bpo-33312: Fix clang ubsan out of bounds warnings in dict. (GH-6537)

Fix clang ubsan (undefined behavior sanitizer) warnings in dictobject.c by
adjusting how the internal struct _dictkeysobject shared keys structure is
declared.

This remains ABI compatible.  We get rid of the union at the end of the
struct being used for conveinence to avoid typecasting in favor of char[]
variable length array at the end of a struct. This is known to clang to be
used for variable sized objects and will not cause an undefined behavior
problem.  Similarly, char arrays do not have strict aliasing undefined
behavior when cast.

PEP-007 does not currently list variable length arrays (VLAs) as allowed
in our subset of C99.  If this turns out to be a problem, the fix to this is
to change the char `dk_indices[]` into `dk_indices[1]` and restore the
three size computation subtractions this change removes:
  `- Py_MEMBER_SIZE(PyDictKeysObject, dk_indices)`

If this works as is I'll make a separate PR to update PEP-007.

1 files changed, 2 insertions, 9 deletions

diff --git a/Objects/dict-common.h b/Objects/dict-common.h
index 3e524686b44..71d6b027442 100644
--- a/Objects/dict-common.h
+++ b/Objects/dict-common.h
@@ -58,15 +58,8 @@ struct _dictkeysobject {
        - 4 bytes if dk_size <= 0xffffffff (int32_t*)
        - 8 bytes otherwise (int64_t*)
 
-       Dynamically sized, 8 is minimum. */
-    union {
-        int8_t as_1[8];
-        int16_t as_2[4];
-        int32_t as_4[2];
-#if SIZEOF_VOID_P > 4
-        int64_t as_8[1];
-#endif
-    } dk_indices;
+       Dynamically sized, SIZEOF_VOID_P is minimum. */
+    char dk_indices[];  /* char is required to avoid strict aliasing. */
 
     /* "PyDictKeyEntry dk_entries[dk_usable];" array follows:
        see the DK_ENTRIES() macro */