| Age | Commit message (Collapse) | Author | Lines |
|---|
|
In some cases (specially when debugging the pipeline), its very
time-consuming to install GCC. With this commit, a `--host-cc' option has
been added to avoid building the C compiler when necessary.
The test to see if `sys/cdefs.h' is available on the system (necessary to
build GCC) has also been moved to the configure script to print a more
visible warning and also use the new `host_cc' variable to let
`dependencies-basic.mk' know that GCC shouldn't be built.
Finally, we are having problems installing M4 from source, so it has been
set as a mandatory dependency.
|
|
On some GNU/Linux distros, the compiler is separated into `multilib' mode
(for 32-bit and 64-bit support) and by default the extra component of the
compiler is not installed! In such systems for now, we are just creating
symbolic links to the host's compiler (similar to Mac).
While testing, I noticed that we weren't passing a "$downloader" option to
the downloading script of `dependencies.mk' and `dependencies-python.mk'.
Also, I noticed that the Cython and Python-pkg-config packages didn't have
setuptools as a dependency! Both have now been fixed. Also, Cython's
tarball name is now all small-caps (as in all the other tarballs).
|
|
Until now, management of the software names and versions in the paper was
done manually (a macro had to be defined in `initialize.mk', then used in
`paper.tex', so they had to be manually set in two places). Managing this
was not easy.
To fix this, with this commit, each software building rule's target is a
text file that contains its human-readable name and its version. In the
end, the configure script sorts them by their name and writes them into a
LaTeX input file that we can easily import as a file into the main paper.
|
|
For the tests, we had just set an absurd value for a test in the GCC build
recipe to always fail, but we had forgot to fix it. It is now corrected.
Also the order of making `g++' and `gfortran' was reversed for easier
readablility (it doesn't matter which one is done first, it only matters
that `gcc' be done last).
|
|
We were developing the build of Numpy and Scipy on Mac in a parallel thread
and things seems to be working relatively nice now. There were only two
problems:
1) GCC still has some random building issues on Mac.
2) ATLAS shared libraries can't be built on Mac (so we used OpenBLAS to
build Numpy and Scipy on both Mac and GNU/Linux).
But for now, none of these problems are critical. So, we can progress in
one branch.
There were only very minor conflicts in the merge.
|
|
We wer not able to build `gcc' on Mac, so we are using links to the host
compilers. In this commit we also found that on Mac the HDF5 library
needs an explicit definition of the compilers.
|
|
Until now, we were using `flock' (file-lock) for downloading the input
datasets in series. But we couldn't do this when downloading the software
tarballs because `flock' wasn't yet available. Generally, unlike
processing, downloading is much better done in series than in parallel.
To enable serial downloads of the software also, with this commit we are
installing `flock' in the configure script (not in a Makefile). As a
result, besides `flock', we can also benefit from the other good features
of the `reproduce/src/bash/download-multi-try' script *(for example
attempting download again after some time).
Some GNU mirrors may have problems at the time of download, so with this
commit, we are using the main GNU FTP server for GNU programs.
|
|
Until now, we were simply using the host's GCC for Mac systems. But we
found that except for a single step (to fixing `rpath'), it works on
Mac!!! So, GCC is now part of the Mac build as well.
However, we are still having some problems in building ATLAS on Mac. It
works on GNU/Linux, but not in Mac. So for the time being (just
temporarily), we are avoiding ATLAS (and thus Scipy) on Mac systems. We
just filed an issue on the ATLAS discussion list to hopefully fix the
problem soon.
|
|
We just noticed that recently the `paste' command on macOS doesn't work
with a pipe. So we are now simply using the `tr' command in reverse to
re-create the PATH (to find where to link to).
|
|
Until now we were using a symbolic link to replace GCC, but Make doesn't
treat symbolic links like files. So it would rebuild the links every
time. With this commit, only for GCC on Mac systems, we are actually
copying the host's GCC executable to avoid this problem.
