Output PUBLIC lines from export symbols on Linux

For cases where we don't have debug symbols, dump_syms currently won't produce
any output on Linux. This patch makes it able to output the contents of the
symbol table as PUBLIC lines instead. I noticed that there aren't currently unit
tests for dump_symbols.cc, and I made a fair amount of changes here, so my next
plan is to write unit tests for that.

[jimb]

Hi, Ted --- looks great! (And the code looks somehow familiar...) I did have
some comments.

Did you run the tests under valgrind?

http://breakpad.appspot.com/275001/diff/1/5
File src/common/linux/elf_symbols_to_module.cc (right):

http://breakpad.appspot.com/275001/diff/1/5#newcode42
Line 42: class SymbolIterator {
Even though this is defined in the .cc file, symbols for its member functions
will have global names in the google_breakpad namespace. So it should be called
ELFSymbolIterator, probably.

http://breakpad.appspot.com/275001/diff/1/5#newcode131
Line 131: return reinterpret_cast<const char *>(strings.start + offset);
If the string section doesn't end with a null byte, couldn't this cause other
code to read off the end of the section when it calls strlen, etc.?  (If so, I
expect this bug is present in the original StabsReader as well.)

One way to fix this would be to have ELFSymbolsToModule scan the string section
for the last null character (memrchr is a GNU extension, sigh), perhaps warn if
there isn't one at the end (I believe there always should be), and then just set
the end of the 'strings' ByteBuffer to only include the null-terminated prefix
of the strings section.

http://breakpad.appspot.com/275001/diff/1/7
File src/common/linux/elf_symbols_to_module_unittest.cc (right):

http://breakpad.appspot.com/275001/diff/1/7#newcode172
Line 172: TEST_P(ELFSymbolsToModuleTest32, OneFunc) {
It might be nice to have a "NoFuncs" test as well, on an empty section, if it's
easy.

[Ted Mielczarek]

On 2011/07/05 18:48:53, jimb wrote:
> Did you run the tests under valgrind?

I hadn't, but I just did and they ran cleanly.

I'll fix the other comments and land this. Thanks!

[Ted Mielczarek]

http://breakpad.appspot.com/275001/diff/1/5
File src/common/linux/elf_symbols_to_module.cc (right):

http://breakpad.appspot.com/275001/diff/1/5#newcode131
Line 131: return reinterpret_cast<const char *>(strings.start + offset);
On 2011/07/05 18:48:53, jimb wrote:
> One way to fix this would be to have ELFSymbolsToModule scan the string
section
> for the last null character (memrchr is a GNU extension, sigh), perhaps warn
if
> there isn't one at the end (I believe there always should be), and then just
set
> the end of the 'strings' ByteBuffer to only include the null-terminated prefix
> of the strings section.

I did this, I used memchr anyway, since this code is in src/common/linux. If
someone wants to make it work on another platform they can worry about it then.
I also added a test for this behavior.

[jimb]

On 2011/07/06 15:37:08, Ted Mielczarek wrote:
> I did this, I used memchr anyway, since this code is in src/common/linux. If
> someone wants to make it work on another platform they can worry about it
then.

Uh, you mean "memrchr", right?

[Ted Mielczarek]

On 2011/07/06 16:01:48, jimb wrote:
> On 2011/07/06 15:37:08, Ted Mielczarek wrote:
> > I did this, I used memchr anyway, since this code is in src/common/linux. If
> > someone wants to make it work on another platform they can worry about it
> then.
> 
> Uh, you mean "memrchr", right?

I managed to type it correctly in the patch, just not here! :)