Cons - Cons: A Software Construction System
The original document was automatically derived from the cons/cons.html by html2pod, thanks to Ulrich Pfiefer. Later revisions were created from the original.
by Bob Sidebotham rns@fore.com
A guide and reference for version 1.7
Copyright (c) 1996-1999 FORE Systems, Inc. All rights reserved.
Permission to use, copy, modify and distribute this software and its documentation for any purpose and without fee is hereby granted, provided that the above copyright notice appear in all copies and that both that copyright notice and this permission notice appear in supporting documentation, and that the name of FORE Systems, Inc. (``FORE Systems'') not be used in advertising or publicity pertaining to distribution of the software without specific, written prior permission.
FORE SYSTEMS DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE, INCLUDING ANY WARRANTIES REGARDING INTELLECTUAL PROPERTY RIGHTS AND ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. IN NO EVENT SHALL FORE SYSTEMS BE LIABLE FOR ANY SPECIAL, INDIRECT OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
Cons is a system for constructing, primarily, software, but is quite different from previous software construction systems. Cons was designed from the ground up to deal easily with the construction of software spread over multiple source directories. Cons makes it easy to create build scripts that are simple, understandable and maintainable. Cons ensures that complex software is easily and accurately reproducible.
Cons uses a number of techniques to accomplish all of this. Construction scripts are just Perl scripts, making them both easy to comprehend and very flexible. Global scoping of variables is replaced with an import/export mechanism for sharing information between scripts, significantly improving the readability and maintainability of each script. Construction environments are introduced: these are Perl objects that capture the information required for controlling the build process. Multiple environments are used when different semantics are required for generating products in the build tree. Cons implements automatic dependency analysis and uses this to globally sequence the entire build. Variant builds are easily produced from a single source tree. Intelligent build subsetting is possible, when working on localized changes. Overrides can be setup to easily override build instructions without modifying any scripts. MD5 cryptographic signatures are associated with derived files, and are used to accurately determine whether a given file needs to be rebuilt.
While offering all of the above, and more, Cons remains simple and easy to use. This will, hopefully, become clear as you read the remainder of this document.
Cons is a make replacement. In the following paragraphs, we look at a few of the undesirable characteristics of make--and typical build environments based on make--that motivated the development of Cons.
Traditional make-based systems of any size tend to become quite complex. The original make utility and its derivatives have contributed to this tendency in a number of ways. Make is not good at dealing with systems that are spread over multiple directories. Various work-arounds are used to overcome this difficulty; the usual choice is for make to invoke itself recursively for each sub-directory of a build. This leads to complicated code, in which it is often unclear how a variable is set, or what effect the setting of a variable will have on the build as a whole. The make scripting language has gradually been extended to provide more possibilities, but these have largely served to clutter an already over extended language. Often, builds are done in multiple passes in order to provide appropriate products from one directory to another directory. This represents a further increase in build complexity.
The bane of all makes has always been the correct handling of dependencies. Most often, an attempt is made to do a reasonable job of dependencies within a single directory, but no serious attempt is made to do the job between directories. Even when dependencies are working correctly, make's reliance on a simple time stamp comparison to determine whether a file is out of date with respect to its dependents is not, in general, adequate for determining when a file should be rederived. If an external library, for example, is rebuilt and then ``snapped'' into place, the timestamps on its newly created files may well be earlier than the last local build, since it was built before it became visible.
Make provides only limited facilities for handling variant builds. With the proliferation of hardware platforms and the need for debuggable vs. optimized code, the ability to easily create these variants is essential. More importantly, if variants are created, it is important to either be able to separate the variants or to be able to reproduce the original or variant at will. With make it is very difficult to separate the builds into multiple build directories, separate from the source. And if this technique isn't used, it's also virtually impossible to guarantee at any given time which variant is present in the tree, without resorting to a complete rebuild.
Make provides only limited support for building software from code that exists in a central repository directory structure. The VPATH feature of GNU make (and some other make implementations) is intended to provide this, but doesn't work as expected: it changes the path of target file to the VPATH name too early in its analysis, and therefore searches for all dependencies in the VPATH directory. To ensure correct development builds, it is important to be able to create a file in a local build directory and have any files in a code repository (a VPATH directory, in make terms) that depend on the local file get rebuilt properly. This isn't possible with VPATH, without coding a lot of complex repository knowledge directly into the makefiles.
A few of the difficulties with make have been cited above. In this and subsequent sections, we shall introduce Cons and show how these issues are addressed.
