Re: libltdl exports no symbols (cygwin)

[Charles Wilson]

Ralf Wildenhues wrote:
> Hello Eric, Charles, all,
>
> * Eric Blake wrote on Tue, Jan 30, 2007 at 06:00:00PM CET:
> > Should we also mention the side effect that you must now mark
> > explicit exports, since you can no longer rely on auto-imports?
>
> This whole issue seems not good for branch-1-5.  I'm pondering backing
> out all related changes from branch-1-5; that is, the NEWS update and
> the LT_GLOBAL_DATA/LT_SCOPE change of 2007-01-28, as well as the
> DLL_EXPORT change of 2006-06-01.  We cannot do such a big change for
> 1.5.24, that will affect all users of --with-included-ltdl. 

And this decision solves the problem for MinGW how?  Remember that part 
of the justification for the 2006-06-01 change was that at least that 
way, we have the same behavior on all win32 variants.  If you revert 
2006-06-01, cygwin != mingw, and mingw still #defines DLL_EXPORT for all 
"PIC" -- which would simply push CVS M4's problems into the future, when 
somebody tries to compile it on MinGW.

And if you try to switch MinGW libtool to not define DLL_EXPORT...hoo 
boy. A *LOT* of MinGW ports rely on that behavior...

So, on one side we have (a) a bunch of MinGW ports, (b) every cygwin 
port *I* know of except *CVS* m4, plus *released* versions of gettext.

On the other...*CVS* m4.

> The patches
> were not introduced to fix a regression, rather something that we can
> see as an improvement, i.e., a /new/ feature, but they cause regressions
> for others.  I would like 1.5.24 to be a no-brainer for users, and I
> would like to do it soon if possible.  If we require changes in user's
> code, then it's better to make the change when going to 2.0; it's easier
> for them if all changes occur at once.  And I'm easier with having
> Gettext require HEAD Libtool than all users of libltdl.

Wait, you're contradicting yourself.  The 2006-06-01 changes were there 
to fix a regression that popped up with gettext -- not "some random 
pseudo-improvement that C. Wilson made up".  By reverting the change 
(e.g. because it causes regressions for *CVS* M4) to "fix" *CVS* M4[*], 
you cause a regression with *released* gettext.  Why is THAT ok?  Why 
can't we say that *CVS* M4 should require CVS libtool, and "fix" this 
issue there?

[*] BTW, I have a longer email I was putting together concerning the 
root cause of M4's problems w.r.t. LT_SCOPE, DLL_EXPORT, and friends. To 
wit: CVS M4 has been borked on cygwin for YEARS. Breakage now is nothing 
new...and reverting cygwin libtool to its broken state vis-a-vis gettext 
simply to help CVS M4 -- which is borked anyway -- is nonsensical.

---
Do what you want, but I am strongly inclined to continue future official 
cygwin releases of libtool RETAINING the 2006-06-01 and 2007-01-28 
changes. I do not look favorably on releasing yet ANOTHER 
backwards-incompatible version of libtool/libltdl on the cygwin 
platform, while pretending that it is still compatible.

BTW -- it's obvious to me that the reason Eric never noticed this 
problem until Ralf committed the change to libtool CVS HEAD, was because 
he has NOT actually been using cygwin official released (e.g. 
libtool-1.5.23a) libtool.  He's been using libtool CVS HEAD to develop 
CVS M4.  So don't use CVS M4 breakage observed with libtool CVS HEAD as 
a justification for re-breaking cygwin libtool 1.5.24.

> For HEAD, guys, if you want this really fixed then may I suggest to take
> another look at Bruno's suggestions:
> <http://lists.gnu.org/archive/html/bug-gnu-utils/2006-05/msg00026.html>

For 2.2, I tentatively agree (but...[**]). For 2.0 -- GOOD GREIF, 
PEOPLE, can't we just release the G-D thing and be done?  I remember in 
Jan 2005 we were gonna release 2.0 "real soon now".  This is freakin' 
ridiculous.

[**] I'm starting to see some support for msvc in libtool -- at least, 
there have been a few patches floating around.  Whatever fancy stuff is 
done to support DLL decorations for mingw/cygwin, if you want that msvc 
stuff to work you'll still need some mechanism for the plain-old-dumb 
msvc linker...and I don't know if Bruno has tried his scheme with that 
toolset (maybe he has; I'm going from memory here).

> Any way we go, we should do something more consistent than "you cannot
> rely on auto-imports if you use libltdl". 

The embedded quote is not an accurate description of the current status. 
 More in my other email.

> And certainly we should
> document user-level-required changes in NEWS.

Well, yeah -- but make sure the proposed documentation is ACCURATE, first.

--
Chuck