Re: [PATCH v2 3/4] perf annotate: configure objdump path at compile time

[Ingo Molnar <mingo@kernel.org>]

* Namhyung Kim <namhyung@kernel.org> wrote:

> On Thu, 27 Sep 2012 14:25:10 +0300, Irina Tirdea wrote:
> >>> The perf built to run on the host needs to use arm-eabi-objdump from
> >>> the toolchain so that it can analyse data recorded on Android. This
> >>> patch is targeting this scenario, not the previous one. In this case,
> >>> the CROSS_COMPILE option would be different than arm-eabi- so using
> >>> $(CROSS_COMPILE)objdump would be wrong. objdump should be overridden
> >>> when running make since there is no connection between the toolchain
> >>> used here and the path for objdump. I am always overriding objdump
> >>> when calling make, so I did not catch this.
> >>>
> >>> I think that I should change DEFAULT_OBJDUMP_PATH=objdump in the
> >>> Makefile to handle the first scenario. I'll also explain this in the
> >>> commit message so that it is more clear and make the same change for
> >>> the addr2line patch.
> >>>
> >>> What do you think?
> >>
> >> I think the right thing to do is finding a correct objdump at runtime in
> >> some way.  Why do you want to make it compile-time configurable?
> >>
> >
> > The correct objdump path can be detected at runtime by setting the
> > toolchain path. But since the name is arm-eabi-objdump and not
> > objdump, it does not know to use it instead.
> >
> > The only way (I can think of) to change objdump at runtime would be to
> > use the --objdump option for perf annotate (and provide a similar
> > option for addr2line). The problem with this approach is that the user
> > has to be aware that perf annotate uses the objdump tool and that he
> > has to use the cross-compiler version instead. Since the user will
> > have perf compiled for host as part of his Android tree, he will
> > expect it to work without these further changes from his part. The
> > path for objdump can be set in the Android Makefile at compile time so
> > that the user doesn't need to be aware of it.
> 
> What I'm thinking is that perf can try to find cross-built 
> binutils when it detects perf.data file is came from other 
> machine/architecture. Fortunately perf_session_env was added 
> recently and it has the arch information from the file so we 
> can use it to find the path.
> 
> Following patch is a proof-of-concept patch and only build 
> tested. What do you think?  Could you play with it for some 
> time? :)

This is a pretty clever idea - much better than hard-coding 
architecture details at build time.

Thanks,

	Ingo