trace over RESUME/RESUME_CHECK_JIT

Author: Fidget-Spinner

Python/bytecodes.c

@@ -3041,7 +3041,7 @@ dummy_func(
             #ifdef _Py_TIER2
             if (IS_JIT_TRACING()) {
                 next_instr = this_instr;
-                goto stop_tracing;
+                goto consider_stop_tracing;
             }
             PyCodeObject *code = _PyFrame_GetCode(frame);
             _PyExecutorObject *executor = code->co_executors->executors[oparg & 255];
@@ -5659,10 +5659,19 @@ dummy_func(
 #endif
         }
 
-        label(stop_tracing) {
+        label(consider_stop_tracing) {
 #if _Py_TIER2
             assert(IS_JIT_TRACING());
             int opcode = next_instr->op.code;
+            PyCodeObject *code = _PyFrame_GetCode(frame);
+            _PyExecutorObject *executor = code->co_executors->executors[oparg & 255];
+            int orig_opcode = executor->vm_data.opcode;
+            // Function entry, trace over it to form a longer trace.
+            if (orig_opcode == RESUME_CHECK_JIT || orig_opcode == RESUME) {
+                oparg = (oparg & ~255) | executor->vm_data.oparg;
+                opcode = orig_opcode;
+                DISPATCH_GOTO_NON_TRACING();
+            }
             _PyJit_translate_single_bytecode_to_trace(tstate, frame, NULL, _EXIT_TRACE);
             LEAVE_TRACING();
             int err = stop_tracing_and_jit(tstate, frame);

Python/generated_cases.c.h

@@ -646,6 +646,29 @@ _PyJit_translate_single_bytecode_to_trace(
     int oparg = _tstate->jit_tracer_state.prev_state.instr_oparg;
     int opcode = this_instr->op.code;
 
+    if (stop_tracing_opcode == _DEOPT) {
+        // gh-143183: It's important we rewind to the last known proper target.
+        // The current target might be garbage as stop tracing usually indicates
+        // we are in something that we can't trace.
+        DPRINTF(2, "Told to stop tracing\n");
+        goto unsupported;
+    }
+    else if (stop_tracing_opcode != 0) {
+        assert(stop_tracing_opcode == _EXIT_TRACE);
+        ADD_TO_TRACE(stop_tracing_opcode, 0, 0, target);
+        goto done;
+    }
+
+    // Function entry executor, trace over it to form a longer trace.
+    // Otherwise, we end up with fragmented loop traces that have bad performance.
+    if (opcode == ENTER_EXECUTOR) {
+        _PyExecutorObject *executor = old_code->co_executors->executors[oparg & 255];
+        int orig_opcode = executor->vm_data.opcode;
+        assert(orig_opcode == RESUME_CHECK_JIT || orig_opcode == RESUME);
+        oparg = (oparg & ~255) | executor->vm_data.oparg;
+        opcode = orig_opcode;
+    }
+
     int rewind_oparg = oparg;
     while (rewind_oparg > 255) {
         rewind_oparg >>= 8;
@@ -699,19 +722,6 @@ _PyJit_translate_single_bytecode_to_trace(
         goto full;
     }
 
-    if (stop_tracing_opcode == _DEOPT) {
-        // gh-143183: It's important we rewind to the last known proper target.
-        // The current target might be garbage as stop tracing usually indicates
-        // we are in something that we can't trace.
-        DPRINTF(2, "Told to stop tracing\n");
-        goto unsupported;
-    }
-    else if (stop_tracing_opcode != 0) {
-        assert(stop_tracing_opcode == _EXIT_TRACE);
-        ADD_TO_TRACE(stop_tracing_opcode, 0, 0, target);
-        goto done;
-    }
-
     assert(stop_tracing_opcode == 0);
 
 #ifdef Py_DEBUG