How can we use perf to generate call-graphs? (Android 4.2.2 (JellyBean) on Sabre SD board)

floriang — Tue, 30 Jul 2013 13:09:01 GMT

Hello!

I would like to generate call-graphs to track optimizations that should be done in the source code of a complex application.

I tried to use oprofile and perf, but both fail to generate a proper call-graph for the part that is related to my library.

perf is able to generate a callgraph for kernel functions if provided with the kernel symbols.

I compiled my native library with the following flags:

LOCAL_CFLAGS := -g -fno-omit-frame-pointer $(LOCAL_CFLAGS)

Then when I get samples for the call graph I use:

perf record -g -ecpu-clock -c50000 -p<mypid>

and output the report via:

perf report -g > perfoutput.txt

When I look at this report, the hierarchy doesn't appear for the functions that are in my library. I got somethign like:

7.43% fec.fecbench libFecBenchNative.so [.] FecBench7_NoOp3

--- 0x78a8ea54

1.84% fec.fecbench libFecBenchNative.so [.] rand

--- 0x78a8ea3c

--100.00%-- 0x0

--- 0x78a8ea54

But actually I call rand() inside the FecBench_NoOp3 function, which is called by FecBench_NoOp2 function.

I expected the output to be something like:

FecBench_NoOp2 ---> FecBench_NoOp3 ---> rand().

But it seems that perf can not see that these functions are actually related.

But for the kernel functions I get a correct graph like:

0.11% fec.fecbench [kernel]

[k] _raw_spin_unlock_irqrestore

	\|
	--- _raw_spin_unlock_irqrestore

	\|
	\|--55.21%-- do_send_sig_info

kill_pid_info

sys_kill

ret_fast_syscall

	\|	\|
	\|	\|--68.97%-- run_timer_softirq

__do_softirq

irq_exit

do_local_timer

__irq_usr

Any ideas? Don't hesitate to suggest another solution if you know of a good way to get a callgraph in android for native code.

I also tried with Oprofile but I think it didn't even give anything useful on the kernel...

topic How can we use perf to generate call-graphs? (Android 4.2.2 (JellyBean) on Sabre SD board) in i.MX Processors