Optimize binary decoding of builtins

[sjakobi]

• Use decodeUtf8ByteArray to avoid UTF16-encoding the scrutinee.

• Optimize the pattern matching by grouping the patterns by length.

  GHC currently doesn't produce static length information for string
  literals. Consequently the pattern matching worked somewhat like this:

  s <- decodeString
  
  let len_s = length s
  
  if len_s == length "Natural/build" && sameBytes s "Natural/build"
      then return NaturalBuild
      else if len_s == length "Natural/fold" && sameBytes s "Natural/fold"
               ...
  

  Decoding Sort, the most extreme case, would involve a total of 32
  conditional jumps as a consequence of length comparisons alone.

  Judging by the Core, we can get that number down to 8 by grouping
  the patterns by length: One to check the length of the decoded string,
  and (unfortunately) still one each for the 7 candidate literals of
  length 4.

  The number of string content comparisons should be unchanged.

The result of these optimizations is that the time to decode the cache for cpkg
is reduced by 7-9%. Decoding time for the Prelude goes down by 13-16%.

This also changes the builtin encoding to use encodeUtf8ByteArray in order
to avoid UTF16-encoding and decoding the builtins strings. I didn't check
the performance implications though.

Context: #1804.

[sjakobi]

Judging by the Core, we can get that number down to 8 by grouping
the patterns by length: One to check the length of the decoded string,
and (unfortunately) still one each for the 7 candidate literals of
length 4.

These completely unnecessary length checks on the literals come from ShortByteString's Eq instance. To avoid them, we could try using the compareByteArrays# primop instead. primitive has a nice backward-compatible wrapper for it, but it's unfortunately not currently exported: haskell/primitive#131

[Gabriella439]

Very nice work! 🙂