diff --git a/Include/internal/pycore_interp_structs.h b/Include/internal/pycore_interp_structs.h
index 3fe1fdaa1589b6..9ab3c89d78c785 100644
--- a/Include/internal/pycore_interp_structs.h
+++ b/Include/internal/pycore_interp_structs.h
@@ -943,7 +943,7 @@ struct _is {
     struct types_state types;
     struct callable_cache callable_cache;
     PyObject *common_consts[NUM_COMMON_CONSTANTS];
-    bool jit;
+    uint8_t jit;
     bool compiling;
     struct _PyExecutorObject *executor_list_head;
     struct _PyExecutorObject *executor_deletion_list_head;
diff --git a/Include/internal/pycore_optimizer.h b/Include/internal/pycore_optimizer.h
index 6a0fc1a59e7965..808cbe49f52385 100644
--- a/Include/internal/pycore_optimizer.h
+++ b/Include/internal/pycore_optimizer.h
@@ -47,6 +47,8 @@ typedef struct _PyExitData {
 typedef struct _PyExecutorObject {
     PyObject_VAR_HEAD
     const _PyUOpInstruction *trace;
+    // The interpreter this executor belongs to.
+    PyInterpreterState *interp;
     _PyVMData vm_data; /* Used by the VM, but opaque to the optimizer */
     uint32_t exit_count;
     uint32_t code_size;
@@ -356,8 +358,6 @@ _PyJit_TryInitializeTracing(PyThreadState *tstate, _PyInterpreterFrame *frame,
 
 void _PyJit_FinalizeTracing(PyThreadState *tstate);
 
-void _PyJit_Tracer_InvalidateDependency(PyThreadState *old_tstate, void *obj);
-
 #ifdef __cplusplus
 }
 #endif
diff --git a/Lib/test/test_capi/test_opt.py b/Lib/test/test_capi/test_opt.py
index ea1606fd5b5f05..19c8f4e244a97b 100644
--- a/Lib/test/test_capi/test_opt.py
+++ b/Lib/test/test_capi/test_opt.py
@@ -9,9 +9,13 @@
 
 import _opcode
 
-from test.support import (script_helper, requires_specialization,
-                          import_helper, Py_GIL_DISABLED, requires_jit_enabled,
-                          reset_code)
+from test.support import (
+    script_helper,
+    import_helper,
+    Py_GIL_DISABLED,
+    requires_jit_enabled,
+    reset_code
+)
 
 _testinternalcapi = import_helper.import_module("_testinternalcapi")
 
@@ -61,8 +65,6 @@ def get_opnames(ex):
     return list(iter_opnames(ex))
 
 
-@requires_specialization
-@unittest.skipIf(Py_GIL_DISABLED, "optimizer not yet supported in free-threaded builds")
 @requires_jit_enabled
 class TestExecutorInvalidation(unittest.TestCase):
 
@@ -130,8 +132,6 @@ def f():
         self.assertIsNone(exe)
 
 
-@requires_specialization
-@unittest.skipIf(Py_GIL_DISABLED, "optimizer not yet supported in free-threaded builds")
 @requires_jit_enabled
 @unittest.skipIf(os.getenv("PYTHON_UOPS_OPTIMIZE") == "0", "Needs uop optimizer to run.")
 class TestUops(unittest.TestCase):
@@ -434,8 +434,6 @@ def testfunc(n, m):
         self.assertIn("_FOR_ITER_TIER_TWO", uops)
 
 
-@requires_specialization
-@unittest.skipIf(Py_GIL_DISABLED, "optimizer not yet supported in free-threaded builds")
 @requires_jit_enabled
 @unittest.skipIf(os.getenv("PYTHON_UOPS_OPTIMIZE") == "0", "Needs uop optimizer to run.")
 class TestUopsOptimization(unittest.TestCase):
@@ -2138,6 +2136,7 @@ def testfunc(n):
         self.assertNotIn("_GUARD_TOS_INT", uops)
         self.assertIn("_POP_TOP_NOP", uops)
 
+    @unittest.skipIf(Py_GIL_DISABLED, "FT build immortalizes constants")
     def test_call_len_known_length_small_int(self):
         # Make sure that len(t) is optimized for a tuple of length 5.
         # See https://github.com/python/cpython/issues/139393.
@@ -2162,6 +2161,7 @@ def testfunc(n):
         self.assertNotIn("_POP_CALL_LOAD_CONST_INLINE_BORROW", uops)
         self.assertNotIn("_POP_TOP_LOAD_CONST_INLINE_BORROW", uops)
 
