gh-137291: Support perf profiler with an evaluation hook (#137292)

Support perf profiler with an evaluation hook

Include/internal/pycore_interp_structs.h

@@ -88,6 +88,7 @@ struct _ceval_runtime_state {
         struct trampoline_api_st trampoline_api;
         FILE *map_file;
         Py_ssize_t persist_after_fork;
+       _PyFrameEvalFunction prev_eval_frame;
 #else
         int _not_used;
 #endif

Lib/test/test_perf_profiler.py

@@ -162,6 +162,7 @@ class TestPerfTrampoline(unittest.TestCase):
 
     @unittest.skipIf(support.check_bolt_optimized(), "fails on BOLT instrumented binaries")
     def test_sys_api(self):
+        for define_eval_hook in (False, True):
             code = """if 1:
                     import sys
                     def foo():
@@ -182,6 +183,12 @@ class TestPerfTrampoline(unittest.TestCase):
                     sys.activate_stack_trampoline("perf")
                     baz()
                     """
+            if define_eval_hook:
+                set_eval_hook = """if 1:
+                                import _testinternalcapi
+                                _testinternalcapi.set_eval_frame_record([])
+"""
+                code = set_eval_hook + code
             with temp_dir() as script_dir:
                 script = make_script(script_dir, "perftest", code)
                 env = {**os.environ, "PYTHON_JIT": "0"}

Misc/NEWS.d/next/Core_and_Builtins/2025-07-31-23-02-02.gh-issue-137291.kIxVZd.rst

@@ -0,0 +1 @@
+The perf profiler can now be used if a previous frame evaluation API has been provided.

Python/perf_trampoline.c

@@ -202,6 +202,7 @@ enum perf_trampoline_type {
 #define perf_map_file _PyRuntime.ceval.perf.map_file
 #define persist_after_fork _PyRuntime.ceval.perf.persist_after_fork
 #define perf_trampoline_type _PyRuntime.ceval.perf.perf_trampoline_type
+#define prev_eval_frame _PyRuntime.ceval.perf.prev_eval_frame
 
 static void
 perf_map_write_entry(void *state, const void *code_addr,
@@ -407,9 +408,12 @@ py_trampoline_evaluator(PyThreadState *ts, _PyInterpreterFrame *frame,
         f = new_trampoline;
     }
     assert(f != NULL);
-    return f(ts, frame, throw, _PyEval_EvalFrameDefault);
+    return f(ts, frame, throw, prev_eval_frame != NULL ? prev_eval_frame : _PyEval_EvalFrameDefault);
 default_eval:
     // Something failed, fall back to the default evaluator.
+    if (prev_eval_frame) {
+        return prev_eval_frame(ts, frame, throw);
+    }
     return _PyEval_EvalFrameDefault(ts, frame, throw);
 }
 #endif  // PY_HAVE_PERF_TRAMPOLINE
@@ -481,18 +485,12 @@ _PyPerfTrampoline_Init(int activate)
 {
 #ifdef PY_HAVE_PERF_TRAMPOLINE
     PyThreadState *tstate = _PyThreadState_GET();
-    if (tstate->interp->eval_frame &&
-        tstate->interp->eval_frame != py_trampoline_evaluator) {
-        PyErr_SetString(PyExc_RuntimeError,
-                        "Trampoline cannot be initialized as a custom eval "
-                        "frame is already present");
-        return -1;
-    }
     if (!activate) {
-        _PyInterpreterState_SetEvalFrameFunc(tstate->interp, NULL);
+        _PyInterpreterState_SetEvalFrameFunc(tstate->interp, prev_eval_frame);
         perf_status = PERF_STATUS_NO_INIT;
     }
-    else {
+    else if (tstate->interp->eval_frame != py_trampoline_evaluator) {
+        prev_eval_frame = _PyInterpreterState_GetEvalFrameFunc(tstate->interp);
         _PyInterpreterState_SetEvalFrameFunc(tstate->interp, py_trampoline_evaluator);
         extra_code_index = _PyEval_RequestCodeExtraIndex(NULL);
         if (extra_code_index == -1) {