How much (quantitatively) is gained by using GCC for compilation? It sure sounds like it causes a lot of trouble. How much better is it than an optimizing assembler?

In most cases very little, in terms of performance, is gained by using gcc as an intermediate. If you use an optimising assembler though, you still need a native code generator (which we have, but GHC's native code generator is missing some things: eg. profiling support and architectures other than sparc & x86). Cheers, Simon