gh-149570: avoid tier-2 self-link in _COLD_EXIT

Include/internal/pycore_optimizer.h

@@ -311,7 +311,8 @@ PyAPI_FUNC(void) _Py_Executors_InvalidateCold(PyInterpreterState *interp);
 int _Py_uop_analyze_and_optimize(
     _PyThreadStateImpl *tstate,
     _PyUOpInstruction *input, int trace_len, int curr_stackentries,
-    _PyUOpInstruction *output, _PyBloomFilter *dependencies);
+    _PyUOpInstruction *output, _PyBloomFilter *dependencies,
+    bool progress_needed);
 
 extern PyTypeObject _PyUOpExecutor_Type;
 

Lib/test/test_capi/test_opt.py

@@ -1,5 +1,6 @@
 import contextlib
 import itertools
+import subprocess
 import sys
 import textwrap
 import unittest
@@ -9,7 +10,7 @@
 
 import _opcode
 
-from test.support import (script_helper, requires_specialization,
+from test.support import (SHORT_TIMEOUT, script_helper, requires_specialization,
                           import_helper, Py_GIL_DISABLED, requires_jit_enabled,
                           reset_code)
 
@@ -5994,6 +5995,51 @@ def __init__(self, name):
              PYTHON_JIT_SIDE_EXIT_INITIAL_VALUE="1")
         self.assertEqual(result[0].rc, 0, result)
 
+    def test_side_exit_to_executor_makes_progress(self):
+        script = textwrap.dedent("""
+        classes = []
+
+        def make_iter():
+            class It:
+                def __init__(self):
+                    self.i = 0
+
+                def __iter__(self):
+                    return self
+
+                def __next__(self):
+                    self.i += 1
+                    if self.i > 1000:
+                        raise StopIteration
+                    return self.i
+
+            classes.append(It)
+            return It()
+
+        def f(n):
+            for outer in range(n):
+                for x in make_iter():
+                    pass
+
+        f(200)
+        """)
+        env = os.environ.copy()
+        env.update({
+            "PYTHON_JIT": "1",
+            "PYTHON_JIT_SIDE_EXIT_INITIAL_VALUE": "1",
+        })
+        try:
+            result = subprocess.run(
+                [sys.executable, "-X", "faulthandler", "-c", script],
+                stdout=subprocess.PIPE,
+                stderr=subprocess.PIPE,
+                env=env,
+                timeout=SHORT_TIMEOUT,
+            )
+        except subprocess.TimeoutExpired as exc:
+            self.fail(f"subprocess timed out: {exc}")
+        self.assertEqual(result.returncode, 0, result)
+
     def test_for_iter_gen_cleared_frame_does_not_crash(self):
         # See: https://github.com/python/cpython/issues/145197
         result = script_helper.run_python_until_end('-c', textwrap.dedent("""

Python/optimizer.c

@@ -1704,7 +1704,7 @@ uop_optimize(
         _PyUOpInstruction *output = &_tstate->jit_tracer_state->uop_array[UOP_MAX_TRACE_LENGTH];
         length = _Py_uop_analyze_and_optimize(
             _tstate, buffer, length, curr_stackentries,
-            output, &dependencies);
+            output, &dependencies, progress_needed);
 
         if (length <= 0) {
             return length;

Python/optimizer_analysis.c

@@ -566,7 +566,8 @@ optimize_uops(
     int trace_len,
     int curr_stacklen,
     _PyUOpInstruction *output,
-    _PyBloomFilter *dependencies
+    _PyBloomFilter *dependencies,
+    bool progress_needed
 )
 {
     assert(!PyErr_Occurred());
@@ -597,9 +598,13 @@ optimize_uops(
 
     _PyUOpInstruction *this_instr = NULL;
     JitOptRef *stack_pointer = ctx->frame->stack_pointer;
+    int bytecode_count = 0;
 
     for (int i = 0; i < trace_len; i++) {
         this_instr = &trace[i];
+        if (this_instr->opcode == _CHECK_VALIDITY) {
+            bytecode_count++;
+        }
         if (ctx->done) {
             // Don't do any more optimization, but
             // we still need to reach a terminator for corrctness.
@@ -629,6 +634,9 @@ optimize_uops(
                 DPRINTF(1, "\nUnknown opcode in abstract interpreter\n");
                 Py_UNREACHABLE();
         }
+        if (progress_needed && bytecode_count < 2) {
+            ctx->out_buffer.next = out_ptr;
+        }
         // If no ADD_OP was called during this iteration, copy the original instruction
         if (ctx->out_buffer.next == out_ptr) {
             *(ctx->out_buffer.next++) = *this_instr;
@@ -797,13 +805,14 @@ _Py_uop_analyze_and_optimize(
     int length,
     int curr_stacklen,
     _PyUOpInstruction *output,
-    _PyBloomFilter *dependencies
+    _PyBloomFilter *dependencies,
+    bool progress_needed
 )
 {
     OPT_STAT_INC(optimizer_attempts);
 
     length = optimize_uops(
-        tstate, buffer, length, curr_stacklen, output, dependencies);
+        tstate, buffer, length, curr_stacklen, output, dependencies, progress_needed);
 
     if (length == 0) {
         return length;