get_cyles()

[Jerry Van Baren]

Shirley Ma wrote:
> 
> Doesn't ppc64 supports get_cycles() correctly?
> I used below function to get_cycles(), it seems not right.
> 
> t1 = get_cycles();
> t2 = get_cycles();
> 
> t2-t1 is huge.
> 
> static inline cycles_t get_cycles(void)
> {
>         cycles_t ret;
> 
>         __asm__ __volatile__("mftb %0" : "=r" (ret) : );
>         return ret;
> }
> 
> Shirley Ma
> IBM Linux Technology Center
> 15300 SW Koll Parkway
> Beaverton, OR 97006-6063
> Phone(Fax): (503) 578-7638

Is (t2-t1) (a) _always_ huge or (b) only sometimes huge?

If (a), are you sure t1 and t2 are executing in the written order?  The 
compiler may be re-ordering the instructions on you.  Use objdump -S to 
disassemble the object file (or executable).

If (b), you are getting burned by carry out of the time base lower 
register.  Every 2^32 counts, the timebase (lower) register will wrap 
around from 0xFFFFFFFF to 0x00000000.  If the wrap occurs between 
reading t1 and t2, you will get a very large jump.

Another possiblility for (b) is that you are getting an interrupt 
between reading t1 and t2, so your delta will include the interrupt 
handling time and potentially more... if your task is pre-empted due to 
the interrupt, the time delta will be (relatively speaking) very large. 
  This is actually the most likely scenario in my mind.

gvb