RE: Libtool head: stresstest.at export of w8$ symbol

[Peter Ekberg]

* Ralf Wildenhues wrote on Thursday, September 08, 2005 09:42 CEST:
> Hi Peter,
> 
> Finally I've managed to do another round of testing on mingw/cygwin..
> Sorry for the delay.

No problem.

> * Peter Ekberg wrote on Fri, Sep 02, 2005 at 09:50:01PM CEST:
> > Ralf Wildenhues wrote:
> > > * Peter Ekberg wrote on Fri, Sep 02, 2005 at 12:03:14PM CEST:
> > > > 
> > > > I'm trying to make sence of the problem with stresstest.at on
> > > > Windows (MinGW, Cygwin and MSVC) where the following symbols
> > > > can't be exported during linking of dlself with -export-symbols
> > > > dlselfsyms:
> *snip*
> 
> > Heh, I searched back in the cvs archives looking for clues,
> > and think the first line of the sed program have been warped
> > in some cleanup...
> > 
> > What's there currently is this (minus m4 quotes):
> > 
> > -e 's/\([ ][.*^$]\)/\\\1/g'
> > 
> > Originally it was:
> > 
> > -e 's/\([][.*^$]\)/\\\1/g'
> > 
> > Which makes much more sense, it's just a general escape for
> > the later grep, to not have any surprises. (if the symfile is
> > constructed badly or something, I don't really see the need
> > here, but that's irrelevant...)
> 
> OK, I remember now why I changed that (but my change was wrong, of
> course):
>   s/\([][.*^$]\)/\\\1/g
> is a sed syntax error.  This is because `[.' in a bracket expression
> starts a character class, so sed expects a `.]' to finish that
> eventually.  Instead, this
>   s/\([].[*^$]\)/\\\1/g
> should work.  This error does show up under cygwin and mingw 
> stresstest
> (both with gcc):
> | /usr/bin/sed: -e expression #1, char 20: unterminated `s' command

Ah! Character classes, didn't think of those... And I even saw the
error without reacting, I was so sure I was right and thought it
was something else. Oh well...

> Looking back, it now makes sense why the original authors did not
> encounter this error: most likely, the sed of seven years ago did
> not know about character classes and parsed this just fine.
> 
> However, with that change, stresstest still eventually fails with a
> | Cannot export w1$: symbol not defined
> | Cannot export w10$: symbol not defined
> | ...
> 
> on mingw, and a
> 
> | grep: Invalid back reference
> | 
> /usr/lib/gcc/i686-pc-cygwin/3.4.4/../../../../i686-pc-cygwin/b
> in/ld: sub3/.libs/dlself.def:8: syntax error
> | 
> /usr/lib/gcc/i686-pc-cygwin/3.4.4/../../../../i686-pc-cygwin/b
in/ld:sub3/.libs/dlself.def: file format not recognized; treating as
linker > script
> | 
> /usr/lib/gcc/i686-pc-cygwin/3.4.4/../../../../i686-pc-cygwin/b
in/ld:sub3/.libs/dlself.def:2: syntax error
> | collect2: ld returned 1 exit status
> 
> on cygwin, both with gcc.  (Yes, I should be implementing better
> stresstest error reporting first..)

Ah, the \\\1 part needs an extra round of escapes. See new patch.

> > Now I just need someone to explain why the w8 symbol has a
> > $ suffix in the symfile. There is definitely something fishy
> > here, as the $ is passed on from the symfile to the grep
> > command, which doesn't seem like the intended design...
> > 
> > So, one of these cases are right:
> 
> > 3. the $ has no meaning to any tool and the $ in w8$ in
> >    the test is to be zapped and the symbols in the symfile
> >    are supposed to be escaped as a precaution before being
> >    used as grep patterns
> 
> This one is mostly right, I believe: the `$' does not belong there.

Good.

> The only thing I could not find out yet is:
> - whether on some systems the symbol file has been used/is in use with
>   wild cards or some kind of regular expressions which either the
>   ltmain script or the linker understands,  
> - or some "active" characters (like `$', `[') need to be specifyable
>   as part of symbol names because some systems use these for mangling
>   and/or versioning.
> 
> Strictly speaking, we are safe to kill the first possibility, as it is
> not mentioned in the documentation, and surely it is not 
> portable as of
> now; but is not explicitly forbidden either, so I only want to step on
> toes if I have to.
> 
> (By the way, similarly to -export-symbols-regex, we _could_ add a new
> interface to libtool with symbol files which allow regexes/wildcards..
> something for the TODO list after 3.x :)
> 
> > I'm putting my money on 3, so here's a patch that fixes the
> > escape, but not the test. Ok to commit?
> 
> Not exactly like this, see above.  And if possible, I'd like to
> eliminate the remaining failure.  :)

Like this?

        * libltdl/config/ltmain.m4sh (func_generate_dlsyms):
        Fix escape expression to actually escape relevant
        characters from the following grep. Fix Cygwin and
        MinGW to use the filtered symbol list when exporting
        from self, and not the symbol filter.
        * tests/stresstest.at: Export w8, not w8$

Cheers,
Peter

head-fix-escape-for-grep-2.patch
Description: head-fix-escape-for-grep-2.patch