We probably should not change the existing --perf output in case existing analysis tools depend on it, but it would be helpful to have a variation that:

• includes the module name in each line of output (to make it easier to grep)
• only shows what's accepted as the best performance estimate ("by_slowest")
• also outputs bytes
• also outputs time (rate, bytes, and time should all be directly related)

This output mode would probably be preferred for future ML analysis tools.