Re: [Valgrind-developers] helgrind

[Nicholas N. <nj...@ca...>]

On 6 Oct 2002, Jeremy Fitzhardinge wrote:

> Can go into a few more details about what's wrong with helgrind?  Does
> it get the Eraser algorithm wrong, or does it just need some more
> refinement?
>
> I read the Eraser paper and the extension suggested in "Visual Threads",
> and it all looks pretty straightforward.  Once I get some profiling
> machinery going, I'm definitely interested in working on helgrind.

That would be really good.  It's been a while since I looked at it.  My
first attempt at the lockset stuff was very simple and probably wasn't
going to scale at all well, but then Julian improved the lockset stuff a
bit (a lot?) so that might not be such a problem yet.

I think the problem is more that I never tried it on any program larger
than about 20 lines.  It could find easy data races in these short
programs.  But it also reported quite a few races in glibc.  One of these
we found to be genuine (an unsafe update of a totally trivial statistics
counter only printed for debugging), but for the others it was unclear
whether they were genuine or if Helgrind was wrong.  I certainly tried to
implement the Eraser algorithm faithfully.

So to summarise:  it probably needs to be robustified to handle large
programs, and it hasn't been verified as to whether it's finding real
races yet.

N