Fix: mark interpreter frames as tier 2 or not (python#34)

* fix: mark interpreter frames as tier 2 or not

* Test: Added tests for python#33

---------

Co-authored-by: Ken Jin <[email protected]>

Author: JuliaPoo

Include/internal/pycore_frame.h

@@ -62,13 +62,14 @@ typedef struct _PyInterpreterFrame {
     int stacktop;  /* Offset of TOS from localsplus  */
     uint16_t yield_offset;
     char owner;
+    bool is_tier2;
     /* Locals and stack and unboxed bit mask */
     PyObject *localsplus[1];
 } _PyInterpreterFrame;
 
 static inline int
 _PyInterpreterFrame_LASTI(_PyInterpreterFrame *f) {
-    if (f->f_code->_tier2_info != NULL) {
+    if (f->is_tier2 && f->f_code->_tier2_info != NULL) {
         return ((int)((f)->prev_instr - f->f_code->_tier2_info->_bb_space->u_code));
     }
     return ((int)((f)->prev_instr - _PyCode_CODE((f)->f_code)));
@@ -130,9 +131,11 @@ _PyFrame_Initialize(
     frame->stacktop = code->co_nlocalsplus;
     frame->frame_obj = NULL;
     if (code->_tier2_info != NULL) {
+        frame->is_tier2 = true;
         frame->prev_instr = code->_tier2_info->_entry_bb->tier2_start - 1;
     }
     else {
+        frame->is_tier2 = false;
         frame->prev_instr = _PyCode_CODE(code) - 1;
     }
     frame->yield_offset = 0;

Python/frame.c

@@ -17,10 +17,13 @@ _PyFrame_Traverse(_PyInterpreterFrame *frame, visitproc visit, void *arg)
     Py_VISIT(frame->f_code);
    /* locals */
     PyObject **locals = _PyFrame_GetLocalsArray(frame);
+    char *unboxed_bitmask = _PyFrame_GetUnboxedBitMask(frame);
     int i = 0;
     /* locals and stack */
     for (; i <frame->stacktop; i++) {
-        Py_VISIT(locals[i]);
+        if (!unboxed_bitmask[i]) {
+            Py_VISIT(locals[i]);
+        }
     }
     return 0;
 }

Python/tier2.c

@@ -2486,6 +2486,7 @@ _Py_CODEUNIT *
 _PyCode_Tier2Warmup(_PyInterpreterFrame *frame, _Py_CODEUNIT *next_instr)
 {
     PyCodeObject *code = frame->f_code;
+    frame->is_tier2 = false;
     if (code->_tier2_warmup != 0) {
         code->_tier2_warmup++;
         if (code->_tier2_warmup >= 0) {
@@ -2494,6 +2495,11 @@ _PyCode_Tier2Warmup(_PyInterpreterFrame *frame, _Py_CODEUNIT *next_instr)
             // just fall back to the tier 1 interpreter.
             _Py_CODEUNIT *next = _PyCode_Tier2Initialize(frame, next_instr);
             if (next != NULL) {
+                assert(!frame->is_tier2);
+                frame->is_tier2 = true;
+                _Py_CODEUNIT *curr = (next_instr - 1);
+                assert(curr->op.code == RESUME || curr->op.code == RESUME_QUICK);
+                curr->op.code = RESUME_QUICK;
                 return next;
             }
         }

tier2_test.py

@@ -519,7 +519,7 @@ def f(a):
 ######################################################################
 # Tests for: Tier 2 recursive functions block generation             #
 ######################################################################
-with TestInfo("tier 2 BB_TEST_POP_IF_FALSE flag setting"):
+with TestInfo("tier 2 recursive functions block generation"):
     # See https://github.com/pylbbv/pylbbv/issues/23 for more information.
     def f(x):
         # Force specialisation
@@ -535,4 +535,19 @@ def f(x):
 
     # As long as it doesn't crash, everything's good.
 
+######################################################################
+# Tests for: Tier 2 specialisation in inner call                     #
+######################################################################
+with TestInfo("tier 2 specialisation in inner call"):
+    # See https://github.com/pylbbv/pylbbv/issues/33 for more information.
+    def f(x):
+        if x == 0: return 0
+        if x%2: return f(x-1)
+        return f(x-1)
+
+    for i in range(15): f(3)
+    f(3)
+
+    # As long as it doesn't crash, everything's good.
+
 print("Tests completed ^-^")