[Dwarf-Discuss] how to find local variable stored in register

[Daniel Berlin]

On Feb 6, 2008 1:30 AM, Jim Blandy <jimb at red-bean.com> wrote:
>
> On Feb 5, 2008 7:40 PM, Daniel Berlin <dberlin at dberlin.org> wrote:
> > On Feb 5, 2008 7:14 PM, Jim Blandy <jimb at red-bean.com> wrote:
> > > On Feb 5, 2008 2:43 PM, Daniel Jacobowitz <drow at false.org> wrote:
> > > > I think 16.1% of executable size is huge.  .debug_frame is normally,
> > > > what, 2-3% of executable size?
> > >
> > > Different organizations have different concerns.  I think it's a
> > > mistake for GCC to assume that a GiB is precious; it's not so for many
> > > real-world organizations.
> >
> > Even at google, which has ungodly amounts of disk space, debug info
> > size is a serious concern.
> > This is more because of the network cost of propagating binaries
> > around than it is because they happen to take up some amount of local
> > disk.
> > I think it would be a huge mistake to assume because local disk is
> > cheap that debug info size doesn't matter.
> > If you ever end up having to propagate your binary to a datacenter
> > full of machines, or build on NFS, it matters a lot.
>
> I'll point out that I was careful not to make the argument that size
> doesn't matter to anyone.  What I did argue is that:
> 1) when I'm working on GDB or Firefox on my own machine, I would (and
> do) happily trade pretty huge factors for more accurate debugging
> info, and
> 2) my situation is neither unusual nor universal.

By arguing that it is a mistake for GCC to care about GiB  ('I think
it's a mistake for GCC to assume that a GiB is precious; it's not so
for many real-world organizations'), you seem to be implying that the
majority of GCC users don't care about the size of the debug info.
I believe both your original claim about GCC, and it's apparent
implication, are at best misleading, and at worst, outright wrong.
Like you, i've spent a lot of time dealing with (and being one of at
times) customers on both sides of the size/precision tradeoff.
I've certainly found people who think precision matters most (and more
who don't), but i've yet to meet anyone who believes increasing the
size massively is reasonable or a good idea.  It is just a fact of
life they have to live with.

In any case, tests we've done at Google show that even running
something like zlib over the debug_* results in massive savings for
almost no cost at all in CPU time or usability (you use various tricks
to make the compression still able to do random access well, etc), so
we have hope that their is room for improvement without loss of
precision.;

PS Off-topic, but out of curiousity, is firefox debuggable with GDB
these days? Last I looked a few years ago, mozilla binaries had 1.6
gig of debug info, and GDB required more than that in memory to truly
debug it.