Unless you have lots and lots of time to spend micro-optimizing everything, you'll get better performance writing in Haskell.
> Maybe there are good C libraries that solve your problem
Maybe, but most of the original post is about how lacking the library ecosystem is. I can well believe that you might have some useful domain-specific libraries for the backend, but that calls for writing a backend service in C (and using something like thrift to call it), not "web programming in C".
> Maybe you just want to learn C better.
If you try and learn on a fake problem you'll pervert your learning. Web development is not amenable to idiomatic C (Cello is a good example - it's not appropriate to the sort of problems C is appropriate for, and if you learn C based on using Cello it will not make you a good C programmer).
If you're an experienced C developer you know what's fast and what not (and you know about things like cache coherency). You don't spend that much time on 'micro optimization' because you know how to layout your memory in the first place, etc.
This is just not true. Haskell is terribly (3x,4x) slow comparing to C for the most trivial of computing tasks. You can probably get it to 2x or 1.5x by giving up all the lists and other default data structures as well as writing everything in procedural (by monads) way. If you do all this you can get close to C performance but you just gave up all the advantages of writing in Haskell in the first place. And you are still slow as hell.
This is stating the case too strongly. Performance wise, the naive, unoptimized solution in Haskell (ex. using the default String type instead of Data.ByteString, number crunching using lots of iteration\recursion) is going to be much slower than than the naive, unoptimized solution in C or Lua. However, for a given subset of the range between "zero" and "micro" levels of optimization, for a given problem, Haskell may still represent the local maximum.