Re: [RFC 00/48] perf tools: Introduce data type profiling (v1)

[Namhyung Kim]

On Thu, Oct 12, 2023 at 2:17 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Wed, Oct 11, 2023 at 08:50:23PM -0700, Namhyung Kim wrote:
>
> > Actually there's a performance issue about getting disassembly from the
> > objdump for kernel.  On my system, GNU objdump was really slower than the
> > one from LLVM for some reason so I had to pass the following option for
> > each perf report and perf annotate.
> >
> >     $ sudo perf report --objdump=llvm-objdump ...
> >
> >     # To save it in the config file and drop the command line option
> >     $ sudo perf config annotate.objdump=llvm-objdump
> >
> > Even with this change, still the most processing time was spent on the
> > objdump to get the disassembly.  It'd be nice if we can get the result
> > without using objdump at all.
>
> So the kernel has an instruction decoder, all we need is something that
> can pretty print the result. IIRC Masami had an early version of that
> somewhere.
>
> With those bits, and some basic ELF parsing (find in objtool for
> instance) you can implement most of objdump yourself.

That would be nice, but I'm a bit afraid of dealing with details
of instruction decoding especially for unusual ones considering
extensibility to user space and other architectures.

Thanks,
Namhyung