Cons is Perl-based. That is, Cons scripts--Conscript and Construct
files, the equivalent to makefiles--are all written in Perl. This provides an immediate benefit: the language
for writing scripts is a familiar one. Even if you don't happen to be a
Perl programmer, it helps to know that Perl is basically just a simple
declarative language, with a well-defined flow of control, and familiar
semantics. It has variables that behave basically the way you would expect
them to, subroutines, flow-of-control, and so on. There is no special
syntax introduced for Cons. The use of Perl as a scripting language
simplifies the task of expressing the appropriate solution to the often
complex requirements of a build.
To ground the following discussion, here's how you could build the Hello, World! C application with Cons:
$env = new cons(); Program $env 'hello', 'hello.c';
If you install this script in a directory, naming the script Construct,
and create the hello.c source file in the same directory, then you can type cons hello to build the application:
% cons hello cc -c hello.c -o hello.o cc -o hello hello.o
A key simplification of Cons is the idea of a construction environment. A construction environment is an object characterized by a set of key/value pairs and a set of methods. In order to tell Cons how to build something, you invoke the appropriate method via an appropriate construction environment. Consider the following example:
$env = new cons(
CC => 'gcc',
LIBS => 'libworld.a'
);
Program $env 'hello', 'hello.c';
In this case, rather than using the default construction environment, as
is, we have overridden the value of CC so that the GNU C Compiler equivalent is used, instead. Since this version
of Hello, World! requires a library,
libworld.a, we have specified that any program linked in this environment should be
linked with that library. If the library exists already, well and good, but
if not, then we'll also have to include the statement:
Library $env 'libworld', 'world.c';
Now if you type cons hello, the library will be built before the program is linked, and, of course, gcc will be used to compile both modules:
% cons hello gcc -c hello.c -o hello.o gcc -c world.c -o world.o ar r libworld.a world.o ar: creating libworld.a ranlib libworld.a gcc -o hello hello.o libworld.a
With Cons, dependencies are handled automatically. Continuing the previous
example, note that when we modify world.c, world.o is recompiled,
libworld.a recreated, and hello relinked:
% touch world.c % cons hello gcc -c world.c -o world.o ar r libworld.a world.o ar: creating libworld.a ranlib libworld.a gcc -o hello hello.o libworld.a
This is a relatively simple example: Cons ``knows'' world.o depends upon
world.c, because the dependency is explicitly set up by the Library
method. It also knows that libworld.a depends upon world.o and that
hello depends upon libworld.a, all for similar reasons.
Now it turns out that hello.c also includes the interface definition file, world.h:
% touch world.h % cons hello gcc -c hello.c -o hello.o gcc -o hello hello.o libworld.a
How does Cons know that hello.c includes world.h, and that hello.o
must therefore be recompiled? For now, suffice it to say that when
considering whether or not hello.o is up-to-date, Cons invokes a scanner for its dependency, hello.c. This scanner enumerates the files included by hello.c to come up with a list of further dependencies, beyond those made explicit
by the Cons script. This process is recursive: any files included by
included files will also be scanned.
Isn't this expensive? The answer is--it depends. If you do a full build of a large system, the scanning time is insignificant. If you do a rebuild of a large system, then Cons will spend a fair amount of time thinking about it before it decides that nothing has to be done (although not necessarily more time than make!). The good news is that Cons makes it very easy to intelligently subset your build, when you are working on localized changes.
Because Cons does full and accurate dependency analysis, and does this globally, for the entire build, Cons is able to use this information to take full control of the sequencing of the build. This sequencing is evident in the above examples, and is equivalent to what you would expect for make, given a full set of dependencies. With Cons, this extends trivially to larger, multi-directory builds. As a result, all of the complexity involved in making sure that a build is organized correctly--including multi-pass hierarchical builds--is eliminated. We'll discuss this further in the next sections.
A larger build, in Cons, is organized by creating a hierarchy of build
scripts. At the top of the tree is a script called Construct. The rest of the scripts, by convention, are each called Conscript. These scripts are connected together, very simply, by the Build, Export, and
Import commands.
The Build command takes a list of Conscript file names, and arranges for them to be included in the build. For example:
Build qw(
drivers/display/Conscript
drivers/mouse/Conscript
parser/Conscript
utilities/Conscript
);
This is a simple two-level hierarchy of build scripts: all the subsidiary
Conscript files are mentioned in the top-level Construct file. Notice that not all directories in the tree necessarily have build
scripts associated with them.
