Why there is the need for the mfence before and after rdtsc?

[zyma-Jasper]

Intel suggests to use a cpuid rdtsc instruction rdtsc cpuid sequence to measure the time.link. Also the test sequence is fence cpuid rdtsc instruction rdtsc cpuid mfence in the software prefetch paper of CCS16. So why here the calibration code use the mfence rdtsc mfence instr mfence rdtsc mfence sequece? I have combine the test process to one specific cpu0 and give it highest priority, Also I have isolated one specific cpu to make sure no other processes can run on it. However, I still get different thresholds after each run. Sometimes 220 cycles and sometimes 130 cycles.