Instead of just forking at each branch in the (binary) tree of tests bugpoint
does (when trying to nail a function being miscompiled), can you turn it into an
n-ary tree? i.e. at the first stage, compile say 4 threads x (1/4 the program
with llc, 3/4 with native) - this will increase the amount of "wasted"
computation but should greatly speed up bugpointing real-world programs on
people's shiny new multicore systems...
when bugpoint is down to testing single instructions, can you make a "work
list" out of these and have however many threads the user says are available
chipping away at it?
(for extra fish) - get 1) and 2) working when trying to find JIT bugs!
email me your address (I'm not joking about the fish ;)