High-level languages for compilation are, 9 times out of 10, a better choice.
2. Having never used OCaml for compiler dev, I can't speak on this.
3. The lack of strong typing has no effect on the ability of Python or Ruby to handle compilation. Like the Lisp family, these are very, very flexible languages with strong benefits to compiler developers, and I wouldn't trade them for anything else I've used.
TBH, I think ML-based languages are poor choices for compiler development. You spend as much time worrying about your own code as you do the code you're compiling, which you just don't do when you write a compiler in Scheme, Ruby, or another 'dynamic' language.
During the compilation? Would you even see the difference if the compiler did 0.1*0.2 by emulating cpu processing, instead of via one instruction? The cost of register allocation for a simple function will be greater than optimising many constant fp expressions. It simply doesn't matter here. If you have a convenient type, you use it, if you don't, you download any libieee754 and interface with its functions...
Seriously - if you're writing a compiler, there are so many more important problems than lack of mapping between the target and local types :( It may not be an elegant solution to handle fp expression evaluation via an external library, but it's only done during the compile time (so only once) and it's easily solvable.
What makes you think everyone is out to write a C compiler?
Another example of performance-sensitive application of compilers is background Java compiler in Eclipse - even for rather small programs it has to be fast. Ruby is 10-100 (actually more than that) times slower than C/Java, So if you optimizer (wriiten in C/Java/Haskell) needs 0.5 second to optimize code, same code will need (by very optimistic measure) 30 sec. Python not much better. Scheme a bit better (especially Stalin compiler), Lisp almost good.
No. You don't need that. You can treat your data however you want until you place it into the final binary. Before that, you can even treat all your numbers as strings if you really want to. If you need some constant expression evaluation, you just have to replicate the target machine's math operations in software - no magic involved here. As long as you get the right result, noone cares what you do internally.
If your compilation machine == target machine, you can even construct a function that calculates the expression and actually run it to get the result.
