diff --git a/utf8.h b/utf8.h
index b7f3b3606eaa..2b38fb2bf878 100644
--- a/utf8.h
+++ b/utf8.h
@@ -461,10 +461,19 @@ uppercase/lowercase/titlecase/fold into.
  */
 #define UTF_START_MARK(len) ((U8) ~(0xFF >> (len)))
 
-/* Masks out the initial one bits in a start byte, leaving the real data ones.
- * Doesn't work on an invariant byte.  'len' is the number of bytes in the
- * multi-byte sequence that comprises the character. */
-#define UTF_START_MASK(len) (UNLIKELY((len) >= 7) ? 0x00 : (0x1F >> ((len)-2)))
+/* Masks out the initial one bits in a start byte, leaving the following 0 bit
+ * and the real data bits.  'len' is the number of bytes in the multi-byte
+ * sequence that comprises the character.
+ *
+ * To illustrate: len = 2 => 0b0011_1111 works on start byte 110xxxxx
+ *                      6 => 0b0000_0011 works on start byte 1111110x
+ *                   >= 7 => There are no data bits in the start byte
+ * Note that on ASCII platforms, this can be passed a len=1 byte; and all the
+ * real data bits will be returned:
+                  len = 1 => 0b0111_1111
+ * This isn't true on EBCDIC platforms, where some len=1 bytes are of the form
+ * 0b101x_xxxx, so this can't be used there on single-byte characters. */
+#define UTF_START_MASK(len) (0xFF >> (len))
 
 /* Adds a UTF8 continuation byte 'new' of information to a running total code
  * point 'old' of all the continuation bytes so far.  This is designed to be