Ref T13164. See PHI766. On PHI766, we saw a slow command on a production host with no ready access to XHProf / --xprofile.
We can reasonably turn --trace into a rough-cut profiler by adding milliseconds-since-start. Then --trace output is good enough to at least roughly ballpark where slowness is happening in many cases. This isn't as good as real profile and won't always be helpful, but at least some of the time it can at least give us enough information to know if, e.g., we should break out the profiler or not.