If you need precision good enough for practical applications, it’s very hard to do. All modern CPUs are pipelined, time they spent on every instruction depends

If you need precision good enough for practical applications, it’s very hard to do. All modern CPUs are pipelined, time they spent on every instruction depends on neighbor instructions + data dependencies between them. All modern CPUs have multi-level caches, static analyzers must also simulate what data is in which levels of caches; for many practical problems RAM access dominates computations and L1 cache is ~2 orders of magnitude faster than main RAM. All modern CPUs have very sophisticated power management, they scale frequency based on temperature, they selectively turn on/off their blocks, this too affects time. All modern CPUs have multiple cores, L3 cache is often shared across them, all caches suffer heavy penalty when a line is shared across cores. And so on. People don’t do it not only because it’s hard, also because running the function and measuring time + memory is much easier, also much more precise. Even if you don’t care about time and memory complexity and only care about these useless O(N), it’s still very hard because complexity depends on input data. E.g. bubble sort is O(N^2) but it can also be O(N) if the data is mostly sorted already and only a single element needs to be moved.