Also, a wrong comment for cURL was removed.
|
|
Conflicts in `gcc' build comments and in mentioning software used in
paper fixed.
|
|
Numpy needs ATLAS as shared libraries. So we also need to build Python with
shared libraries. We also need to define site.cfg for numpy and scipy so we
define a master template:
`reproduce/config/pipeline/dependency-numpy-scipy.cfg'
Also `Openssl' did not have rpath so we added with this commit.
|
|
An initial installation of atlas is now included in the pipeline,
but we are still trying to make it compile and build smoothly. In
the process, we found that GCC also needs some modifications
(for example rpath issues).
|
|
Until recently, there was no problem with the `makelink' script of
`dependencies-basic.mk' because it was called on separate recipe lines (and
thus separate shells). But recently we added a call to it within a single
shell (for GCC on Mac OS systems). So a previous call to it would effect
the next call. To fix this, in this commit, we are re-setting PATH to its
original value after each call finishes.
|
|
Bzip2 has a special/separate Makefile to build shared libraries which
didn't work on a macOS. So with this commit, we are allowing Bzip2 shared
libraries only on macOS systems.
Also, I noticed that macOS's `sed' doesn't have the `-i' option (to do the
change in place within the same file). So we are using `-e' to write the
changed Makefile in a temporary directory, then rename that.
|
|
We still have a few problems with building GCC on a MacOS system. To allow
using the pipeline on this operating system, until we find the solution,
GCC is only built on non-Mac systems. On Mac, we'll just make a symbolic
link to the host's executables.
|
|
With the help of Raul, we were able to build many higher-level Python
packages to enable the installation of packages like Matplotlib and
Astroquery. With this commit, that work is being merged into the master
branch.
|
|
Until this commit, we had some of the python packages intalled
but they did not work properly because of the `PYTHONPATH' variables.
That is, the pipeline's `python' was the `python' of the system
instead of the pipeline's `python'.
With this commit this issue has been fixed by setting the correct
`PYTHONPATH'. In this commit we also modify the installation of
`bzip2' because `CMake' was complaining about some libraries built
statically.
|
|
Until now, the pipeline was not installing its own `gcc' but using the
system one by making a symbolic link.
With this commit, GNU GCC has been added into the pipeline. Right now
the installation does not work on Mac OS system beause of some conflicts
with `clang', but in principle it should work on GNU Linux distributions.
|
|
In an attempt to test the GCC build rule (without Binutils, because its too
architecture dependent), all the necessary dependencies were moved to GCC
(from `ld'). Also `fortran' was also added to the languages supported by
GCC. This rule built GCC 8.2.0 nicely on my GNU/Linux system. But `gcc' is
still not a final target to built, so the rule is being ignored for now.
|
|
As in all programs, the build process of ncurses depends on the running
shell (Bash) and AWK. At the start of the building of ncurses, we remove
its library. But Bash and AWK depend on ncurses to run (this creates a
circular dependency). Therefore its necessary to remove the Bash and AWK
executables when re-building ncurses.
This bug was found by Raul Infante Sainz.
|
|
Until now, the check to see if the patchelf program should be used or not
(for GNU/Linux vs. Mac installations) was mistakenly added over the step
that we define the `sh' symbolic link, not over the call to patchelf. This
is corrected with this commit.
|
|
After testing the built of Metastore on a server, I noticed that because
its `/etc/passwd' doesn't have the list of users, the `getpwuid' call
within metastore failed and wouldn't let it finish.
So I looked into the code and was able to implement a solution to this
problem by adding two options to it for default values for the user and
group. Also, file attributes are not necessary in our (current) use case of
metastore and caused crashes on our server, so they are also disabled.
|
|
With the current build system, Bash and AWK don't write RPATH into the
executables. This causes many problems in the pipeline (for example when
using the `$(shell)' function in Make which doesn't have
`LD_LIBRARY_PATH').
