Indicate that _PyGC_Head is only 8-byte aligned. (closes bpo-33374)

fweimer · benjaminp · commit 0b91f8a66820 · 2018-04-29T12:18:33.000-07:00
By spec, the "long double" in _PyGC_Head requires the union to always be 16-byte aligned. However, obmalloc only yields 8-byte alignment. Compilers including GCC 8 are starting to use alignment information to do store-merging. So, the "long double" needs to be changed to a simple "double" as was long ago done in Python 3 by e348c8d. For 2.7, we need to add some dummy padding to make sure _PyGC_Head stays the same size.
diff --git a/Include/objimpl.h b/Include/objimpl.h
@@ -248,14 +248,29 @@ PyAPI_FUNC(PyVarObject *) _PyObject_GC_Resize(PyVarObject *, Py_ssize_t);
 /* for source compatibility with 2.2 */
 #define _PyObject_GC_Del PyObject_GC_Del
 
+/*
+ * Former over-aligned definition of PyGC_Head, used to compute the size of the
+ * padding for the new version below.
+ */
+union _gc_head;
+union _gc_head_old {
+    struct {
+        union _gc_head_old *gc_next;
+        union _gc_head_old *gc_prev;
+        Py_ssize_t gc_refs;
+    } gc;
+    long double dummy;
+};
+
 /* GC information is stored BEFORE the object structure. */
 typedef union _gc_head {
     struct {
         union _gc_head *gc_next;
         union _gc_head *gc_prev;
         Py_ssize_t gc_refs;
     } gc;
-    long double dummy;  /* force worst-case alignment */
+    double dummy; /* Force at least 8-byte alignment. */
+    char dummy_padding[sizeof(union _gc_head_old)];
 } PyGC_Head;
 
 extern PyGC_Head *_PyGC_generation0;
diff --git a/Misc/NEWS.d/next/Core and Builtins/2018-04-29-12-07-00.bpo-33374.-xegL6.rst b/Misc/NEWS.d/next/Core and Builtins/2018-04-29-12-07-00.bpo-33374.-xegL6.rst
@@ -0,0 +1,3 @@
+Tweak the definition of PyGC_Head, so compilers do not believe it is always
+16-byte aligned on x86. This prevents crashes with more aggressive
+optimizations present in GCC 8.

Original file line number	Diff line number	Diff line change
`@@ -0,0 +1,3 @@`
	`1`	`+Tweak the definition of PyGC_Head, so compilers do not believe it is always`
	`2`	`+16-byte aligned on x86. This prevents crashes with more aggressive`
	`3`	`+optimizations present in GCC 8.`