gh-136396: Include instrumentation when creating new copies of the bytecode (#136525)

Previously, we assumed that instrumentation would happen for all copies of
the bytecode if the instrumentation version on the code object didn't match
the per-interpreter instrumentation version. That assumption was incorrect:
instrumentation will exit early if there are no new "events," even if there
is an instrumentation version mismatch.

To fix this, include the instrumented opcodes when creating new copies of
the bytecode, rather than replacing them with their uninstrumented variants.
I don't think we have to worry about races between instrumentation and creating
new copies of the bytecode: instrumentation and new bytecode creation cannot happen
concurrently. Instrumentation requires that either the world is stopped or the
code object's per-object lock is held and new bytecode creation requires holding
the code object's per-object lock.

Lib/test/test_free_threading/test_monitoring.py

@@ -2,10 +2,12 @@
 environment to verify things are thread-safe in a free-threaded build"""
 
 import sys
+import threading
 import time
 import unittest
 import weakref
 
+from contextlib import contextmanager
 from sys import monitoring
 from test.support import threading_helper
 from threading import Thread, _PyRLock, Barrier
@@ -192,6 +194,16 @@ class SetProfileMultiThreaded(InstrumentationMultiThreadedMixin, TestCase):
         self.set = not self.set
 
 
+class TraceBuf:
+    def __init__(self):
+        self.traces = []
+        self.traces_lock = threading.Lock()
+
+    def append(self, trace):
+        with self.traces_lock:
+            self.traces.append(trace)
+
+
 @threading_helper.requires_working_threading()
 class MonitoringMisc(MonitoringTestMixin, TestCase):
     def register_callback(self, barrier):
@@ -246,6 +258,135 @@ class MonitoringMisc(MonitoringTestMixin, TestCase):
         finally:
             sys.settrace(None)
 
+    def test_toggle_setprofile_no_new_events(self):
+        # gh-136396: Make sure that profile functions are called for newly
+        # created threads when profiling is toggled but the set of monitoring
+        # events doesn't change
+        traces = []
+
+        def profiler(frame, event, arg):
+            traces.append((frame.f_code.co_name, event, arg))
+
+        def a(x, y):
+            return b(x, y)
+
+        def b(x, y):
+            return max(x, y)
+
+        sys.setprofile(profiler)
+        try:
+            a(1, 2)
+        finally:
+            sys.setprofile(None)
+        traces.clear()
+
+        def thread_main(x, y):
+            sys.setprofile(profiler)
+            try:
+                a(x, y)
+            finally:
+                sys.setprofile(None)
+        t = Thread(target=thread_main, args=(100, 200))
+        t.start()
+        t.join()
+
+        expected = [
+            ("a", "call", None),
+            ("b", "call", None),
+            ("b", "c_call", max),
+            ("b", "c_return", max),
+            ("b", "return", 200),
+            ("a", "return", 200),
+            ("thread_main", "c_call", sys.setprofile),
+        ]
+        self.assertEqual(traces, expected)
+
+    def observe_threads(self, observer, buf):
+        def in_child(ident):
+            return ident
+
+        def child(ident):
+            with observer():
+                in_child(ident)
+
+        def in_parent(ident):
+            return ident
+
+        def parent(barrier, ident):
+            barrier.wait()
+            with observer():
+                t = Thread(target=child, args=(ident,))
+                t.start()
+                t.join()
+                in_parent(ident)
+
+        num_threads = 5
+        barrier = Barrier(num_threads)
+        threads = []
+        for i in range(num_threads):
+            t = Thread(target=parent, args=(barrier, i))
+            t.start()
+            threads.append(t)
+        for t in threads:
+            t.join()
+
+        for i in range(num_threads):
+            self.assertIn(("in_parent", "return", i), buf.traces)
+            self.assertIn(("in_child", "return", i), buf.traces)
+
+    def test_profile_threads(self):
+        buf = TraceBuf()
+
+        def profiler(frame, event, arg):
+            buf.append((frame.f_code.co_name, event, arg))
+
+        @contextmanager
+        def profile():
+            sys.setprofile(profiler)
+            try:
+                yield
+            finally:
+                sys.setprofile(None)
+
+        self.observe_threads(profile, buf)
+
+    def test_trace_threads(self):
+        buf = TraceBuf()
+
+        def tracer(frame, event, arg):
+            buf.append((frame.f_code.co_name, event, arg))
+            return tracer
+
+        @contextmanager
+        def trace():
+            sys.settrace(tracer)
+            try:
+                yield
+            finally:
+                sys.settrace(None)
+
+        self.observe_threads(trace, buf)
+
+    def test_monitor_threads(self):
+        buf = TraceBuf()
+
+        def monitor_py_return(code, off, retval):
+            buf.append((code.co_name, "return", retval))
+
+        monitoring.register_callback(
+            self.tool_id, monitoring.events.PY_RETURN, monitor_py_return
+        )
+
+        monitoring.set_events(
+            self.tool_id, monitoring.events.PY_RETURN
+        )
+
+        @contextmanager
+        def noop():
+            yield
+
+        self.observe_threads(noop, buf)
+
 
 if __name__ == "__main__":
     unittest.main()

Misc/NEWS.d/next/Core_and_Builtins/2025-07-10-15-53-16.gh-issue-136525.xAko0e.rst

@@ -0,0 +1,2 @@
+Fix issue where per-thread bytecode was not instrumented for newly created
+threads.

Objects/codeobject.c

@@ -3330,12 +3330,29 @@ _PyCodeArray_New(Py_ssize_t size)
     return arr;
 }
 
+// Get the underlying code unit, leaving instrumentation
+static _Py_CODEUNIT
+deopt_code_unit(PyCodeObject *code, int i)
+{
+    _Py_CODEUNIT *src_instr = _PyCode_CODE(code) + i;
+    _Py_CODEUNIT inst = {
+        .cache = FT_ATOMIC_LOAD_UINT16_RELAXED(*(uint16_t *)src_instr)};
+    int opcode = inst.op.code;
+    if (opcode < MIN_INSTRUMENTED_OPCODE) {
+        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);
+    return inst;
+}
+
 static void
 copy_code(_Py_CODEUNIT *dst, PyCodeObject *co)
 {
     int code_len = (int) Py_SIZE(co);
     for (int i = 0; i < code_len; i += _PyInstruction_GetLength(co, i)) {
-        dst[i] = _Py_GetBaseCodeUnit(co, i);
+        dst[i] = deopt_code_unit(co, i);
     }
     _PyCode_Quicken(dst, code_len, 1);
 }