"Data.BufferBuilder" is an efficient library for incrementally building up 'ByteString's, one chunk at a time. Early benchmarks show it is over twice as fast as ByteString Builder, primarily because 'BufferBuilder' is built upon an ST-style restricted monad and mutable state instead of ByteString Builder's monoidal AST.
Internally, BufferBuilder is backed by a few C functions. Examination of GHC's output shows nearly optimal code generation with no intermediate thunks -- and thus, continuation passing and its associated indirect jumps and stack traffic only occur when BufferBuilder is asked to append a non-strict ByteString.
I benchmarked four approaches with a URL encoding benchmark:
State monad, concatenating ByteStrings: 6.98 us
State monad, ByteString Builder: 2.48 us
Crazy explicit RealWorld baton passing with unboxed state: 28.94 us (GHC
generated really awful code for this, but see the revision history for the technique)
Using BufferBuilder is very simple:
> import qualified Data.BufferBuilder as BB > > let byteString = BB.runBufferBuilder $ do > BB.appendBS "http" > BB.appendChar8 '/' > BB.appendBS "//"
This package also provides "Data.BufferBuilder.Utf8" for generating UTF-8 buffers and "Data.BufferBuilder.Json" for encoding data structures into JSON.