Re: libtool-2.0 release
* Gary V. Vaughan wrote on Fri, Feb 03, 2006 at 03:43:11PM CET: [Is the personal Cc: okay? The list lag is so long that I've gotten into the habit of Cc:ing you back in so you don't have to wait half a day to get this.] Yes, surely. There was one point in time where I fixed my gnu.org subscriptions to do what I want for my mail setup, and since then I could stop bothering about this for mails sent to me. List lag is interesting though: it can vary between several hours and less than a minute within the time frame of 12 hours. I am at a loss how to explain this, maybe it's fast when the US-based spammers go to sleep. ;- It means that LT_WITH_LTDL in configure.ac that mentions neither LTDL_CONVENIENCE nor LTDL_INSTALLABLE doesn't build libltdl at all. I have a start to a fix for this. Well, so is that really a bug? I originally wrote LT_WITH_LTDL as a convenience wrapper for AC_LIB_LTDL in CVS M4, and realised that it was useful enough to almost all clients of libltdl that it should probably belong in the Libtool distribution... There is definitely a documentation bug (added to RoadMap) that it is still undocumented. It's not LT_WITH_LTDL that is undocumented. LTDL_INIT is. The real question then is whether LT_WITH_LTDL alone should be equivalent to LTDL_CONVENIENCE plus LTDL_INIT (overridable with LTDL_INSTALLABLE) or whether it should be *in addition* to all that. Right. I'm leaning toward the former, but either way the current situation of being like LTDL_INIT plus --with-ltdl processing is counter-intuitive. I'll post a documentation patch to help us define the semantics clearly, and then fix the code to implement what we decide upon. Good. - I know about a couple more tweaks necessary for HEAD libtoolize and Bob has a couple of failures I (or somebody else) need to track down. Agreed. If you put them on the RoadMap, I might get to them before you. Well, this comment of Bob completely mystifies me: http://article.gmane.org/gmane.comp.gnu.libtool.patches/6582 I believe in that thread there were more issues mentioned. Can you transcribe them to the RoadMap once we've got clarification on the issue from Bob please? OK. I'll ask him to test again after patch-6 is in (or test myself). Ahh, and there was another one: the breakage on need_lib_prefix systems. I think we did not mark it release-blocking, but I still would like to test Pierre's patch extensively and use it if it turns out good: otherwise libltdl will be completely useless on e.g. BeOS. Okay. Is that a long standing bug, or a regression? Please mark the RoadMap accordingly :-) Well, both. Apparently dlpreloading has never worked on need_lib_prefix systems, so that is long-standing. Now that libltdl requires working dlpreloading, it fails to work completely, while in 1.5.x it at least worked in some cases. From a user POV, that's a regression, from a Libtool developer POV it's long-standing. ;-) Just discovered yet another in my patch queue (needs another round of testing before I post): make installcheck currently always fails in trees that used 'libtoolize --ltdl' in some modes. (Added to RoadMap). OK. Cheers, Ralf
libtool-2.0 release [WAS per-deplib static/dynamic flags]
Hallo Ralf, Bob, Peter, Eric,
Is this patch really necessary for 2.0? I think that introducing so
much code churn in to libtool at this stage is going to bring yet more
release delays. Surely the feature is useful and desireable, but I
really *really* want to avoid more delays for 2.0.
Now is the time to branch! Either a feature branch for developing the
per-deplib feature for merging after 2.0, or else a 2.0 branch that we
can keep stable.
My preference is to make a feature branch now, and not branch for 2.0
at least until the release blockers are resolved, so that we can commit
any bugfixes we discover during testing just once (we went through the
mess of porting to three branches during the last branch-2-0 debacle,
and I don't want to do that again!).
According to: http://tkd.kicks-ass.net/GnuLibtoolProject/RoadMap, the
three remaining release blockers for 2.0 are:
* ! libtool.m4 macro ordering/requirement audit. pending
* ! LT_WITH_LTDL should build libltdl by default; currently
nothing happens unless LTDL_CONVENIENCE or LTDL_INSTALLABLE
is also given.
* ! fix potential denial of service with compilers that do not
understand -c -o.
* not very likely to happen
* http://lists.gnu.org/archive/html/libtool-patches/2005-03/msg00252.html
And I would even dispute whether the first of those is really a release
blocker -- the code *does* need an audit, but it's been in that state for
several existing releases already. lists.gnu.org and mail.gnu.org appear
to be down, so I can't read the post attached to the final blocker to
remind myself of that issue.
So here are the action points, initialed where I plan to take them
pending agreement from (most of) you guys.
1. strike the macro ordering audit from the release blocker list. GVV
2. fix LTDL_{CONVENIENCE,INSTALLABLE} bug on CVS HEAD. GVV
3. make a per-deblib-branch for these changes.
4. evaluate whether -c -o DoS is a release blocker too.
