STORE_FAST
 #define _STORE_GLOBAL STORE_GLOBAL
 #define _STORE_NAME STORE_NAME
 #define _STORE_SLICE STORE_SLICE
-#define _STORE_SUBSCR 448
+#define _STORE_SUBSCR 449
 #define _STORE_SUBSCR_DICT STORE_SUBSCR_DICT
 #define _STORE_SUBSCR_LIST_INT STORE_SUBSCR_LIST_INT
 #define _SWAP SWAP
-#define _TIER2_RESUME_CHECK 449
-#define _TO_BOOL 450
+#define _TIER2_RESUME_CHECK 450
+#define _TO_BOOL 451
 #define _TO_BOOL_BOOL TO_BOOL_BOOL
 #define _TO_BOOL_INT TO_BOOL_INT
 #define _TO_BOOL_LIST TO_BOOL_LIST
@@ -280,13 +281,14 @@ extern "C" {
 #define _UNARY_NEGATIVE UNARY_NEGATIVE
 #define _UNARY_NOT UNARY_NOT
 #define _UNPACK_EX UNPACK_EX
-#define _UNPACK_SEQUENCE 451
+#define _UNPACK_SEQUENCE 452
 #define _UNPACK_SEQUENCE_LIST UNPACK_SEQUENCE_LIST
 #define _UNPACK_SEQUENCE_TUPLE UNPACK_SEQUENCE_TUPLE
 #define _UNPACK_SEQUENCE_TWO_TUPLE UNPACK_SEQUENCE_TWO_TUPLE
 #define _WITH_EXCEPT_START WITH_EXCEPT_START
+#define _YIELD_OFFSET 453
 #define _YIELD_VALUE YIELD_VALUE
-#define MAX_UOP_ID 451
+#define MAX_UOP_ID 453
 
 #ifdef __cplusplus
 }
diff --git a/Include/internal/pycore_uop_metadata.h b/Include/internal/pycore_uop_metadata.h
index 470e95e2b3b041..36911b533244e7 100644
--- a/Include/internal/pycore_uop_metadata.h
+++ b/Include/internal/pycore_uop_metadata.h
@@ -260,6 +260,8 @@ const uint16_t _PyUop_Flags[MAX_UOP_ID+1] = {
     [_DEOPT] = 0,
     [_ERROR_POP_N] = HAS_ARG_FLAG,
     [_TIER2_RESUME_CHECK] = HAS_DEOPT_FLAG,
+    [_RETURN_OFFSET] = 0,
+    [_YIELD_OFFSET] = 0,
 };
 
 const uint8_t _PyUop_Replication[MAX_UOP_ID+1] = {
@@ -464,6 +466,7 @@ const char *const _PyOpcode_uop_name[MAX_UOP_ID+1] = {
     [_REPLACE_WITH_TRUE] = "_REPLACE_WITH_TRUE",
     [_RESUME_CHECK] = "_RESUME_CHECK",
     [_RETURN_GENERATOR] = "_RETURN_GENERATOR",
+    [_RETURN_OFFSET] = "_RETURN_OFFSET",
     [_SAVE_RETURN_OFFSET] = "_SAVE_RETURN_OFFSET",
     [_SETUP_ANNOTATIONS] = "_SETUP_ANNOTATIONS",
     [_SET_ADD] = "_SET_ADD",
@@ -509,6 +512,7 @@ const char *const _PyOpcode_uop_name[MAX_UOP_ID+1] = {
     [_UNPACK_SEQUENCE_TUPLE] = "_UNPACK_SEQUENCE_TUPLE",
     [_UNPACK_SEQUENCE_TWO_TUPLE] = "_UNPACK_SEQUENCE_TWO_TUPLE",
     [_WITH_EXCEPT_START] = "_WITH_EXCEPT_START",
+    [_YIELD_OFFSET] = "_YIELD_OFFSET",
     [_YIELD_VALUE] = "_YIELD_VALUE",
 };
 int _PyUop_num_popped(int opcode, int oparg)
@@ -996,6 +1000,10 @@ int _PyUop_num_popped(int opcode, int oparg)
             return oparg;
         case _TIER2_RESUME_CHECK:
             return 0;
+        case _RETURN_OFFSET:
+            return 0;
+        case _YIELD_OFFSET:
+            return 0;
         default:
             return -1;
     }
diff --git a/Python/bytecodes.c b/Python/bytecodes.c
index b2a0dc030e20cc..2e9b98370133df 100644
--- a/Python/bytecodes.c
+++ b/Python/bytecodes.c
@@ -2589,8 +2589,7 @@ dummy_func(
                     _PyErr_Clear(tstate);
                 }
                 /* iterator ended normally */
-                assert(next_instr[oparg].op.code == END_FOR ||
-                       next_instr[oparg].op.code == INSTRUMENTED_END_FOR);
+                assert(base_opcode(_PyFrame_GetCode(frame), INSTR_OFFSET() + oparg) == END_FOR);
                 Py_DECREF(iter);
                 STACK_SHRINK(1);
                 /* Jump forward oparg, then skip following END_FOR and POP_TOP instruction */
@@ -4351,6 +4350,14 @@ dummy_func(
             assert(tstate->tracing || eval_breaker == FT_ATOMIC_LOAD_UINTPTR_ACQUIRE(_PyFrame_GetCode(frame)->_co_instrumentation_version));
         }
 