+    @unittest.skipIf(Py_GIL_DISABLED, "FT build immortalizes constants")
     def test_call_len_known_length(self):
         # Make sure that len(t) is not optimized for a tuple of length 2048.
         # See https://github.com/python/cpython/issues/139393.
@@ -2848,6 +2848,7 @@ def testfunc(n):
 
         self.assertIn("_POP_TOP_NOP", uops)
 
+    @unittest.skipIf(Py_GIL_DISABLED, "FT might immortalize this.")
     def test_pop_top_specialize_int(self):
         def testfunc(n):
             for _ in range(n):
@@ -2861,6 +2862,7 @@ def testfunc(n):
 
         self.assertIn("_POP_TOP_INT", uops)
 
+    @unittest.skipIf(Py_GIL_DISABLED, "FT might immortalize this.")
     def test_pop_top_specialize_float(self):
         def testfunc(n):
             for _ in range(n):
diff --git a/Misc/NEWS.d/next/Core_and_Builtins/2025-11-15-16-30-46.gh-issue-141594.PSsC5J.rst b/Misc/NEWS.d/next/Core_and_Builtins/2025-11-15-16-30-46.gh-issue-141594.PSsC5J.rst
new file mode 100644
index 00000000000000..5db6330ca540b0
--- /dev/null
+++ b/Misc/NEWS.d/next/Core_and_Builtins/2025-11-15-16-30-46.gh-issue-141594.PSsC5J.rst
@@ -0,0 +1,3 @@
+Add free-threading support to the JIT. The JIT is only enabled on
+single-threaded code in free-threading, and is disabled when multiple
+threads are spawned. Patch by Ken Jin.
diff --git a/Modules/_testinternalcapi.c b/Modules/_testinternalcapi.c
index a7fbb0f87b6e9c..3afd7c630cd140 100644
--- a/Modules/_testinternalcapi.c
+++ b/Modules/_testinternalcapi.c
@@ -1240,8 +1240,7 @@ add_executor_dependency(PyObject *self, PyObject *args)
 static PyObject *
 invalidate_executors(PyObject *self, PyObject *obj)
 {
-    PyInterpreterState *interp = PyInterpreterState_Get();
-    _Py_Executors_InvalidateDependency(interp, obj, 1);
+    _Py_Executors_InvalidateDependency(_PyInterpreterState_GET(), obj, 1);
     Py_RETURN_NONE;
 }
 
diff --git a/Objects/codeobject.c b/Objects/codeobject.c
index 3aea2038fd17e7..b61d86d110fb1b 100644
--- a/Objects/codeobject.c
+++ b/Objects/codeobject.c
@@ -2432,7 +2432,7 @@ code_dealloc(PyObject *self)
         PyMem_Free(co_extra);
     }
 #ifdef _Py_TIER2
-    _PyJit_Tracer_InvalidateDependency(tstate, self);
+    _Py_Executors_InvalidateDependency(tstate->interp, self, 1);
     if (co->co_executors != NULL) {
         clear_executors(co);
     }
@@ -3363,8 +3363,12 @@ deopt_code_unit(PyCodeObject *code, int i)
         inst.op.code = _PyOpcode_Deopt[opcode];
         assert(inst.op.code < MIN_SPECIALIZED_OPCODE);
     }
-    // JIT should not be enabled with free-threading
-    assert(inst.op.code != ENTER_EXECUTOR);
+    if (inst.op.code == ENTER_EXECUTOR) {
+        _PyExecutorObject *exec = code->co_executors->executors[inst.op.arg];
+        assert(exec != NULL);
+        inst.op.code = exec->vm_data.opcode;
+        inst.op.arg = exec->vm_data.oparg;
+    }
     return inst;
 }
 
diff --git a/Objects/frameobject.c b/Objects/frameobject.c
index 3c0b454503be66..086ac5b1d6a178 100644
--- a/Objects/frameobject.c
+++ b/Objects/frameobject.c
@@ -262,7 +262,6 @@ framelocalsproxy_setitem(PyObject *self, PyObject *key, PyObject *value)
 
 #if _Py_TIER2
         _Py_Executors_InvalidateDependency(_PyInterpreterState_GET(), co, 1);
-        _PyJit_Tracer_InvalidateDependency(_PyThreadState_GET(), co);
 #endif
 
