Boxed versus Unboxed

Hal Daume III hdaume@ISI.EDU
Wed, 2 Oct 2002 11:20:14 -0700 (PDT)


We can think of it like this:

  When we have a value x :: Int, then 'x' is a computation
  which when evaluated will return either an Int or will
  be bottom (undefined).

When the program runs and x is evaluated, suppose it evalutes to an
actual Int (not bottom).  Then in the future any time x is evaluated,
instead of redoing the entire computation, we only want to get the value
out which we previously calculated.

What we do to accomplish this is to replace to thunk (computation) which
calculates x with a thunk which simply returns the value which was
computed before.

The problem is that every time you need to get x in the future, you have
to follow this pointer to the (trivial) code which returns a value.  THis
gets expensive if you need these values frequently.

Enter unboxed values.  An unboxed value is just that low-level value, not
wrapped inside a thunk.  This means that it is strict in the sense that it
cannot be undefined without your program necessarily dying.

In GHC, the type of an unboxed integer is Int#.  In fact, the Int type is
defined in terms of this:

   data Int = Int Int#

You can get more information that you really want from SPJ's paper
"Implementing lazy functional languages on stock hardware: the Spineless
Tagless G-machine" which talks about this issue.

--
Hal Daume III

 "Computer science is no more about computers    | hdaume@isi.edu
  than astronomy is about telescopes." -Dijkstra | www.isi.edu/~hdaume

On Wed, 2 Oct 2002, Shawn P. Garbett wrote:

> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> I've been reading on this list about boxed/unboxed structures. I'm not quite 
> clear what this terminology means. Is there a good paper that covers this 
> topic in detail, or a book reference?
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v1.0.7 (GNU/Linux)
> 
> iEYEARECAAYFAj2a8OEACgkQDtpPjAQxZ6DUSgCeIRgAl3dpOPavJhrmwFxal5O1
> Y8IAnRDzmdag9lyljlacKtE4aZMujfiG
> =h2Zn
> -----END PGP SIGNATURE-----
> 
> _______________________________________________
> Haskell-Cafe mailing list
> Haskell-Cafe@haskell.org
> http://www.haskell.org/mailman/listinfo/haskell-cafe
>