gh-115999: Specialize loading attributes from modules in free-threaded builds (#127711)

We use the same approach that was used for specialization of LOAD_GLOBAL in free-threaded builds:

_CHECK_ATTR_MODULE is renamed to _CHECK_ATTR_MODULE_PUSH_KEYS; it pushes the keys object for the following _LOAD_ATTR_MODULE_FROM_KEYS (nee _LOAD_ATTR_MODULE). This arrangement avoids having to recheck the keys version.

_LOAD_ATTR_MODULE is renamed to _LOAD_ATTR_MODULE_FROM_KEYS; it loads the value from the keys object pushed by the preceding _CHECK_ATTR_MODULE_PUSH_KEYS at the cached index.

7 files changed, 222 insertions, 91 deletions

diff --git a/Python/bytecodes.c b/Python/bytecodes.c
index f0eb540..772b46d 100644
--- a/Python/bytecodes.c
+++ b/Python/bytecodes.c
@@ -2070,7 +2070,7 @@ dummy_func(
         };
 
         specializing op(_SPECIALIZE_LOAD_ATTR, (counter/1, owner -- owner)) {
-            #if ENABLE_SPECIALIZATION
+            #if ENABLE_SPECIALIZATION_FT
             if (ADAPTIVE_COUNTER_TRIGGERS(counter)) {
                 PyObject *name = GETITEM(FRAME_CO_NAMES, oparg>>1);
                 next_instr = this_instr;
@@ -2079,7 +2079,7 @@ dummy_func(
             }
             OPCODE_DEFERRED_INC(LOAD_ATTR);
             ADVANCE_ADAPTIVE_COUNTER(this_instr[1].counter);
-            #endif  /* ENABLE_SPECIALIZATION */
+            #endif  /* ENABLE_SPECIALIZATION_FT */
         }
 
         op(_LOAD_ATTR, (owner -- attr, self_or_null if (oparg & 1))) {
@@ -2158,33 +2158,43 @@ dummy_func(
             _LOAD_ATTR_INSTANCE_VALUE +
             unused/5;  // Skip over rest of cache
 
-        op(_CHECK_ATTR_MODULE, (dict_version/2, owner -- owner)) {
+        op(_CHECK_ATTR_MODULE_PUSH_KEYS, (dict_version/2, owner -- owner, mod_keys: PyDictKeysObject *)) {
             PyObject *owner_o = PyStackRef_AsPyObjectBorrow(owner);
             DEOPT_IF(Py_TYPE(owner_o)->tp_getattro != PyModule_Type.tp_getattro);
             PyDictObject *dict = (PyDictObject *)((PyModuleObject *)owner_o)->md_dict;
             assert(dict != NULL);
-            DEOPT_IF(dict->ma_keys->dk_version != dict_version);
-        }
-
-        op(_LOAD_ATTR_MODULE, (index/1, owner -- attr, null if (oparg & 1))) {
-            PyObject *owner_o = PyStackRef_AsPyObjectBorrow(owner);
-            PyDictObject *dict = (PyDictObject *)((PyModuleObject *)owner_o)->md_dict;
-            assert(dict->ma_keys->dk_kind == DICT_KEYS_UNICODE);
-            assert(index < dict->ma_keys->dk_nentries);
-            PyDictUnicodeEntry *ep = DK_UNICODE_ENTRIES(dict->ma_keys) + index;
-            PyObject *attr_o = ep->me_value;
+            PyDictKeysObject *keys = FT_ATOMIC_LOAD_PTR_ACQUIRE(dict->ma_keys);
+            DEOPT_IF(FT_ATOMIC_LOAD_UINT32_RELAXED(keys->dk_version) != dict_version);
+            mod_keys = keys;
+        }
+
+        op(_LOAD_ATTR_MODULE_FROM_KEYS, (index/1, owner, mod_keys: PyDictKeysObject * -- attr, null if (oparg & 1))) {
+            assert(mod_keys->dk_kind == DICT_KEYS_UNICODE);
+            assert(index < FT_ATOMIC_LOAD_SSIZE_RELAXED(mod_keys->dk_nentries));
+            PyDictUnicodeEntry *ep = DK_UNICODE_ENTRIES(mod_keys) + index;
+            PyObject *attr_o = FT_ATOMIC_LOAD_PTR_RELAXED(ep->me_value);
+            DEAD(mod_keys);
+            // Clear mod_keys from stack in case we need to deopt
+            POP_DEAD_INPUTS();
             DEOPT_IF(attr_o == NULL);
-            STAT_INC(LOAD_ATTR, hit);
+            #ifdef Py_GIL_DISABLED
+            int increfed = _Py_TryIncrefCompareStackRef(&ep->me_value, attr_o, &attr);
+            if (!increfed) {
+                DEOPT_IF(true);
+            }
+            #else
             Py_INCREF(attr_o);
             attr = PyStackRef_FromPyObjectSteal(attr_o);
+            #endif
+            STAT_INC(LOAD_ATTR, hit);
             null = PyStackRef_NULL;
-            DECREF_INPUTS();
+            PyStackRef_CLOSE(owner);
         }
 
