fix a bug with specializzation

Author: Fidget-Spinner

Include/internal/pycore_opcode_metadata.h

@@ -359,10 +359,12 @@ extern void _Py_ClearExecutorDeletionList(PyInterpreterState *interp);
 int
 _PyJIT_translate_single_bytecode_to_trace(
     PyThreadState *tstate,
+    _PyInterpreterFrame *frame,
     _Py_CODEUNIT *this_instr,
     _Py_CODEUNIT *next_instr,
     PyCodeObject *code,
     PyFunctionObject *func,
+    int opcode,
     int oparg);
 
 void

Include/internal/pycore_optimizer.h

@@ -3050,18 +3050,20 @@ dummy_func(
             #endif /* _Py_TIER2 */
         }
 
-        op(_POP_JUMP_IF_FALSE, (cond -- )) {
+        replaced op(_POP_JUMP_IF_FALSE, (cond -- )) {
             assert(PyStackRef_BoolCheck(cond));
             int flag = PyStackRef_IsFalse(cond);
             DEAD(cond);
-            JUMPBY(flag ? oparg : 0);
+            RECORD_BRANCH_TAKEN(this_instr[1].cache, flag);
+            JUMPBY(flag ? oparg : next_instr->op.code == NOT_TAKEN);
         }
 
-        op(_POP_JUMP_IF_TRUE, (cond -- )) {
+        replaced op(_POP_JUMP_IF_TRUE, (cond -- )) {
             assert(PyStackRef_BoolCheck(cond));
             int flag = PyStackRef_IsTrue(cond);
             DEAD(cond);
-            JUMPBY(flag ? oparg : 0);
+            RECORD_BRANCH_TAKEN(this_instr[1].cache, flag);
+            JUMPBY(flag ? oparg : next_instr->op.code == NOT_TAKEN);
         }
 
         op(_IS_NONE, (value -- b)) {

Include/internal/pycore_uop_metadata.h

@@ -975,12 +975,12 @@ _PyObjectArray_Free(PyObject **array, PyObject **scratch)
 
 // 1 for trace full, 0 for successful write.
 static int
-add_to_code_trace(PyThreadState *tstate, _PyInterpreterFrame *frame, PyCodeObject *old_code, _Py_CODEUNIT *this_instr, _Py_CODEUNIT *next_instr, int oparg)
+add_to_code_trace(PyThreadState *tstate, _PyInterpreterFrame *frame, PyCodeObject *old_code, _Py_CODEUNIT *this_instr, _Py_CODEUNIT *next_instr, int opcode, int oparg)
 {
     assert(frame != NULL);
     assert(tstate->interp->jit_tracer_code_curr_size < UOP_MAX_TRACE_LENGTH);
     PyFunctionObject *func = (PyFunctionObject *)PyStackRef_AsPyObjectBorrow(frame->f_funcobj);
-    return !_PyJIT_translate_single_bytecode_to_trace(tstate, this_instr, next_instr, old_code, func, oparg);
+    return !_PyJIT_translate_single_bytecode_to_trace(tstate, frame, this_instr, next_instr, old_code, func, opcode, oparg);
 }
 
 /* _PyEval_EvalFrameDefault is too large to optimize for speed with PGO on MSVC.

Python/bytecodes.c

@@ -144,7 +144,7 @@
     DISPATCH();
 #  define RECORD_TRACE() do { \
         frame->instr_ptr = next_instr; \
-        if (add_to_code_trace(tstate, frame, old_code, this_instr, next_instr, oparg)) { \
+        if (add_to_code_trace(tstate, frame, old_code, this_instr, next_instr, opcode, oparg)) { \
             BAIL_TRACING(); \
         } \
     } while (0);
@@ -407,6 +407,7 @@ do {                                                   \
     }                                                  \
     if (keep_tracing_bit) { \
         ENTER_TRACING(); \
+        _PyJIT_InitializeTracing(tstate, frame, next_instr, STACK_LEVEL(), 0); \
     } \
     else { \
         LEAVE_TRACING(); \