NAME¶
makepp_repositories -- How to use repositories for variant builds, for
maintaining a central set of sources, and other things
DESCRIPTION¶
A
repository is a directory or directory hierarchy outside of the default
directory that contains files which the makefile needs in the current
directory tree. Makepp can automatically link files from the repository into
the current directory tree if they are needed. Repositories provide similar
functionality to the "VPATH" variable, but (unlike "VPATH"
in other versions of make) you do not have to do anything special to your
makefile to get them to work.
Repositories are specified with the "-R" or "--repository"
command line option or with the "repository" statement in the
makefile. Note that if you have a habit of calling makepp in different
subdirectories of your build tree, it is easy to accidentally reimport a
repository somewhere else. As a safeguard against this, if you use
RootMakeppfile, makepp will refuse to start if it finds one above or
below where it would be imported.
This is somewhat comparable to operating system union filesystems (unionfs...)
The current directory is like the highest level writable layer. All
repositories are like lower read-only layers.
Repositories are useful in several different situations:
- •
- When you want to place your object and executable files in a separate
directory, but the makefile is written to place them in the same directory
as the sources.
- •
- When you want to build the same program two different ways (e.g., with two
different sets of compilation options, or for two different
architectures).
- •
- When you don't have write access to all or part of the source tree.
- •
- When several developers are working on the same project, and there is a
common source repository containing all the sources for the project. Each
developer can modify only the files he needs to change in his local
directory without affecting the other developers, and makepp will
automatically fetch the unmodified files from the source repository.
Makepp's implementation of repositories does not require rewriting of the build
commands at all, unlike (for example) repositories in cons. Makepp puts a
symbolic link into the directory where the command is expecting it. As long as
the command does not refer to absolute directories, the exact same shell
command will work with files from a repository. This means that it works not
only for compilation commands, but any kind of command you can think to put in
your makefile.
Makepp has another kind of mechanism called a
build cache which solves
some of the same sorts of problems as repositories in a different way.
Depending on your problem, a build cache may be more useful than a repository.
See makepp_build_cache for information about build caches and a comparison of
build caches and repositories.
Examples¶
Repositories are best explained by several examples of what you can do.
Different compilation options
Suppose you have a simple program with a makefile that looks something like
this:
CFLAGS = -O2
OBJECTS = a.o b.o c.o
my_program: $(OBJECTS)
cc $(inputs) -o $(output)
%.o: %.c
cc $(CFLAGS) -c $(input) -o $(output)
This makefile places the files "a.o", "b.o",
"c.o", and "my_program" in the same directory as the
source files.
Sometimes you want to place the binary files into a separate directory. For
example, you might build your program on several different architectures, and
you don't want the binary files on one architecture to be replaced with the
binary files on the other. Or you might want to make a temporary change and
recompile without wiping out the previous compilation results. Without
repositories, you would have to modify your makefile to place the objects
elsewhere.
With a repository, however, you don't have to touch your makefile at all.
Consider the following sequence of commands:
% cd my_program_source
% makepp # Builds using the above makefile, and
# object files go into the directory
# my_program_source.
% cd ..
% mkdir binary-debug # Make a clean directory for building the
% cd binary-debug # same program with different options.
% makepp -R ../my_program_source CFLAGS=-g
# Now objects go into binary-debug.
The first makepp command compiles the source files with optimization and puts
the objects into the directory "my_program_source", because that's
what the makefile is supposed to do. Now we want to rebuild the program, but
we want to change the value of "CFLAGS" to compile for debug. We
specify the new value of "CFLAGS" on the command line, and we also
tell makepp that the "my_program_source" directory is a repository
using the "-R" option.
Every time makepp realizes that it needs a file that it doesn't already have in
current directory, it looks in the repository. In this case, it first looks
for the makefile, which doesn't exist in the "binary-debug"
subdirectory. So it creates a symbolic link to it from the makefile in
"my_program_source", and then reads in the makefile. Then it notices
that it needs the file "a.c" in order to build "a.o", and
so it links in "a.c" from the repository. If "a.c"
includes any files contained in "my_program_source", then these will
be automatically linked in as well. Note: Those links are useful for things
like debugging, but if you don't like them, "makeppclean -R" can
remove them.
Running the build command in "binary-debug" won't touch any of the
files in "my_program_source". Thus from the same set of source
files, you now have two different copies of the program, one compiled with
optimization and one compiled for debug. And this happened without touching
the makefile at all.