         macro(LOAD_ATTR_MODULE) =
             unused/1 +
-            _CHECK_ATTR_MODULE +
-            _LOAD_ATTR_MODULE +
+            _CHECK_ATTR_MODULE_PUSH_KEYS +
+            _LOAD_ATTR_MODULE_FROM_KEYS +
             unused/5;
 
         op(_CHECK_ATTR_WITH_HINT, (owner -- owner)) {
@@ -4963,6 +4973,21 @@ dummy_func(
             null = PyStackRef_NULL;
          }
 
+        tier2 op(_LOAD_ATTR_MODULE, (index/1, owner -- attr, null if (oparg & 1))) {
+            PyObject *owner_o = PyStackRef_AsPyObjectBorrow(owner);
+            PyDictObject *dict = (PyDictObject *)((PyModuleObject *)owner_o)->md_dict;
+            assert(dict->ma_keys->dk_kind == DICT_KEYS_UNICODE);
+            assert(index < dict->ma_keys->dk_nentries);
+            PyDictUnicodeEntry *ep = DK_UNICODE_ENTRIES(dict->ma_keys) + index;
+            PyObject *attr_o = ep->me_value;
+            DEOPT_IF(attr_o == NULL);
+            STAT_INC(LOAD_ATTR, hit);
+            Py_INCREF(attr_o);
+            attr = PyStackRef_FromPyObjectSteal(attr_o);
+            null = PyStackRef_NULL;
+            DECREF_INPUTS();
+        }
+
         /* Internal -- for testing executors */
         op(_INTERNAL_INCREMENT_OPT_COUNTER, (opt --)) {
             _PyCounterOptimizerObject *exe = (_PyCounterOptimizerObject *)PyStackRef_AsPyObjectBorrow(opt);
diff --git a/Python/executor_cases.c.h b/Python/executor_cases.c.h
index 19ba67a..55e9c3a 100644
--- a/Python/executor_cases.c.h
+++ b/Python/executor_cases.c.h
@@ -2641,8 +2641,9 @@
 
         /* _LOAD_ATTR_INSTANCE_VALUE is split on (oparg & 1) */
 
-        case _CHECK_ATTR_MODULE: {
+        case _CHECK_ATTR_MODULE_PUSH_KEYS: {
             _PyStackRef owner;
+            PyDictKeysObject *mod_keys;
             owner = stack_pointer[-1];
             uint32_t dict_version = (uint32_t)CURRENT_OPERAND0();
             PyObject *owner_o = PyStackRef_AsPyObjectBorrow(owner);
@@ -2652,33 +2653,51 @@
             }
             PyDictObject *dict = (PyDictObject *)((PyModuleObject *)owner_o)->md_dict;
             assert(dict != NULL);
-            if (dict->ma_keys->dk_version != dict_version) {
+            PyDictKeysObject *keys = FT_ATOMIC_LOAD_PTR_ACQUIRE(dict->ma_keys);
+            if (FT_ATOMIC_LOAD_UINT32_RELAXED(keys->dk_version) != dict_version) {
                 UOP_STAT_INC(uopcode, miss);
                 JUMP_TO_JUMP_TARGET();
             }
+            mod_keys = keys;
+            stack_pointer[0].bits = (uintptr_t)mod_keys;
+            stack_pointer += 1;
+            assert(WITHIN_STACK_BOUNDS());
             break;
         }
 
