Use of language extensions
(undecidable instances, overlapping instances)
(add another mildly serious drawback of using experimental extensions)
Revision as of 23:22, 23 December 2008
Haskell compilers are full of experimental features. However, when it comes to designing libraries one should carefully think about which extensions to use and which not. Every required language extension and every imported library which on its own depends on such an extension, causes some of the following problems:
- It limits the range of Haskell compilers you can use.
- According to the availability of compilers on different platforms, it will also limit the range of machines it can be run on.
- It complicates the installation, if a user has to install a particular compiler to use your library.
- Many extensions are complicated enough to produce even more incomprehensible error messages, than errorneous Haskell 98 programs.
- Even with the same compiler, in a year or two your code might need fixing if the extension or implementation is sufficiently experimental (for example, the rules for what type-signatures GADTs require, has changed from GHC 6.4 to 6.6 to 6.8 to 6.10). People who depend on your library may be impatient to upgrade to the newest version of GHC.
We suggest the following hierarchy of complexity with respect to types:
- Simple algebraic types
- H98 type classes
- Fancier instance heads (but still single-parameter, non-overlapping)
- Existential and local universal quantification
- GADTs (I'm still uncertain about this judgement, but they seem to be less troublesome than...)
- Multiparameter type classes with functional dependencies
- Multiparameter type classes including undecidable instances, overlapping instances