I had made a change to make 64 bit builds the default on those platforms that
support 64 bit binaries. However, the expected behavior for other mozilla
components is that you have to explicitly use --enable-64bit to make a 64
bit binary, and default to 32 bit binaries. So, I made mozldap work like the
other components.
Allow build on Mac OSX
1) There is no RPATH on Darwin, so undefine the RPATH macros before linking
2) Have to use g++ to link
3) The iconv functions are in libiconv
This fix allows the trunk ldap c sdk build to be used to build the mozilla client. I was able to use this to build the mozilla 1_8 branch of thunderbird on Fedora Core 4. Addressbook and typedown addressing work fine.
The gist of the fix is to completely ignore svrcore if not specified, and to imply --with-nspr if --with-mozilla is specified. If not specified, the client build uses $(DIST)/include/nspr and $(DIST)/lib for the locations of the NSPR files.
Enable autoconf build on Windows with cygwin and free MSVC compiler + SDK
1) Don't set the -lsvrcore in SVRCORE_LIBS in configure - instead, do it the
way we do the NSPR and NSS libs, in build.mk
2) For the component LINK macros, use name.lib instead of -lname on Windows
3) The new free MSVC compiler does not have lib.exe - add an autoconf test for
the lib program and use link /lib if missing
4) If using MSVC (which expects DOS style absolute paths) and cygwin (which
uses unix style abs. paths), use cygpath -m on all user supplied paths to make
sure they are in the correct format for cl and link. It's better to do this
during configure rather than depend on cygwin_wrapper which is much, much
slower.
5) Don't link with the odbc libs (Why did we ever need these!?!??!?)
6) The free MSVC doesn't include afxwin.h, so use the other header files
instead
7) Add libutil to the Windows build, including getting rid of the old Makefile
and adding a new Makefile.in for the autoconf build (the Makefile.client
remains)
8) getopt.c doesn't need lber.h
I verified that these changes also build on Win2k with MSVC 6 SP 3. I also changed the copyright in the new Makefile.in as suggested by Mark Smith.
When the malloc fails in the 'v' or 'V' ber_scanf case, we need to break out of the loop and return with an LBER_DEFAULT code. Also, in the realloc case, we need to save a copy of the old pointer and use either ber_svecfree ('v' case) or ber_bvecfree ('V' case) to clean up the memory.
bug 324525
If the malloc is done to allocate the *bv, but then an lber error is
encountered while doing the ber_skip_tag (due to a bad or malicious client),
the bv->bv_val is set to a random value. If you then try to use ber_bvfree to
free the bv, it may call free on this uninitialized value, and badness ensues.
I think the proper fix is for ber_get_stringal to set (*bv)->bv_val to NULL and
(*bv)->bv_len to 0 just after the malloc.
1) +DSblended can be omitted entirely - it's the default on IA, and ignored on
PA
2) +DD64 is supported on both IA and PA for B.11.23
3) The $ORIGIN linker flag should be added to any B.11.11 or later linker - the
patch to the linker to support this on B.11.11 is pretty old and should be
applied by any customer wanting a supported HP-UX system. If for some reason
they don't have the patch, the presence of $ORIGIN just causes a couple of
extra path lookups that fail - no big deal.
There were only a few places that needed to be changed. I needed to change the
order of c++ and cc detection around in configure - it seems that if you do the
cross compilation stuff first, it messes up the "regular" case, which is the
more important one. In order for the GNU_CC tests to work, you must define
CC=gcc and CXX=g++ on the configure command line. If building 64 bit, you must
also specify CFLAGS=-m64 and CXXFLAGS=-m64. I also removed the extra arguments
for static library building as they were not only not necessary but they were
also making the build abort.
Mem and resource leaks in ldapsinit.c
1) I'm pretty sure the reference to socket should be fd instead. I don't know
why it compiled with socket in there . . . note to self: read compiler warnings
. . .
2) The cert returned by SSL_PeerCertificate is "dup"ed, so we need to call
CERT_DestroyCertificate to release it otherwise we'll leak references.
3) The hostname returned by SSL_RevealURL is strduped and must be freed.
Reviewed by Mark and Pete (Thanks!)
It also seems to be important in certain cases when using g++
to build shared libraries.
I was having weird errors building the CLUs on linux x86_64. These turned out to be due to the incorrect ordering of objects on the link line.
The LINK_EXE macro already has everything needed - it is not necessary to
supply extra libs in the form of the LDTOOLS_LIBS. Just specify them in the
correct order in EXTRA_LIBS. You don't need to set EXTRA_LIBS for every
platform - on *nix platforms, they are all the same, so just define them in a
common place in the correct order. Use PLATFORMLIBS to specify platform
dependent libs.
The one tricky thing is that LINK_EXE expects OBJS to have all of the objects.
This works if you only build one executable per makefile, but this makefile
builds several, so you need to get the executable specific .obj file in the
link line as well. I suppose I could have used target specific variable
assignment to OBJS, but that seemed messy to me, so I just added a $(filter ..)
to the LINK_EXE command which will add all of the .obj files specified in the
executable link target dependency list to the link line. The filter is to
filter out libraries and other dependencies, which are usually specified
elsewhere. This seems to work on linux x86_64 and HP IPF 64 just fine.
Reviewed by Mark, Dan, Chris (Thanks!)
The newly added README.rpm explains how to build an RPM.
Running ./configure will print the list of flags and what they do.
Description: Replace ORO Regexp with java 1.4 Pattern/Matcher
Fix Description: I removed any and all references to oro from build.properties and build.xml. For the code replacement, the conversion was pretty straightforward. I had to slightly modify the regular expressions but for the most part they just work the same as the oro ones. ORO uses a PatternInput to hold the state of the iteration over the string to match, but the Matcher class does that implicitly. For the most part, it didn't matter because the code didn't really use the state, it would just always reset the PatternInput to the beginning of the string, so I just used the String instead. There was one place where I needed to keep track of state, so I used the Matcher object instead of a Perl5Pattern + PatternInput. Thanks to nkinder@redhat.com for the review.
benjamin%smedbergs.us
richm%stanfordalumni.org
jag%tty.nl
bryner%brianryner.com
bugzilla%standard8.demon.co.uk
timeless%mozdev.org
gerv%gerv.net
cls%seawood.org
dmose%mozilla.org