On 18.10.2013 19:54, janI wrote:
On 18 October 2013 16:52, Andre Fischer <awf....@gmail.com> wrote:
On 18.10.2013 15:58, janI wrote:
On 18 October 2013 15:00, Andre Fischer <awf....@gmail.com> wrote:
On 18.10.2013 14:02, janI wrote:
sd
On 18 October 2013 13:36, Andre Fischer <awf....@gmail.com> wrote:
On 18.10.2013 11:32, janI wrote:
Hi.
due to the discussion in thread "Mentor a new build system", I have
made a
proposal for a central Makefile located in main.
Hi Jan,
it is great that you are going to improve this part of the build
system.
But I think that we need more details about how the proposed build
system
works. Without them I can not really evaluate the proposal.
First of all, I agree with juergens remarks that this should be
discussed
before implemented, hence the wiki page.
Secondly this has nothing directly to do with the proposed build system,
its a simple replacement of build.pl in the current system.
Yes, that is how I understood it. I just did not know how to call the
build.pl replacement.
I know that build.pl works, but having a Makefile in main, would make
us
one step closer on being compatible with the distros. To me this job is
a
simple cleanup, not something we deadly need, but nice to have.
Some remarks regarding the missing options:
--from <module>
This is one of the more important options and one that I use
frequently
(also in the form --all:<module>).
Note that if you are in <moduleA> and call 'make --from <moduleB>'
then
all modules are built
a) which <moduleA> depends on
b) but not those that <moduleB> depends on
c) Both <moduleA> and <moduleB> are built.
I have changed the documentation.
I use the --all:<module> myself very often, and have changed the
documentation, because it is of course supported.
The difference is that you do the call in main, but that is a minor
detail
that can be easily corrected (have <module>/Makefile calling
main/Makefile.
I have also changed documentation on --html due to juergens comments.
I am not sure that we understand --from and --since in the same way so
I
will try to explain what I think they do.
Let's imagine that we have a simple project with modules A, B, C, D and
E.
where B depends on A, C on B, D on C, and E on D.
A ' make all' would mean 'make E'. The dependencies would then lead to
building modules A, B, C, D, E in this order.
If I am in E and call 'make --from C' then only C, D, and E should be
built. A 'make --since C' would only build D and E.
If I am in D and call 'make --from B' then modules B, C, and D are built.
Call 'make --since B' to build only C and D.
Note that 'make --from' accepts more than one module name (while 'make
--all:<module>' does not).
Note also that in the above case (stand in D, call 'make --from B')
module
A is not built, regardless of whether there are changes in A or not.
Whereas a simple call to make (still standing in D) would build all
modules that D depends on, directly or indirectly. Thus the options
'--from' and '--since' exist to actively exclude modules from being
built.
The whole thing becomes a little bit more complicated with multiple
options to '--from' (I never use '--since' and also don't know a valid
use
case so I will ignore it for now) and more complex dependencies then in
the
simple example above. Let's say that if we stand in instsetoo_native and
call 'make --from svx sfx2'. Note that svx depends on sfx2. This would
build svx, sfx2 and all modules that depend (directly or indirectly) on
svx
OR sfx2.
got it, now I just have one problem, why would you not build the
dependent
modules, if they needed to be built, thats a scenario I dont understand.
With a central makefile, <module>/makefile will not be called so we do not
waste cpu cycles.
With the .done files, we know when a module was last built and all modules
that depend it should be rebuilt which the rule
<module>.done : <module_depend>.done
will ensure, so If we have A -> B -> C -> D
I go in B, and call make, then when I go in D and make, B,C,D will be
made.
If we have A -> B -> D C -> D
and do the same then only D will be made.
So --from is not really saving anything ?
a) In your example you first go into B then, in a second step, into D.
The '--from' option lets you do the same (well, not really the same, but
see below) just from D.
b) You go first to B and call make. This makes A, if necessary, then B.
The making of A is exactly the thing that you want to prevent with the
'--from' option. Go into D and call 'make --from B'. A is not built.
Actually I go to main and say "make D", that will cause B but not A to be
built.
