To get significant results benchmarking a function with the x86_64 RDTSCP instruction it's necessary to reduce jitter caused by interrupts, running the program as a kernel module permit to disable preemption and hardware interrupts during the measurement.
From the BIOS settings disable CPU frequency scaling, turbo, hyperthreading and power saving features to have less indeterminism.
Run $ make all to build the program.
Then run # insmod cycles_count.ko to insert the module in the kernel and # rmmod cycles_count to remove it.
The statistics are logged in /var/log/syslog.
If insmod cycles_count.ko and rmmod cycles_count are run from a console the statistics are dispayed directly in the console.
Running the program without an user-provided function to be measured will report the overhead of the measurement itself, the statistics from this run are to be subtracted from the respective statistics resulted from a run when a function to be measured is provided by the user.
Edit cycles_count.c to define a function to be measured and call it in count_cycles().
https://www.kernel.org/doc/Documentation/preempt-locking.txt