This could also be written as a multi-level script. For example, the
Construct file might contain this command:
Build qw(
parser/Conscript
drivers/Conscript
utilities/Conscript
);
and the Conscript file in the drivers directory might contain this:
Build qw(
display/Conscript
mouse/Conscript
);
Experience has shown that the former model is a little easier to understand, since the whole construction tree is laid out in front of you, at the top-level. Hybrid schemes are also possible. A separately maintained component that needs to be incorporated into a build tree, for example, might hook into the build tree in one place, but define its own construction hierarchy.
You may have noticed that the file names specified to the Build command are relative to the location of the script it is invoked from. This is generally true for other filename arguments to other commands, too, although we might as well mention here that if you begin a file name with a hash mark, ``#'', then that file is interpreted relative to the top-level directory (where the Construct file resides). And, not surprisingly, if you begin it with ``/'', then it is considered to be an absolute pathname. This is true even on systems which use a back slash rather than a forward slash to name absolute paths.
Each Conscript file, and also the top-level Construct file, begins life in a separate Perl package. Except for the Construct file, which gets some of the command line arguments, the symbol table for
each script is empty. All of the variables that are set, therefore, are set
by the script itself--not by some external script. Variables can be
explicitly imported
by a script from its parent script. To import a variable, it must have been
exported by the parent and initialized (otherwise an error will occur). It is
therefore possible to determine, from looking at a single script, exactly
where each variable in that script is set.
The Export command is used as in the following example:
$ENV = new cons(); $INCLUDE = "#export/include"; $LIB = "#export/lib"; Export qw( ENV INCLUDE LIB ); Build qw( util/Conscript );
The values of the simple variables mentioned in the Export list will be squirreled away by any subsequent Build commands. The Export command will only export Perl scalar
variables, that is, variables whose name begins with $. Other variables, objects, etc. can be exported by reference--but all
scripts will refer to the same object, and this object should be considered
to be read-only by the subsidiary scripts and by the original exporting
script. It's acceptable, however, to assign a new value to the exported
scalar variable--that won't change the underlying variable referenced. This
sequence, for example, is OK:
$ENV = new cons(); Export qw( ENV INCLUDE LIB ); Build qw( util/Conscript ); $ENV = new cons(CFLAGS => '-O'); Build qw( other/Conscript );
It doesn't matter whether the variable is set before or after the Export
command. The important thing is the value of the variable at the time the
Build command is executed. This is what gets squirreled away. Any subsequent Export commands, by the way, invalidate the first: you must mention all the
variables you wish to export on each Export command.
Variables exported by the Export command can be imported into subsidiary scripts by the Import command. The subsidiary script always imports variables directly from the
superior script. Consider this example:
Import qw( ENV INCLUDE );
This is only legal if the parent script exported both $ENV and
$INCLUDE. It also must have given each of these variables values. It is OK for the
subsidiary script to only import a subset of the exported variables (in
this example, $LIB, which was exported by the previous example, is not imported).
All the imported variables are automatically re-exported, so the sequence:
Import qw ( ENV INCLUDE ); Build qw ( beneath-me/Conscript );
will supply both $ENV and $INCLUDE to the subsidiary file. If only
$ENV is to be exported, then the following will suffice:
Import qw ( ENV INCLUDE ); Export qw ( ENV ); Build qw ( beneath-me/Conscript );
Needless to say, the variables may be modified locally before invoking Build on the subsidiary script.
The only constraint on the ordering of build scripts is that superior
scripts are evaluated before their inferior scripts. The top-level
Construct file, for instance, is evaluated first, followed by any inferior scripts.
This is all you really need to know about the evaluation order, since order
is generally irrelevant. Consider the following Build
command:
Build qw(
drivers/display/Conscript
drivers/mouse/Conscript
parser/Conscript
utilities/Conscript
);
We've chosen to put the script names in alphabetical order, simply because that's the most convenient for maintenance purposes. Changing the order will make no difference to the build.
In any complex software system, a method for sharing build products needs to be established. We propose a simple set of conventions which are trivial to implement with Cons, but very effective.
The basic rule is to require that all build products which need to be
shared between directories are shared via an intermediate directory. We
have typically called this export, and, in a C environment, provided conventional sub-directories of this
directory, such as include, lib,
bin, etc.
