Index: kern/kern_lock.c
===================================================================
--- kern/kern_lock.c	(revision 196271)
+++ kern/kern_lock.c	(working copy)
@@ -334,6 +334,9 @@
 	int iflags;
 
 	MPASS((flags & ~LK_INIT_MASK) == 0);
+	ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock,
+            ("%s: lockmgr not aligned for %s: %p", __func__, wmesg,
+            &lk->lk_lock));
 
 	iflags = LO_SLEEPABLE | LO_UPGRADABLE;
 	if (flags & LK_CANRECURSE)
Index: kern/kern_mutex.c
===================================================================
--- kern/kern_mutex.c	(revision 196271)
+++ kern/kern_mutex.c	(working copy)
@@ -783,8 +783,9 @@
 
 	MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
 		MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE)) == 0);
-	ASSERT_ATOMIC_LOAD(m->mtx_lock, ("%s: mtx_lock not aligned for %s: %p",
-	    __func__, name, &m->mtx_lock));
+	ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock,
+	    ("%s: mtx_lock not aligned for %s: %p", __func__, name,
+	    &m->mtx_lock));
 
 #ifdef MUTEX_DEBUG
 	/* Diagnostic and error correction */
Index: kern/kern_rwlock.c
===================================================================
--- kern/kern_rwlock.c	(revision 196271)
+++ kern/kern_rwlock.c	(working copy)
@@ -174,8 +174,9 @@
 
 	MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
 	    RW_RECURSE)) == 0);
-	ASSERT_ATOMIC_LOAD(rw->rw_lock, ("%s: rw_lock not aligned for %s: %p",
-	    __func__, name, &rw->rw_lock));
+	ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
+	    ("%s: rw_lock not aligned for %s: %p", __func__, name,
+	    &rw->rw_lock));
 
 	flags = LO_UPGRADABLE;
 	if (opts & RW_DUPOK)
Index: kern/kern_sx.c
===================================================================
--- kern/kern_sx.c	(revision 196271)
+++ kern/kern_sx.c	(working copy)
@@ -205,8 +205,9 @@
 
 	MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 	    SX_NOPROFILE | SX_NOADAPTIVE)) == 0);
-	ASSERT_ATOMIC_LOAD(sx->sx_lock, ("%s: sx_lock not aligned for %s: %p",
-	    __func__, description, &sx->sx_lock));
+	ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
+	    ("%s: sx_lock not aligned for %s: %p", __func__, description,
+	    &sx->sx_lock));
 
 	flags = LO_SLEEPABLE | LO_UPGRADABLE;
 	if (opts & SX_DUPOK)
Index: sys/systm.h
===================================================================
--- sys/systm.h	(revision 196271)
+++ sys/systm.h	(working copy)
@@ -89,9 +89,18 @@
 #define	__CTASSERT(x, y)	typedef char __assert ## y[(x) ? 1 : -1]
 #endif
 
-#define	ASSERT_ATOMIC_LOAD(var,msg)					\
-	KASSERT(sizeof(var) <= sizeof(uintptr_t) &&			\
-	    ALIGN(&(var)) == (uintptr_t)&(var), msg)
+/*
+ * Assert that a pointer can be loaded atomically with a direct memory
+ * access, even if the value is stale.
+ *
+ * On some architecture this condition is stronger than necessary as
+ * the atomic operation can be invalidated only by overflowing
+ * cache-lines boundaries, but this is a more robust and portable
+ * definition.
+ */
+#define	ASSERT_ATOMIC_LOAD_PTR(var, msg)				\
+	KASSERT(sizeof(var) == sizeof(void *) &&			\
+	    ((uintptr_t)&(var) & (sizeof(void *) - 1)) == 0, msg)
 
 /*
  * XXX the hints declarations are even more misplaced than most declarations