Pass 2 libstdc++ still links to libgcc.a even with LDFLAGS_FOR_TARGET=,
despite this libgcc.a is from pass 2 instead of pass 1.
The difference between pass 2 libgcc and pass 1 libgcc is Glibc wasn't
installed when the pass 1 libgcc was built. This difference causes both
consequences (1) pass 1 libgcc lacks shared library and (2) pass 1
libgcc cannot support C++ EH, but it's not (1) causing (2).
This option makes ld use DT_RUNPATH instead of DT_RPATH. DT_RPATH is
generally considered bad because it takes precedence over
LD_LIBRARY_PATH. For example, eog is linked with -rpath /usr/lib/eog,
and with DT_RPATH if an old eog is already installed we are basically
impossible to debug a new eog build w/o overwriting the system
installation first or explicitly using "ld.so --inhibit-rpath" to
invoke it.
This "new" actually means "new in 2000," it's 24 years ago and all other
distros has enabled it. Thus I guess some unexplainable "test suite
uses installed library instead of the just built one" issues in BLFS are
actually caused by this difference: the package author just assumes
everyone is using DT_RUNPATH thus they just set LD_LIBRARY_PATH and
consider it enough to test with the just built libraries, but DT_RPATH
breaks this expectation.
Let's eliminate the difference as it seems not doing anything good and
doing so just takes one switch.
GCC 14 libsanitizer no longer depends on crypt.h. But let's keep this
option for reducing build time, just update the explanation.
Also remove libxcrypt from GCC depedency list.
Update to vim-9.1.0405.
Update to util-linux-2.40.1.
Update to linux-6.8.9.
Update to jinja2-3.1.4 (Python mpdule).
Update to iana-etc-20240502.
Update to gcc-14.1.0.
The effect will not change, but with symlinks ld can save some time
invoking open(), read(), etc. syscalls and parsing the linker scripts.
Note that I've also removed "libcursesw" symlink because this library
has never existed. Instead libcurses.so is created as a symlink
direct to libncursesw.so.
instead of the 8-bit ncurses.
We don't provide the 8-bit ncurses library and we are "faking" it using
ncursesw. Thus innocent package may be compiled with the 8-bit ABI
(because it does not know what we are doing and so it does not use
the "expected" preprocessor definitions to enable the wide ABI) but
linked against ncursesw, causing a potential ABI mismatch.
Well, I was blaming libtool too much. If the entire Binutils tree uses
libtool this won't happen. The problem is Binutils building system is
using libtool-style idiom on non-libtool components.
And this issue is not related to cross compiling, at all. A native
build can exploit the issue as well (see the updated comment).
Maybe I'll submit a patch to GCC (yes, not a typo, GCC is the upstream
of Binutils building system) to fix the issue when I have the mood...
- Update to systemd-253
- Update to bc-6.3.1
- Update to linux-6.2.2
- Update to procps-ng-4.0.3
- Update to iproute2-6.2.0
- Update to meson-1.0.1
- Update to make-4.4.1
- Update to elfutils-0.189
We need to enable decimal float here or MPFR will be built w/o decimal
float support. Then 2 of 183 tests will be skipped, and this will also
cause an ICA issue.
Q: Why we need decimal float in pass 1?
A: We need pass-1 GCC with decimal float support to build decimal float
routines in pass-2 libgcc.
Fix make-4.4 bug.
Update to wheel-0.38.4 (Python Module).
Update to texinfo-7.0.
Update to sysvinit-3.05.
Update to shadow-4.13.
Update to sed-4.9.
Update to meson-0.64.0.
Update to linux-6.0.7.
Update to elfutils-0.188.
Update to bc-6.1.1.
Committing only the commands for now, so that others can test the
build. TODO:
- add command explanations
- add changelog
- comment on failing tests in binutils and gcc
They are really harmful. In Binutils pass 2, libstdc++.la caused the
building system to use host /usr/lib/libstdc++.so for gprofng. We now
has disabled gprofng for pass 2, but the similar issue also exists in
GCC pass 2. In a normal LFS build, the building system silently uses
/usr/lib/libstdc++.so (I guess it does not blow up simply because some
blind luck); in a real cross build (x86 -> ARM for example) the build
will fail.
Remove the .la files to fix this issue. Instead of only modifying
clfs-ng, it makes more sense to apply the change for trunk: though
the build does not fail, using host library is still a contamination.
Presently we let the build system generate static C++ bindings, and
then we remove them. Note that we could also prevent generating
any C++ binding, since nothing in LFS/BLFS use them, but it seems to
me that generating the shared ones is closer to what is done for
other packages.
This is the issue preventing us from cross-compiling libstdc++ in
Chapter 6. By fixing this issue we can remove a seperate pass 2 for
libstdc++ and simplify the instruction.
The upstream fix will be released in 11.3 and 12.0, so we can remove the
first sed upgrading gcc next time.
A very old libtool copy (2009-11-29) is shipped in binutils tarball. It
does not support sysroot, so the cross-built binutils binaries may link
to libraries from the host distro, if certain libraries are available.
The ideal solution should be updating libtool, as libtool-2.4.6 (in LFS)
has sysroot support. However, updating libtool for binutils is not
trivial: it would require to rerun autoconf and binutils building system
sticks to autoconf-2.69. Another issue is the sysroot support for
libtool has introduced a configure option "--with-sysroot", which
conflicts with an already existing option with the same name in
GCC and binutils building system (we are using the GCC/binutils version
of --with-sysroot in chapter 5).
GCC building system has --with-build-sysroot (we are using this for GCC
pass 2) for this issue. Binutils copied GCC building system, but it
does not respect --with-build-sysroot.
So for now we just edit libtool code to prevent "-L/usr/lib" in
$LFS_TGT_gcc command line. It should fix the issue about host libiberty
(reported in #lfs-support) as well, but it still need to be confirmed by
someone having such a host.
Tested with a jhalfs run on LFS.
