TIL: Return Stack Buffer, a component of the CPU branch predictor. When we make function calls, the return address is pushed onto the stack. When the CPU encounters a return instruction, it will return to that address. Accessing data on the stack can require fetching from main memory (if the requested address is not in the cache), which can take hundreds of cycles. Thus, returning from function calls can be very expensive. To optimize this, the hardware includes a Return Stack Buffer (RSB), a small, specialized buffer that stores the return addresses of the recently executed functions. As the CPU executes instructions ahead of time and out-of-order, it pops the top of the RSB to predict the return address of the current function, and starts fetching instructions from that address. If the prediction is wrong, the CPU has to flush the pipeline and start fetching instructions from the correct address. The ARM documentation notes that in a code sequence involving deeply nested function calls, the RSB will store only the first few return addresses (limited by the RSB size). That means the inner function calls will not benefit from it. So I'm guessing having such a deep chain of calls is costly, or at the very least misses out on the benefits of the RSB?

I made a short video giving an overview of profiling in Python. It covers tracing vs sampling profilers, and also gives a quick demo of a few profilers, including cProfile, py-spy, and perf! blog.codingconfessions.com/p…