Re: Demangling of method names and addr2line/c++filt

[Andrew Haley <aph at redhat dot com>]

Ranjit Mathew writes:
 > >  > The final logic then simply looks like:
 > >  >
 > >  >   if ( usingAddr2Line && (USER_LABEL_PREFIX[0]=='_'))
 > >  >     /* Explicitly pass "-n" to c++filt */
 > >  >
 > >  > IMHO, this looks like the safest approach.
 > >  >
 > >  > If this is OK, I can work towards a patch for this.
 > > 
 > > Excellent.
 > 
 > I spoke too soon...
 > 
 > First off, USER_LABEL_PREFIX is not easily made available to libgcj
 > and I could not see an easy way to get GCC's config.h (which
 > includes target-specific headers that really define
 > USER_LABEL_PREFIX, that too guarded by an "IN_GCC" definition)
 > included in, say, natNameFinder.cc.
 > 
 > More seriously though, the c++filt process is created only once in
 > NameFinder.java for an entire stack of addresses which could
 > potentially come from various sources (addr2line, addr2name.awk,
 > interpreter and dladdr).
 > 
 > Either we can spawn c++filt individually for each address (totally
 > gross!), or we can prefix USER_LABEL_PREFIX to symbols given by
 > addr2line before passing it to c++filt (somewhat less gross).
 > 
 > It looks gross either way to me and not worth the effort - your
 > reflection data based scheme is much more reliable and would work
 > even if the binary is stripped. Except of course that Win32 would
 > be using SJLJ and wouldn't benefit from it until somehow DWARF2 is
 > brought back in... :-(
 > 
 > I could just distribute a MinGW-specific binary with "-n" passed to
 > c++filt as default...

Oh, I give up.

Let's pass "-n" always, but only check this in to the trunk, not 3.3.
If any other system breaks we'll find out about it.

Andrew.