These directories are defined by the top-level Construct file. A simple
Construct file for a Hello, World! application, organized using multiple directories, might look like this:
# Construct file for Hello, World!
# Where to put all our shared products. $EXPORT = '#export';
Export qw( CONS INCLUDE LIB BIN );
# Standard directories for sharing products. $INCLUDE = "$EXPORT/include"; $LIB = "$EXPORT/lib"; $BIN = "$EXPORT/bin";
# A standard construction environment.
$CONS = new cons (
CPPPATH => $INCLUDE, # Include path for C Compilations
LIBPATH => $LIB, # Library path for linking programs
LIBS => '-lworld', # List of standard libraries
);
Build qw(
hello/Conscript
world/Conscript
);
The world directory's Conscript file looks like this:
# Conscript file for directory world Import qw( CONS INCLUDE LIB );
# Install the products of this directory Install $CONS $LIB, 'libworld.a'; Install $CONS $INCLUDE, 'world.h';
# Internal products Library $CONS 'libworld.a', 'world.c';
and the hello directory's Conscript file looks like this:
# Conscript file for directory hello Import qw( CONS BIN );
# Exported products Install $CONS $BIN, 'hello';
# Internal products Program $CONS 'hello', 'hello.c';
To construct a Hello, World! program with this directory structure, go to the top-level directory, and
invoke cons with the appropriate arguments. In the following example, we tell Cons to
build the directory
export. To build a directory, Cons recursively builds all known products within
that directory (only if they need rebuilding, of course). If any of those
products depend upon other products in other directories, then those will
be built, too.
% cons export Install world/world.h as export/include/world.h cc -Iexport/include -c hello/hello.c -o hello/hello.o cc -Iexport/include -c world/world.c -o world/world.o ar r world/libworld.a world/world.o ar: creating world/libworld.a ranlib world/libworld.a Install world/libworld.a as export/lib/libworld.a cc -o hello/hello hello/hello.o -Lexport/lib -lworld Install hello/hello as export/bin/hello
You'll note that the two Conscript files are very clean and to-the-point. They simply specify products of the
directory and how to build those products. The build instructions are
minimal: they specify which construction environment to use, the name of
the product, and the name of the inputs. Note also that the scripts are
location-independent: if you wish to reorganize your source tree, you are
free to do so: you only have to change the Construct file (in this example), to specify the new locations of the Conscript files. The use of an export tree makes this goal easy.
Note, too, how Cons takes care of little details for you. All the export
directories, for example, were made automatically. And the installed files
were really hard-linked into the respective export directories, to save
space and time. This attention to detail saves considerable work, and makes
it even easier to produce simple, maintainable scripts.
It's often desirable to keep any derived files from the build completely separate from the source files. This makes it much easier to keep track of just what is a source file, and also makes it simpler to handle variant builds, especially if you want the variant builds to co-exist.
Cons provides a simple mechanism that handles all of these requirements.
The
Link command is invoked as in this example:
Link 'build' => 'src';
The specified directories are ``linked'' to the specified source directory.
Let's suppose that you setup a source directory, src, with the sub-directories world and hello below it, as in the previous example. You could then substitute for the
original build lines the following:
Build qw(
build/world/Conscript
build/hello/Conscript
);
Notice that you treat the Conscript file as if it existed in the build directory. Now if you type the same
command as before, you will get the following results:
% cons export Install build/world/world.h as export/include/world.h cc -Iexport/include -c build/hello/hello.c -o build/hello/hello.o cc -Iexport/include -c build/world/world.c -o build/world/world.o ar r build/world/libworld.a build/world/world.o ar: creating build/world/libworld.a ranlib build/world/libworld.a Install build/world/libworld.a as export/lib/libworld.a cc -o build/hello/hello build/hello/hello.o -Lexport/lib -lworld Install build/hello/hello as export/bin/hello
Again, Cons has taken care of the details for you. In particular, you will
notice that all the builds are done using source files and object files
from the build directory. For example, build/world/world.o is compiled from
build/world/world.c, and export/include/world.h is installed from
build/world/world.h. This is accomplished on most systems by the simple expedient of ``hard''
linking the required files from each source directory into the appropriate
build directory.
The links are maintained correctly by Cons, no matter what you do to the source directory. If you modify a source file, your editor may do this ``in place'' or it may rename it first and create a new file. In the latter case, any hard link will be lost. Cons will detect this condition the next time the source file is needed, and will relink it appropriately.
You'll also notice, by the way, that no changes were required to the underlying Conscript files. And we can go further, as we shall see in the next section.
