It's implausible to me that this really improves performance by 50%!!!!! unless we have another issue somewhere else. Let me poke at this...

I can't reproduce this effect locally in isolation. Here's an unmodified binary:

epriestley@orbital ~/dev/libphutil $ time ./xhpast.with-whitespace < ../phabricator/src/__phutil_library_map__.php | wc -c
  965744

real	0m0.142s
user	0m0.139s
sys	0m0.006s

Here's a modified binary:

epriestley@orbital ~/dev/libphutil $ time ./xhpast.no-whitespace < ../phabricator/src/__phutil_library_map__.php | wc -c
  912648

real	0m0.143s
user	0m0.140s
sys	0m0.007s

The runtime is identical and the output size differs by only 5%.

Can you reproduce the effect by running the binary directly?

It's quite possible that my thinking is wrong, but I figured that this might be faster because there is less data being passed through ExecFuture rather than directly improving XHPAST. I agree that the observed speed up seems unwarranted though.