gh-157468: Validate correct builtins used under JIT

Include/internal/pycore_optimizer_types.h

@@ -140,6 +140,7 @@ typedef union {
 
 typedef struct _Py_UOpsAbstractFrame {
     bool globals_watched;
+    bool builtins_checked;
     // The version number of the globals dicts, once checked. 0 if unchecked.
     uint32_t globals_checked_version;
     // Max stacklen

Lib/test/test_capi/test_opt.py

@@ -1,3 +1,4 @@
+import builtins
 import contextlib
 import dis
 import itertools
@@ -5054,6 +5055,86 @@ def jitted(funcs):
         with self.assertRaises(NameError):
             jitted([f, f_with_bad_globals])
 
+    def test_jitted_code_sees_changed_copied_builtins(self):
+        # Trace-time check.  The traced function's builtins is a copy of the
+        # canonical dict with the same keys version, so a version check
+        # cannot tell them apart.  The optimizer must see that func_builtins
+        # is not interp->builtins and keep _LOAD_GLOBAL_BUILTINS, which reads
+        # the frame's own dict, rather than fold a constant from the
+        # canonical one.  No runtime guard is involved.
+
+        def f(n):
+            return [len("hello") for _ in range(n)]
+
+        copied_builtins = vars(builtins).copy()
+        f = types.FunctionType(f.__code__, {"__builtins__": copied_builtins})
+
+        f(TIER2_THRESHOLD)
+        ex = get_first_executor(f)
+        self.assertIsNotNone(ex)
+        # Not folded: the load must still consult the frame's builtins.
+        self.assertIn("_LOAD_GLOBAL_BUILTINS", get_opnames(ex))
+
+        # Replacing an existing value does not change the keys version.
+        copied_builtins["len"] = lambda s: 42
+        self.assertEqual(f(8), [42] * 8)
+
+    def test_jitted_code_sees_different_builtins(self):
+        # Runtime check.  The traced function's builtins IS the canonical
+        # dict, so folding len to a constant is correct at trace time.
+        # A second function sharing the code object then enters the same
+        # executor with other builtins, so only the runtime guard on the
+        # executing frame's builtins can catch it.
+        def f(n):
+            return [len("hello") for _ in range(n)]
+
+        namespace = {"__builtins__": builtins}
+        f_canonical = types.FunctionType(f.__code__, namespace)
+        copied_builtins = vars(builtins).copy()
+        namespace["__builtins__"] = copied_builtins
+        f_copied = types.FunctionType(f.__code__, namespace)
+
+
+        f_canonical(TIER2_THRESHOLD)
+        ex = get_first_executor(f_canonical)
+        self.assertIsNotNone(ex)
+        self.assertIn("_GUARD_BUILTINS_IS_CANONICAL", get_opnames(ex))
+
+        copied_builtins["len"] = lambda s: 42
+        # The executor's owner still sees the canonical len.
+        self.assertEqual(f_canonical(8), [5] * 8)
+        # A different function enters the same executor with other builtins.
+        self.assertEqual(f_copied(8), [42] * 8)
+
+    def test_builtins_guard_emitted_once_per_frame(self):
+        # A frame's builtins cannot change once the frame is pushed, so
+        # repeated builtin loads in one frame share a single guard, just as
+        # they already share a single _GUARD_GLOBALS_VERSION.
+
+        def warmup(n):
+            x = 0
+            for _ in range(n):
+                x += len("ab")
+            return x
+
+        def one_frame(n):
+            x = 0
+            for _ in range(n):
+                x += len("ab") + abs(-1) + ord("c")
+            return x
+
+        # The optimizer context is reused for every compilation, so compile an
+        # unrelated trace first: state that is not reset per frame leaks here.
+        warmup(TIER2_THRESHOLD)
+        self.assertIsNotNone(get_first_executor(warmup))
+
+        _, ex = self._run_with_optimizer(one_frame, TIER2_THRESHOLD)
+        self.assertIsNotNone(ex)
+        uop_names = get_opnames(ex)
+        self.assertNotIn("_LOAD_GLOBAL_BUILTINS", uop_names)  # all folded
+        self.assertEqual(uop_names.count("_GUARD_BUILTINS_IS_CANONICAL"), 1)
+        self.assertEqual(uop_names.count("_GUARD_GLOBALS_VERSION"), 1)
+
     def test_reference_tracking_across_call_doesnt_crash(self):
 
         def f1():

Misc/NEWS.d/next/Core_and_Builtins/2026-09-18-20-24-43.gh-issue-157468.zVIAva.rst

@@ -0,0 +1,7 @@
+Fix the JIT optimizer folding a builtin name to a constant taken from the
+interpreter's builtins dictionary even when the running function uses a
+different ``__builtins__`` mapping, such as one created with
+``vars(builtins).copy()``. The optimizer now only folds when the function's
+builtins is the interpreter's, and the folded constant is guarded at runtime so
+that another function sharing the same code object but a different builtins
+mapping does not use it.

Python/bytecodes.c

@@ -2357,6 +2357,10 @@ dummy_func(
             STAT_INC(LOAD_GLOBAL, hit);
         }
 
+        tier2 op(_GUARD_BUILTINS_IS_CANONICAL, (--)) {
+            DEOPT_IF(BUILTINS() != tstate->interp->builtins);
+        }
+
         macro(LOAD_GLOBAL_MODULE) =
             unused/1 + // Skip over the counter
             NOP + // For guard insertion in the JIT optimizer

Python/optimizer_bytecodes.c

@@ -1,3 +1,5 @@
+#include <stdint.h>
+
 #include "Python.h"
 #include "pycore_long.h"
 #include "pycore_opcode_utils.h"
@@ -2519,13 +2521,24 @@ dummy_func(void) {
         else if (interp->rare_events.builtin_dict >= _Py_MAX_ALLOWED_BUILTINS_MODIFICATIONS) {
             /* Do nothing */
         }
+        else if (ctx->frame->func == NULL ||
+                 ctx->frame->func->func_builtins != builtins) {
+        }
         else {
             if (!ctx->builtins_watched) {
                 PyDict_Watch(BUILTINS_WATCHER_ID, builtins);
                 ctx->builtins_watched = true;
             }
-            if (ctx->frame->globals_checked_version != 0 && ctx->frame->globals_watched) {
+            if (ctx->frame->globals_checked_version != 0 &&
+                ctx->frame->globals_watched &&
+                uop_buffer_remaining_space(&ctx->out_buffer) >= 2)
+            {
                 cnst = convert_global_to_const(this_instr, builtins);
+                if (cnst != NULL && !ctx->frame->builtins_checked) {
+                    ctx->frame->builtins_checked = true;
+                    ADD_OP(_GUARD_BUILTINS_IS_CANONICAL, 0, 0);
+                    ADD_OP(this_instr->opcode, 0, (uintptr_t)cnst);
+                }
             }
         }
         if (cnst == NULL) {

Python/optimizer_symbols.c

@@ -1378,6 +1378,7 @@ _Py_uop_frame_new(
     frame->stack_pointer = frame->stack;
     frame->globals_checked_version = 0;
     frame->globals_watched = false;
+    frame->builtins_checked = false;
     frame->func = NULL;
     frame->caller = false;
     frame->is_c_recursion_checked = false;