The advantage of using repositories instead of simply recompiling and
overwriting the original binaries is that now if we fix our bugs and want to
go back to the optimized version, we don't have to recompile everything. Since
the original object files are still around, and most of them are still valid,
we can save a lot of time on recompilation. This does not make a big
difference when only three source files are involved, but for a larger build
that takes minutes or hours to complete, the savings in programmer time and
frustration can be significant.
Rebuilding one file with a minor modification to the compilation commands
Makepp doesn't fetch only source files from the repository. If the object files
in the repository don't need rebuilding, it will use them. For example,
consider a slight modification to the above makefile:
CFLAGS := -O2
A_CFLAGS := -O6 -funroll-loops
OBJECTS := a.o b.o c.o
my_program: $(OBJECTS)
cc $(inputs) -o $(output)
%.o: %.c
cc $(CFLAGS) -c $(input) -o $(output)
a.o: a.c
cc $(A_CFLAGS) -c $(input) -o $(output)
The idea is that "a.o" contains the time-critical code, so it is
compiled with higher optimization than the rest of the objects. Now suppose we
want to test just how different the timing is with different compile options.
A repository can help with this, too:
% cd my_program_source
% makepp # Builds using the above makefile, and
# object files go into the directory
# my_program_source.
% cd ..
% mkdir no-unrolling # Make a clean directory for building the
% cd no-unrolling # same program with different options.
% makepp -R ../my_program_source A_CFLAGS=-O2
% cd ..
% time no-unrolling/my_program # Benchmark the two versions of the program.
% time my_program_source/my_program
Makepp proceeds as before, linking in a copy of the makefile and then examining
the object files. Now only the "a.o" module needs recompiling, since
the options for "b.o" and "c.o" haven't changed. Makepp
notices that it can use "b.o" and "c.o" from the
repository, so it just links those in. However, it will recompile
"a.o" in the "no-unrolling" directory. Once the
compilation is finished, the two different versions of the program can be
benchmarked.
Rebuilding with a minor modification to the source
Now suppose we want to make a change to "a.c" and benchmark the
program before and after the change. Repositories can help again. Consider
this sequence of commands:
% mkdir modified-a
% cp my_program_source/a.c modified-a
% cd modified-a
% emacs a.c # Make some modifications just to this module.
% makepp -R ../my_program_source
Here we have created a new directory that just contains the single source file
we want to modify. Makepp now takes "a.c" from the
"modified-a" subdirectory, but uses the copies of "b" and
"c" from the "my_program_source" directory. Without
changing any of the binary files in "my_program_source", we have
created a separate copy of the program that incorporates our changes to
"a.c". If there are other developers using the sources in
"my_program_source", they will be unaffected by our changes.
Repositories can thus be used as a quick way to build variants of a program,
without adding complicated conditions to the makefile. None of the files in
the original directory are modified; they are used as needed.
Using a directory hierarchy
A repository is actually not just a single directory, it's a whole directory
hierarchy. Suppose you use
/our/library as a repository. Now
/our/library may well contain many subdirectories, e.g.,
/our/library/gui and
/our/library/network. Consider this
command:
% makepp -R /our/library
Any commands in the makefile that refer to files in the directory
./network will actually get files from
/our/library/network, and
similarly for
./gui. Makepp automatically creates any directories that
exist in the repository but not in the current directory.
Linking to any place in the file system
All of the above examples show files from a repository being linked into the
current directory or its subdirectories, but you can actually have makepp link
them into any place in the file system that you have write access to. This is
done by specifying "-R new-location=old-location".
For example, sometimes it's a little tedious to type the following:
mkdir alternate-build
cd alternate-build
makepp -R ..
You can do it all with one command, like this:
makepp -R alternate-build=. -F alternate-build
"-F" or "-makeppfile" changes to that directory before
loading the makefile. You must specify "-R" before "-F".
Note that this example puts the new build tree inside the repository. That
will not work if you use a
RootMakeppfile because makepp safeguards
against nested trees. It's also not a good idea if you use
**, because
if you ever build in the repository it will also find edited and generated
files in this subtree.
Assigning a different location in the file system may be also useful for more
complicated builds, where there are several library subdirectories. For
example, here's a command I have used to build variants of one of my programs:
% makepp -R test-build/seescape=/src/seescape \
-R test-build/HLib=/src/HLib \
-R test-build/H5pp=/src/H5pp \
-R qwt=/src/external_libraries/qwt \
-F test-build/seescape
This command loads in files from four different repositories, and then cds to
the
./test-build/seescape directory and executes the makefile there.