         _PyLocals_Kind kind = _PyLocals_GetKind(co->co_localspluskinds, i);
diff --git a/Objects/funcobject.c b/Objects/funcobject.c
index b659ac8023373b..e11db261bc23f3 100644
--- a/Objects/funcobject.c
+++ b/Objects/funcobject.c
@@ -11,7 +11,7 @@
 #include "pycore_setobject.h"     // _PySet_NextEntry()
 #include "pycore_stats.h"
 #include "pycore_weakref.h"       // FT_CLEAR_WEAKREFS()
-#include "pycore_optimizer.h"     // _PyJit_Tracer_InvalidateDependency
+#include "pycore_optimizer.h"     // _Py_Executors_InvalidateDependency
 
 static const char *
 func_event_name(PyFunction_WatchEvent event) {
@@ -298,7 +298,7 @@ functions is running.
 
 */
 
-#ifndef Py_GIL_DISABLED
+#if _Py_TIER2
 static inline struct _func_version_cache_item *
 get_cache_item(PyInterpreterState *interp, uint32_t version)
 {
@@ -315,11 +315,13 @@ _PyFunction_SetVersion(PyFunctionObject *func, uint32_t version)
     // This should only be called from MAKE_FUNCTION. No code is specialized
     // based on the version, so we do not need to stop the world to set it.
     func->func_version = version;
-#ifndef Py_GIL_DISABLED
+#if _Py_TIER2
     PyInterpreterState *interp = _PyInterpreterState_GET();
+    FT_MUTEX_LOCK(&interp->func_state.mutex);
     struct _func_version_cache_item *slot = get_cache_item(interp, version);
     slot->func = func;
     slot->code = func->func_code;
+    FT_MUTEX_UNLOCK(&interp->func_state.mutex);
 #endif
 }
 
@@ -330,13 +332,15 @@ func_clear_version(PyInterpreterState *interp, PyFunctionObject *func)
         // Version was never set or has already been cleared.
         return;
     }
-#ifndef Py_GIL_DISABLED
+#if _Py_TIER2
+    FT_MUTEX_LOCK(&interp->func_state.mutex);
     struct _func_version_cache_item *slot =
         get_cache_item(interp, func->func_version);
     if (slot->func == func) {
         slot->func = NULL;
         // Leave slot->code alone, there may be use for it.
     }
+    FT_MUTEX_UNLOCK(&interp->func_state.mutex);
 #endif
     func->func_version = FUNC_VERSION_CLEARED;
 }
@@ -358,8 +362,9 @@ _PyFunction_ClearVersion(PyFunctionObject *func)
 void
 _PyFunction_ClearCodeByVersion(uint32_t version)
 {
-#ifndef Py_GIL_DISABLED
+#if _Py_TIER2
     PyInterpreterState *interp = _PyInterpreterState_GET();
+    FT_MUTEX_LOCK(&interp->func_state.mutex);
     struct _func_version_cache_item *slot = get_cache_item(interp, version);
     if (slot->code) {
         assert(PyCode_Check(slot->code));
@@ -369,15 +374,17 @@ _PyFunction_ClearCodeByVersion(uint32_t version)
             slot->func = NULL;
         }
     }
+    FT_MUTEX_UNLOCK(&interp->func_state.mutex);
 #endif
 }
 
 PyFunctionObject *
 _PyFunction_LookupByVersion(uint32_t version, PyObject **p_code)
 {
-#ifdef Py_GIL_DISABLED
-    return NULL;
-#else
+#if _Py_TIER2
+    // This function does not need locking/atomics as it can only be
+    // called from the optimizer, which is currently disabled
+    // when there are multiple threads.
     PyInterpreterState *interp = _PyInterpreterState_GET();
     struct _func_version_cache_item *slot = get_cache_item(interp, version);
     if (slot->code) {
@@ -395,12 +402,18 @@ _PyFunction_LookupByVersion(uint32_t version, PyObject **p_code)
         return slot->func;
     }
     return NULL;
+#else
+    return NULL;
 #endif
 }
 
 uint32_t
 _PyFunction_GetVersionForCurrentState(PyFunctionObject *func)
 {
+    // This function does not need locking/atomics as it can only be
+    // called from the specializing interpreter or optimizer.
+    // The specializing interpreter holds a strong reference to the function.
+    // The optimizer is currently disabled when there are multiple threads.
     return func->func_version;
 }
 
