I always thought that story was a bit unfortunate -- mostly because I don't
believe a good solution emerged. I also find it odd that they felt the need
to continually _remove_ Eval annotations. Leaving an unused extraneous Eval
annotation in place should be mostly harmless, An extra dictionary being
passed around here or there shouldn't destroy performance too badly -- you
only need to pass them in when you are polymorphic in the argument type, and
how often is your code really polymorphic _in something you want to
arbitrarily seq_ in a performance sensitive part of your TCP/IP stack?
Asking to seq a polymorphic argument these days is generally taken as a sign
that you are sprinkling seq's around without understanding why. We have
strategies now for a reason.
-Edward Kmett
On Fri, Nov 6, 2009 at 1:38 AM, Stefan Holdermans
Henning,
If I understood it correctly, the problem was more general than just
debugging. Every introduction of seq in a function could result in the requirement to also adapt the type signatures of calling functions.
Sure, but why was this a problem? Because they had to re-arrange a lot,
and had to change the signature each time. But once that re-arrangement settles, it would be nice to have the Seq type constraint, right?
I cannot tell whether *I* would find it problematic in practice. Hudak et al. write:
"However, the limitations of this solution soon became apparent. Inspired by the Fox project at CMU, two of Hughes’s students implemented a TCP/IP stack in Haskell, making heavy use of polymorphism in the different layers. Their code turned out to contain serious space leaks, which they attempted to fix using seq. But whenever they inserted a call of seq on a type variable, the type signature of the enclosing function changed to require an Eval instance for that variable—just as the designers of Haskell 1.3 intended. But often, the type signatures of very many functions changed as a consequence of a single seq. This would not have mattered if the type signatures were inferred by the compiler—but the students had written them explicitly in their code. Moreover, they had done so not from choice, but because Haskell’s monomorphism restriction required type signatures on these particular definitions [...]. As a result, each insertion of a seq became a nightmare, requiring repeated compilations to find affected type signatures and manual correction of each one. Since space debugging is to some extent a question of trial and error, the students needed to insert and remove calls of seq time and time again. In the end they were forced to conclude that fixing their space leaks was simply not feasible in the time available to complete the project—not because they were hard to find, but because making the necessary corrections was simply too heavyweight. This experience provided ammunition for the eventual removal of class Eval in Haskell 98."
Cheers,
Stefan_______________________________________________
Libraries mailing list Libraries@haskell.org http://www.haskell.org/mailman/listinfo/libraries