Messages in this thread | | | From | Vineet Gupta <> | Subject | Re: 'branches' perf event mapping differs on ARC and ARM | Date | Tue, 27 Nov 2018 17:06:10 +0000 |
| |
On 11/27/18 6:36 AM, Eugeniy Paltsev wrote: > Hi, > > While playing with perf tool on ARMv7 and ARCv2 processors and profiling the > same application I got interesting results. Even if we got pretty > similar total > execution time and instructions number the number of branches on ARC is about > three times more then on ARM. > > I dug into architecture > specific perf sources and found that we map different > HW counters into generic 'branches' event on ARC and ARM. > - We use "ijmp" event on ARC which > counts all jump and branch instructions (regardless > of real execution flow - even if no real jump happens)
That doesn't seem correct IMO. A NOT taken conditional branch doesn't change control flow, so semantically doesn't qualify as a branch. On ARC, the generic branches event should be mapped to "actually taken branches" condition i.e. ijmptak
> - We use "pc_write_retired" event on ARM > which counts only taken branches (Instruction > architecturally executed, condition check pass - software change of the PC)
That seems correct.
> I guess counting all jump and branch instructions is correct because we use > 'branches' event value to calculate relative value of 'branch-misses' using > > following formula: > ----------------------------8---------------------------- > branch-misses-ration = 'branch-misses' / 'branches' * 100.0 > ---------------- > ------------8---------------------------- > And using only taken branches here is incorrect IMHO.
Why ? branch-misses is a CPU specific micro-arch state where the a changed control flow was NOT predicted. If an implementation mispredicts NOT taken branches, those should actually get counted and be fed to hardware folks to improve the micro-arch.
> So I guess we should > map 'br_immed_retired' instead of > "pc_write_retired" into generic 'branches' > event on ARM.
| |