glibc installs some bash-scripts, but uses the path to the buildtool
bash as interpreter (on the shebang line). This is only a symlink to
the real bash, and thus is not available at runtime.
Fix that by assuming that bash on the target *will* be /bin/bash.
In C, the proper syntax for a bit-wise OR is a single '|', not two.
It worked so far because all was well:
- X_OK == 1
- R_OK||X_OK == 1
- the file we searched for had the x-bit set
-> access( file, R_OK||X_OK ) worked
- inicidentally, the file we searched for also had the r-bit set,
but we were not testing that in fact.
Accept a local tarball name as the source of the Linux kernel headers,
rather than forcing the user to use either an upstream tarball, or a
local pre-installed headers tree.
Here, we implement a highly ugly hack. I'm not proud of that one...
To build the libstdc++ library, the compiler requires the C library. In
case we build for non-baremetal, this is normally handled by the final
step, later.
But in the case of bare-metal, we never go through the final step (because
it does not work, and it seems complex enough to make it work), so the
baremetal compilers are issued out of the core step.
A few facts:
- building the C library requires a proper core compiler
- core compiler is issued from one of the core passes
- the C library is required to build libstdc++
- newlib is only built for baremetal
- in bare metal, the final compiler is issued from one of the core passes
So we need to build the C library between core pass 1 and core pass 2.
The only place is eithe libc_headers() or libc_start_files(). The most
pertinent seems to be libc_start_files().
So we build newlib from libc_start_files(), and leave libc() empty.
static linking is not possible on MacOS, and unnessecary on other systems.
The old optimization and warning flags crash the gcc on MacOS
and (imho) are a bit overdone for this software.
This patch adds support for installing the gcc test suite. A helper
Makefile is provided for building and running the gcc tests.
The default configuration runs all gcc tests and requires automatic
ssh/scp login access to a networked target board. See README for
more details.
Note: Current feature is tested with the powerpc-unknown-linux-gnu
sample but it should work with others as well.
Signed-off-by: Martin Lund <mgl@doredevelopment.dk>
The shell wrapper script uses a nonportable call to readlink.
Thus, always use the binary wrapper under BSD/MacOS.
yann.morin.1998@anciens.enib.fr:
Use 'case' instead of 'if'.
On 64bit MacOS `gcc -dumpmachine` gives i686 for the host machine.
This conflicts with the expectations of some following configure scripts
that a 64bit x86 is given as x86_64; i686 is understood as a 32 bit machine.
config.guess sets the host machine in CT_BUILD correctly.
yann.morin.1998@anciens.enib.fr:
As suggested by Khem RAJ on the ML, always use config.guess.
Call to get the directory mode depending on $CT_SYS_OS
yann.morin.1998@anciens.enib.fr:
CT_SYS_OS has changed on Linuxsystem, it only gets the kernel name "Linux",
and not the system name, 'GNU/'.
The call to stat to find out if a file is a symlink works only on GNU systems,
and the replacing portable call to readlink is also shorter and more concise code.
yann.morin.1998@anciens.enib.fr:
Apply simpler test, after discussion with author and Arnaud LACOMBE on the ML.
On non-GNU systems (BSD/MacOS) there is no uname -o.
Suppress the failure message on these systems in the
call to set CT_SYS_OS (uname -s actually sets this variable).
yann.morin.1998@anciens.enib.fr: remove 'uname -o' altogether.
g++ is only needed when building additonal libs on the HOST,
so check wheter g++ is available for the HOST compiler only
Signed-off-by: Bart vdr. Meulen <bartvdrmeulen@gmail.com>
[Yann E. MORIN: fix space damage]
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@anciens.enib.fr>
While compiling a canadian toolchain for host=mingw32, build=linux,
target=m68k-elf the build fails because in this step of the gcc build
the Host compiler is used in this stage with the build-flags for the
build system. This results in an error where the header <sys/wait.h>
cannot be found.
