On Sun, Oct 16, 2016 at 10:47 AM, Gerardo Santana Gómez Garrido gerardo.santana@gmail.com wrote:
I would like to get more feedback regarding the summary I posted a week ago. Reverse engineering configure to get configure.ac has meant a lot of work (see my progress attached) yet I still believe this step is crucial to tidy up our build system and add new platforms more easily, by adding more and *better* tests to configure.ac.
Please let me know if there's any interest on this or whether you want to follow a different approach.
On Sat, Oct 8, 2016 at 12:27 AM, Gerardo Santana Gómez Garrido gerardo.santana@gmail.com wrote:
To summarize it:
- use simple, hand tailored, GNU Makefiles for building
- generate and distribute a config.h once per platform
- re-generate them only when a new platform is added or a current platform has had significant changes
For #3 to happen we need a configure script to do it. New platforms or changes to current ones will certainly mean adding tests to the configure script and so we also need configure.ac, for maintaining it.
Sadly, it seems that the configure script was being generated and checked into the source tree without the corresponding configure.ac. FYI, I'm working on reverse engineering it to get the original configure.ac. After that it will be easy to move to where we want to be.
On Wed, Oct 5, 2016 at 8:38 PM, Gerardo Santana Gómez Garrido gerardo.santana@gmail.com wrote:
Yes, makes sense. Thanks. I will work on simplify it, based on your directives.
On Tue, Oct 4, 2016 at 2:16 PM, Eliot Miranda eliot.miranda@gmail.com wrote:
Hi Gerardo,
On Tue, Oct 4, 2016 at 12:32 AM, Gerardo Santana Gómez Garrido gerardo.santana@gmail.com wrote:
Yes, energy and will.
Actually I would like to keep autoconf. I found out some duplicated code in the build.* directories that could probably be solved by using the autotools properly.
Are you strongly opposed to using autotools? May I try to make it work?
No, I'm not. I'm opposed to running it on each build. The scheme we've agreed upon is as follows (and I'm cc'ing vm-dev to include others working on this):
The correct realm for autoconf is in determining what facilities the platform provides, /not/ for deciding how to configure the VM, /not/ for generating Makefiles, etc. So autoconf (or CMake) should be run once per platform (updating after significant third-party software is installed, C compiler is updated, etc), and produces a config.h in the relevant build directory, e.g. build.linux32x86/config.h.
the VM is to be configured from config.h (which informs us as to what third-party libraries, os facilities, word sizes, etc are available for the current platform), the Makefile (e.g. build.macos32x86/pharo.cog.spur/Makefile), and plugins.int and plugins.ext (to determine the set of plugins to attempt to build; their makefiles may further filter-out plugins depending on available third-party libraries).
So the build scheme we're trying to move towards is
a) either autoconf or CMake is used to generate a config.h file in each build directory that defines platform facilities (#define HAS_EPOLL 1 et al). This is run occasionally, /not/ on every build
b) use conventional Gnu Make makefiles, avoiding all the mkmf scripts in platforms/unix/conf, to build specific VMs
Does that make sense? For me this is very important. See
http://lists.pharo.org/pipermail/pharo-dev_lists.pharo.org/2016-February/119...
and this is from a Squeak Oversight Board discussion before we moved to githug:
"It was decided to leave the build system as-is using GNU Makefiles where available with a commitment to move to GNU Makefiles on Linux. We will use CMake to produce per-platform config files that identify platform facilities (such as epoll(2) vs kqueue(2) vs poll(s) vs select(3))."
Is the CMake approach certain, or yet to be proven? Maybe CMake fits in better than autoconf with the third party libs used by Pharo ?
From my naive reading CMake seems a better cross platform choice than
autoconf, but it seems that CMake can't separate out generating just a config.h file from generating the build makefiles, a part that seems not wanted. Does anyone know any different? I've signed up to the CMake mail list to try to clarify this.
On the flip side, most of the stuff I read about autoconf is that it is very unix based and not suited for cross-platform with native MSWindows MSVC. Except an enabler might be CCCL, a gcc compatibility wrapper for the MSVC command-line.... https://github.com/swig/cccl#autotools-and-msvc https://folti.blogs.balabit.com/2009/08/compiling-autoconfmake-projects-unde...
What is the future of (IIUC) Pharo's dynamic generation of Cmake files?
On Mon, Oct 3, 2016 at 8:26 PM, Eliot Miranda eliot.miranda@gmail.com wrote:
Hi Gerardo,
welcome!On Sun, Oct 2, 2016 at 8:45 AM, Gerardo Santana Gómez Garrido gerardo.santana@gmail.com wrote: > > Hi Eliot, > > what is the proper way to ask questions regarding building Clog? > > I just found that in opensmalltalk-vm/platforms/unix/config, configure is not in sync with configure.ac
I guess this observation comes from generating configure from configure.ac and finding a difference with the committed configure.
But I see the history date on these files seems in sync https://github.com/OpenSmalltalk/opensmalltalk-vm/commits/Cog/platforms/unix... https://github.com/OpenSmalltalk/opensmalltalk-vm/commits/Cog/platforms/unix...
so maybe its down to different environments?
I see your version on the right here... https://www.diffchecker.com/jAQEeQOn (valid 1 month only) has this different... AC_DEFUN([AC_ICONV], [ AC_CHECK_FUNC(_dyld_present,[],[ AC_CHECK_LIB(iconv, iconv_open, ac_cv_iconv=yes, [ AC_CHECK_LIB(iconv, libiconv_open, ac_cv_iconv=yes, ac_cv_iconv=no) ]) if test $ac_cv_iconv = yes; then LIBS="$LIBS -liconv" fi] )])
which I find in... https://github.com/OpenSmalltalk/opensmalltalk-vm/blob/Cog/platforms/unix/vm... which is included by... https://github.com/OpenSmalltalk/opensmalltalk-vm/blob/Cog/platforms/unix/co... which seems referenced by... https://github.com/OpenSmalltalk/opensmalltalk-vm/blob/Cog/platforms/unix/co... but I don't really know how they all fit together.
btw Eliot, I see the following... https://github.com/OpenSmalltalk/opensmalltalk-vm/blob/Cog/platforms/unix/do... describes running configure, but I wonder how much of that from 2010 is still relevant today?
In fact I learnt something new in that doc, reading... "by running the config.status script ... is much faster than running configure all over again. (In fact, make should detect any changes to the plugin configuration and re-run config.status for you automatically.) Note: `configure' doesn't actually create any files. The last thing it does is run `config.status' to create the configured files in blddir from the corresponding file.ins in the unix/config directory.
It may or may not be. Currently I can only reliably run make to generate a new configure script on CentOS 5.3. If I try on e.g. 6.x I get an invalid configure.
Eliot, How do you determine its an invalid configure file? I just ran 'make' in platforms/unix/config without error, although the diff lines is 40k.
We hope to eliminate configure soon and use a makefile style similar to the Mac and Windows platforms. Do you have energy to contribute to this?
> > > Thanks in advance. > > P. S. I have made squeak.clog.spur compile on OpenBSD, but I'm looking for a better way to do it.