-        case _LOAD_ATTR_MODULE: {
+        case _LOAD_ATTR_MODULE_FROM_KEYS: {
+            PyDictKeysObject *mod_keys;
             _PyStackRef owner;
             _PyStackRef attr;
             _PyStackRef null = PyStackRef_NULL;
             oparg = CURRENT_OPARG();
-            owner = stack_pointer[-1];
+            mod_keys = (PyDictKeysObject *)stack_pointer[-1].bits;
+            owner = stack_pointer[-2];
             uint16_t index = (uint16_t)CURRENT_OPERAND0();
-            PyObject *owner_o = PyStackRef_AsPyObjectBorrow(owner);
-            PyDictObject *dict = (PyDictObject *)((PyModuleObject *)owner_o)->md_dict;
-            assert(dict->ma_keys->dk_kind == DICT_KEYS_UNICODE);
-            assert(index < dict->ma_keys->dk_nentries);
-            PyDictUnicodeEntry *ep = DK_UNICODE_ENTRIES(dict->ma_keys) + index;
-            PyObject *attr_o = ep->me_value;
+            assert(mod_keys->dk_kind == DICT_KEYS_UNICODE);
+            assert(index < FT_ATOMIC_LOAD_SSIZE_RELAXED(mod_keys->dk_nentries));
+            PyDictUnicodeEntry *ep = DK_UNICODE_ENTRIES(mod_keys) + index;
+            PyObject *attr_o = FT_ATOMIC_LOAD_PTR_RELAXED(ep->me_value);
+            // Clear mod_keys from stack in case we need to deopt
+            stack_pointer += -1;
+            assert(WITHIN_STACK_BOUNDS());
             if (attr_o == NULL) {
                 UOP_STAT_INC(uopcode, miss);
                 JUMP_TO_JUMP_TARGET();
             }
-            STAT_INC(LOAD_ATTR, hit);
+            #ifdef Py_GIL_DISABLED
+            int increfed = _Py_TryIncrefCompareStackRef(&ep->me_value, attr_o, &attr);
+            if (!increfed) {
+                if (true) {
+                    UOP_STAT_INC(uopcode, miss);
+                    JUMP_TO_JUMP_TARGET();
+                }
+            }
+            #else
             Py_INCREF(attr_o);
             attr = PyStackRef_FromPyObjectSteal(attr_o);
+            #endif
+            STAT_INC(LOAD_ATTR, hit);
             null = PyStackRef_NULL;
             PyStackRef_CLOSE(owner);
             stack_pointer[-1] = attr;
@@ -5928,6 +5947,35 @@
             break;
         }
 