I guess that calling 'make D' from main/ would be equivalent to call it
from instsetoo_native/. But the --from option also works from inside
other directories. It can then be used to build a 'range' of modules,
ie. go into module Y/ and call 'make --from X' and all modules from X/
to Y/ will be built. A real life example for using this method could be:
1. Make some changes to documentation but not code in sfx2/.
2. Make some changes in svx/.
3. Then go to sw/ and call 'make --from svx'
This prevents the changes in sfx2/ to trigger recompilation. It builds
only modules svx/ and everything that depends on it until sw/ is
reached. The costly recreation of install sets (especially when there
are multiple languages) is omitted.
With .done files you could do this in main/ with
make done_before_sfx2
make sw
The 'done_before_svx' could be implemented similar to the regular svx
target like this:
done_before_svx : done_l10n ... done_libxslt
# nothing to be done for done_before...
done_l10n : ...
touch <some-path>/l10n.done
c) After the discussion with you I am not sure if we still need --from
because the two reasons I know for its existence my not be relevant with
the new approach.
Thx, I think the discussion is important. I am always afraid (still feeling
newbee) to miss something, and at the same time many other say "we are used
to". I think the right thing to do is somewhere in the middle, and we can
only reach it by discussing it.
I am now working for about 13 years on OpenOffice but due to its size
and complexity I know only a small part (mostly Impress UI and
slideshow). Everywhere else I am a newbee, too.
And there is always the possibility to miss something. I think it is
important to prepare as good as possible but at some point you just have
to implement your changes/features and handle the consequences. That
means that there will be errors. But those can be fixed. Part of being
a developer in an open source project is to find this point when
planning stops and development starts and when development is finished
to 'weather the storm' by fixing bugs and handling criticism.
I find discussion where e.g. you say "I need option X because" very
fruitfull, because they are based on actual need, and not just "nice to
have".
c1) With the '--from' option you can tweak the dependency rules at runtime
(a bit). This allows you to exclude projects from being built when you
know that that is not necessary. But from experience I know that can lead
to very subtle errors. Letting the system determine what to built is
usually more reliable.
you are a pro. and I am getting there, but all newcommers are not, and to
attracts them we need to make it easy and fault tolerant !
Yes, absolutely. I am convinced that providing a system that does
everything the pros need is more easy then make the same system
reliable, fault tolerant and easy to use.
Therefore focusing on the last three is a good strategy.
I am sure, I dont need to tell you how to tweek .done file to achieve what
you want :-)
c2) With build.pl a 'build --all' still builds every module on which the
one you are standing in depends on. When those modules have been built
previously, then no compilation takes place. But calling dmake for a
couple of directories for close to 200 modules (when you stand in
instsetoo_native) takes a lot of time (several minutes on Windows), even
when no file has to be compiled. This wasteful way of doing nothing can be
prevented with the --from option.
However, with the new approach and its .done flag files you can determine
which modules need to be built much faster. You don't have to call dmake
on directories that where already built. Hm, but this again, does only
work if your .done files have dependencies on all relevant source files.
That is something that is missing at least from my script.
Hmmm, I had a different idea, let me try that on you. If someone changes a
source file, but does not do a make, its not intented to be public. Meaning
we do not depend on the source files, but the actual run of make. Is that a
wrong idea ?
I am not quite sure that I understand you. Let's look at an (abstract)
example. There is a file code.cxx in module X/. I can build the whole
module by calling 'make X' in main/. The question is, what happens on
'make all'? Currently, build.pl uses a brute force method and goes into
every module, and in case of dmake modules, into every of their
directories, and calls 'make'. This takes a couple of minutes (on
Windows) when no files has been changed. That is why the --from option,
and others like it, where created.
The alternative is to use dependencies to let 'make' figure out what to
build. These dependencies do not yet exist. There are dependencies
between files, even when they are in different modules. So, when I
change a file in sfx2/ and then build sw/ then some (or many) files in
sw/ will be compiled. But there are no dependencies that tell us that
sw/ has to be built. For this we need dependencies of .done files on
the source file. When a source file changes then these dependencies will
trigger the corresponding .done file to be made. This would then
include building the module and touching the .done file.