Files contained in the directory tree beginning with
/src/seescape are
linked into
./test-build/seescape. In other words, makepp will
temporarily link the file
/src/seescape/gui/image_canvas.cxx to
./test-build/seescape/gui/image_canvas.cxx when it is needed. This
command will work even if the "test-build" directory doesn't exist
yet; makepp will create it for you. (But you must specify the "-R"
options before the "-F" option on the command line.)
Multiple equivalent repositories
Say your project is maintained by several fairly autonomous groups. You could
have one complete repository with all the sources as they are in production or
at least successfully tested. Every group can have a mostly empty repository
with (part of) the same structure, containing the files group members have
finished developing.
Developers' current directories will have the files they are still working on.
The group repository will be the first one given and the production repository
the last one, so that it furnishes the files not found in the group
repository:
$ makepp -R/path/to/group/repository -R/path/to/production/repository
Since this is probably fairly static for that directory, you may want to put a
file
.makepprc at its root with the following content:
-R/path/to/group/repository -R/path/to/production/repository
Or, presuming that it has a fixed path, you could write into your makefile:
repository /path/to/production/repository
and, because options are seen before makefiles are read, you can then call just
$ makepp -R/path/to/group/repository
Repositories as fixed part of your build system
If you know you always use some repository you can use the
"repository" or "vpath" statements in your makefile.
Caveats with repositories¶
When the links get in the way
For finding your way around your file hierarchy and for allowing the debugger to
find the sources it is useful to have the links used while building. But when
you want to edit a file or resync it with your version control, the links can
get in the way. That is because the system traverses the link and writes to
the file in the repository. Unless it's your personal repository used just for
keeping things apart, that may not be what you want.
As a safeguard against inadvertent overwriting of public files it is suggested
to make the sources in the repository unwritable. It might even not be enough
to remove the write bit, because a version control system which insists on
your locking the files for editing might also do that, but temporarily make
the file writable while resyncing it. If that is the case for you, the
repository should actually belong to a different user.
There are a few tactics to surmount this:
- •
- Keep the sources you edit in a repository, separate from your build tree.
Whenever you put a file, which was previously fetched from another
repository, into this editing repository, makepp will notice and fetch it
from there, provided it is the first repository you specify.
- •
- Remember to delete any file, before you create a copy for writing. If you
follow the safeguard suggestion above, forgetting to do this will give an
error message when writing. To help you, the following function
"delink" will replace one link by a copy of the linked file. The
first variant is for all kinds of Bournish Shells, the second one for csh
(or at least tcsh):
$ delink() { { rm $1 && cat >$1; } <$1; }
% alias delink '( rm \!:1 && cat >\!:1; ) <\!:1'
- •
- If you feel you don't need them, you can delete them all, whenever you
want, e.g. after every makepp run, possibly backgrounded (either short or
long form):
makeppclean --recurse --only-repository-links
mppc -rR
Don't build in a repository during use
A repository is meant to be read-only while it is being used as a repository.
Makepp will
not work properly if you change files in your repository
during the course of a build. Nightly builds may be ok for you, if no one else
uses the repository at that time. Before it starts the build, makepp gets a
list of all the files that exist in the repository, and never updates its
list, except for files it expects to appear.
If you need a repository that's changing as you build, you might want to
consider makepp's build cache mechanism (see makepp_build_cache).
Alternatively, you can use a "poor man's repository": you can put
explicit rules into your makefile to create the soft links, like this:
%.c : $(directory_I_wish_was_a_repository)/%.c
&ln -fs $(input) $(output)
This works only for source files; you can't easily use this to link a file if it
is already built in the repository, but build it here if it's not already
built, since there is only allowed to be one way to build a file.
Use only relative filenames
Repositories work completely transparently
if the makefiles use only
relative filenames. In the above example, it's ok if the makefile in
/src/seescape refers to
../HLib, but the above command will not
work as expected if it refers to
/src/HLib. If you need to use absolute
file names, you can put them into make variables and then override them on the
command line, like this:
% makepp -R test-build/seescape=/src/seescape SEESCAPE=/home/holt/test-build/seescape \
-R test-build/HLib=/src/HLib HLIB=/home/holt/test-build/HLib \
-R test-build/H5pp=/src/H5pp H5pp=/home/holt/test-build/H5pp \
-R qwt=/src/external_libraries/qwt QWT=/home/holt/test-build/qwt \
-F test-build/seescape
The above will work as long as the "HLib" directory is referred to as
"$(HLIB)" in all the makefiles. Note that you have to specify
absolute paths for the directories, because makepp cd's to
"test-build/seescape" before reading the makefile. This leads to
long and complicated make commands; use relative paths when possible.
