4 files changed, 16 insertions, 10 deletions
diff --git a/include/grpc/support/alloc.h b/include/grpc/support/alloc.h
index c758065576..09ea97565b 100644
--- a/include/grpc/support/alloc.h
+++ b/include/grpc/support/alloc.h
@@ -46,8 +46,8 @@ void *gpr_malloc(size_t size);
 void gpr_free(void *ptr);
 /* realloc, never returns NULL */
 void *gpr_realloc(void *p, size_t size);
-/* aligned malloc, never returns NULL, alignment must be power of 2 */
-void *gpr_malloc_aligned(size_t size, size_t alignment);
+/* aligned malloc, never returns NULL, will align to 1 << alignment_log */
+void *gpr_malloc_aligned(size_t size, size_t alignment_log);
 /* free memory allocated by gpr_malloc_aligned */
 void gpr_free_aligned(void *ptr);
 
diff --git a/include/grpc/support/atm_win32.h b/include/grpc/support/atm_win32.h
index acacf12013..9bb1cfec35 100644
--- a/include/grpc/support/atm_win32.h
+++ b/include/grpc/support/atm_win32.h
@@ -93,11 +93,13 @@ static __inline gpr_atm gpr_atm_no_barrier_fetch_add(gpr_atm *p,
 static __inline gpr_atm gpr_atm_full_fetch_add(gpr_atm *p, gpr_atm delta) {
   /* Use a CAS operation to get pointer-sized fetch and add */
   gpr_atm old;
+#ifdef GPR_ARCH_64
   do {
     old = *p;
-#ifdef GPR_ARCH_64
   } while (old != (gpr_atm)InterlockedCompareExchange64(p, old + delta, old));
 #else
+  do {
+    old = *p;
   } while (old != (gpr_atm)InterlockedCompareExchange(p, old + delta, old));
 #endif
   return old;
diff --git a/include/grpc/support/port_platform.h b/include/grpc/support/port_platform.h
index 27efa29448..0a651757bc 100644
--- a/include/grpc/support/port_platform.h
+++ b/include/grpc/support/port_platform.h
@@ -147,16 +147,18 @@
 #include <stdint.h>
 
 /* Cache line alignment */
-#ifndef GPR_CACHELINE_SIZE
+#ifndef GPR_CACHELINE_SIZE_LOG
 #if defined(__i386__) || defined(__x86_64__)
-#define GPR_CACHELINE_SIZE 64
+#define GPR_CACHELINE_SIZE_LOG 6
 #endif
-#ifndef GPR_CACHELINE_SIZE
+#ifndef GPR_CACHELINE_SIZE_LOG
 /* A reasonable default guess. Note that overestimates tend to waste more
    space, while underestimates tend to waste more time. */
-#define GPR_CACHELINE_SIZE 64
-#endif /* GPR_CACHELINE_SIZE */
-#endif /* GPR_CACHELINE_SIZE */
+#define GPR_CACHELINE_SIZE_LOG 6
+#endif /* GPR_CACHELINE_SIZE_LOG */
+#endif /* GPR_CACHELINE_SIZE_LOG */
+
+#define GPR_CACHELINE_SIZE (1 << GPR_CACHELINE_SIZE_LOG)
 
 /* scrub GCC_ATOMIC if it's not available on this compiler */
 #if defined(GPR_GCC_ATOMIC) && !defined(__ATOMIC_RELAXED)
diff --git a/include/grpc/support/slice.h b/include/grpc/support/slice.h
index 261e3baabe..8a2129028f 100644
--- a/include/grpc/support/slice.h
+++ b/include/grpc/support/slice.h
@@ -165,7 +165,9 @@ gpr_slice gpr_slice_split_head(gpr_slice *s, size_t split);
 
 gpr_slice gpr_empty_slice(void);
 
-/* Returns <0 if a < b, ==0 if a == b, >0 if a > b */
+/* Returns <0 if a < b, ==0 if a == b, >0 if a > b
+   The order is arbitrary, and is not guaranteed to be stable across different
+   versions of the API. */
 int gpr_slice_cmp(gpr_slice a, gpr_slice b);
 int gpr_slice_str_cmp(gpr_slice a, const char *b);