+        case _LOAD_ATTR_MODULE: {
+            _PyStackRef owner;
+            _PyStackRef attr;
+            _PyStackRef null = PyStackRef_NULL;
+            oparg = CURRENT_OPARG();
+            owner = stack_pointer[-1];
+            uint16_t index = (uint16_t)CURRENT_OPERAND0();
+            PyObject *owner_o = PyStackRef_AsPyObjectBorrow(owner);
+            PyDictObject *dict = (PyDictObject *)((PyModuleObject *)owner_o)->md_dict;
+            assert(dict->ma_keys->dk_kind == DICT_KEYS_UNICODE);
+            assert(index < dict->ma_keys->dk_nentries);
+            PyDictUnicodeEntry *ep = DK_UNICODE_ENTRIES(dict->ma_keys) + index;
+            PyObject *attr_o = ep->me_value;
+            if (attr_o == NULL) {
+                UOP_STAT_INC(uopcode, miss);
+                JUMP_TO_JUMP_TARGET();
+            }
+            STAT_INC(LOAD_ATTR, hit);
+            Py_INCREF(attr_o);
+            attr = PyStackRef_FromPyObjectSteal(attr_o);
+            null = PyStackRef_NULL;
+            PyStackRef_CLOSE(owner);
+            stack_pointer[-1] = attr;
+            if (oparg & 1) stack_pointer[0] = null;
+            stack_pointer += (oparg & 1);
+            assert(WITHIN_STACK_BOUNDS());
+            break;
+        }
+
         case _INTERNAL_INCREMENT_OPT_COUNTER: {
             _PyStackRef opt;
             opt = stack_pointer[-1];
diff --git a/Python/generated_cases.c.h b/Python/generated_cases.c.h
index 51227c9..94343f9 100644
--- a/Python/generated_cases.c.h
+++ b/Python/generated_cases.c.h
@@ -5200,7 +5200,7 @@
                 owner = stack_pointer[-1];
                 uint16_t counter = read_u16(&this_instr[1].cache);
                 (void)counter;
-                #if ENABLE_SPECIALIZATION
+                #if ENABLE_SPECIALIZATION_FT
                 if (ADAPTIVE_COUNTER_TRIGGERS(counter)) {
                     PyObject *name = GETITEM(FRAME_CO_NAMES, oparg>>1);
                     next_instr = this_instr;
@@ -5211,7 +5211,7 @@
                 }
                 OPCODE_DEFERRED_INC(LOAD_ATTR);
                 ADVANCE_ADAPTIVE_COUNTER(this_instr[1].counter);
-                #endif  /* ENABLE_SPECIALIZATION */
+                #endif  /* ENABLE_SPECIALIZATION_FT */
             }
             /* Skip 8 cache entries */
             // _LOAD_ATTR
@@ -5553,10 +5553,11 @@
             INSTRUCTION_STATS(LOAD_ATTR_MODULE);
             static_assert(INLINE_CACHE_ENTRIES_LOAD_ATTR == 9, "incorrect cache size");
             _PyStackRef owner;
+            PyDictKeysObject *mod_keys;
             _PyStackRef attr;
             _PyStackRef null = PyStackRef_NULL;
             /* Skip 1 cache entry */
-            // _CHECK_ATTR_MODULE
+            // _CHECK_ATTR_MODULE_PUSH_KEYS
             {
                 owner = stack_pointer[-1];
                 uint32_t dict_version = read_u32(&this_instr[2].cache);
@@ -5564,21 +5565,29 @@
                 DEOPT_IF(Py_TYPE(owner_o)->tp_getattro != PyModule_Type.tp_getattro, LOAD_ATTR);
                 PyDictObject *dict = (PyDictObject *)((PyModuleObject *)owner_o)->md_dict;
                 assert(dict != NULL);
-                DEOPT_IF(dict->ma_keys->dk_version != dict_version, LOAD_ATTR);
+                PyDictKeysObject *keys = FT_ATOMIC_LOAD_PTR_ACQUIRE(dict->ma_keys);
+                DEOPT_IF(FT_ATOMIC_LOAD_UINT32_RELAXED(keys->dk_version) != dict_version, LOAD_ATTR);
+                mod_keys = keys;
             }
-            // _LOAD_ATTR_MODULE
+            // _LOAD_ATTR_MODULE_FROM_KEYS
             {
                 uint16_t index = read_u16(&this_instr[4].cache);
-                PyObject *owner_o = PyStackRef_AsPyObjectBorrow(owner);
-                PyDictObject *dict = (PyDictObject *)((PyModuleObject *)owner_o)->md_dict;
-                assert(dict->ma_keys->dk_kind == DICT_KEYS_UNICODE);
-                assert(index < dict->ma_keys->dk_nentries);
-                PyDictUnicodeEntry *ep = DK_UNICODE_ENTRIES(dict->ma_keys) + index;
-                PyObject *attr_o = ep->me_value;
+                assert(mod_keys->dk_kind == DICT_KEYS_UNICODE);
+                assert(index < FT_ATOMIC_LOAD_SSIZE_RELAXED(mod_keys->dk_nentries));
+                PyDictUnicodeEntry *ep = DK_UNICODE_ENTRIES(mod_keys) + index;
+                PyObject *attr_o = FT_ATOMIC_LOAD_PTR_RELAXED(ep->me_value);
+                // Clear mod_keys from stack in case we need to deopt
                 DEOPT_IF(attr_o == NULL, LOAD_ATTR);
-                STAT_INC(LOAD_ATTR, hit);
+                #ifdef Py_GIL_DISABLED
+                int increfed = _Py_TryIncrefCompareStackRef(&ep->me_value, attr_o, &attr);
+                if (!increfed) {
+                    DEOPT_IF(true, LOAD_ATTR);
+                }
+                #else
                 Py_INCREF(attr_o);
                 attr = PyStackRef_FromPyObjectSteal(attr_o);
+                #endif
+                STAT_INC(LOAD_ATTR, hit);
                 null = PyStackRef_NULL;
                 PyStackRef_CLOSE(owner);
             }
diff --git a/Python/optimizer_analysis.c b/Python/optimizer_analysis.c
index a4a0472..0ef15c6 100644
--- a/Python/optimizer_analysis.c
+++ b/Python/optimizer_analysis.c
@@ -95,7 +95,7 @@ type_watcher_callback(PyTypeObject* type)
 static PyObject *
 convert_global_to_const(_PyUOpInstruction *inst, PyObject *obj)
 {
-    assert(inst->opcode == _LOAD_GLOBAL_MODULE || inst->opcode == _LOAD_GLOBAL_BUILTINS || inst->opcode == _LOAD_ATTR_MODULE);
+    assert(inst->opcode == _LOAD_GLOBAL_MODULE || inst->opcode == _LOAD_GLOBAL_BUILTINS || inst->opcode == _LOAD_ATTR_MODULE_FROM_KEYS);
     assert(PyDict_CheckExact(obj));
     PyDictObject *dict = (PyDictObject *)obj;
     assert(dict->ma_keys->dk_kind == DICT_KEYS_UNICODE);
diff --git a/Python/optimizer_bytecodes.c b/Python/optimizer_bytecodes.c
index 42bdbd9..0b8aff0 100644
--- a/Python/optimizer_bytecodes.c
+++ b/Python/optimizer_bytecodes.c
@@ -492,8 +492,9 @@ dummy_func(void) {
         (void)owner;
     }
 
