[TUHS] looking for a quote

Theodore Ts'o tytso at mit.edu
Fri Dec 8 07:06:53 AEST 2017


On Thu, Dec 07, 2017 at 05:45:00PM +0000, Ralph Corderoy wrote:
> Hi Ted,
> 
> > The reason why it's useful is that it's much more likely to find
> > relevant lines of code by using "grep st_size *.[ch]".  If you grep
> > for "size", there will be way too many false matches.
> 
> True, but hasn't static analysis come on since then?  I don't use them,
> but tools like http://ffevotte.github.io/clang-tags/ build on clang's
> work to allow queries like finding all mentions of st_size AFAIK.
> They'll even cope with two different struct definitions with a st_size
> kicking about and use the one you give as a seed reference.

First of all, these tools didn't exist in the days of BSD 4.x.

Second of all, those tools aren't integrated into ctags and etags, so
for the needs of working programmers, the need to run some modified
clang and then parse clang output when trying to figure out where all
the places that use some structure field would be unnecessarily
complex and destructive of a working programmer's productivity.

So arguably the tools *still* don't exist in a useful form for
developers to use.

> I'm surprised with things like DWARF in ELF that compilers weren't used
> to generate a super `ctags' longer ago.  We had cscope, that was Bell
> Labs IIRC.  And then there's been GNU GLOBAL, GNU's idutils, ...

DWARF is incredibly bloated for large projects, because unless the
header files which are #included in each .c file is exactly the same
(the exact same header files, included in the exact same order, with
no other local structure or type definitions in the C source file),
there will N different copies of the structure information in the .o
files, and this duplication will also exist in combined executable's
DWARF information.

I complained about this a decade ago, when this defect essentially
made Systemtap completely unusable by real kernel developers.
Compiling with DWARF exploded the compile time, and size of the build
tree, and the size of the insatlled kernel, by a factor of at least
ten.  We now have a way of generated reduced debuginfo files, but it
works by eliding all of the structure definitions, and only keeping
the file/line number information.  So that won't work for your "super
ctags" proposal.

If you want to try to fix this, feel free --- but we don't actually
have a solution that works in the real world today.  It's something
that only really works in the Powerpoint Slideware world inhabited by
IBM executives....

						- Ted


More information about the TUHS mailing list