The main thing I wish someone would write is a profiler that (a) had sufficiently low overhead that it could run in production code and (b) could be installed without huge changes to ac.scm and/or the underlying MzScheme.
Oof. Tough. I and raymyers did some work back on a profiler when we were optimizing treeparse, but it didn't quite work well in my use-case.
Certainly it seems to involve deeper hackery skills than I seem to have right now ^^.
As an aside, it might be slightly easier to actually insert profiling code in arc2c. Each function has its own ID number in arc2c and all we need is some sort of index from ID number to original function name (given that a function in arc2c may be split by CPS conversion, so a function with a name may end up being split into several functions),