-    op(_CHECK_ATTR_MODULE, (dict_version/2, owner -- owner)) {
+    op(_CHECK_ATTR_MODULE_PUSH_KEYS, (dict_version/2, owner -- owner, mod_keys)) {
         (void)dict_version;
+        mod_keys = sym_new_not_null(ctx);
         if (sym_is_const(owner)) {
             PyObject *cnst = sym_get_const(owner);
             if (PyModule_CheckExact(cnst)) {
@@ -515,12 +516,12 @@ dummy_func(void) {
         self_or_null = sym_new_unknown(ctx);
     }
 
-    op(_LOAD_ATTR_MODULE, (index/1, owner -- attr, null if (oparg & 1))) {
+    op(_LOAD_ATTR_MODULE_FROM_KEYS, (index/1, owner, mod_keys -- attr, null if (oparg & 1))) {
         (void)index;
         null = sym_new_null(ctx);
         attr = NULL;
         if (this_instr[-1].opcode == _NOP) {
-            // Preceding _CHECK_ATTR_MODULE was removed: mod is const and dict is watched.
+            // Preceding _CHECK_ATTR_MODULE_PUSH_KEYS was removed: mod is const and dict is watched.
             assert(sym_is_const(owner));
             PyModuleObject *mod = (PyModuleObject *)sym_get_const(owner);
             assert(PyModule_CheckExact(mod));
@@ -530,6 +531,9 @@ dummy_func(void) {
                 this_instr[-1].opcode = _POP_TOP;
                 attr = sym_new_const(ctx, res);
             }
+            else {
+                this_instr->opcode = _LOAD_ATTR_MODULE;
+            }
         }
         if (attr == NULL) {
             /* No conversion made. We don't know what `attr` is. */
diff --git a/Python/optimizer_cases.c.h b/Python/optimizer_cases.c.h
index f77a5aa..f4fbe8c 100644
--- a/Python/optimizer_cases.c.h
+++ b/Python/optimizer_cases.c.h
@@ -1134,61 +1134,74 @@
             break;
         }
 
-        case _CHECK_ATTR_MODULE: {
+        case _CHECK_ATTR_MODULE_PUSH_KEYS: {
             _Py_UopsSymbol *owner;
+            _Py_UopsSymbol *mod_keys;
             owner = stack_pointer[-1];
             uint32_t dict_version = (uint32_t)this_instr->operand0;
             (void)dict_version;
+            mod_keys = sym_new_not_null(ctx);
             if (sym_is_const(owner)) {
                 PyObject *cnst = sym_get_const(owner);
                 if (PyModule_CheckExact(cnst)) {
                     PyModuleObject *mod = (PyModuleObject *)cnst;
                     PyObject *dict = mod->md_dict;
+                    stack_pointer[0] = mod_keys;
+                    stack_pointer += 1;
+                    assert(WITHIN_STACK_BOUNDS());
                     uint64_t watched_mutations = get_mutations(dict);
                     if (watched_mutations < _Py_MAX_ALLOWED_GLOBALS_MODIFICATIONS) {
                         PyDict_Watch(GLOBALS_WATCHER_ID, dict);
                         _Py_BloomFilter_Add(dependencies, dict);
                         this_instr->opcode = _NOP;
                     }
+                    stack_pointer += -1;
+                    assert(WITHIN_STACK_BOUNDS());
                 }
             }
+            stack_pointer[0] = mod_keys;
+            stack_pointer += 1;
+            assert(WITHIN_STACK_BOUNDS());
             break;
         }
 
-        case _LOAD_ATTR_MODULE: {
+        case _LOAD_ATTR_MODULE_FROM_KEYS: {
             _Py_UopsSymbol *owner;
             _Py_UopsSymbol *attr;
             _Py_UopsSymbol *null = NULL;
-            owner = stack_pointer[-1];
+            owner = stack_pointer[-2];
             uint16_t index = (uint16_t)this_instr->operand0;
             (void)index;
             null = sym_new_null(ctx);
             attr = NULL;
             if (this_instr[-1].opcode == _NOP) {
-                // Preceding _CHECK_ATTR_MODULE was removed: mod is const and dict is watched.
+                // Preceding _CHECK_ATTR_MODULE_PUSH_KEYS was removed: mod is const and dict is watched.
                 assert(sym_is_const(owner));
                 PyModuleObject *mod = (PyModuleObject *)sym_get_const(owner);
                 assert(PyModule_CheckExact(mod));
                 PyObject *dict = mod->md_dict;
-                stack_pointer[-1] = attr;
-                if (oparg & 1) stack_pointer[0] = null;
-                stack_pointer += (oparg & 1);
+                stack_pointer[-2] = attr;
+                if (oparg & 1) stack_pointer[-1] = null;
+                stack_pointer += -1 + (oparg & 1);
                 assert(WITHIN_STACK_BOUNDS());
                 PyObject *res = convert_global_to_const(this_instr, dict);
                 if (res != NULL) {
                     this_instr[-1].opcode = _POP_TOP;
                     attr = sym_new_const(ctx, res);
                 }
-                stack_pointer += -(oparg & 1);
+                else {
+                    this_instr->opcode = _LOAD_ATTR_MODULE;
+                }
+                stack_pointer += 1 - (oparg & 1);
                 assert(WITHIN_STACK_BOUNDS());
             }
             if (attr == NULL) {
                 /* No conversion made. We don't know what `attr` is. */
                 attr = sym_new_not_null(ctx);
             }
-            stack_pointer[-1] = attr;
-            if (oparg & 1) stack_pointer[0] = null;
-            stack_pointer += (oparg & 1);
+            stack_pointer[-2] = attr;
+            if (oparg & 1) stack_pointer[-1] = null;
+            stack_pointer += -1 + (oparg & 1);
             assert(WITHIN_STACK_BOUNDS());
             break;
         }
@@ -2528,6 +2541,18 @@
             break;
         }
 