@@ -1153,7 +1166,6 @@ func_dealloc(PyObject *self)
     }
 #if _Py_TIER2
     _Py_Executors_InvalidateDependency(_PyInterpreterState_GET(), self, 1);
-    _PyJit_Tracer_InvalidateDependency(_PyThreadState_GET(), self);
 #endif
     _PyObject_GC_UNTRACK(op);
     FT_CLEAR_WEAKREFS(self, op->func_weakreflist);
diff --git a/Objects/listobject.c b/Objects/listobject.c
index 4a98c8e54ab03f..20092e122cafc1 100644
--- a/Objects/listobject.c
+++ b/Objects/listobject.c
@@ -79,7 +79,9 @@ ensure_shared_on_resize(PyListObject *self)
     // We can't use _Py_CRITICAL_SECTION_ASSERT_OBJECT_LOCKED here because
     // the `CALL_LIST_APPEND` bytecode handler may lock the list without
     // a critical section.
-    assert(Py_REFCNT(self) == 1 || PyMutex_IsLocked(&_PyObject_CAST(self)->ob_mutex));
+    assert(Py_REFCNT(self) == 1 ||
+        (_Py_IsOwnedByCurrentThread((PyObject *)self) && !_PyObject_GC_IS_SHARED(self)) ||
+        PyMutex_IsLocked(&_PyObject_CAST(self)->ob_mutex));
 
     // Ensure that the list array is freed using QSBR if we are not the
     // owning thread.
diff --git a/Objects/typeobject.c b/Objects/typeobject.c
index 7f5149aeece12b..43b45689eaa477 100644
--- a/Objects/typeobject.c
+++ b/Objects/typeobject.c
@@ -1148,8 +1148,9 @@ static void
 set_version_unlocked(PyTypeObject *tp, unsigned int version)
 {
     assert(version == 0 || (tp->tp_versions_used != _Py_ATTR_CACHE_UNUSED));
-#ifndef Py_GIL_DISABLED
+#if _Py_TIER2
     PyInterpreterState *interp = _PyInterpreterState_GET();
+    BEGIN_TYPE_LOCK();
     // lookup the old version and set to null
     if (tp->tp_version_tag != 0) {
         PyTypeObject **slot =
@@ -1157,6 +1158,8 @@ set_version_unlocked(PyTypeObject *tp, unsigned int version)
             + (tp->tp_version_tag % TYPE_VERSION_CACHE_SIZE);
         *slot = NULL;
     }
+#endif
+#ifndef Py_GIL_DISABLED
     if (version) {
         tp->tp_versions_used++;
     }
@@ -1166,13 +1169,14 @@ set_version_unlocked(PyTypeObject *tp, unsigned int version)
     }
 #endif
     FT_ATOMIC_STORE_UINT_RELAXED(tp->tp_version_tag, version);
-#ifndef Py_GIL_DISABLED
+#if _Py_TIER2
     if (version != 0) {
         PyTypeObject **slot =
             interp->types.type_version_cache
             + (version % TYPE_VERSION_CACHE_SIZE);
         *slot = tp;
     }
+    END_TYPE_LOCK();
 #endif
 }
 
@@ -1357,9 +1361,12 @@ _PyType_SetVersion(PyTypeObject *tp, unsigned int version)
 PyTypeObject *
 _PyType_LookupByVersion(unsigned int version)
 {
-#ifdef Py_GIL_DISABLED
+#ifndef _Py_TIER2
     return NULL;
 #else
+    // This function does not need locking/atomics as it can only be
+    // called from the optimizer, which is currently disabled
+    // when there are multiple threads.
     PyInterpreterState *interp = _PyInterpreterState_GET();
     PyTypeObject **slot =
         interp->types.type_version_cache
diff --git a/Python/bytecodes.c b/Python/bytecodes.c
index ea09c0645aa39c..04b743ac6829fd 100644
--- a/Python/bytecodes.c
+++ b/Python/bytecodes.c
@@ -2937,9 +2937,9 @@ dummy_func(
         };
 
         specializing tier1 op(_SPECIALIZE_JUMP_BACKWARD, (--)) {
-        #if ENABLE_SPECIALIZATION
+        #if ENABLE_SPECIALIZATION_FT
             if (this_instr->op.code == JUMP_BACKWARD) {
-                uint8_t desired = tstate->interp->jit ? JUMP_BACKWARD_JIT : JUMP_BACKWARD_NO_JIT;
+                uint8_t desired = FT_ATOMIC_LOAD_UINT8(tstate->interp->jit) ? JUMP_BACKWARD_JIT : JUMP_BACKWARD_NO_JIT;
                 FT_ATOMIC_STORE_UINT8_RELAXED(this_instr->op.code, desired);
                 // Need to re-dispatch so the warmup counter isn't off by one:
                 next_instr = this_instr;
@@ -3282,11 +3282,9 @@ dummy_func(
 