Variant builds require just another simple extension. Let's take as an
example a requirement to allow builds for both the baNaNa and peAcH
operating systems. In this case, we are using a distributed file system,
such as NFS to access the particular system, and only one or the other of
the systems has to be compiled for any given invocation of cons. Here's one way we could set up the Construct file for our Hello, World!
application:
# Construct file for Hello, World!
die qq(OS must be specified) unless $OS = $ARG{OS};
die qq(OS must be "peach" or "banana")
if $OS ne "peach" && $OS ne "banana";
# Where to put all our shared products. $EXPORT = "#export/$OS";
Export qw( CONS INCLUDE LIB BIN );
# Standard directories for sharing products. $INCLUDE = "$EXPORT/include"; $LIB = "$EXPORT/lib"; $BIN = "$EXPORT/bin";
# A standard construction environment.
$CONS = new cons (
CPPPATH => $INCLUDE, # Include path for C Compilations
LIBPATH => $LIB, # Library path for linking programs
LIBS => '-lworld', # List of standard libraries
);
# $BUILD is where we will derive everything. $BUILD = "#build/$OS";
# Tell cons where the source files for $BUILD are. Link $BUILD => 'src';
Build (
"$BUILD/hello/Conscript",
"$BUILD/world/Conscript",
);
Now if we login to a peAcH system, we can build our Hello, World! application for that platform:
% cons export OS=peach Install build/peach/world/world.h as export/peach/include/world.h cc -Iexport/peach/include -c build/peach/hello/hello.c -o build/peach/hello/hello.o cc -Iexport/peach/include -c build/peach/world/world.c -o build/peach/world/world.o ar r build/peach/world/libworld.a build/peach/world/world.o ar: creating build/peach/world/libworld.a ranlib build/peach/world/libworld.a Install build/peach/world/libworld.a as export/peach/lib/libworld.a cc -o build/peach/hello/hello build/peach/hello/hello.o -Lexport/peach/lib -lworld Install build/peach/hello/hello as export/peach/bin/hello
Other variations of this model are possible. For example, you might decide
that you want to separate out your include files into platform dependent
and platform independent files. In this case, you'd have to define an
alternative to $INCLUDE for platform-dependent files. Most Conscript
files, generating purely platform-independent include files, would not have
to change.
You might also want to be able to compile your whole system with debugging
or profiling, for example, enabled. You could do this with appropriate
command line options, such as DEBUG=on. This would then be translated into the appropriate platform-specific
requirements to enable debugging (this might include turning off
optimization, for example). You could optionally vary the name space for
these different types of systems, but, as we'll see in the next section,
it's not essential to do this, since Cons is pretty smart about rebuilding things when you
change options.
Whenever Cons creates a derived file, it stores a signature for that file. The signature is stored in a separate file, one per
directory. After the previous example was compiled, the .consign file in the
build/peach/world directory looked like this:
world.o:834179303 23844c0b102ecdc0b4548d1cd1cbd8c6 libworld.a:834179304 9bf6587fa06ec49d864811a105222c00
The first number is a timestamp--for a UNIX systems, this is typically the
number of seconds since January 1st, 1970. The second value is an MD5
checksum. The Message Digest Algorithm is an algorithm that, given an input string, computes a strong
cryptographic signature for that string. The MD5 checksum stored in the .consign file is, in effect, a digest of all the dependency information for the
specified file. So, for example, for the file world.o, this includes at least the file world.c, and also any header files that Cons knows about that are included,
directly or indirectly by world.c. Not only that, but the actual command line that was used to generate world.o is also fed into the computation of the signature. Similarly, libworld.a gets a signature which ``includes'' all the signatures of its constituents
(and hence, transitively, the signatures of their constituents), as well as the command line that created the file.
The signature of a non-derived file is computed, by default, by taking the
current modification time of the file and the file's entry name (unless
there happens to be a current .consign entry for that file, in which case that signature is used).
Notice that there is no need for a derived file to depend upon any
particular Construct or Conscript file--if changes to these files affect the file in question, then this will
be automatically reflected in its signature, since relevant parts of the
command line are included in the signature. Unrelated changes will have no
effect.
When Cons considers whether to derive a particular file, then, it first
computes the expected signature of the file. It then compares the file's
last modification time with the time recorded in the .consign entry, if one exists. If these times match, then the signature stored in
the
.consign file is considered to be accurate. If the file's previous signature does
not match the new, expected signature, then the file must be rederived.
Notice that a file will be rederived whenever anything about a dependent file changes. In particular, notice that any change to the modification time of a dependent (forward or backwards in time) will force recompilation of the derived file.