+        case _LOAD_ATTR_MODULE: {
+            _Py_UopsSymbol *attr;
+            _Py_UopsSymbol *null = NULL;
+            attr = sym_new_not_null(ctx);
+            null = sym_new_null(ctx);
+            stack_pointer[-1] = attr;
+            if (oparg & 1) stack_pointer[0] = null;
+            stack_pointer += (oparg & 1);
+            assert(WITHIN_STACK_BOUNDS());
+            break;
+        }
+
         case _INTERNAL_INCREMENT_OPT_COUNTER: {
             stack_pointer += -1;
             assert(WITHIN_STACK_BOUNDS());
diff --git a/Python/specialize.c b/Python/specialize.c
index fd182e7..6eb2982 100644
--- a/Python/specialize.c
+++ b/Python/specialize.c
@@ -738,22 +738,16 @@ unspecialize(_Py_CODEUNIT *instr)
 }
 
 static int function_kind(PyCodeObject *code);
+#ifndef Py_GIL_DISABLED
 static bool function_check_args(PyObject *o, int expected_argcount, int opcode);
 static uint32_t function_get_version(PyObject *o, int opcode);
+#endif
 static uint32_t type_get_version(PyTypeObject *t, int opcode);
 
 static int
-specialize_module_load_attr(
-    PyObject *owner, _Py_CODEUNIT *instr, PyObject *name
-) {
+specialize_module_load_attr_lock_held(PyDictObject *dict, _Py_CODEUNIT *instr, PyObject *name)
+{
     _PyAttrCache *cache = (_PyAttrCache *)(instr + 1);
-    PyModuleObject *m = (PyModuleObject *)owner;
-    assert((Py_TYPE(owner)->tp_flags & Py_TPFLAGS_MANAGED_DICT) == 0);
-    PyDictObject *dict = (PyDictObject *)m->md_dict;
-    if (dict == NULL) {
-        SPECIALIZATION_FAIL(LOAD_ATTR, SPEC_FAIL_NO_DICT);
-        return -1;
-    }
     if (dict->ma_keys->dk_kind != DICT_KEYS_UNICODE) {
         SPECIALIZATION_FAIL(LOAD_ATTR, SPEC_FAIL_ATTR_NON_STRING);
         return -1;
@@ -773,19 +767,35 @@ specialize_module_load_attr(
                             SPEC_FAIL_OUT_OF_RANGE);
         return -1;
     }
-    uint32_t keys_version = _PyDictKeys_GetVersionForCurrentState(
-            _PyInterpreterState_GET(), dict->ma_keys);
+    uint32_t keys_version = _PyDict_GetKeysVersionForCurrentState(
+            _PyInterpreterState_GET(), dict);
     if (keys_version == 0) {
         SPECIALIZATION_FAIL(LOAD_ATTR, SPEC_FAIL_OUT_OF_VERSIONS);
         return -1;
     }
     write_u32(cache->version, keys_version);
     cache->index = (uint16_t)index;
-    instr->op.code = LOAD_ATTR_MODULE;
+    specialize(instr, LOAD_ATTR_MODULE);
     return 0;
 }
 
-
+static int
+specialize_module_load_attr(
+    PyObject *owner, _Py_CODEUNIT *instr, PyObject *name)
+{
+    PyModuleObject *m = (PyModuleObject *)owner;
+    assert((Py_TYPE(owner)->tp_flags & Py_TPFLAGS_MANAGED_DICT) == 0);
+    PyDictObject *dict = (PyDictObject *)m->md_dict;
+    if (dict == NULL) {
+        SPECIALIZATION_FAIL(LOAD_ATTR, SPEC_FAIL_NO_DICT);
+        return -1;
+    }
+    int result;
+    Py_BEGIN_CRITICAL_SECTION(dict);
+    result = specialize_module_load_attr_lock_held(dict, instr, name);
+    Py_END_CRITICAL_SECTION();
+    return result;
+}
 