         // Only used by Tier 2
         op(_GUARD_NOT_EXHAUSTED_LIST, (iter, null_or_index -- iter, null_or_index)) {
-#ifndef Py_GIL_DISABLED
             PyObject *list_o = PyStackRef_AsPyObjectBorrow(iter);
             assert(Py_TYPE(list_o) == &PyList_Type);
             EXIT_IF((size_t)PyStackRef_UntagInt(null_or_index) >= (size_t)PyList_GET_SIZE(list_o));
-#endif
         }
 
         replaced op(_ITER_NEXT_LIST, (iter, null_or_index -- iter, null_or_index, next)) {
@@ -5203,6 +5201,19 @@ dummy_func(
         }
 
         tier2 op(_CHECK_VALIDITY, (--)) {
+            // For FT:
+            // This doesn't need atomics (for now) as there is only a single time
+            // where a write from another thread is possible:
+            // when a new thread is spawned and it invalidates all current
+            // executors.
+            // The new thread can only be created by an executing uop prior to the
+            // _CHECK_VALIDITY check. New thread creation is synchronized by
+            // locking of the runtime, and the current thread is naturally
+            // paused/waiting for the new thread to be created. Thus,
+            // there is a strict happens-before relation between that
+            // uop's invalidation of validity and this check.
+            // So for now, while the JIT does not run on multiple threads,
+            // it is safe for this to be non-atomic.
             DEOPT_IF(!current_executor->vm_data.valid);
         }
 