Makepp must know about all dependencies
Repositories will not work if there are hidden dependencies that makepp doesn't
know about. (In fact, doing a build using repositories, is one way of checking
for forgotten dependencies. But, just for this check, don't combine it with a
build cache, since fetching something there, instead of building it, might
hide a forgotten dependency.) Sometimes these dependencies can be fairly
subtle. For example, the
libtool command will not only create
".lo" and ".la" files as listed on the command line, but
it also may create a subdirectory called ".libs" which contains the
actual object files. To prevent build mistakes, makepp refuses to link in a
".la" file from a repository. Hopefully in the future libtool will
be better supported.
Many hidden dependencies related to compilation are caught by the command line
scanner. If your compiler uses the common Unix compilation flags (e.g.,
"-I", "-D", etc.), then makepp will usually figure out
where all your include files are. You may have to be careful if you have any
homegrown scripts that create files that makepp doesn't know about. For
correct builds, it is vitally important to list
all targets and
dependencies (or determine them automatically by scanning).
Putting absolute filenames into programs
Repositories will also not work if any of the files built contain absolute file
names in them (e.g., if any of your build commands write out an absolute
filename). For example, it turns out that the ".la" files produced
by
libtool have this property. (If you look at the contents of the
".la" file you'll see that the dependency list contains absolute
filenames.) In order to solve this particular problem, makepp will not link
".la" files from a repository; it will insist on rebuilding them.
Avoid linking in unnecessary directories
Repositories can be slow on startup and use a lot of memory if there are a lot
of unnecessary files in the repository. For example, if you use an automatic
HTML documentation generator which makes thousands of ".html" files
from your source code, you may not want to put them in a subdirectory of a
directory that's used as a repository. It's better to put them in a different
directory tree entirely, so the repository mechanism won't load in their
names.
Too Many Files
The disadvantage of repositories is that symbolic links, which the repository
mechanism uses, are individual files (though they use almost no disk space).
This is unlike real links, but those can't cross file system boundaries. In
extreme cases the presence of very many symbolic links can lead to exhaustion
of the number of foreseen files (so called inodes), even though there is
plenty of space left. In this case the sysadmin will need to tune the file
system.
Overriding repository copies¶
If you make any modifications to a file locally, makepp will ordinarily realize
this and recompile the file using the local copy rather than the repository
copy.
If you're using a repository to maintain a central code base, and you have
developers working on local copies which contain only the files they have
modified, one problem that comes up is: what if a developer wants to remove a
file from his local build but the repository still contains it? If the
developer removes the local copy, makepp will happily put in the copy from the
repository, and the build will proceed as if the file existed.
One technique (alas not for user root) for this problem is to make the file that
you want not to include in the build process unreadable, like this:
chmod a-rw file-to-be-excluded
This will prevent makepp from incorporating it from the repository. Makepp also
includes special code so that unreadable files do not match wildcards or
pattern rules.
Similarly, to prevent makepp from incorporating an entire subdirectory, make a
local directory that has the same name but is unwritable. If you want makepp
to ignore the directory entirely, then make it unreadable too. (Read-only
directories are searched but targets in them are usually not built.)
The other way to do this is calling makepp with one or more exclusion options:
mpp -R /path/to/rep --dont-read=/path/to/rep/file-to-be-excluded
Don't use repositories for files which can change!¶
Don't try to use a repository for a file which is part of your build. For
example, you might be tempted to try to use repositories to put all of your
public .h files in the same directory, like this:
# top level makefile
repository include=module1/include
repository include=module2/include
repository include=module3/include
repository include=module4/include
This is probably not a good idea if
any of the
.h files are
themselves outputs of a program (e.g., yacc or some other program that spits
out C source code), because makepp assumes that files in repositories never
change. If the build needs
include/xyz.h, and
module2/include/xyz.h actually needs to be produced by some program,
makepp will not know to run the program. It's better to use a technique like
this to put all of your
.h files into a common include directory:
# module1/Makeppfile
../include/%.h : include/%.h
&cp $(input) $(output)
# You could also (more efficiently but problematic on Windows) do the following:
# &ln -r $(input) $(output)
Makepp might still try to build files that happen to be in a repository if
something asks for them directly, but it won't build them
on behalf of
the local directory. The result of this can be quite confusing, because it can
lead to a repository symbolic link being used while its repository target is
out-of-date, but that target might get updated later in the build. You can
prevent this from happening either by making sure that the repository is
referred to
only through the repository path, or by making sure that
there is also a local rule for all the generated repository files.
Another way to avoid recompiling identical files in different directories is to
use a build cache (see makepp_build_cache for details). A build cache does not
have the restriction that the file may not change.