 /* Attribute specialization */
 
@@ -968,7 +978,7 @@ specialize_dict_access(
         }
         write_u32(cache->version, type->tp_version_tag);
         cache->index = (uint16_t)offset;
-        instr->op.code = values_op;
+        specialize(instr, values_op);
     }
     else {
         PyDictObject *dict = _PyObject_GetManagedDict(owner);
@@ -992,11 +1002,12 @@ specialize_dict_access(
         }
         cache->index = (uint16_t)index;
         write_u32(cache->version, type->tp_version_tag);
-        instr->op.code = hint_op;
+        specialize(instr, hint_op);
     }
     return 1;
 }
 
+#ifndef Py_GIL_DISABLED
 static int specialize_attr_loadclassattr(PyObject* owner, _Py_CODEUNIT* instr, PyObject* name,
     PyObject* descr, DescriptorClassification kind, bool is_method);
 static int specialize_class_load_attr(PyObject* owner, _Py_CODEUNIT* instr, PyObject* name);
@@ -1093,7 +1104,7 @@ specialize_instance_load_attr(PyObject* owner, _Py_CODEUNIT* instr, PyObject* na
             write_u32(lm_cache->type_version, type->tp_version_tag);
             /* borrowed */
             write_obj(lm_cache->descr, fget);
-            instr->op.code = LOAD_ATTR_PROPERTY;
+            specialize(instr, LOAD_ATTR_PROPERTY);
             return 0;
         }
         case OBJECT_SLOT:
@@ -1117,7 +1128,7 @@ specialize_instance_load_attr(PyObject* owner, _Py_CODEUNIT* instr, PyObject* na
             assert(offset > 0);
             cache->index = (uint16_t)offset;
             write_u32(cache->version, type->tp_version_tag);
-            instr->op.code = LOAD_ATTR_SLOT;
+            specialize(instr, LOAD_ATTR_SLOT);
             return 0;
         }
         case DUNDER_CLASS:
@@ -1126,7 +1137,7 @@ specialize_instance_load_attr(PyObject* owner, _Py_CODEUNIT* instr, PyObject* na
             assert(offset == (uint16_t)offset);
             cache->index = (uint16_t)offset;
             write_u32(cache->version, type->tp_version_tag);
-            instr->op.code = LOAD_ATTR_SLOT;
+            specialize(instr, LOAD_ATTR_SLOT);
             return 0;
         }
         case OTHER_SLOT:
@@ -1162,7 +1173,7 @@ specialize_instance_load_attr(PyObject* owner, _Py_CODEUNIT* instr, PyObject* na
             /* borrowed */
             write_obj(lm_cache->descr, descr);
             write_u32(lm_cache->type_version, type->tp_version_tag);
-            instr->op.code = LOAD_ATTR_GETATTRIBUTE_OVERRIDDEN;
+            specialize(instr, LOAD_ATTR_GETATTRIBUTE_OVERRIDDEN);
             return 0;
         }
         case BUILTIN_CLASSMETHOD:
@@ -1186,6 +1197,7 @@ specialize_instance_load_attr(PyObject* owner, _Py_CODEUNIT* instr, PyObject* na
             if (shadow) {
                 goto try_instance;
             }
+            set_counter((_Py_BackoffCounter*)instr + 1, adaptive_counter_cooldown());
             return 0;
     }
     Py_UNREACHABLE();
@@ -1197,14 +1209,14 @@ try_instance:
     }
     return -1;
 }
+#endif //  Py_GIL_DISABLED
 
 void
 _Py_Specialize_LoadAttr(_PyStackRef owner_st, _Py_CODEUNIT *instr, PyObject *name)
 {
-    _PyAttrCache *cache = (_PyAttrCache *)(instr + 1);
     PyObject *owner = PyStackRef_AsPyObjectBorrow(owner_st);
 