This problem happens at least in the GCC-4.3.x and GCC-4.4.x range.
This is solved by passing the proper compilers on the Make cmd-line
Signed-off-by: Remy Bohmer <linux@bohmer.net>
Previous addition of the canadian cross compiler did not allow
to build a baremetal only variant, no reason why this is not
allowed
Signed-off-by: Bart vdr. Meulen <bartvdrmeulen@gmail.com>
When building a cross-compiler for a host which depends
on file extensions the symlink for cc was not installed correctly
Signed-off-by: Bart vdr. Meulen <bartvdrmeulen@gmail.com>
[Yann E. MORIN: style fixes, enhancements, code prettying]
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@anciens.enib.fr>
Insight seems to be very slow to follow up on mainstreram gdb.
Latest snapshots are more than 6 months old.
Moreover, I don't have time to maintain insight support in crosstool-NG;
and, because I don't use it, I am unable to find any breakage.
For uClibc, the name of the Blackfin architecture is 'bfin'. Actually,
the naming of the architecture is quite messy: for toolchain tuples
and uClibc, it's bfin, but for the kernel, it's blackfin. We've
arbitraly choosen to name it "blackfin" in Crosstool-NG.
Add Blackfin-related uClibc patch to fix a build failure related to
fork() being used in unistd/daemon.c.
Yann E. MORIN:
Apply the patch to the kernel/linux build script to use 'linux'
in the noMMU tuples. See:
http://sourceware.org/ml/crossgcc/2010-04/msg00010.html
Saving and restoring the steps requires saving/restoring multiple
directories. Depending on the configuration, some may not exist.
Add a wrapper that checks before creating/extracting the tarballs.
When building for bare-metal the core-gcc compiler is delivered
as final compiler, so the version info and bugurl is useful
in the core compiler as well.
Signed-off-by: Remy Bohmer <linux@bohmer.net>
Not all target tuples consist of an VENDOR, KERNEL and SYSTEM part, build up the
tuple in such a way to no extra or trailing dashes are added to CT_TARGET
Signed-off-by: Bart vdr Meulen <bartvdrmeulen@gmail.com>
In some exotic case the autoreconf step of mpfr is not executed (correctly)
leaving an incorrect version number for libtool in the configure script.
After extracting the sources files, force autoreconf to be executed.
Signed-off-by: Bart vdr. Meulen <bartvdrmeulen@gmail.com>
If threads are disabled in libc, we don't want to enable them in the
final compiler. Doing so pass the configure stage, but fails latter on
a missing <pthread.h>.
Moreover, we don't want to build libgomp if threads are disabled; its
configure script would fails anyway.
Signed-off-by: Arnaud Lacombe <lacombar@gmail.com>
- don't list samples in the main help screen
- improve the samples listing in list-samples
- don't document the 'config' action, it's long dead
- document the 'V' environment variable
- improve on START, STOP and PREFIX environment variables
- add PREFIX and V to autocomplete
- advertise auto-complete at install time
Curently, populate will iterate over all ELF (shared objects|executables)
to look for missing NEEDED DSOs, adding to the list at every iterations
of the search loop.
Instead of looking again at previously handled ELF files, recursively
resolve every ELf files.
Also, in case there are a whole lot of files (more than the shell can
accept as arguments list, or creating a command line longer than the
shell can cope with), use a temporary file with the list of files
to search for missing dependencies.
- it's a POSIX compliant shell script: drop bash, use /bin/sh
- fix help text
- use an absolute path for sysroot
- replace "echo" with "printf"
- replace "stat -c '%i'" with "ls -1id"
- replace "pushd / popd" with "cd / cd -"
- remove superfluous break
- bail out if required lib not found, except if forced
If a list-file is used, then each library in the file will be handled
twice (not a real issue, as the second iteration will find the library
already present, just avoid doing the job twice).