4a. fix it if it is
4b. strike it from the blocker list if it isn't
5. test like crazy. and fix any platform issues uncovered GVV
6. release libtool-2.0 already! GVV
Once 2.0 is finally out, I will take a back seat with libtool for a while
to work on m4-2.0 and my new book.
Ralf Wildenhues wrote:
Hi Bob, Albert,
* Bob Friesenhahn wrote on Wed, Feb 01, 2006 at 07:47:51AM CET:
On Tue, 31 Jan 2006, Albert Chin wrote:
On Mon, Jan 30, 2006 at 10:28:52PM +0100, Ralf Wildenhues wrote:
- Should the corresponding libtool flags be named `-Bstatic' resp.
`-Bdynamic'? Those were the most common names I could find, but IMHO
they are not very self-explanatory for users not used to them.
-prefer-static, -prefer-dynamic/-prefer-shared? The -Bxxx doesn't seem
similar with current libtool options.
At least for the static case, I would prefer the link to entirely fail
if a static library is requested but one can not be found. Usually
there is a very important reason to use a static library if it has
specifically been requested. So the wording should not be 'prefer'
for the static case. I agree that the -B syntax does not fit the
style of other libtool options (but does match many linkers).
Several different issues here:
- the option naming, which I will not delve into in this post, and
- whether what I'll call -Bstatic for now should fail if it does not
find a static library to link against.
The latter point is intricate, and requires much more elaboration to be
completely specified. That's the main reason I kept the semantics so
vague in my proposal. The issues are: library search algorithm,
difference between libtool and non-libtool libraries, linker capability.
1) The linker may not allow to specify per-deplib linkage at all, or may
not have a flag to force static linkage (as opposed to only _prefer_ it).
Examples for the former are Darwin; I haven't found an example for the
latter yet (good!). But this means, that for non-libtool libraries we
cannot guarantee a certain linking mode, unless we were to change our
search algorithm quite drastically (which I don't see as an option ATM).
2) Difference between libtool libraries (*.la) and non-libtool ones, and
libtool search algorithm and native linker search algorithm.
The current libtool library search algorithm looks a bit like this
(before my patch):
- given `path/to/libfoo.la', that is taken and nothing else.
- given `path/to/libfoo.a' ($libext), that is taken and nothing else.
- given `-lfoo'
for all searchdirs that we look at,
for the extensions `.la', `$std_shrext', `.so', `.a' [1]
check whether there exists a file libfoo.$extension
if yes, exit search algorithm
My patch skips `$std_shrext' and `.so' when in Bstatic mode, but it
still happily picks the first `.la' file it can find, even iff that one
was shared-only, for example. Currently it then links shared
Solaris/64: wrong sys_lib_search_path_spec (was: libtool-2.0 release)
Hi Bob, * Bob Friesenhahn wrote on Thu, Feb 02, 2006 at 07:38:13PM CET: On Thu, 2 Feb 2006, Bob Friesenhahn wrote: I have also noticed some disturbing leakage of default compiler (GCC) path information into the build which causes -m64 to not work properly (at least under Solaris, but seems likely to impact other targets as well). This bug did not used to exist on the 2.0 branch. In my test case, the compiler is always run via a shell script wrapper (gcc-64) which always runs gcc with -m64 so I am not sure what libtool did to encourage gcc to undo the effect of that option. In order to clarify this issue some more, the problem only happens when linking a C++ library and occurs because while libtool does link using the C++ compiler, it supplies the options '-shared -nostdlib' and then proceeds to tell the compiler what it should already have done by default except that it gets something wrong and passes the path to the 32-bit version of libstdc++.so rather than the 64-bit version the compiler would have used when given the -m64 option. So I think that it is wrong for libtool to be specifying -nostdlib and that it would be more reliable to allow the compiler to find its own crt objects and libraries. The operation may be a vestige of when C++ libraries were linked directly using ld, except that this problem did not occur a year ago so it has somehow been added since then. Nope, but thank you very much for both your patience and your work on this issue. AFAICS this is really Solaris specific: gcc -m64 -print-search-dirs will happily print the 32bit search directories, but not the names of the `sparcv9' subdirectories. This means we have to fix this snippet of libtool.m4: | if test $withGCC = yes; then | sys_lib_search_path_spec=`$CC -print-search-dirs | $GREP ^libraries: | $SED -e s/^libraries:// -e s,=/,/,g` | if $ECHO $sys_lib_search_path_spec | $GREP ';' /dev/null ; then | # if the path contains ; then we assume it to be the separator | # otherwise default to the standard path separator (i.e. :) - it is | # assumed that no part of a normal pathname contains ; but that should | # okay in the real world where ; in dirpaths is itself problematic. | sys_lib_search_path_spec=`$ECHO $sys_lib_search_path_spec | $SED -e 's/;/ /g'` | else | sys_lib_search_path_spec=`$ECHO $sys_lib_search_path_spec | $SED -e s/$PATH_SEPARATOR/ /g` | fi | else | sys_lib_search_path_spec=/lib /usr/lib /usr/local/lib | fi or add a Solaris specific addon to the system-specific stuff below. Do you know if GCC will search _only_ $dir/sparcv9 in the list or first all $dir/sparcv9 directories and then then above? The fix should probably not change the directory name if it already ends in sparcv9 -- GCC might fix this eventually. I even vaguely remember that this may be a reported bug; it would be nice if somebody could report it otherwise. And BTW, not passing `-lstdc++' won't work as that will kill dlopening. Cheers, Ralf
Re: libtool-2.0 release
Hi Gary,
Gary V. Vaughan writes:
Is this patch really necessary for 2.0?