After consulting the Bash and Make mailing lists, so far, the best solution
was to use the Patchelf program to manually write RPATH in these
executables. With this commit, Patchelf is now installed in the pipeline
and used in Bash and AWK to fix this problem.
|
|
The build of bash has been made a little cleaner to help in readability and
management of the code.
|
|
In the previous commit, the copyright year and owner were mistakenly
modified. They are corrected now.
|
|
While working on a pipeline based on this, I noticed many linking errors of
our installed Bash, complaining that it can't link with libreadline. This
was while readline was present in the proper directory and the Bash within
a recipe would work properly.
After some investigation, I found out that this is because Make's `foreach'
function (which was used to define the targets) was apparently calling Bash
without setting `LD_LIBRARY_PATH', causing this error.
To avoid such sitations, Bash now uses its internal build of readline and
we no longer ask it to link with the installed readline.
|
|
The targets of the links to have the extra common `ncurses' packages were
previously just `pkgconfig/*.pc'! But this would only work when run within
the `installed/lib' directory, not any other! So the targets for these
packages now use an absolute address.
|
|
Wget and cURL depend on many network related libraries by default and if
they are present on the host operating system, they will be linked
with. This causes problems for the pipeline when these libraries are
updated on the host system.
With this commit, I went through the configure time options of both Wget
and cURL and removed any library that didn't seem related to merely
downloading of files (possibly with SSL, because we do build OpenSSL in the
pipeline).
Also, I noticed a new version of cURL has come, so that is also updated.
|
|
In a previous implementation, we were using a `target' variable to define
the final target of several links, but with the new `sov' solution, we just
used its base name. However, we had forgot to correct two instances of
`target'. This is corrected now.
Also, the step to clean all already built outputs of the NCURSES library
has been simplified to a platform independent wildcard.
|
|
On Mac OS systems, the full version number is not used in the filename
given to libncurses. For example for version 6.1, it is called
`libncursesw.6.dylib'. So a more generic and easier to maintain and read
script is now used to be able to make links for both Mac and GNU/Linux
systems.
In short, instead of checking if we are in Mac every time, we just set the
suffixes at the start based on the machine once as variables and use those
to define the links.
|
|
Readline is a prerequisite of Bash and AWK, while NCURSES is a prerequisite
of Readline. With the recent update of GNU Bash (and thus GNU Readline) on
my host operating system, the pipeline crashed and I noticed this hole in
the pipeline. In particular, AWK (which linked with Readline 7.0) would
complain about not finding it and abort.
|
|
ccache is a super annoying program in the context of the reproduction
pipeline. On systems that use it, the `gcc' and `g++' that are found in
PATH are actually calls to `ccache' (so it can manage their call)!
Two steps have been taken to ignore and disable ccache (if it isn't ignored
properly!): 1) when making symbolic links to compilers, if a directory
containing `ccache' is present in the PATH, it is first removed, then we
look for the low-level programs that we won't be building. 2) The
`CCACHE_DISABLE' environment variable is set to 1 where necessary (with the
other environment variables).
|
|
After installing Bash, we would just blindly try to build the $(ibdir)/sh'
symbolic link. But that could fail if it already existed. To make things
clean, we now remove any link first before attempting to make a new one.
|
|
Since the current implementation of this pipeline officially started in
2018, all the files only had 2018 in their copyright years. This has now
been corrected to 2018-2019.
|
|
By giving this option specifically at the build time of Pkg-config, we'll
ensure that any package that uses pkg-config will first look into our
locally installed build.
|
|
Both Gzip and Gnuastro were being bootstrapped personally from their Git
repository until now. But fortunately a new release of both came out last
week and so to make things standard we are now using their standard
tarballs.
I also noticed that we weren't checking the version of Gzip or mentioning
it in the acknowledgement section. This was also corrected.
|
|
While checking the build of the previous commit, a failure happened when
linking `reproduce/build/dependencies/installed/bin/sh' with the built Bash
(because the symbolic link already existed!). So a `-f' flag was added to
`ln' to just change it without complaining.