This fixes two problems:
- the sysroot might be in a sub-directory (think SYSROOT_DIR_PREFIX)
- it is not needed to have the target tuple to properly detect the sysroot
and the required tools
As a side effect, this script is now no longer dependent on the target
tuple, and in the future, we might be able to share it across many
toolchains (when/if we can install all of them in the same place).
Add a new command line option, "-r", which allows the user to specify
an alternate sysroot location to copy libraries from. This is useful
when using the toolchain in combination with a separate root filesystem,
or when working with multiple different root filesystems.
Signed-off-by: Nate Case <ncase@xes-inc.com>
sstrip is now alone in its 'tools' menu, and we will probably never gain
any other 'tool'. Besides, sstrip is just strip, but a little bit more
agressive, so it deserves going to the 'binary utilities' menu.
The native 'tic' will _always_ be run on the build
machine, so no need to handle canadian/native/...
Reported by: Trevor Woerner
http://sourceware.org/ml/crossgcc/2010-03/msg00055.html
(transplanted from 26e89d367ea11660fd3a0bf0bcad8763e4fa21cf)
ltrace uses i386 and x86_64, whereas crosstool-NG use x86 for both cases.
Fix that by detecting what bitness we're building for, and pass appropriate
i386 or x86_64 to ltrace's configure.
The companion libraries on the target are required only for internal use by
binutils and gdb. The user should not have to know about this, so hide the
option.
On some systems (eg. *BSD and Darwin), date does not support nanoseconds
(%N) precision. Instead of printing '%N' in this case, it just prints 'N'.
Fix the sed expression to handle this case.
Add a git wrapper to retrieve components from their git tree.
Add a git wrapper to create a working copy (in our tarballs dir).
Recognise git trees when searching for local copies.
We can not rely on the user-provided version string (be it via the
choice, or manually entered), so fallback to reading version.h,
which is both reliable and always present.
It's now been a while that glibc switched to git from cvs.
Get rid of cvs to download glibc; this will make for a good
cleanup before we add git support! :-)
It's broken anyway. Eg.:
- user is already niced at 10
- user configures to renice at 5
- breaks because user is not allowed to 'boost' his/her nice value
Bette let the user handle the renice with:
nice -XX ct-ng 'action'
Signed-off-by: Remy Bohmer <linux@bohmer.net>
[yann.morin.1998@anciens.enib.fr: use defaults for CT_TARGET_ARCH]
Signed-off-by: "Yann E. MORIN <yann.morin.1998@anciens.enib.fr>
We only build the static ncurses, to be used to build the native gdb,
and it needs not be available for anyone but us. So install it into
a temporary place, and get rid of it once gdb is built.
Initial version of adding autoconf as a companion tool.
Signed-off-by: Richard Strand <richard.strand@icomera.com>
[yann.morin.1998@anciens.enib.fr: use generic overide tools dir]
[yann.morin.1998@anciens.enib.fr: update menu entries]
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@anciens.enib.fr>
Change the overide bin dir so it can be used by companion tools
Signed-off-by: Richard Strand <richard.strand@icomera.com>
[yann.morin.1998@anciens.enib.fr: more generic overide tools dir]
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@anciens.enib.fr>
By default curl doesn't folow redirects. This breaks sourceforge downloads.
Add the -L option to curl to fix this.
Curl also downloads the html as a file even when it gets a 404. This breaks
http downloads when using the failback system. Add the -f option to curl to fix
this.
Signed-off-by: Richard Strand <richard.strand@icomera.com>
If the selected ARCH is dual-bitness (eg. supports 32- and 64-bit),
then we need to know the correct place where to fetch some headers.
Currently, this applies only to x86 variants: i386 and x86_64.
Trying to download every extension in turn does not work.
The Debian server returns a friendly 404-page that is
saved as the orig.tar.bz2 file. Help the helper by giving
it the extension to retrieve.
From this version of ltrace the maintainer has removed support for
GNU Autotools, so the patch sets needed to be reworked.