The use of these signatures is an extremely simple, efficient, and effective method of improving--dramatically--the reproducibility of a system.
We'll demonstrate this with a simple example:
# Simple "Hello, World!" Construct file
$CFLAGS = '-g' if $ARG{DEBUG} eq 'on';
$CONS = new cons(CFLAGS => $CFLAGS);
Program $CONS 'hello', 'hello.c';
Notice how Cons recompiles at the appropriate times:
% cons hello cc -c hello.c -o hello.o cc -o hello hello.o % cons hello cons: "hello" is up-to-date. % cons DEBUG=on hello cc -g -c hello.c -o hello.o cc -o hello hello.o % cons DEBUG=on hello cons: "hello" is up-to-date. % cons hello cc -c hello.c -o hello.o cc -o hello hello.o
Many software development organizations will have one or more central repository directory trees containing the current source code for one or more projects, as well as the derived object files, libraries, and executables. In order to reduce unnecessary recompilation, it is useful to use files from the repository to build development software--assuming, of course, that no newer dependency file exists in the local build tree.
Cons provides a mechanism to specify a list of code repositories that will be searched, in-order, for source files and derived files not found in the local build directory tree.
The following lines in a Construct file will instruct Cons to look first under the /usr/experiment/repository directory and then under the
/usr/product/repository directory:
Repository qw (
/usr/experiment/repository
/usr/product/repository
);
The repository directories specified may contain source files, derived files (objects, libraries and executables), or both. If there is no local file (source or derived) under the directory in which Cons is executed, then the first copy of a same-named file found under a repository directory will be used to build any local derived files.
Cons maintains one global list of repositories directories. Cons will eliminate the current directory, and any non-existent directories, from the list.
Cons will also search for Construct and Conscript files in the repository tree or trees. This leads to a chicken-and-egg
situation, though: how do you look in a repository tree for a Construct file if the
Construct file tells you where the repository is? To get around this, repositories
may be specified via -R options on the command line:
% cons -R /usr/experiment/repository -R /usr/product/repository .
Any repository directories specified in the Construct or Conscript
files will be appended to the repository directories specified by
command-line -R options.
If the source code (include the Conscript file) for the library version of the Hello, World! C application is in a repository (with no derived files), Cons will use the
repository source files to create the local object files and executable
file:
% cons -R /usr/src_only/repository hello gcc -c /usr/src_only/repository/hello.c -o hello.o gcc -c /usr/src_only/repository/world.c -o world.o ar r libworld.a world.o ar: creating libworld.a ranlib libworld.a gcc -o hello hello.o libworld.a
Creating a local source file will cause Cons to rebuild the appropriate derived file or files:
% touch world.c % cons -R /usr/src_only/repository hello gcc -c world.c -o world.o ar r libworld.a world.o ar: creating libworld.a ranlib libworld.a gcc -o hello hello.o libworld.a
And removing the local source file will cause Cons to revert back to building the derived files from the repository source:
% touch -R /usr/src_only/repository hello world.c % cons hello gcc -c /usr/src_only/world.c/repository -o world.o ar r libworld.a world.o ar: creating libworld.a ranlib libworld.a gcc -o hello hello.o libworld.a
If a repository tree contains derived files (usually object files,
libraries, or executables), Cons will perform its normal signature
calculation to decide whether the repository file is up-to-date or a
derived file must be built locally. This means that, in order to ensure
correct signature calculation, a repository tree must also contain the .consign
files that were created by Cons when generating the derived files.
This would usually be accomplished by building the software in the repository (or, alternatively, in a build directory, and then copying the result to the repository):
% cd /usr/all/repository % cons hello gcc -c hello.c -o hello.o gcc -c world.c -o world.o ar r libworld.a world.o ar: creating libworld.a ranlib libworld.a gcc -o hello hello.o libworld.a
(This is safe even if the Construct file lists the /usr/all/repository
directory in a Repository command because Cons will remove the current directory from the repository
list.)
Now if we want to build a copy of the application with our own hello.c
file, we only need to create the one necessary source file, and use the
-R option to have Cons use other files from the repository:
% mkdir $HOME/build1 % cd $HOME/build1 % touch hello.c % cons -R /usr/all/repository hello gcc -c hello.c -o hello.o gcc -o hello hello.o /usr/all/repository/libworld.a
Notice that Cons has not bothered to recreate a local libworld.a library (or recompile the world.o module), but instead uses the already-compiled version from the repository.
