diff --git a/sim_atomic.h b/sim_atomic.h
index 1ad301525..ebcc7639a 100644
--- a/sim_atomic.h
+++ b/sim_atomic.h
@@ -172,12 +172,18 @@ static SIM_INLINE sim_atomic_type_t sim_atomic_add(sim_atomic_value_t *p, sim_at
             (void) __atomic_fetch_add_8(&p->value, x, __ATOMIC_ACQ_REL);
 #      endif
             /* Return the updated value. */
-            retval = sim_atomic_get(p);
+            retval = p->value;
 #    else
 #      error "sim_atomic_add: No __ATOMIC_ACQ_REL intrinsic?"
 #    endif
 #  elif defined(_WIN32) || defined(_WIN64)
+#    if defined(InterlockedAdd)
         retval = InterlockedAdd(&p->value, x);
+#    else
+        /* Older Windows InterlockedExchangeAdd, which returns the original value in p->value. */
+        InterlockedExchangeAdd(&p->value, x);
+        retval = p->value;
+#    endif
 #  else
 #    error "sim_atomic_add: No intrinsic?"
 #  endif
@@ -206,13 +212,19 @@ static SIM_INLINE sim_atomic_type_t sim_atomic_sub(sim_atomic_value_t *p, sim_at
             __atomic_fetch_sub_8(&p->value, x, __ATOMIC_ACQ_REL);
 #      endif
             /* Return the updated value. */
-            retval = sim_atomic_get(p);
+            retval = p->value;
 #    else
 #      error "sim_atomic_sub: No __ATOMIC_ACQ_REL intrinsic?"
 #    endif
 #  elif defined(_WIN32) || defined(_WIN64)
-        /* No InterlockedSub() -- use the math identity. */
-        retval = sim_atomic_add(p, -x);
+    /* There isn't a InterlockedSub function. Revert to basic math(s). */
+#    if defined(InterlockedAdd)
+        retval = InterlockedAdd(&p->value, -x);
+#    else
+        /* Older Windows InterlockedExchangeAdd, which returns the original value in p->value. */
+        InterlockedExchangeAdd(&p->value, -x);
+        retval = p->value;
+#    endif
 #  else
 #    error "sim_atomic_sub: No intrinsic?"
 #  endif
diff --git a/sim_fio.c b/sim_fio.c
index 6707d9ec1..ed1f54deb 100644
--- a/sim_fio.c
+++ b/sim_fio.c
@@ -129,15 +129,17 @@ sim_byte_swap_data (bptr, size, count);
 
 void sim_byte_swap_data (void *bptr, size_t size, size_t count)
 {
-uint32 j;
-int32 k;
+size_t j;
 unsigned char by, *sptr, *dptr;
 
 if (sim_end || (count == 0) || (size == sizeof (char)))
     return;
-for (j = 0, dptr = sptr = (unsigned char *) bptr;       /* loop on items */
-     j < count; j++) {
-    for (k = (int32)(size - 1); k >= (((int32) size + 1) / 2); k--) {
+dptr = sptr = (unsigned char *) bptr;
+for (j = 0; j < count; j++) {                           /* loop on items */
+    size_t k;
+    const size_t midpoint = (size + 1) / 2;
+
+    for (k = size - 1; k >= midpoint; k--) {
         by = *sptr;                                     /* swap end-for-end */
         *sptr++ = *(dptr + k);
         *(dptr + k) = by;