Included is the latest Debian patch, by the Debian ltrace maintainer
Juan Cespedes <cespedes@debian.org>, the OpenEmbedded patches for cross
compiling, by Khem Raj <raj.khem@gmail.com> and a further set of patches
by Joachim Nilsson <jocke@vmlinux.org> for crosstool-NG.
Using this: tar cf - -C "/some/place" |tar xf - -C "/some/other/place"
to copy a directory to another place does not properly fail (when it does).
Using this instead: cp -av "/some/place" "/some/other/place"
makes it easy to see why and how it failed.
Impacted:
libc/uClibc
debug/ltrace
tools/sstrip
scripts/populate
In case the remote file does not exist (and probably for some
other reasons as well), aria2 nonetheless creates an empty file
(or not empty for some other reasons).
The solution is to delete the file whenever aria2 fails.
aria2 is a powerfull downloader that is capable of chunking and
parallel retrieval.
Due to li;itations in crosstool-NG retrieval facilities, it's not possible
to take fully advantage of aria2. It might happen that, in the future,
those limitations get lifted away, so we can take use features such as
parallel downloading from more than one server at the same time. For now,
it should still speed up downloads thanks to parallel downloading of chunks.
The user shall provide a directory to install the toolchain into.
If he/she does not, this is an error, and shall be detected properly,
rather than relying on failure down the road.
Thanks to "Pedro I. Sanchez" <psanchez@colcan.ca> for pointing out
the issue:
http://sourceware.org/ml/crossgcc/2009-12/msg00011.html
Warn about a missing local tarball directory, only if it was configured.
Avoid the spurious message:
Directory '${CT_LOCAL_TARBALLS_DIR}' does not exist.
Will not save downloaded tarballs to local storage.
Thanks to "Pedro I. Sanchez" <psanchez@colcan.ca> for pointing out the
issue:
http://sourceware.org/ml/crossgcc/2009-12/msg00011.html
Don't select unneeded config knobs. Don't select non-existing config knobs.
Use the "no patch" config knob, instead of pointing to an non-exiting local
patch dir. Simplify the tuple-related scripts. Update the samples.
When renumbering patches, the original patches get removed and replaced
with the new ones. This can be annoying to loose the original patches.
Fix this by putting the new patchs in a directory of their own.
Add config option to build wtarget code with THUMB interworking.
This is used to build the C library as well as all other code
that runs on the target.
The newlib "team" rolls new releases about once a year (december).
This is quite a long time between releases, in case code was fixed.
So, allow user to use a CVS snapshot to benefit early from fixes
and enhancements to newlib.
Some projects' module (eg. newlib) are checked-out into a sudirectory
rather than into their own directory. Handle this case in the CT_GetCVS
function.
newlib handles the build/host/target a bit differently as one would expect:
build : not used
host : the nachine that builds newlib
target : the machine on which newlib will run
Some patchsets have superfluous members in their names (eg. the ones coming
from Gentoo), so it can come in handy to pass a sed RE to strip them out of
the final patch name.
Also add a 'fake' mode, where the command will only be printed and not
executed, so we can check beforehand if the rename will be OK.
ncurses is built solely for the sake of building a native gdb.
The user should not rely on this library to build his/her userland,
but should rather build his/her own. So we remove it from the
sysroot after we successfully build the native gdb.
The option to retrieve snapshots is already handled by
the generic 'specific date' and 'use latest' entries.
No need for a special case, as there's no code for it.
For CLooG/PPL 0.15.3, the directory name was simply cloog-ppl.
For any later versions, the driectory name does have the version, such as
cloog-ppl-0.15.4.
For CLooG/PPL 0.15.3, the directory name was simply cloog-ppl.
For any later versions, the driectory name does have the version, such as
cloog-ppl-0.15.4.
Add the WRAPPER_NEEDED silent config option, that can be selected by
components that require it (companion libs so far).
Rely on this config option when deciding to install the wrapper,
instead of checking GMP/MPFR or PPL/CLoog/MPC.
