i#3699 ARM: Add ALIGN_VAR(8) to _dr_simd_t.

2dfd7cb made the size of priv_mcontext_t a multiple of 8, enabling
alignment when the struct is pushed onto the stack by assembly code.
This patch adds ALIGN_VAR(8) to the definition of _dr_simd_t so that
alignment is also enforced by the compiler.

Change-Id: I3afd0a9babfe4091f414fb72a569ffd480db5985

api/docs/release.dox

@@ -126,9 +126,9 @@ clients.
 
 The changes between version \DR_VERSION and 11.3.0 include the following compatibility
 changes:
- - The size of #dr_mcontext_t on 32-bit Arm has been increased by 4 so that
-   the struct can be pushed onto the 8-byte aligned stack without additional
-   padding. The offset of the field "simd" has changed.
+ - On 32-bit Arm the size of #dr_mcontext_t has been increased by 4 and the
+   struct is now required to be 8-byte aligned. The offset of the field "simd"
+   has changed.
 
 Further non-compatibility-affecting changes include:
  - Added support for reading a single drmemtrace trace file from stdin

core/lib/globals_api.h

@@ -744,7 +744,7 @@ typedef union _dr_svep_t {
  */
 typedef dr_svep_t dr_ffr_t;
 #    else
-typedef union _dr_simd_t {
+typedef union ALIGN_VAR(8) _dr_simd_t {
     uint s[4];   /**< Representation as 4 32-bit Sn elements. */
     uint d[4];   /**< Representation as 2 64-bit Dn elements: d[3]:d[2]; d[1]:d[0]. */
     uint u32[4]; /**< The full 128-bit register. */