An amazing performance tool is the ‘Processor Trace’ instrument. It’s M4 only unfortunately, I think due to hardware capability. It allows a detailed timeline of which functions were called in which order and how long they took. This can reveal all sorts of things that statistical profilers cannot – sometimes a call is much faster than other times; sometimes you do something slow and stupid between two steps when you would rather they run together; sometimes you are calling the same function again and again instead of caching its result. Unlike older methods, this is low overhead and doesn’t require recombination or knowing which functions are ‘interesting’. It is similar to intel processor trace and I think based on Arm CoreSight (both IPT and coresight tracing (I’m a bit unclear on the exact features that ‘coresight’ corresponds to) are supported by perf on Linux; not sure about asahi on Apple silicon).