Today, the cheapest primitive to measure durations is the FastClockSource, which still requires running hundreds of instructions. We can introduce a new timing primitive that utilizes CPU timestamps to measure elapsed time. The following, for example, is how that can be done on x86 and AArch64:

MONGO_COMPILER_ALWAYS_INLINE uint64_t getCPUTimestamp {
#if defined(__x86_64__)
    unsigned int lo, hi;
    __asm__ __volatile__("rdtsc" : "=a"(lo), "=d"(hi));
    __asm__ __volatile__("lfence" ::: "memory");
    return ((uint64_t)hi << 32) | lo;
#else
    asm volatile("isb" : : : "memory");
    uint64_t tsc;
    asm volatile("mrs %0, cntvct_el0" : "=r"(tsc));
    return tsc;
#endif
    return 0;
}

On AArch64, there is also an instruction to read the CPU frequency. Considering there is no dynamic frequency scaling on Graviton processors, we can use that to convert the timestamps/cycles into durations (e.g. nanoseconds).