I also noticed that the Git build was also not in verbose mode. So this has
also been corrected.
|
|
While we were testing this pipeline on a Mac OS system, we found and
reported a problem in Gzip's build (bug #33689). However, since the Gzip
build is not verbose, it was necessary to run its `make' with
`V=1'. Generally, since almost all the programs are built in verbose mode
(where you can see the compilation commands), we have also set this flag in
any build to be clear and make it easier to spot bugs in the future.
|
|
Some problems with using the number of threads in dependency building were
fixed.
|
|
Some host Make systems may not allow automatic passing of the number of
threads to sub-Makes. So while building the basic dependencies, we'll need
to explicity add the `-j' option to the Make files that can benefit most
from it: those that are dependencies of many others (Tar & Make), or are
the last to build (Coreutils).
|
|
The build systems of Libgit2 and WCSLIB on Mac OS does not account for
installation in non-standard addresses: `Libgit2' keeps the absolute
address of its build directory (not the installation directory) and WCSLIB
doesn't write any absolute address at all (so the system uses the first one
it finds).
To address these issues, we are now using Mac OS's `install_name_tool'
program to fix the absolute path within the installed shared library.
Since the version of the library is actually present in its shared library
name, in `dependency-versions.mk' we have also separated these two
libraries so later when their version is changed, we are careful in
correcting the shared library name also.
|
|
`ln' will complain about a link already existing. So to avoid having to
rely on the `-f' option (which may not be portable across systems), when we
are making symbolic links to the OS tools that we won't be building, we now
remove the file if it exists, then make a new symbolic link.
|
|
Until now the low-level links that we put in our internal installation from
the operating system were a prerequisite of essentially all the basic
dependencies. So a change in them would mean a full re-build of all the
basic dependencies. But in building the basic dependencies, we already have
the operating system's PATH and other environment variables. So unlike the
higher-level dependencies, they don't need these links at all!
With this commit, the `low-level-links' file is placed in `installed/bin'
and is a top-level target of the basic dependencies build. In this way, if
it is necessary to update/change to use something from the host operating
system, we can simply delete it and run `./configure' again wihout having
to re-build all the basic dependencies.
|
|
On Mac OS systems, the `sw_vers' executable prints information about the
operating system. It is used by TeX Live to determine the necessary builds
to download and install. We are thus importing it as a low-level tool in
`dependency-basics.mk'.
|
|
Until now, we were checking the existance of the `configure' file and if it
wasn't present, we would check for `config' (for OpenSSL which also has a
lower-level "Configure" script that is called by the `config' script). But
after two tests on Mac machines by Raul Infante Sainz and Cristina MartÃnez
Lombilla, we found out that Mac Os's file names aren't case sensitive and
thus the build wouldn't use `config', but `Configure'.
Now, the exact configuration script can be specified as the 7th argument to
the `gbuild' script. If it isn't given, the standard `configure' name will
be used, but when it is, the given name will be used.
|
|
OpenSSL can't automatically detect the architecture of Mac OS systems, so
as it suggests on its Wiki, it needs some help for doing that. With this
commit, we are checking the build on Mac OS with the presence of `otool'
(Mac OS's linker). If it's there, we'll add the OpenSSL configuration
options suggested by OpenSSL's Wiki.
|
|
Until now, we weren't including the `rpath' linking options to the basic
dependencies. They are now added. Also, when the download of an input file
fails for any reason, an empty file won't be replaced there any more.
|
|
We were missing a `\' at the end of the `$(call' function of Coreutils to
connect the two lines. It has been fixed now.
|
|
To enable easy downloading of HTTPS links with Wget (this pipeline's defaut
downloader), we need a set of trusted CA certificates. Until the time that
we can generate one ourselves, one generic set of trusted CA certificates
is now downloaded like a tarball and placed in the OpenSSL configuration
directory.
With these CA certificates, within the pipeline we can now safely use the
pipeline's own installed Wget.
|