But that would create a lot of dependencies, at least one for each of
our source files. And they all have to be loaded whenever 'make' is
called. Loading the dependencies might take a few seconds. Determining
the time stamps of all files involved might take longer. But gbuild
when used Office wide would introduce more dependencies. But we still
have to find out if the time it takes is acceptable when spend on every
call of 'make' or if we need some shortcuts for the daily work.
So, reasons for the existence of '--from' are a result of old/slow
computers, slow files systems (still valid on Windows), missing global
dependencies (which we now have for gbuild) and impatient developers.
While this is easy to do with eg Perl I am not sure how to handle this
with just a Makefile. The straightforward approach with handling
<module>.done files does not work. And that is one of the reasons why I
don't think that (GNU) makefiles are a good solution for any problem.
Most
of us are used to program object oriented/imperative. Makefiles require
a
declarative approach. Maybe the use of Perl is not such a bad idea.
Maybe
it would be better to reimplement build.pl with a lot fewer options and
with better readable code.
I agree that makefiles are nowhere near a good solution to many of these
problems, but its like windows, I dont like it, but everybody uses it.
We could easily write a new build.pl, that also took care of the local
makefiles, but our build system would not be in the mainstream, and e.g.
the distros would not like to integrate AOO.
I have over the last years followed research in building systems, and
there
are (sadly enough) nobody that tries a real object oriented aproach. Also
if you look at packages like visual studio, QT, eclipse they all use the
principle of makefiles with declarative approach.
So my simple question is, do we want to approach the main road (makefiles
for unix, visual studio for windows/mac) or do we want to have better but
non standard system.
Good analysis. Maybe I should answer with Faust: "Zwei Seelen wohnen,
ach! in meiner Brust" (two souls alas! are dwelling in my breast). The
pragmatist says to use the make. It is good enough for others, it is good
enough for use.
On the other hand, when I start a new project I usually start with the
question of what are the best tools for the problem at hand. And make does
not seem to be the first or the best answer. Look at our dmake or gbuild
system. Both don't use make in a standard way. gmake even defines its own
language, object oriented and imperative, on top of the GNU make macros.
That is, for me, an act of desperation.
I just wonder did you see "gütterdämmerungen" (I hope I spelled it right,
almost, it is Götterdämmerung
its 2 years since I lived in Vienna so its a bit rusty), I saw it on
danube, 2 nights in a row.
I have not yet have much exposure to Wagner. But seeing this in Vienna
on the Danube must have been an unforgettable spectacle.
I have made experiments with an alternative approach, a domain specific
language somewhat similar to Java. I personally like this approach because
a) it uses the paradigm that I already use when writing C++ code. That
means that I can apply my existing knowledge to the build process and that
I don't have to remember all the tricks and pitfalls of makefiles.
b) as expected it was much easier to handle file dependencies and parallel
processing of build jobs in Java than adding object orientation and
imperative control flow to makefiles.
I like your approach, and would like to help on that, BUT I think we
benefit as a community to stay on the main road.
Good point.
If I had the time and if I would be the one working on it then I would
prefer an non-Makefile approach. But maybe I am just suffering too much
from one of the 'three great virtues of a programmer': hubris.
No you are like me, just not understaning why the world isnt perfect :-)
:-)
You are the one who leads the build project changes, so you have to decide
which approach to use. I am not trying to make your life harder (than
necessary), I am only trying to point out some of the pitfalls that I have
encountered in the past. And to prevent you from removing features that I
use :-)
I do not decide things, I think about solutions, throw them at the
community, and see if the float. If they float I help implement them.
I am a doer, but also a listener. Implementing what I think is right, would
be dead wrong, but I pave the way and make us discuss/think sometimes by
provoking a bit, which I think is good.
A good approach in general, but I don't think that it would be wrong
when you implemented what you think is right, if you do it after a
discussion.
-Andre
rgds
jan I.
Ps. its always refreshing to discuss with you, you often have an
alternative approach, which is not just a dream but doable.
Thanks. That makes two of us.
Have a nice weekend,
same to you !
rgds
jan I.
Andre
-Andre
--prepare
Also one option that is important for our every day work. Use
case:
You make changes in <module> and are not sure if these changes are
compatible/incompatible. To be on the safe side you discard the output
of
all depending modules. To save time you keep the output of all other
modules.