-    assert(ENABLE_SPECIALIZATION);
+    assert(ENABLE_SPECIALIZATION_FT);
     assert(_PyOpcode_Caches[LOAD_ATTR] == INLINE_CACHE_ENTRIES_LOAD_ATTR);
     PyTypeObject *type = Py_TYPE(owner);
     bool fail;
@@ -1219,22 +1231,24 @@ _Py_Specialize_LoadAttr(_PyStackRef owner_st, _Py_CODEUNIT *instr, PyObject *nam
         fail = specialize_module_load_attr(owner, instr, name);
     }
     else if (PyType_Check(owner)) {
+        #ifdef Py_GIL_DISABLED
+        SPECIALIZATION_FAIL(LOAD_ATTR, SPEC_FAIL_EXPECTED_ERROR);
+        fail = true;
+        #else
         fail = specialize_class_load_attr(owner, instr, name);
+        #endif
     }
     else {
+        #ifdef Py_GIL_DISABLED
+        SPECIALIZATION_FAIL(LOAD_ATTR, SPEC_FAIL_EXPECTED_ERROR);
+        fail = true;
+        #else
         fail = specialize_instance_load_attr(owner, instr, name);
+        #endif
     }
 
     if (fail) {
-        STAT_INC(LOAD_ATTR, failure);
-        assert(!PyErr_Occurred());
-        instr->op.code = LOAD_ATTR;
-        cache->counter = adaptive_counter_backoff(cache->counter);
-    }
-    else {
-        STAT_INC(LOAD_ATTR, success);
-        assert(!PyErr_Occurred());
-        cache->counter = adaptive_counter_cooldown();
+        unspecialize(instr);
     }
 }
 
@@ -1339,6 +1353,7 @@ success:
     cache->counter = adaptive_counter_cooldown();
 }
 
+#ifndef Py_GIL_DISABLED
 
 #ifdef Py_STATS
 static int
@@ -1422,10 +1437,10 @@ specialize_class_load_attr(PyObject *owner, _Py_CODEUNIT *instr,
             write_obj(cache->descr, descr);
             if (metaclass_check) {
                 write_u32(cache->keys_version, Py_TYPE(cls)->tp_version_tag);
-                instr->op.code = LOAD_ATTR_CLASS_WITH_METACLASS_CHECK;
+                specialize(instr, LOAD_ATTR_CLASS_WITH_METACLASS_CHECK);
             }
             else {
-                instr->op.code = LOAD_ATTR_CLASS;
+                specialize(instr, LOAD_ATTR_CLASS);
             }
             return 0;
 #ifdef Py_STATS
@@ -1461,7 +1476,7 @@ PyObject *descr, DescriptorClassification kind, bool is_method)
             return 0;
         }
         write_u32(cache->keys_version, keys_version);
-        instr->op.code = is_method ? LOAD_ATTR_METHOD_WITH_VALUES : LOAD_ATTR_NONDESCRIPTOR_WITH_VALUES;
+        specialize(instr, is_method ? LOAD_ATTR_METHOD_WITH_VALUES : LOAD_ATTR_NONDESCRIPTOR_WITH_VALUES);
     }
     else {
         Py_ssize_t dictoffset;
@@ -1476,7 +1491,7 @@ PyObject *descr, DescriptorClassification kind, bool is_method)
             }
         }
         if (dictoffset == 0) {
-            instr->op.code = is_method ? LOAD_ATTR_METHOD_NO_DICT : LOAD_ATTR_NONDESCRIPTOR_NO_DICT;
+            specialize(instr, is_method ? LOAD_ATTR_METHOD_NO_DICT : LOAD_ATTR_NONDESCRIPTOR_NO_DICT);
         }
         else if (is_method) {
             PyObject *dict = *(PyObject **) ((char *)owner + dictoffset);
@@ -1490,7 +1505,7 @@ PyObject *descr, DescriptorClassification kind, bool is_method)
             dictoffset -= MANAGED_DICT_OFFSET;
             assert(((uint16_t)dictoffset) == dictoffset);
             cache->dict_offset = (uint16_t)dictoffset;
-            instr->op.code = LOAD_ATTR_METHOD_LAZY_DICT;
+            specialize(instr, LOAD_ATTR_METHOD_LAZY_DICT);
         }
         else {
             SPECIALIZATION_FAIL(LOAD_ATTR, SPEC_FAIL_ATTR_CLASS_ATTR_SIMPLE);
@@ -1516,6 +1531,9 @@ PyObject *descr, DescriptorClassification kind, bool is_method)
     return 1;
 }
 
+#endif //  Py_GIL_DISABLED
+
+
 static void
 specialize_load_global_lock_held(
     PyObject *globals, PyObject *builtins,
@@ -1661,6 +1679,7 @@ function_kind(PyCodeObject *code) {
     return SIMPLE_FUNCTION;
 }
 
+#ifndef Py_GIL_DISABLED
 /* Returning false indicates a failure. */
 static bool
 function_check_args(PyObject *o, int expected_argcount, int opcode)
@@ -1693,6 +1712,7 @@ function_get_version(PyObject *o, int opcode)
     }
     return version;
 }
+#endif  // Py_GIL_DISABLED
 
 /* Returning 0 indicates a failure. */
 static uint32_t