Re: 2 problems with sprof

["H . J . Lu" <hjl at lucon dot org>]

On Sun, Sep 23, 2001 at 09:39:47AM -0700, Ulrich Drepper wrote:
> "H . J . Lu" <hjl@lucon.org> writes:
> 
> > Did you mean the number of text segments in the program header is
> > different between DSO and executable? I don't think it is the case.
> 
> No, the total number of text segments you have to profile.
> 
> > I don't think they are the same problem.
> 
> It's only not the same with regard to the data structures created and
> the dynamic nature of loading and unloading.  The rest is the same.

Did you mean profiling executable plus a DSO doesn't work or just
profiling executable doesn't work? I agree profiling executable plus
a DSO is almost the same as profiling 2 DSOs. But I don't think it
is worth fixing profiling executable alone if we still can't profile
DSO at the same time. As I said, when I make the change, I will take
profiling executable into account.

> 
> > May I suggest we do
> > 
> > LD_PROFILE=dso1:dso2...		# profile a list of DSOs
> > LD_PROFILE_EXEC=		# profile executable
> > LD_PROFILE_ALL=			# profile all, including executable
> 
> I don't see why, and how.  ld.so cannot profile the executable with
> the same mechanisms.  And it's the wrong place to start.

Do we need a way to tell ld.so to profile executable? If yes, how?
Also does sprof support executable?

> 
> I think what you want is profiling with dynamic linking just as I
> mentioned.  Then leave ld.so out.
> 
> > Can we change HZ in kernel to get high-resolution timer?
> 
> No, that's no solution.  What you need is POSIX timer support in the
> kernel.  Anoter of these real-world things the kernel people ignore
> since they don't need it themselves.

Are there any patches for kernel POSIX timer?


H.J.