@dysfun there's also a _nofence variant:

#define rdtsc_nofence(lo, hi) \
__asm__ volatile("rdtsc" : "=a" (lo), "=d" (hi))

static inline uint64_t
rdtsc64_nofence(void)
{
uint64_t lo, hi;
rdtsc_nofence(lo, hi);
return ((hi) << 32) | (lo);
}