Re: [Haskell-cafe] Possible Improvements

On Mon, 2007-12-03 at 10:48 +0100, Ketil Malde wrote:

"Johan Tibell" writes:

...

It would be great if someone could exemplify these "rules of thumb", e.g. "Primitive types such as Int should be strict unless in the three canonical examples X, Y and Z." My strictness radar is still quite poor and I feel I can't make informed decisions on when I need to make something more strict or lazy.

I find that I often need to add strictness when:

left thumb) parsing [Char] into something more compact, i.e. almost all cases. right thumb) storing data into maps, especially when the values are produced by multiple updates - i.e. doing word frequency counts.

Indeed, this generalizes fairly well. In general when going from a "large" structure (especially recursive types or arrays) to a "small" one (especially base types or small non-recursive types e.g. a vector type) you want strictness. Cale Gibbard argues that this is the only case where strictness is desirable. In the other three cases, "small" to "large", "large" to "large" and "small" to "small" either laziness is preferable or there is not a big difference between them. http://www.haskell.org/haskellwiki/Stack_overflow gives some advice on how to choose strictness for avoiding stack overflows. On that page you can see the above rule in action in, for example, the difference between concat :: [[a]] -> [a] and sum :: Num a => [a] -> a. The techniques sidebar on the Performance page, http://www.haskell.org/haskellwiki/Performance also contains some bits of advice. For example, the widely known advice about making accumulating parameters strict. This is related to Ketil's "right rule of thumb." Oftentimes the best way to get good behaviour is to use strict (in the appropriate places) data constructors. This will often eliminate most or all of the need for (other) strictness annotations. For example, one way of solving the issue with the scanl code on the Stack Overflow wiki page is by using a head strict list type (which, incidentally, Clean has native support for.) In fact, I suspect most of the time a head strict list type is either comparable or what is desired (though certainly not all of the time).