Starting offset not necessary

[ekmett]

GHC >= 8.2 exports plusForeignPtr, which allows you to switch the address targeted by a ForeignPtr. The finalizers remain attached. However, this appears to be implementable, by hand, using the public contents of GHC.ForeignPtr clear back to at least GHC 7.0!

Removing the offset would save an entire word per bytestring!

It may be worth including a backwards compatibility shim on modern enough GHCs to support bidirectional pattern synonyms. This would enable most consumers that use the Data.ByteString.Internal module to proceed unpatched.

e.g. something like

module Data.ByteString.Internal 
  ( ByteString(BS, PS) -- add some c preprocessing to handle older GHCs that don't support data-type associated patterns
  , ...
  ) where

...

data ByteString = BS !(ForeignPtr Word8) !Int

pattern PS :: ForeignPtr Word -> Int -> Int -> ByteString
pattern PS fp 0 len <- BS fp len where
  PS fp o len = BS (plusForeignPtr fp o) len

With sufficient rewrite rules for replacing plusForeignPtr fp 0 with fp it should give almost all of the benefits of a complete rewrite of user code.

This would be a major version bump as the behavior of round-tripping through the pattern synonym would change.

I don't have any compilers older than GHC 7.0 to test with, so I'm not sure how much further back into antiquity this can be ported. The current lower bound on base for bytestring is set at GHC 6.12, so this doesn't appear to slide our compatibility window much if at all.

[ekmett]

Update: I'm implementing this idea in

https://github.com/ekmett/bytestring/tree/no-offset

[ekmett]

Addressed by #175.