Messages in this thread | | | Date | Thu, 19 Nov 2015 12:28:50 +0100 | From | Peter Zijlstra <> | Subject | Re: [BUG REPORT] perf tools: x86_64: Broken calllchain when sampling taken at 'callq' instruction |
| |
On Thu, Nov 19, 2015 at 06:43:14PM +0800, Wangnan (F) wrote: > Also, why not fixing rsp in kernel if that instruction is a 'callq'? > For avoiding instruction decoding?
Why would the kernel do that? It doesn't need this.
Have userspace do it when trying to interpret the data.
If PERF_EFLAGS_EXACT is set, you know the IP points to the actual instruction and you can attempt to undo whatever parts of it you need at post processing time.
| |