Laziness is not always good
Generally, since Haskell is a non-strict language, you should try to make a function least strict.
This is in many cases the best semantics and the most efficient implementation.
However, here is an important exception from the rule:
Consider the Monoid
instance of the null type ()
:
mempty = ()
mappend _ _ = ()
These functions are least strict, but have a subtle problem:
They do not generally satisfy the monoid laws.
Remind you: mempty
must be the identity element with respect to mappend
:
forall a. mappend mempty a = a
forall a. mappend a mempty = a
You find that it is not mappend mempty undefined = undefined
, but mappend mempty undefined = mempty
.
Is this academic nitpicking or practically relevant?
I think it is the latter one, because a Monoid
instance implicitly promises that monoid laws can be applied in every case.
A programmer expects that every occurence of mappend mempty a
can be safely replaced by a
.
You might even create an optimizer rule doing this.
The above implementation of mappend
however evaluates its operands lazily, and this gets lost when the optimization is applied.
The solution of this issue is to define
mempty = ()
mappend () () = ()
force :: () -> ()
force _ = ()
and write
mappend (force a) (force b)
instead of mappend a b
.
If you find that example too academic, you can choose any other data type with one constructor instead.
Exercise
Find out whether it would help to define mempty = undefined
.
See also
- Haskell-Cafe on Laws and partial values
- Haskell-Cafe on a space leak caused by the garbage collector that did not recognize a selector-like function call
- Maintaining laziness
- Amanda Clare has written a paper on various methods for twisting the evaluation order: http://users.aber.ac.uk/afc/stricthaskell.html.