Because the MD5 signatures that Cons puts in the .consign file contain timestamps for the derived files, the signature timestamps
must match the file timestamps for a signature to be considered valid.
Some software systems may alter the timestamps on repository files (by copying them, e.g.), in which case Cons will, by default, assume the repository signatures are invalid and rebuild files unnecessarily. This behavior may be altered by specifying:
Repository_Sig_Times_OK 0;
This tells Cons to ignore timestamps when deciding whether a signature is
valid. (Note that avoiding this sanity check means there must be proper
control over the repository tree to ensure that the derived files cannot be
modified without updating the .consign signature.)
If the repository tree contains the complete results of a build, and we try to build from the repository without any files in our local tree, something moderately surprising happens:
% mkdir $HOME/build2 % cd $HOME/build2 % cons -R /usr/all/repository hello cons: "hello" is up-to-date.
Why does Cons say that the hello program is up-to-date when there is no
hello program in the local build directory? Because the repository (not the local
directory) contains the up-to-date hello program, and Cons correctly determines that nothing needs to be done to
rebuild this up-to-date copy of the file.
There are, however, many times in which it is appropriate to ensure that a
local copy of a file always exists. A packaging or testing script, for
example, may assume that certain generated files exist locally Instead of
making these subsidiary scripts aware of the repository directory, the
Local command may be added to a Construct or Conscript file to specify that a certain file or files must appear in the local build
directory:
Local qw(
hello
);
Then, if we re-run the same command, Cons will make a local copy of the program from the repository copy (telling you that it is doing so):
% cons -R /usr/all/repository hello Local copy of hello from /usr/all/repository/hello cons: "hello" is up-to-date.
Notice that, because the act of making the local copy is not considered a
``build'' of the hello file, Cons still reports that it is up-to-date.
Creating local copies is most useful for files that are being installed
into an intermediate directory (for sharing with other directories) via the
Install command. Accompanying the Install command for a file with a companion Local command is so common that Cons provides a
Install_Local command as a convenient way to do both:
Install_Local $env, '#export', 'hello';
is exactly equivalent to:
Install $env '#export', 'hello'; Local '#export/hello';
Both the Local and Install_Local commands update the local .consign
file with the appropriate file signatures, so that future builds are
performed correctly.
Due to its built-in scanning, Cons will search the specified repository
trees for included .h files. Unless the compiler also knows about the repository trees, though,
it will be unable to find .h files that only exist in a repository. If, for example, the hello.c file includes the
hello.h file in its current directory:
% cons -R /usr/all/repository hello gcc -c /usr/all/repository/hello.c -o hello.o /usr/all/repository/hello.c:1: hello.h: No such file or directory
Solving this problem forces some requirements onto the way construction
environments are defined and onto the way the C #include preprocessor directive is used to include files.
In order to inform the compiler about the repository trees, Cons will add
appropriate -I flags to the compilation commands. This means that the
CPPPATH variable in the construct environment must explicitly specify all
subdirectories which are to be searched for included files, including the
current directory. Consequently, we can fix the above example by changing
the environment creation in the Construct file as follows:
$env = new cons(
CC => 'gcc',
CPPPATH => '.',
LIBS => 'libworld.a',
);
Due to the definition of the CPPPATH variable, this yields, when we re-execute the command:
% cons -R /usr/all/repository hello gcc -c -I. -I/usr/all/repository /usr/all/repository/hello.c -o hello.o gcc -o hello hello.o /usr/all/repository/libworld.a
The order of the -I flags replicates, for the C preprocessor, the same repository-directory
search path that Cons uses for its own dependency analysis. If there are
multiple repositories and multiple CPPPATH
directories, Cons will append the repository directories to the beginning
of each CPPPATH directory, rapidly multiplying the number of -I flags. As an extreme example, a Construct file containing:
Repository qw(
/u1
/u2
);
$env = new cons(
CPPPATH => 'a:b:c',
);
Would yield a compilation command of:
cc -Ia -I/u1/a -I/u2/a -Ib -I/u1/b -I/u2/b -Ic -I/u1/c -I/u2/c -c hello.c -o hello.o
Because Cons relies on the compiler's -I flags to communicate the order in which repository directories must be
searched, Cons' handling of repository directories is fundamentally
incompatible with using double-quotes on the #include directives in your C source code:
#include "file.h" /* DON'T USE DOUBLE-QUOTES LIKE THIS */
This is because most C preprocessors, when faced with such a directive,
will always first search the directory containing the source file. This
undermines the elaborate -I options that Cons constructs to make the preprocessor conform to its
preferred search path.