Add an initial wrapper:
- find the realpath of the tool being called
- add the '.' in front of the tool name
- add the '/lib' dir to the base dir of the tool
- set and export LD_LIBRARY_PATH
- execve the real tool
Downoading a non-existing file from sourceforge gives you a "200 OK"
and an index.html. As we try to retrieve a .tar.bz2 first, and duma
is bundled in a .tar.gz, we won't get appropriate content, so
just force the extension to avoid the problem.
Thanks to Ingmar Schraub <is@eseco.de> for pointing out the issue.
During the conversion to using bash arrays, the glibc build script
was improperly converted, and contains an incorrect variable
assignment to the config_options array.
For every components where it makes sense, use bash arrays (instead
of a string with space-separated values) to store the options pased
to ./configure.
Rewrite part of the code to better match the rest.
Most notably, rewrite handling of:
if [ ... ] && [ ... ]
to:
if [ ... -a ... ]
This has the positive side effect of calling "[" only once, although
"[" is probably a shell built-in.
To test for existing files, use "[ -f blabla ]", not "[ -a blabla ]"
Checking for a file exsitence with "-a" is a bashism.
Althoug we _are_ using bash, it's disturbing as it can be misread as
the 'and' operator. Fix by using "-f".
The tmul test uses a compiled-in input file in $(srcdir).
The problem is that the Makefile passes it unquoted. The C code
tries to stringify it using clever macros, which may *usually* work.
In my case the source directory was named:
.../toolchain-powerpc-e500v2-linux-gnuspe-1.0-2.fc10/.../tests
And guess what? During testing I found out the program fails because
it tries to open:
.../toolchain-powerpc-e500v2-1-gnuspe-1.0-2.fc10/.../tests
Yes, CPP tokenized the macro before stringifying it and not surprisingly
the 'linux' part was converted to 1.
[on Fedora-10: cpp (GCC) 4.3.2 20081105 (Red Hat 4.3.2-7)]
So the attached patch simplify the macros and pass the path as string
from the Makefile.
In case the shell the user wants to use as CONFIG_SHELL is located in a
weird place (eg. /opt/bash/bin/bash), or is weirdly named (eg. bash-4),
let the user enter the patch to the shell.
The wrapper script placed around the target binaries when
using the companion libraries does not work for symbolic links
The wrapper scripts needs to follow the links before calling the
actual binary
Signed-off-by: Bart vdr. Meulen <bartvdrmeulen@gmail.com>
---
- save the canadian smples in their own way, so as not to
mix non-canadian samples with canadian ones
- list canadian samples with the host information
Add implementation for a candadian build option already
present in crosstool in order to build a cross-compiler
where build != host != target
Signed-off-by: Bart van der Meulen <bartvdrmeulen@gmail.com>
Prepare saving canadian samples by making the saveSample.sh script
a little bit more generic, using conditional code-paths (even if
there's currently a single code-path).
Collect the build tools in a seperate folder in order to prevent accidental
calling our newly build tools.
Signed-off-by: Bart van der Meulen <bartvdrmeulen@gmail.com>
Once we have canadian in place, Mingw32 can be a legitimate host,
so we have to recognise that along with Cygwin.
Also fix recognising Cygwin hosts.
Signed-off-by: Bart van der Meulen <bartvdrmeulen@gmail.com>
The symbol link that is created in the sysroot directory only needs
to be made when the cross compiler is build with the sysroot option
Signed-off-by: Bart van der Meulen <bartvdrmeulen@gmail.com>
In the non-sysroot-ed case, the debuf-root directory would not be set;
debug tools would have been installed God-only-knows-where...
Spotted by Bart van der Meulen <bartvdrmeulen@gmail.com>.
Spaces are evil in paths. Print the path that contains a sapce.
Don't print computed paths, as they'rebased on provided paths,
and don't get space added into them.
The dynamic linker, ld.so, needs the execute bit to be set.
Detect tht the library being installed is in fact ld.so and
install it with 0755 instead of 0644.