Often used together with '--from' like 'make --prepare --from
svx' to
prepare a build after making changes in svx.
Documentation changed, funny thing is that svx does not clear
correctly
on
my ubuntu build.
--since <module>
A variant of '--from'. The only difference is that <module>
itself
is
not built.
If your proposed approach is similar to what my script produces
then
it
is not too difficult to support --from/--since. I made some
experiments
in
this direction but was to lazy to finish them.
My approach is very similar, but I failed to see how --since is
supported.
And question is if its real important.
--job
--pre_job
--post_job
These are sometimes handy to run a non-standard command for all
modules.
I have added them, they are by the way a good example why we need a
discussion I have never used them.
However maybe the real discussion is "do we want to replace build and
have
a main/Makefile instead?"
- I have not used the rest of the unsupported options and would not
miss
them. Others may have other sets of options that are important to
them.
Some general remarks:
- Why keep one makefile per module? Why not put all the inter-module
dependencies into one file (like my script does)?
Ups, I did not explain that correctly, I propose 1 Makefile
"main/Makefile"
with all inter-module and 1 Makefile "<module>/Makefile" that today just
will call the old makefiles as described in prj/build.lst
- Why not use the oportunity to move (a part of) the build environment
out
of the way to, say, build/ ?
You have guessed my next step.
- How are dependencies between modules handled (just the manual
dependencies from prj/build.lst or also the file dependencies
introduced
by
gmake).
See doc. on --from. Its done with <module>.done files
- How is the output of the individual calls to dmake or GNU make
handled/made accessible. Ie. if there is a build error, how can I look
up
the corresponding build output?
see doc. script make_log
- Are the gmake makefiles included (run in the same process) or is GNU
make started for them it its own process?
For a start they would be called (own process), but its something
where
I
have no strong opinions.
Please (just to be sure), this proposal has nothing to do with the
students
work, its simply because I saw a positive discussion on removing
build.pl
,
and spent a couple of hours looking at it. If there is a preference not
to
remove build.pl I will simply forget it.
rgds
jan I.
Regards,
Andre
It has been roughly tested it, thanks to a clever utility from andre.
As discussed build.pl contains a lot of options, which need to be
considered in a makefile.
My suggestion is on
http://wiki.openoffice.org/******wiki/Build_System_Analysis:**<http://wiki.openoffice.org/****wiki/Build_System_Analysis:**>
**<http://wiki.openoffice.org/****wiki/Build_System_Analysis:**<http://wiki.openoffice.org/**wiki/Build_System_Analysis:**>
build.pl_versus_makefile<http:****//wiki.openoffice.org/wiki/****<http://wiki.openoffice.org/wiki/**>
Build_System_Analysis:build.****pl_versus_makefile<http://**
wiki.openoffice.org/wiki/**Build_System_Analysis:build.**
pl_versus_makefile<http://wiki.openoffice.org/wiki/Build_System_Analysis:build.pl_versus_makefile>
Please feel free to edit/comment on the page. I have reduced to
options
a
lot, and some of them might be in use.
thanks in advance for your comments.
------------------------------******--------------------------**
--**
--**---------
To unsubscribe, e-mail: dev-unsubscribe@openoffice.****a**pache.org<
http://apache.org**>
<dev-unsubscribe@**openoffice.**apache.org<http://openoffice.apache.org>
<dev-unsubscribe@**openoffice.apache.org<dev-unsubscr...@openoffice.apache.org>
For additional commands, e-mail: dev-h...@openoffice.apache.org
------------------------------****----------------------------**
--**---------
To unsubscribe, e-mail:
dev-unsubscribe@openoffice.**a**pache.org<http://apache.org>
<dev-unsubscribe@**openoffice.apache.org<dev-unsubscr...@openoffice.apache.org>
For additional commands, e-mail: dev-h...@openoffice.apache.org
------------------------------**------------------------------**---------
To unsubscribe, e-mail:
dev-unsubscribe@openoffice.**apache.org<dev-unsubscr...@openoffice.apache.org>
For additional commands, e-mail: dev-h...@openoffice.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@openoffice.apache.org
For additional commands, e-mail: dev-h...@openoffice.apache.org