No.
I think that introducing so
much code churn in to libtool at this stage is going to bring yet more
release delays. Surely the feature is useful and desireable, but I
really *really* want to avoid more delays for 2.0.
OK.
Now is the time to branch! Either a feature branch for developing the
per-deplib feature for merging after 2.0, or else a 2.0 branch that we
can keep stable.
No way, without me. I outright refuse to maintain another branch.
It's insane. It makes more work for us and causes the resulting code
to receive less outside testers per branch.
However, I have absolutely no problem with delaying the application of
the per-deplibs-flags patch to shortly after 2.0.0 or 2.0.2. Although
we should still commit both the `-static' hardcoding fix and the
static.at test patch (the latter is written so that it works also
without the per-deplibs-flags patch; it needs only -static-libtool-libs).
My preference is to make a feature branch now, and not branch for 2.0
at least until the release blockers are resolved, so that we can commit
any bugfixes we discover during testing just once (we went through the
mess of porting to three branches during the last branch-2-0 debacle,
and I don't want to do that again!).
Right. So we put this off, and don't apply it anywhere right now.
According to: http://tkd.kicks-ass.net/GnuLibtoolProject/RoadMap, the
three remaining release blockers for 2.0 are:
The list is outdated.
* ! libtool.m4 macro ordering/requirement audit. pending
This is as good as done. Forget about it. I have a list of remaining
issues (not with me right now), and I will get to it eventually, but am
pretty sure those issues are not important to many users.
* ! LT_WITH_LTDL should build libltdl by default; currently
nothing happens unless LTDL_CONVENIENCE or LTDL_INSTALLABLE
is also given.
I don't even know what this means. I'd guess we should ignore it?
* ! fix potential denial of service with compilers that do not
understand -c -o.
* not very likely to happen
I don't mind postponing that to after 2.0, iff that is _not_ responsible
for bugs like this one: http://bugs.debian.org/350989 . I have replied
there, BTW.
HOWEVER.
- The regressions that happened in 1.5.22 need to be fixed in 2.0, IMHO
(most notably the OpenBSD failure; and good would be the
application of the `-static-libtool-libs' patch for users that
need the other `-static' semantics; together with the hardlinking
regression that I also found for `-static').
- And my outstanding patch-6 needs to be split up and applied.
It fixes most but not all remaining issues with HEAD libtoolize.
- I know about a couple more tweaks necessary for HEAD libtoolize
and Bob has a couple of failures I (or somebody else) need to track
down.
- AC_PROG_SED may not be AC_DEFUNed (for aclocal 1.6?), and
LT_AC_PROG_SED should be catered for.
- Makefile.am rules somewhere use GNU make 3.80 features. I have
encountered many difficulties preventing autotools reruns on other
systems, and am quite fed up with hunting these down.
- I have a pending Solaris/whole_archive_flag_spec patch, fixing a
regression, and to make libtool work on Solaris 10.
So here are the action points, initialed where I plan to take them
pending agreement from (most of) you guys.
1. strike the macro ordering audit from the release blocker list. GVV
2. fix LTDL_{CONVENIENCE,INSTALLABLE} bug on CVS HEAD. GVV
OK. Whatever that may be. You webpage does not seem accessible at the
moment.
3. make a per-deblib-branch for these changes.
Do not make a branch. Pretty please.
4. evaluate whether -c -o DoS is a release blocker too.
4a. fix it if it is
4b. strike it from the blocker list if it isn't
See above. I have a half-baked fix lying around somewhere, but will
post only after many tests.
5. test like crazy. and fix any platform issues uncovered GVV
If I haven't made it clear yet: that is what I've *really* been doing.
The static test was the aim, and the bulk of the work, the per-deplibs
flags and the `-static' hardcoding fix were the fallout.
And I have several more tests which I would like to write or have
already started. For example: I would like comprehensive exposure of -L
path issues in order to fix the OpenBSD link `-L path order issue' right,
so that it does not turn into another regression on another system.
I know you would rather release. There is a trade off: better test
exposure vs. release delay. My being fed up with working on bootstrap
and similar issues that were mostly introduced by the great code
reorganizations, and also there being very few test suite contributors
and people working on bug reports, has led me to work on the former.
Libtool will only get consistently better with a testsuite that exposes
much more issues. And if you then