+        tier2 op(_RETURN_OFFSET, (--)) {
+            frame->instr_ptr += frame->return_offset;
+        }
+
+        tier2 op(_YIELD_OFFSET, (--)) {
+            frame->instr_ptr += 1 + INLINE_CACHE_ENTRIES_SEND;
+        }
+
 // END BYTECODES //
 
     }
diff --git a/Python/ceval.c b/Python/ceval.c
index 0d02a9887bef7a..e0f8b8e1fc4fa5 100644
--- a/Python/ceval.c
+++ b/Python/ceval.c
@@ -655,6 +655,18 @@ extern const struct _PyCode_DEF(8) _Py_InitCleanup;
 
 #ifdef Py_DEBUG
 extern void _PyUOpPrint(const _PyUOpInstruction *uop);
+
+static int
+base_opcode(PyCodeObject *code, int offset)
+{
+    int opcode = _Py_GetBaseOpcode(code, offset);
+    if (opcode == ENTER_EXECUTOR) {
+        int oparg = _PyCode_CODE(code)[offset].op.arg;
+        _PyExecutorObject *ex = code->co_executors->executors[oparg];
+        return ex->vm_data.opcode;
+    }
+    return opcode;
+}
 #endif
 
 
diff --git a/Python/executor_cases.c.h b/Python/executor_cases.c.h
index 5f15f67324292b..6cf901d2b72ebd 100644
--- a/Python/executor_cases.c.h
+++ b/Python/executor_cases.c.h
@@ -4487,4 +4487,14 @@
             break;
         }
 
+        case _RETURN_OFFSET: {
+            frame->instr_ptr += frame->return_offset;
+            break;
+        }
+
+        case _YIELD_OFFSET: {
+            frame->instr_ptr += 1 + INLINE_CACHE_ENTRIES_SEND;
+            break;
+        }
+
 #undef TIER_TWO
diff --git a/Python/generated_cases.c.h b/Python/generated_cases.c.h
index 87098b0506522f..769527d55bd22b 100644
--- a/Python/generated_cases.c.h
+++ b/Python/generated_cases.c.h
@@ -2788,8 +2788,7 @@
                         _PyErr_Clear(tstate);
                     }
                     /* iterator ended normally */
-                    assert(next_instr[oparg].op.code == END_FOR ||
-                       next_instr[oparg].op.code == INSTRUMENTED_END_FOR);
+                    assert(base_opcode(_PyFrame_GetCode(frame), INSTR_OFFSET() + oparg) == END_FOR);
                     Py_DECREF(iter);
                     STACK_SHRINK(1);
                     /* Jump forward oparg, then skip following END_FOR and POP_TOP instruction */
diff --git a/Python/optimizer.c b/Python/optimizer.c
index 8be2c0ffbd78e9..4ddf3bcb1d4d4b 100644
--- a/Python/optimizer.c
+++ b/Python/optimizer.c
@@ -211,6 +211,9 @@ _PyOptimizer_Optimize(
     _PyInterpreterFrame *frame, _Py_CODEUNIT *start,
     PyObject **stack_pointer, _PyExecutorObject **executor_ptr)
 {
+    if (start->op.code == INTERPRETER_EXIT || start->op.code == EXIT_INIT_CHECK) {
+        return 0;
+    }
     PyCodeObject *code = _PyFrame_GetCode(frame);
     assert(PyCode_Check(code));
     PyInterpreterState *interp = _PyInterpreterState_GET();
@@ -747,17 +750,6 @@ translate_bytecode_to_trace(
                     // Reserve space for nuops (+ _SET_IP + _EXIT_TRACE)
                     int nuops = expansion->nuops;
                     RESERVE(nuops + 1); /* One extra for exit */
-                    int16_t last_op = expansion->uops[nuops-1].uop;
-                    if (last_op == _POP_FRAME || last_op == _RETURN_GENERATOR || last_op == _YIELD_VALUE) {
-                        // Check for trace stack underflow now:
-                        // We can't bail e.g. in the middle of
-                        // LOAD_CONST + _POP_FRAME.
-                        if (trace_stack_depth == 0) {
-                            DPRINTF(2, "Trace stack underflow\n");
-                            OPT_STAT_INC(trace_stack_underflow);
-                            goto done;
-                        }
-                    }
                     uint32_t orig_oparg = oparg;  // For OPARG_TOP/BOTTOM
                     for (int i = 0; i < nuops; i++) {
                         oparg = orig_oparg;
@@ -811,6 +803,19 @@ translate_bytecode_to_trace(
                         }
 
