cohttp.headers: faster faster comparison

see
  mirage#778 (comment)

http/src/http.ml

@@ -5,21 +5,50 @@ module Transfer = struct
 end
 
 module Header = struct
+  (* [caseless_equal a b] must be equivalent to
+     [String.equal (String.lowercase_ascii a) (String.lowercase_ascii b)]. *)
   let caseless_equal a b =
     if a == b then true
     else
       let len = String.length a in
       len = String.length b
+      (* Note: at this point we konw that [a] and [b] have the same length. *)
       &&
-      let stop = ref false in
-      let idx = ref 0 in
-      while (not !stop) && !idx < len do
-        let c1 = String.unsafe_get a !idx in
-        let c2 = String.unsafe_get b !idx in
-        if Char.lowercase_ascii c1 <> Char.lowercase_ascii c2 then stop := true;
-        incr idx
-      done;
-      not !stop
+      (* [word_loop a b i len] compares strings [a] and [b] from
+         offsets [i] (included) to [len] (excluded), one word at a time.
+         [i] is a world-aligned index into the strings.
+      *)
+      let rec word_loop a b i len =
+        if i = len then true
+        else
+          let i' = i + 8 in
+          (* If [i' > len], what remains to be compared is strictly
+             less than a word long, use byte-per-byte comparison. *)
+          if i' > len then byte_loop a b i len
+          else begin
+            if String.get_int64_ne a i = String.get_int64_ne b i
+            then word_loop a b i' len
+            else
+              (* If the words at [i] differ, it may due to a case
+                 difference; we check the individual bytes of this
+                 work, and then we continue checking the other
+                 words. *)
+              byte_loop a b i i'
+              && word_loop a b i' len
+          end
+      (* [byte_loop a b i len] compares the strings [a] and [b] from
+         offsets [i] (included) to [len] (excluded), one byte at
+         a time.
+
+         This function assumes that [i < len] holds -- its only called
+         by [word_loop] when this is known to hold. *)
+      and byte_loop a b i len =
+        let c1 = String.unsafe_get a i in
+        let c2 = String.unsafe_get b i in
+        Char.lowercase_ascii c1 = Char.lowercase_ascii c2
+        && (let i' = i + 1 in i' = len || byte_loop a b i' len)
+      in
+      word_loop a b 0 len
 
   type t = (string * string) list