diff --git a/Python/ceval.c b/Python/ceval.c
index 924afaa97443cb..040ffbd8fbb378 100644
--- a/Python/ceval.c
+++ b/Python/ceval.c
@@ -1719,7 +1719,7 @@ _PyTier2Interpreter(
     for (;;) {
         uopcode = next_uop->opcode;
 #ifdef Py_DEBUG
-        if (frame->lltrace >= 3) {
+        if (frame->lltrace >= 4) {
             dump_stack(frame, stack_pointer);
             printf("    cache=[");
             dump_cache_item(_tos_cache0, 0, current_cached_values);
diff --git a/Python/executor_cases.c.h b/Python/executor_cases.c.h
index 4e8fa34c0b2c0d..ea9fb0b20ac347 100644
--- a/Python/executor_cases.c.h
+++ b/Python/executor_cases.c.h
@@ -10322,7 +10322,6 @@
             _PyStackRef iter;
             null_or_index = stack_pointer[-1];
             iter = stack_pointer[-2];
-            #ifndef Py_GIL_DISABLED
             PyObject *list_o = PyStackRef_AsPyObjectBorrow(iter);
             assert(Py_TYPE(list_o) == &PyList_Type);
             if ((size_t)PyStackRef_UntagInt(null_or_index) >= (size_t)PyList_GET_SIZE(list_o)) {
@@ -10330,7 +10329,6 @@
                 SET_CURRENT_CACHED_VALUES(0);
                 JUMP_TO_JUMP_TARGET();
             }
-            #endif
             _tos_cache1 = null_or_index;
             _tos_cache0 = iter;
             SET_CURRENT_CACHED_VALUES(2);
@@ -10348,7 +10346,6 @@
             _PyStackRef _stack_item_0 = _tos_cache0;
             null_or_index = _stack_item_0;
             iter = stack_pointer[-1];
-            #ifndef Py_GIL_DISABLED
             PyObject *list_o = PyStackRef_AsPyObjectBorrow(iter);
             assert(Py_TYPE(list_o) == &PyList_Type);
             if ((size_t)PyStackRef_UntagInt(null_or_index) >= (size_t)PyList_GET_SIZE(list_o)) {
@@ -10357,7 +10354,6 @@
                 SET_CURRENT_CACHED_VALUES(1);
                 JUMP_TO_JUMP_TARGET();
             }
-            #endif
             _tos_cache1 = null_or_index;
             _tos_cache0 = iter;
             SET_CURRENT_CACHED_VALUES(2);
@@ -10376,7 +10372,6 @@
             _PyStackRef _stack_item_1 = _tos_cache1;
             null_or_index = _stack_item_1;
             iter = _stack_item_0;
-            #ifndef Py_GIL_DISABLED
             PyObject *list_o = PyStackRef_AsPyObjectBorrow(iter);
             assert(Py_TYPE(list_o) == &PyList_Type);
             if ((size_t)PyStackRef_UntagInt(null_or_index) >= (size_t)PyList_GET_SIZE(list_o)) {
@@ -10386,7 +10381,6 @@
                 SET_CURRENT_CACHED_VALUES(2);
                 JUMP_TO_JUMP_TARGET();
             }
-            #endif
             _tos_cache1 = null_or_index;
             _tos_cache0 = iter;
             SET_CURRENT_CACHED_VALUES(2);
@@ -10404,7 +10398,6 @@
             _PyStackRef _stack_item_2 = _tos_cache2;
             null_or_index = _stack_item_2;
             iter = _stack_item_1;
-            #ifndef Py_GIL_DISABLED
             PyObject *list_o = PyStackRef_AsPyObjectBorrow(iter);
             assert(Py_TYPE(list_o) == &PyList_Type);
             if ((size_t)PyStackRef_UntagInt(null_or_index) >= (size_t)PyList_GET_SIZE(list_o)) {
@@ -10415,7 +10408,6 @@
                 SET_CURRENT_CACHED_VALUES(3);
                 JUMP_TO_JUMP_TARGET();
             }
-            #endif
             _tos_cache2 = null_or_index;
             _tos_cache1 = iter;
             _tos_cache0 = _stack_item_0;
diff --git a/Python/generated_cases.c.h b/Python/generated_cases.c.h
index e63852aee1134c..62eb10984cb239 100644
--- a/Python/generated_cases.c.h
+++ b/Python/generated_cases.c.h
@@ -7247,9 +7247,9 @@
             /* Skip 1 cache entry */
             // _SPECIALIZE_JUMP_BACKWARD
             {
-                #if ENABLE_SPECIALIZATION
+                #if ENABLE_SPECIALIZATION_FT
                 if (this_instr->op.code == JUMP_BACKWARD) {
-                    uint8_t desired = tstate->interp->jit ? JUMP_BACKWARD_JIT : JUMP_BACKWARD_NO_JIT;
+                    uint8_t desired = FT_ATOMIC_LOAD_UINT8(tstate->interp->jit) ? JUMP_BACKWARD_JIT : JUMP_BACKWARD_NO_JIT;
                     FT_ATOMIC_STORE_UINT8_RELAXED(this_instr->op.code, desired);
                     next_instr = this_instr;
                     DISPATCH_SAME_OPARG();
diff --git a/Python/instrumentation.c b/Python/instrumentation.c
index 28bbe1d82a3b88..632b17632ef0ed 100644
--- a/Python/instrumentation.c
+++ b/Python/instrumentation.c
@@ -1785,7 +1785,6 @@ force_instrument_lock_held(PyCodeObject *code, PyInterpreterState *interp)
         _PyCode_Clear_Executors(code);
     }
     _Py_Executors_InvalidateDependency(interp, code, 1);
-    _PyJit_Tracer_InvalidateDependency(PyThreadState_GET(), code);
 #endif
     int code_len = (int)Py_SIZE(code);
     /* Exit early to avoid creating instrumentation
diff --git a/Python/optimizer.c b/Python/optimizer.c
index b497ac629960ac..c25f6c070977eb 100644
--- a/Python/optimizer.c
+++ b/Python/optimizer.c
@@ -140,7 +140,6 @@ _PyOptimizer_Optimize(
     }
     assert(!interp->compiling);
     assert(_tstate->jit_tracer_state.initial_state.stack_depth >= 0);
-#ifndef Py_GIL_DISABLED
     assert(_tstate->jit_tracer_state.initial_state.func != NULL);
     interp->compiling = true;
     // The first executor in a chain and the MAX_CHAIN_DEPTH'th executor *must*
@@ -198,9 +197,6 @@ _PyOptimizer_Optimize(
     }
     interp->compiling = false;
     return 1;
-#else
-    return 0;
-#endif
 }
 
 static _PyExecutorObject *
@@ -473,7 +469,11 @@ static PyMethodDef uop_executor_methods[] = {
 static int
 executor_is_gc(PyObject *o)
 {
+#ifdef Py_GIL_DISABLED
+    return 1;
+#else
     return !_Py_IsImmortal(o);
+#endif
 }
 
