"Parse, don't validate" using ViewPatterns

In 2019, Alexis King wrote the Parse, don't validate article that catches some of the essences of type-driven software development and design. Several books were released on this matter, including Alexander Granin's Functional Design and Architecture, Sandy Maguire's Thinking with Types: Type-Level Programming in Haskell, or even Edwin Brady's Type-Driven Development with Idris for a book about the future.

This blog post is about where Alexis' three words got me today.

I refactor a Haskell library that handles Ethereum VM opcodes, and in particular, the part of the library that lets you convert a Word8 into an Opcode.

The first iteration of the code was derived from Hevm, which interprets binary opcodes using this function:

readOp :: Word8 -> ByteString -> Op
readOp x _  | x >= 0x80 && x <= 0x8f = OpDup (x - 0x80 + 1)
readOp x _  | x >= 0x90 && x <= 0x9f = OpSwap (x - 0x90 + 1)
readOp x _  | x >= 0xa0 && x <= 0xa4 = OpLog (x - 0xa0)
readOp x xs | x >= 0x60 && x <= 0x7f =
  let n   = x - 0x60 + 1
      xs' = BS.take (num n) xs
  in OpPush (word xs')
readOp x _ = case x of
  0x00 -> OpStop
  ...
  _ -> OpUnknown x

I would like for that function to live in evm-opcodes, but it needs cleaning up:

The x >= 0x80 && x <= 0x8f stuff is hard to read, but it also sort of ties together the hex codes and the OpDup constructor. If the hex codes were only mentioned here, that'd be fine, but they are littered throughout the codebase, making the interpreter harder to read.
Instead of inventing an OpUnknown error case, use Nothing. One may argue that OpUnknown x is better represented as Left x, but the caller already has x.

The first rewrite looks like this:

isDUP, isSWAP, isLOG, isPUSH :: Word8 -> Bool
isDUP b  = b >= 0x80 && b <= 0x8f
isSWAP b = b >= 0x90 && b <= 0x9f
isLOG b  = b >= 0xa0 && b <= 0xa4
isPUSH b = b >= 0x60 && b <= 0x7f

-- | Parse an 'Opcode' from a 'Word8'. In case of 'PUSH' instructions, read the
-- constant being pushed from a subsequent 'ByteString'.
readOp :: Word8 -> ByteString -> Maybe Opcode
readOp b bs
  | isDUP b   = pure . DUP  . fromWord8 $ b - 0x80
  | isSWAP b  = pure . SWAP . fromWord8 $ b - 0x90
  | isLOG b   = pure . LOG  . fromWord8 $ b - 0xa0
  | isPUSH b  = let n = fromIntegral (b - 0x60 + 1)
                in PUSH <$> word256 (BS.take n bs)
  | otherwise = readOp' b

readOp' :: Word8 -> Maybe Opcode
readOp' word = case word of
  0x00 -> pure STOP
  ...
  _    -> Nothing

This fixes some things and makes other things worse. While isDUP can be used elsewhere, removing many magical constants in many places, it preserves some magical constants in readOp that are now no longer on the same line. This means they may drift apart.

In the case of readOp, a better programming pattern is needed. It is time to parse, rather than validate. Omitting a few imports and helper functions,

readDUP, readSWAP, readLOG :: Word8 -> Maybe Opcode
readDUP b = do
  guard (b >= 0x80 && b <= 0x8f)
  pure (DUP (fromWord8 (b - 0x80)))

readSWAP b = do
  guard (b >= 0x90 && b <= 0x9f)
  pure (SWAP (fromWord8 (b - 0x90)))

readLOG b = do
  guard (b >= 0xa0 && b <= 0xa4)
  pure (LOG (fromWord8 (b - 0xa0)))

readPUSH :: Word8 -> ByteString -> Maybe Opcode
readPUSH b bs = do
  guard (b >= 0x60 && b <= 0x7f)
  let n = fromIntegral (b - 0x60 + 1)
  PUSH <$> word256 (BS.take n bs)

The is... functions may come handy since their definitions repeat throughout another codebase. They can be written without duplicating the hex ranges:

isDUP, isSWAP, isLOG  :: Word8 -> Bool
isDUP = isJust . readDUP
isSWAP = isJust . readSWAP
isLOG = isJust . readLOG

isPUSH :: Word8 -> ByteString -> Bool
isPUSH b bs = isJust (readPUSH b bs)

As for using these read... parsers instead of their is... predecessors, GHC's extension called ViewPatterns enables the following syntax:

readOp :: Word8 -> ByteString -> Maybe Opcode
readOp (readDUP -> Just dup) _ = Just dup
readOp (readSWAP -> Just swap) _ = Just swap
readOp (readLOG -> Just log) _ = Just log
readOp b (readPUSH b -> Just push) = Just push
readOp b _bs = case b of
  0x00 -> pure STOP
  ...
  _    -> Nothing

At this point, a hierarchy of is... and read... helper functions are composed to form readOp in a way that does not leak hex codes and does not duplicate those hex codes. The readOp (read...) and readOp ... Just ... = Just ... repetitions make it seem like there could be a better way.

But I'm happy for now.

Blog

"Parse, don't validate" using ViewPatterns

Simon Shine

Join Our Newsletter. No Spam, Only the good stuff.

Related