I believe I can use bugpoint to help you folks track down an LLVM bug using the bugpoint mode where it runs the program and looks at its output.
However there is a slight complication: to distinguish between correct and incorrect runs I need to run the program under test in our hacked valgrind. After reading all the bugpoint docs I can't figure out how I would do that. What I would like is an option telling bugpoint to run some other program that can evaluate the program that was compiled.
Note that this option will be broadly useful for people developing cross compilers since it permits specifying the simulator that runs the compiled code.