 PyTypeObject _PyUOpExecutor_Type = {
@@ -1647,8 +1647,11 @@ unlink_executor(_PyExecutorObject *executor)
         prev->vm_data.links.next = next;
     }
     else {
-        // prev == NULL implies that executor is the list head
-        PyInterpreterState *interp = PyInterpreterState_Get();
+        // prev == NULL often implies that executor is the list head
+        // Note that we should *not* get the current interpreter, as
+        // that may not always correspond to the interpreter this executor
+        // belongs to.
+        PyInterpreterState *interp = executor->interp;
         assert(interp->executor_list_head == executor);
         interp->executor_list_head = next;
     }
@@ -1663,6 +1666,7 @@ _Py_ExecutorInit(_PyExecutorObject *executor, const _PyBloomFilter *dependency_s
     for (int i = 0; i < _Py_BLOOM_FILTER_WORDS; i++) {
         executor->vm_data.bloom.bits[i] = dependency_set->bits[i];
     }
+    executor->interp = _PyInterpreterState_GET();
     link_executor(executor);
 }
 
@@ -1777,12 +1781,20 @@ _Py_Executor_DependsOn(_PyExecutorObject *executor, void *obj)
     _Py_BloomFilter_Add(&executor->vm_data.bloom, obj);
 }
 
+static void  jit_tracer_invalidate_dependency(PyThreadState *tstate, void *obj);
+
 /* Invalidate all executors that depend on `obj`
  * May cause other executors to be invalidated as well
  */
 void
 _Py_Executors_InvalidateDependency(PyInterpreterState *interp, void *obj, int is_invalidation)
 {
+
+    // It doesn't matter if we don't invalidate all threads.
+    // If more threads are spawned, we force the jit not to compile anyways
+    // so the trace gets abandoned.
+    jit_tracer_invalidate_dependency(_PyThreadState_GET(), obj);
+
     _PyBloomFilter obj_filter;
     _Py_BloomFilter_Init(&obj_filter);
     _Py_BloomFilter_Add(&obj_filter, obj);
@@ -1820,8 +1832,8 @@ _Py_Executors_InvalidateDependency(PyInterpreterState *interp, void *obj, int is
     _Py_Executors_InvalidateAll(interp, is_invalidation);
 }
 
-void
-_PyJit_Tracer_InvalidateDependency(PyThreadState *tstate, void *obj)
+static void
+jit_tracer_invalidate_dependency(PyThreadState *tstate, void *obj)
 {
     _PyBloomFilter obj_filter;
     _Py_BloomFilter_Init(&obj_filter);
@@ -2029,4 +2041,4 @@ _PyExecutor_Free(struct _PyExecutorObject *self)
     Py_UNREACHABLE();
 }
 
-#endif /* _Py_TIER2 */
+#endif /* _Py_TIER2 */
\ No newline at end of file
diff --git a/Python/optimizer_analysis.c b/Python/optimizer_analysis.c
index 81479a8f28e319..318a7813d680ad 100644
--- a/Python/optimizer_analysis.c
+++ b/Python/optimizer_analysis.c
@@ -114,7 +114,7 @@ convert_global_to_const(_PyUOpInstruction *inst, PyObject *obj, bool pop)
     if (res == NULL) {
         return NULL;
     }
-    if (_Py_IsImmortal(res)) {
+    if (_Py_IsImmortal(res) || _PyObject_HasDeferredRefcount(res)) {
         inst->opcode = pop ? _POP_TOP_LOAD_CONST_INLINE_BORROW : _LOAD_CONST_INLINE_BORROW;
     }
     else {
@@ -243,14 +243,14 @@ eliminate_pop_guard(_PyUOpInstruction *this_instr, bool exit)
 
 static JitOptRef
 lookup_attr(JitOptContext *ctx, _PyBloomFilter *dependencies, _PyUOpInstruction *this_instr,
-            PyTypeObject *type, PyObject *name, uint16_t immortal,
+            PyTypeObject *type, PyObject *name, uint16_t deferred_refcount,
             uint16_t mortal)
 {
     // The cached value may be dead, so we need to do the lookup again... :(
     if (type && PyType_Check(type)) {
         PyObject *lookup = _PyType_Lookup(type, name);
         if (lookup) {
-            int opcode = _Py_IsImmortal(lookup) ? immortal : mortal;
+            int opcode = _Py_IsImmortal(lookup) || _PyObject_HasDeferredRefcount(lookup) ? deferred_refcount : mortal;
             REPLACE_OP(this_instr, opcode, 0, (uintptr_t)lookup);
             PyType_Watch(TYPE_WATCHER_ID, (PyObject *)type);
             _Py_BloomFilter_Add(dependencies, type);
diff --git a/Python/pylifecycle.c b/Python/pylifecycle.c
index bb663db195c089..4a30b67814d002 100644
--- a/Python/pylifecycle.c
+++ b/Python/pylifecycle.c
@@ -1364,7 +1364,7 @@ init_interp_main(PyThreadState *tstate)
             } else
 #endif
             {
-                interp->jit = true;
+                interp->jit = 1;
             }
         }
     }
@@ -1721,7 +1721,7 @@ finalize_modules(PyThreadState *tstate)
     PyInterpreterState *interp = tstate->interp;
 
