was not created before trying to create files in it.
The ``install_deps()`` method was also dead. It looks like that
should have been named ``mk_install_deps()`` but even if we give it
the right name it does the wrong thing. That method is supposed
emit target names (i.e. dependencies) not commands.
".NET" bindings. This file is required for Monodevelop to find the
bindings because Monodevelop uses pkg-config to find packages
(it doesn't use the GAC).
For lack of a better name the GAC (and pkg-config) package name is now
``Microsoft.Z3.Sharp``. I don't want to call it ``Microsoft.Z3`` because
someone may want to create a ``Microsoft.Z3.pc`` file in the future for
the native Z3 library (i.e. C++ or C bindings).
In addition there is a new utility function ``configure_file()``
which reads a template file, applies some substitutions and writes
the result to another file. This very similar to what CMake does.
There is a new environment variable ``Z3_INSTALL_PKGCONFIG_DIR``
which allows pkgconfig directory to be controlled for the install.
under non Windows systems (i.e. Using mono).
Building these bindings is unfortunately on by default because
I didn't want to change the command line interface (i.e. ``--nodotnet``)
which people might be relying on. This should really be changed to
match the other binding flags (e.g. ``--java``) but I will leave
this for now.
To perform the build a C# compiler and the GAC utility are required.
The script will try to automatically detect them but the user can
override this by setting the ``CSC`` and ``GACUTIL`` environment
variables.
In order for the ".NET bindings" to be installed the assembly
(``Microsoft.Z3.dll``) needs to have a strong name which means
we need a Strong name key file which is what the
``Microsoft.Z3.mono.snk`` is for. This is the public and private
key so this key **must never** be used for checking integrity. Instead its
only purpose is to avoid any name clashes in the GAC.
It is also worth noting that slightly different flags needs to
be passed to the C# compiler on non Windows platforms. I don't
understand why some of the flags are being used on Windows but I left
a comment there that hopefully someone can fix...
variables:
Z3_INSTALL_BIN_DIR - defaults to "bin"
Z3_INSTALL_LIB_DIR - defaults to "lib"
Z3_INSTALL_INCLUDE_DIR - defaults to "include"
This has two advantages
* We no longer hard code strings like "bin" all over the place
* Packagers can easily control where things get installed.
create_relative_symbolic_link(out, '/usr/lib64/libz3.so',
'/usr/lib/python3.5/site-package/libz3.so') would create an incorrect relative
path because it would consider ``/usr/lib`` to a be a path prefix of
``/usr/lib64``.
which would try to uninstall components that were never installed.
This bug would cause the following line to be emitted in the
``Makefile`` under the ``uninstall`` rule even though there was
no corresponding rule to install the file under the ``install`` rule.
```
@rm -f $(DESTDIR)$(PREFIX)/bin/test-z3$(EXE_EXT)
```
duplicated in too many places by refactoring the installation and
removal of the Python bindings to use the ``MakeRuleCmd`` class.
In order to make this change:
* ``PYTHON_PACKAGE_DIR`` no longer contains the text ``$(PREFIX)``
* ``PYTHON_PACKAGE_DIR`` **MUST BE** inside the installation prefix
duplicated in too many places and being worried that someone might
forget to use it when installing additional components.
To acheive this the new ``MakeRuleCmd`` class provides
several class methods to generate commonly needed commands used in
make file rules.
Most of the build system has been changed to use these helper methods
apart from stuff related to the Python bindings. This can't be changed
until we fix how PYTHON_PACKAGE_DIR is handled. Right it not guaranteed
to live under the install prefix but this is a requirement when using
the ``MakeRuleCmd`` methods.
There were several problems with the existing implementation.
* When specifying ``--prefix`` the implementation would assume the
path was ``$(PREFIX)/lib/python-<VERSION>/dist-packages``. This
is incorrect. ``dist-packages`` is Debian (and its derivatives,
i.e Ubuntu) specific and won't work on other Linux distributions
such as Arch Linux.
* When generating the ``Makefile``, ``$(PREFIX)`` was only emitted
during the Python installation when ``--prefix`` was passed on
the command line. When ``--prefix`` was not passed the absolute
path to the Python package directory was emitted. This is not
very consistent.
This patch checks that the detected Python package directory lives
under the install prefix and emits an error if it does not as this
indicates that the installation will be broken. If the Python package
directory does live under the install prefix it replaces that prefix
with the ``$(PREFIX)`` variable when emitting the ``Makefile`` for
consistency with the other install commands.
If a user really wants to install to a particular Python package
directory they can force it with the newly added ``--pypkgdir=``
flag.
when installing the Python bindings.
If ``DESTDIR`` is set the bindings will now be installed under this
path but ``$(PREFIX)`` only appears in the ``Makefile`` if ``--prefix``
was set which seems a little broken (we'll fix this soon).
The creation of the Python ``site-packages`` (and ``__pycache__`` for Python
3) directories has been moved to build time instead of configure time
because we don't know what ``DESTDIR`` will be set to at configure time.
DESTDIR make file variable (https://www.gnu.org/prep/standards/html_node/DESTDIR.html)
for ``install`` and ``uninstall`` targets.
Typically packagers build packages like so
```
$ ./configure --prefix=/usr/
$ make
$ make DESTDIR=/some/path/ install
```
Doing this installs the files into a directory ``/some/path`` but places
files inside that directory using the layout in ``--prefix`` (e.g.
``/some/path/usr/bin/z3``). The ``/some/path`` directory can then be
packaged (e.g. tarballed) for later installation.
The ``DESTDIR`` is not set in the Makefile and thus is empty by default
which maintains the existing ``make install`` behaviour.
Unfortunately this situation isn't fixed for the Python bindings (and
possibly others) yet as more invasive changes are needed here. I'll fix
this in later commits.