                         if (uop == _POP_FRAME || uop == _RETURN_GENERATOR || uop == _YIELD_VALUE) {
+                            if (trace_stack_depth == 0) {
+                                DPRINTF(2, "Trace stack underflow\n");
+                                OPT_STAT_INC(trace_stack_underflow);
+                                ADD_TO_TRACE(uop, oparg, 0, target);
+                                if (uop == _POP_FRAME || uop == _RETURN_GENERATOR) {
+                                    ADD_TO_TRACE(_RETURN_OFFSET, 0, 0, 0);
+                                }
+                                else {
+                                    ADD_TO_TRACE(_YIELD_OFFSET, 0, 0, 0);
+                                }
+                                ADD_TO_TRACE(_DYNAMIC_EXIT, 0, 0, 0);
+                                goto done;
+                            }
                             TRACE_STACK_POP();
                             /* Set the operand to the function or code object returned to,
                              * to assist optimization passes. (See _PUSH_FRAME below.)
diff --git a/Python/optimizer_bytecodes.c b/Python/optimizer_bytecodes.c
index 928bc03382b8fb..37166fe475e84f 100644
--- a/Python/optimizer_bytecodes.c
+++ b/Python/optimizer_bytecodes.c
@@ -641,7 +641,9 @@ dummy_func(void) {
     op(_POP_FRAME, (retval -- res)) {
         SYNC_SP();
         ctx->frame->stack_pointer = stack_pointer;
-        frame_pop(ctx);
+        if (frame_pop(ctx)) {
+            goto done;
+        }
         stack_pointer = ctx->frame->stack_pointer;
         res = retval;
 
@@ -663,7 +665,9 @@ dummy_func(void) {
     op(_RETURN_GENERATOR, ( -- res)) {
         SYNC_SP();
         ctx->frame->stack_pointer = stack_pointer;
-        frame_pop(ctx);
+        if (frame_pop(ctx)) {
+            goto done;
+        }
         stack_pointer = ctx->frame->stack_pointer;
         OUT_OF_SPACE_IF_NULL(res = sym_new_unknown(ctx));
 
diff --git a/Python/optimizer_cases.c.h b/Python/optimizer_cases.c.h
index 2a4efd73d794df..4b6ef160d0c348 100644
--- a/Python/optimizer_cases.c.h
+++ b/Python/optimizer_cases.c.h
@@ -629,7 +629,9 @@
             retval = stack_pointer[-1];
             stack_pointer += -1;
             ctx->frame->stack_pointer = stack_pointer;
-            frame_pop(ctx);
+            if (frame_pop(ctx)) {
+                goto done;
+            }
             stack_pointer = ctx->frame->stack_pointer;
             res = retval;
             /* Stack space handling */
@@ -1904,7 +1906,9 @@
         case _RETURN_GENERATOR: {
             _Py_UopsSymbol *res;
             ctx->frame->stack_pointer = stack_pointer;
-            frame_pop(ctx);
+            if (frame_pop(ctx)) {
+                goto done;
+            }
             stack_pointer = ctx->frame->stack_pointer;
             OUT_OF_SPACE_IF_NULL(res = sym_new_unknown(ctx));
             /* Stack space handling */
@@ -2199,3 +2203,11 @@
             break;
         }
 
+        case _RETURN_OFFSET: {
+            break;
+        }
+
+        case _YIELD_OFFSET: {
+            break;
+        }
+
diff --git a/Python/optimizer_symbols.c b/Python/optimizer_symbols.c
index 4aeb04fe0405d2..3fea7c9b8d109c 100644
--- a/Python/optimizer_symbols.c
+++ b/Python/optimizer_symbols.c
@@ -384,9 +384,11 @@ _Py_uop_frame_pop(_Py_UOpsContext *ctx)
     _Py_UOpsAbstractFrame *frame = ctx->frame;
     ctx->n_consumed = frame->locals;
     ctx->curr_frame_depth--;
-    assert(ctx->curr_frame_depth >= 1);
+    if (ctx->curr_frame_depth == 0) {
+        ctx->frame = NULL;
+        return -1;
+    }
     ctx->frame = &ctx->frames[ctx->curr_frame_depth - 1];
-
     return 0;
 }
 
diff --git a/Python/specialize.c b/Python/specialize.c
index 9ac428c3593f56..2266adb3345183 100644
--- a/Python/specialize.c
+++ b/Python/specialize.c
@@ -2357,7 +2357,8 @@ _Py_Specialize_ForIter(PyObject *iter, _Py_CODEUNIT *instr, int oparg)
     }
     else if (tp == &PyGen_Type && oparg <= SHRT_MAX) {
         assert(instr[oparg + INLINE_CACHE_ENTRIES_FOR_ITER + 1].op.code == END_FOR  ||
-            instr[oparg + INLINE_CACHE_ENTRIES_FOR_ITER + 1].op.code == INSTRUMENTED_END_FOR
+            instr[oparg + INLINE_CACHE_ENTRIES_FOR_ITER + 1].op.code == INSTRUMENTED_END_FOR ||
+            instr[oparg + INLINE_CACHE_ENTRIES_FOR_ITER + 1].op.code == ENTER_EXECUTOR
         );
         if (_PyInterpreterState_GET()->eval_frame) {
             SPECIALIZATION_FAIL(FOR_ITER, SPEC_FAIL_OTHER);