Fix detecting src == dst.
Use a simpler command to copy src -> dst.
Also change echo to printf, get rid of 'echo -n', which is
highly non-portable.
-------- diffstat follows --------
/trunk/scripts/populate.in | 76 43 33 0 +++++++++++++++++++++++++++++-----------------------
1 file changed, 43 insertions(+), 33 deletions(-)
They have nothing to do in here, just let the user
configure his/her system appropriately.
-------- diffstat follows --------
/trunk/scripts/build/libc/eglibc.sh | 1 0 1 0 -
/trunk/scripts/functions | 100 0 100 0 -----------------------------
/trunk/config/global/download.in | 148 0 148 0 -------------------------------------------
3 files changed, 249 deletions(-)
- introduce the config dir, where components can store their config files
- move the munged uClibc config file to the config dir
- now, the state dir really is an indication that a build can be restarted
Thanks to Groleo Marius <groleo@gmail.com> for spotting the inconsistency
of the state dir usage, and suggesting this change.
/trunk/scripts/build/libc/uClibc.sh | 6 3 3 0 +++---
/trunk/scripts/crosstool-NG.sh.in | 9 7 2 0 +++++++--
/trunk/scripts/functions | 15 12 3 0 ++++++++++++---
3 files changed, 22 insertions(+), 8 deletions(-)
Initial patch by Dmitry PLOTNIKOV: http://sourceware.org/ml/crossgcc/2009-03/msg00053.html
It [the toolchain] uses current ct-ng (nightly snapshot 20090324, latest
release 1.3.2 work also), glibc 2.9 (from CVS), binutils 2.19 and latest
snapshot of GCC 4.4.0 (as of March 20, 2009).
We have successfully built linux kernel 2.6.29 and a lot of other stuff
with this toolchain.
Here's the patch that adds GCC 4.4.0 to the ct-ng menu and enables it to
download a 4.4.0 snapshot from ftp.
Patch was adpated by me, mostly to better fit the configuration layout.
/trunk/scripts/build/cc/gcc.sh | 34 22 12 0 ++++++++++++++++++++++------------
/trunk/config/cc/gcc.in | 35 30 5 0 ++++++++++++++++++++++++++++++-----
2 files changed, 52 insertions(+), 17 deletions(-)
- find the executables extension (needed under some OS, like Winblows)
- build tic in //
- simplify the make and install command lines
/trunk/scripts/build/debug/300-gdb.sh | 10 7 3 0 +++++++---
1 file changed, 7 insertions(+), 3 deletions(-)
Seems ncurses 5.7 need build host stage for tic step; if use host tic
(ubuntu) the build process hang in the below step.
So I guess need to build ncurses host stage to build new tic
and provided a patch to that efect.
And in fact, we do need "tic" to run on the _build_ system to properly
generate the terminfo database.
Note: this is fully functional, but still requires a litle bit of
tweaking so that ${CT_BUILD}-gcc gets used instead of plain gcc.
But that's a minor problem for now...
/trunk/scripts/build/debug/300-gdb.sh | 35 33 2 0 +++++++++++++++++++++++++++++++++--
/trunk/scripts/build/internals.sh | 1 1 0 0 +
2 files changed, 34 insertions(+), 2 deletions(-)
- recently, tarballs for glibc 2.8 and 2.9 have appeared on the GNU ftp site
- always use a dot in version strings (eg. 2.9, not 2_9)
/trunk/scripts/build/libc/glibc.sh | 135 76 59 0 +++++++++++++++++++++++++-------------------
/trunk/config/libc/glibc.in | 71 45 26 0 +++++++++++++++--------
2 files changed, 121 insertions(+), 85 deletions(-)
"chmod u+w" the full src tree: because of nochdir and cvs snapshots, we can't reliably know were we are...
/trunk/scripts/functions | 11 3 8 0 +++--------
1 file changed, 3 insertions(+), 8 deletions(-)
