I suppose it should have gone like this: * You need to unmarshal hardware floats * You visit Hackage, and use the IEEE754 package. http://hackage.haskell.org/

I suppose it should have gone like this: * You need to unmarshal hardware floats * You visit Hackage, and use the IEEE754 package. http://hackage.haskell.org/package/data-binary-ieee754 http://hackage.haskell.org/package/data-binary-ieee754 If that is fast enough, you stop. You are done. * IEEE754 is too slow for some reason, so you decide to do your own unsafeCoerce of bytestring * You look at the source for IEE754 and decide to marshal memory yourself: http://hackage.haskell.org/packages/archive/data-binary-ieee754/0.4.4/doc/html/src/Data-Binary-IEEE754.html#wordToDouble http://hackage.haskell.org/packages/archive/data-binary-ieee... * You read the bytestring package docs: http://hackage.haskell.org/packages/archive/bytestring/0.9.1.4/doc/html/Data-ByteString-Internal.html http://hackage.haskell.org/packages/archive/bytestring/0.9.1... and pay attention to the "unsafe" functions that cast raw pointers. * That lets you re-implement the IEEE754 package your own way The mistake here was not using the highly-optimized existing library - http://hackage.haskell.org/packages/archive/data-binary-ieee754/0.4.4/doc/html/src/Data-Binary-IEEE754.html#wordToDouble http://hackage.haskell.org/packages/archive/data-binary-ieee... - which already does the correct thing. You jumped straight to "have to roll my own low level binary parsing" and then started a reddit thread instead. None of that was necessary.