GitLab

The main goal is to reduce the overhead of things like:

data Bloated =
  Bloated {-# UNPACK #-} !Word8
          {-# UNPACK #-} !Int8
          {-# UNPACK #-} !Bool

Assuming 64-bit architecture, currently those fields will take 8 bytes each! So for this example we'd need: 8 bytes for header + 3 * 8 bytes for fields = 32 bytes. But we should be able to pack the fields into a single word (a word is 8 bytes and each field really only needs 1 byte) for a total of 16 bytes (8 bytes header + 8 bytes for fields, with the 5 bytes being "overhead" due to heap alignment).

My understanding is that we need a few things to make this happen:

• Ability to refer to fields that are packed into a single word (currently everything in GHC assumes that each field occupies a single word). Simon Marlow started working on this in https://phabricator.haskell.org/D38
• Introduce primitives like Word8#, Int8#, ... (currently WordX and IntX are defined as wrappers of Word# and Int# respectively) and change WordX/IntX definitions to use those primitives.
• ~~Figure out what to do with Bool (should it be just Word8#? should we have Bool#?) and change its definition (using pattern synonyms for True/False)~~ Bool should be handled by #605 (see ticket:13825#comment:138006)

Some additional info:

• Thread on ghc-devs: https://mail.haskell.org/pipermail/ghc-devs/2017-June/014304.html