Fix test_atomic failure caused unaligned AO_double_t access on x86 (VC++)

Passing an unaligned AO_double_t pointer to AO double-wide primitives
results in an undefined behavior of the latter when running on x86 (or
violation of the corresponding assertion on the alignment).
MS VC++/x86 aligns AO_double_t values on a 4-byte boundary by default.
For the proper alignment, __declspec(align(8)) should be applied to
variables defined in the client code (which uses double-wide AO
primitives).  Unfortunately, the attribute cannot be added to
AO_double_t definition itself because the compiler does not allow the
attribute for function arguments.

This patch introduces AO_DOUBLE_ALIGN attribute for use by clients of
the double-wide AO primitives (and, thus, AO_stack clients).
Matters only Visual Studio compiler for X86.
The inner clients (atomic_ops_malloc, test_atomic, test_stack) are
updated to use this attribute.

* doc/README_win32.txt [x86] (AO_DOUBLE_ALIGN): Document.
* src/atomic_ops/sysdeps/generic_pthread.h (AO_DOUBLE_ALIGN): Define
(as empty).
* src/atomic_ops/sysdeps/standard_ao_double_t.h [!_WIN64 && _WIN32
&& !__GNUC__ && _MSC_VER] (AO_DOUBLE_ALIGN): Define as declspec
align(8); document it.
* src/atomic_ops/sysdeps/standard_ao_double_t.h [!AO_DOUBLE_ALIGN]
(AO_DOUBLE_ALIGN): Define as empty (otherwise).
* src/atomic_ops_stack.h [AO_USE_ALMOST_LOCK_FREE && !AO_DOUBLE_ALIGN]
(AO_DOUBLE_ALIGN): Likewise.
* src/atomic_ops_malloc.c (AO_free_list): Use AO_DOUBLE_ALIGN attribute.
* tests/test_stack.c (the_list): Likewise.
* src/atomic_ops_stack.h [!AO_USE_ALMOST_LOCK_FREE] (AO_stack_t):
Document AO_DOUBLE_ALIGN usage (by clients).
* tests/test_atomic_include.template (test_atomicXX): Use
AO_DOUBLE_ALIGN attribute for old_w and w double-wide local variables
(to avoid alignment assertion violation or AO primitives undefined
behavior on x86 if the test code is compiled by VC++).

2 files changed, 15 insertions, 0 deletions

diff --git a/src/atomic_ops/sysdeps/generic_pthread.h b/src/atomic_ops/sysdeps/generic_pthread.h
index 3c65624..2e70068 100644
--- a/src/atomic_ops/sysdeps/generic_pthread.h
+++ b/src/atomic_ops/sysdeps/generic_pthread.h
@@ -367,6 +367,7 @@ typedef struct {
         AO_t AO_val2;
 } AO_double_t;
 #define AO_HAVE_double_t
+#define AO_DOUBLE_ALIGN /* empty */
 
 #define AO_DOUBLE_T_INITIALIZER { (AO_t)0, (AO_t)0 }
 
diff --git a/src/atomic_ops/sysdeps/standard_ao_double_t.h b/src/atomic_ops/sysdeps/standard_ao_double_t.h
index 7d85c9d..636ba46 100644
--- a/src/atomic_ops/sysdeps/standard_ao_double_t.h
+++ b/src/atomic_ops/sysdeps/standard_ao_double_t.h
@@ -50,11 +50,25 @@
   typedef __m128 double_ptr_storage;
 #elif defined(_WIN32) && !defined(__GNUC__)
   typedef unsigned __int64 double_ptr_storage;
+# ifdef _MSC_VER
+    /* VC++/x86 does not align __int64 properly by default, thus,       */
+    /* causing an undefined behavior or assertions violation in         */
+    /* the double-wide atomic primitives.  For the proper alignment,    */
+    /* all variables of AO_double_t type (in the client code) those     */
+    /* address is passed to an AO primitive should be defined with the  */
+    /* given attribute.  Not a part of double_ptr_storage because the   */
+    /* attribute cannot be applied to function parameters.              */
+#   define AO_DOUBLE_ALIGN __declspec(align(8))
+# endif
 #else
   typedef unsigned long long double_ptr_storage;
 #endif
 # define AO_HAVE_DOUBLE_PTR_STORAGE
 
+#ifndef AO_DOUBLE_ALIGN
+# define AO_DOUBLE_ALIGN /* empty */
+#endif
+
 typedef union {
     struct { AO_t AO_v1; AO_t AO_v2; } AO_parts;
         /* Note that AO_v1 corresponds to the low or the high part of   */