Angstrom: sliding window parser

Learning
1

I’m receiving chunks of data containing messages with a possible garbage between them. Each message is

<message> ::=  <constant prefix> <port> <length> <body of length>

So I have to skip some garbage, find a prefix and length, cut out a body and repeat that til the end of the chunk. One message could begin in one chunk and end in another, so I have to keep the ending of the previous chunk. I’ve written a simple Angstrom parser which do the job, but unfortunately it’s about 4 times slower than a very straightforward naive handwritten parser:

let prefix = 0x44BB

let msg_parser =
  let open Angstrom in
  let prefix_parser = BE.uint16 prefix in
  let len_parser =
    any_uint8 >>= fun port' ->
    any_uint8 >>= fun length ->
    let parity    = port' land 0x10 > 0 in
    let port      = port' land 0xF in
    let len       = (length * 2) - (if parity then 1 else 0) in
    return (port,len)
  in
  let header_parser =
    prefix_parser *> len_parser
  in
  let msg_parser' = 
    header_parser >>= fun (port,len) ->
    take_bigstring len >>= fun msg ->
    return (port, Cstruct.of_bigarray msg)
  in
  many @@ fix (fun p ->
              msg_parser' <|> (any_char *> p))

Is there any way to retrieve messages with a garbage in between in a more optimal manner?