Fix data race on CodeExtra::calls for FT Python

Summary:
Fix TSAN-reported data race on extra->calls by introducing atomic accessor functions for FT-Python.

ThreadSanitizer: data race fbcode/cinderx/Interpreter/3.14/Includes/generated_cases.c.h:4249 in Ci_EvalFrame
==================
```
- After
No race warnings.

Reviewed By: DinoV

Differential Revision: D93088007

fbshipit-source-id: 086c562d3bc73fd0286eaabfd17d87443666d4f1

Author: yoney

cinderx/Common/code_extra.h

@@ -4,6 +4,10 @@
 
 #include <stdint.h>
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 // Extra data attached to a code object.
 typedef struct CodeExtra {
   union {
@@ -13,3 +17,35 @@ typedef struct CodeExtra {
     struct CodeExtra* next;
   };
 } CodeExtra;
+
+// Thread-safe accessors for CodeExtra::calls.
+// Under FT-Python, these use atomics to avoid data races.
+#ifdef Py_GIL_DISABLED
+
+// Note: _Py_atomic_add_uint64 uses seq_cst ordering, which might be stronger
+// than needed for the calls counter. On x86-64, this is the same cost as
+// relaxed (both emit lock xaddq). On ARM, a relaxed variant would be cheaper
+// but there is no _Py_atomic_add_uint64_relaxed.
+static inline void Ci_code_extra_incr_calls(CodeExtra* extra) {
+  _Py_atomic_add_uint64(&extra->calls, 1);
+}
+
+static inline uint64_t Ci_code_extra_get_calls(const CodeExtra* extra) {
+  return _Py_atomic_load_uint64_relaxed(&extra->calls);
+}
+
+#else
+
+static inline void Ci_code_extra_incr_calls(CodeExtra* extra) {
+  extra->calls += 1;
+}
+
+static inline uint64_t Ci_code_extra_get_calls(const CodeExtra* extra) {
+  return extra->calls;
+}
+
+#endif
+
+#ifdef __cplusplus
+}
+#endif

cinderx/Interpreter/3.14/Includes/ceval_macros.h

@@ -453,7 +453,7 @@ do { \
                 if (extra == NULL) { \
                     adaptive_enabled = false; \
                 } else { \
-                    extra->calls += 1; \
+                    Ci_code_extra_incr_calls(extra); \
                     adaptive_enabled = is_adaptive_enabled(extra); \
                 } \
             } \

cinderx/Interpreter/3.14/interpreter.c

@@ -38,7 +38,7 @@ bool Ci_DelayAdaptiveCode = false;
 uint64_t Ci_AdaptiveThreshold = 80;
 
 bool is_adaptive_enabled(CodeExtra *extra) {
-    return !Ci_DelayAdaptiveCode || extra->calls > Ci_AdaptiveThreshold;
+    return !Ci_DelayAdaptiveCode || Ci_code_extra_get_calls(extra) > Ci_AdaptiveThreshold;
 }
 
 #endif

cinderx/Interpreter/3.15/interpreter.c

@@ -35,7 +35,7 @@ bool Ci_DelayAdaptiveCode = false;
 uint64_t Ci_AdaptiveThreshold = 80;
 
 bool is_adaptive_enabled(CodeExtra *extra) {
-    return !Ci_DelayAdaptiveCode || extra->calls > Ci_AdaptiveThreshold;
+    return !Ci_DelayAdaptiveCode || Ci_code_extra_get_calls(extra) > Ci_AdaptiveThreshold;
 }
 
 #endif
@@ -454,7 +454,7 @@ Py_ssize_t load_method_static_cached_oparg_slot(int oparg) {
         if (extra == NULL) {                                                 \
           adaptive_enabled = false;                                          \
         } else {                                                             \
-          extra->calls += 1;                                                 \
+          Ci_code_extra_incr_calls(extra);                                   \
           adaptive_enabled = is_adaptive_enabled(extra);                     \
         }                                                                    \
       }                                                                      \
@@ -547,7 +547,7 @@ void **opcode_targets = opcode_targets_table;
 
     // Suppress unused variable warning because it's too hard to improve the
     // variable's scope to avoid an unused-but-set-variable warning.
-    (void)adaptive_enabled; 
+    (void)adaptive_enabled;
 
     /* support for generator.throw() */
     if (throwflag) {

cinderx/Jit/pyjit.cpp

@@ -102,7 +102,7 @@ uint64_t countCalls(PyCodeObject* code) {
   return code->co_mutable->ncalls;
 #else
   auto extra = codeExtra(code);
-  return extra != nullptr ? extra->calls : 0;
+  return extra != nullptr ? Ci_code_extra_get_calls(extra) : 0;
 #endif
 }
 

cinderx/RuntimeTests/backend_test.cpp

@@ -297,7 +297,7 @@ for i in range(30):
 #else
   auto extra = codeExtra(code);
   ASSERT_NE(extra, nullptr) << "Failed to load code object extra data";
-  uint64_t ncalls = extra->calls;
+  uint64_t ncalls = Ci_code_extra_get_calls(extra);
 #endif
 
   // TASK(T190615535): This is waiting on the 3.12 custom interpreter loop.