Use lazy allocation

Author: corona10

Include/internal/pycore_tstate.h

@@ -78,7 +78,7 @@ typedef struct _PyThreadStateImpl {
 #endif
 
 #ifdef _Py_TIER2
-    struct _PyUOpInstruction buffer[UOP_MAX_TRACE_LENGTH];
+    struct _PyUOpInstruction *buffer;
 #endif
 
 } _PyThreadStateImpl;

Python/optimizer.c

@@ -1281,6 +1281,13 @@ uop_optimize(
     _PyBloomFilter dependencies;
     _Py_BloomFilter_Init(&dependencies);
     _PyThreadStateImpl *tstate = (_PyThreadStateImpl *)_PyThreadState_GET();
+    if (tstate->buffer == NULL) {
+        tstate->buffer = (_PyUOpInstruction *)PyMem_RawMalloc(UOP_MAX_TRACE_LENGTH*sizeof(_PyUOpInstruction));
+        if (tstate->buffer == NULL) {
+            PyErr_NoMemory();
+            return -1;
+        }
+    }
     _PyUOpInstruction *buffer = tstate->buffer;
     OPT_STAT_INC(attempts);
     char *env_var = Py_GETENV("PYTHON_UOPS_OPTIMIZE");

Python/pylifecycle.c

@@ -1704,6 +1704,11 @@ finalize_modules(PyThreadState *tstate)
     interp->jit = false;
 #ifdef _Py_TIER2
     _Py_Executors_InvalidateAll(interp, 0);
+    _PyThreadStateImpl *_tstate = (_PyThreadStateImpl *)tstate;
+    if (_tstate->buffer != NULL) {
+        PyMem_RawFree(_tstate->buffer);
+        _tstate->buffer = NULL;
+    }
 #endif
 
     // Stop watching __builtin__ modifications