     // Invalidate all executors and turn off JIT:
-    interp->jit = false;
+    FT_ATOMIC_STORE_UINT8(interp->jit, 0);
     interp->compiling = false;
 #ifdef _Py_TIER2
     _Py_Executors_InvalidateAll(interp, 0);
diff --git a/Python/pystate.c b/Python/pystate.c
index cf55297cf8d94e..c6c8f6db7a7aa0 100644
--- a/Python/pystate.c
+++ b/Python/pystate.c
@@ -566,7 +566,7 @@ init_interpreter(PyInterpreterState *interp,
         interp->monitoring_tool_versions[t] = 0;
     }
     interp->_code_object_generation = 0;
-    interp->jit = false;
+    interp->jit = 0;
     interp->compiling = false;
     interp->executor_list_head = NULL;
     interp->executor_deletion_list_head = NULL;
@@ -1545,6 +1545,12 @@ add_threadstate(PyInterpreterState *interp, PyThreadState *tstate,
 {
     assert(interp->threads.head != tstate);
     if (next != NULL) {
+#if defined(_Py_TIER2) && defined(Py_GIL_DISABLED)
+        FT_ATOMIC_STORE_UINT8(interp->jit, 0);
+        // There's more than one thread. In FT mode,
+        // disable the JIT completely for now.
+        _Py_Executors_InvalidateAll(interp, 1);
+#endif
         assert(next->prev == NULL || next->prev == tstate);
         next->prev = tstate;
     }
@@ -1846,6 +1852,13 @@ tstate_delete_common(PyThreadState *tstate, int release_gil)
         _PyObject_VirtualFree(_tstate->jit_tracer_state.code_buffer, UOP_BUFFER_SIZE);
         _tstate->jit_tracer_state.code_buffer = NULL;
     }
+#ifdef Py_GIL_DISABLED
+    // There's only one thread. Re-enable JIT.
+    PyThreadState *curr = interp->threads.head;
+    if (curr != NULL && curr->prev == NULL && curr->next == NULL) {
+       FT_ATOMIC_STORE_UINT8(interp->jit, 1);
+    }
+#endif
 #endif
 
     HEAD_UNLOCK(runtime);
diff --git a/Python/sysmodule.c b/Python/sysmodule.c
index 94eb3164ecad58..ce54bc96221be9 100644
--- a/Python/sysmodule.c
+++ b/Python/sysmodule.c
@@ -2365,7 +2365,7 @@ sys_activate_stack_trampoline_impl(PyObject *module, const char *backend)
 {
 #ifdef PY_HAVE_PERF_TRAMPOLINE
 #ifdef _Py_JIT
-    if (_PyInterpreterState_GET()->jit) {
+    if (FT_ATOMIC_LOAD_UINT8(_PyInterpreterState_GET()->jit)) {
         PyErr_SetString(PyExc_ValueError, "Cannot activate the perf trampoline if the JIT is active");
         return NULL;
     }
diff --git a/Tools/jit/template.c b/Tools/jit/template.c
index 3537c74a820365..90a4668f610275 100644
--- a/Tools/jit/template.c
+++ b/Tools/jit/template.c
@@ -118,6 +118,16 @@ do {                                                                      \
 #define ASSERT_WITHIN_STACK_BOUNDS(F, L) (void)0
 #endif
 
+// For now, the FT JIT only supports single-threaded code.
+#undef LOCK_OBJECT
+#undef UNLOCK_OBJECT
+#define LOCK_OBJECT(op) (1)
+#define UNLOCK_OBJECT(op) ((void)0)
+
+#ifdef Py_GIL_DISABLED
+#undef Py_GIL_DISABLED
+#endif
+
 __attribute__((preserve_none)) _Py_CODEUNIT *
 _JIT_ENTRY(
     _PyExecutorObject *executor, _PyInterpreterFrame *frame, _PyStackRef *stack_pointer, PyThreadState *tstate,