On 6 May 2015 at 15:05, Peter Crosthwaite <peter.crosthwa...@xilinx.com> wrote: > What I am most worried about (and I need to run some tests to really > confirm facts) is what happens if a CPU WFIs. Should the PMCCNTR > continue on or hold its value? If we match instruction execution to > PMCCNTR to the PMCCNTR will freeze.
See the v8 ARM ARM D5.1.1: this doesn't count PE clock cycles, it's linked to the hardware processor clock. The exact relationship is IMPDEF so we have some leeway for doing whatever seems reasonable here. Permitted things: * counter can stop on WFI (see D5.1.3) * counter can continue to run on WFI (it's impdef whether the PE clock is gated when in WFI, though I would expect that to be a popular implementation) * counter can read the same value if read twice in a row * counter can run forward a lot even if no insns executed (the example given is of a hyperthreading implementation) So we should do whatever seems most convenient implementation-wise I think... -- PMM