Consequently, when using repository trees in Cons, always use angle-brackets for included files:
#include <file.h> /* USE ANGLE-BRACKETS INSTEAD */
Cons provides a Repository_List command to return a list of all repository directories in their current search order. This can be used for debugging, or to do more complex Perl stuff:
@list = Repository_List;
print join(' ', @list), "\n";
Cons' handling of repository trees interacts correctly with other Cons features--which is to say, it generally does what you would expect.
Most notably, repository trees interact correctly, and rather powerfully,
with the 'Link' command. A repository tree may contain one or more
subdirectories for version builds established via Link to a source subdirectory. Cons will search for derived files in the
appropriate build subdirectories under the repository tree.
Cons provides two methods for reducing the size of given build. The first is by specifying targets on the command line, and the second is a method for pruning the build tree. We'll consider target specification first.
Like make, Cons allows the specification of ``targets'' on the command line. Cons targets may be either files or directories. When a directory is specified, this is simply a short-hand notation for every derivable product--that Cons knows about--in the specified directory and below. For example:
cons build/hello/hello.o
means buildhello.o and everything that hello.o might need. This is from a previous version of the Hello, World! program in which hello.o
depended upon export/include/world.h. If that file is not up-to-date (because someone modified src/world/world.h), then it will be rebuilt, even though it is in a directory remote from build/hello.
In this example:
cons build
Everything in the build directory is built, if necessary. Again, this may cause more files to be
built. In particular, both export/include/world.h
and export/lib/libworld.a are required by the build/hello directory, and so they will be built if they are out-of-date.
If we do, instead:
cons export
then only the files that should be installed in the export directory will
be rebuilt, if necessary, and then installed there. Note that cons build
might build files that cons export doesn't build, and vice-versa.
With Cons, make-style ``special'' targets are not required. The simplest
analog with Cons is to use special export directories, instead. Let's suppose, for example, that you have a whole
series of unit tests that are associated with your code. The tests live in
the source directory near the code. Normally, however, you don't want to
build these tests. One solution is to provide all the build instructions
for creating the tests, and then to install the tests into a separate part
of the tree. If we install the tests in a top-level directory called tests, then:
cons tests
will build all the tests.
cons export
will build the production version of the system (but not the tests), and:
cons build
should probably be avoided (since it will compile tests unecessarily).
If you want to build just a single test, then you could explicitly name the
test (in either the tests directory or the build directory). You could also aggregate the tests into a convenient hierarchy
within the tests directory. This hierarchy need not necessarily match the
source hierarchy, in much the same manner that the include hierarchy
probably doesn't match the source hierarchy (the include hierarchy is
unlikely to be more than two levels deep, for C programs).
If you want to build absolutely everything in the tree (subject to whatever options you select), you can use:
cons .
This is not particularly efficient, since it will redundantly walk all the trees, including the source tree. The source tree, of course, may have buildable objects in it--nothing stops you from doing this, even if you normally build in a separate build tree.
In conjunction with target selection, build pruning can be used to reduce the scope of the build. In the previous peAcH and
baNaNa example, we have already seen how script-driven build pruning can be
used to make only half of the potential build available for any given
invocation of cons. Cons also provides, as a convenience, a command line convention that
allows you to specify which Conscript files actually get ``built''--that is, incorporated into the build tree.
For example:
cons build +world
The + argument introduces a Perl regular expression. This must, of course, be
quoted at the shell level if there are any shell meta-characters within the
expression. The expression is matched against each Conscript
file which has been mentioned in a Build statement, and only those scripts with matching names are actually
incorporated into the build tree. Multiple such arguments are allowed, in
which case a match against any of them is sufficient to cause a script to
be included.
In the example, above, the hello program will not be built, since Cons will have no knowledge of the script hello/Conscript. The libworld.a
archive will be built, however, if need be.
There are a couple of uses for build pruning via the command line. Perhaps the most useful is the ability to make local changes, and then, with sufficient knowledge of the consequences of those changes, restrict the size of the build tree in order to speed up the rebuild time. A second use for build pruning is to actively prevent the recompilation of certain files that you know will recompile due to, for example, a modified header file. You may know that either the changes to the header file are immaterial, or that the changes may be safely ignored for most of the tree, for testing purposes.With Cons, the view is that it is pragmatic to admit this type of behavior, with the understanding that on the next full build everything that needs to be rebuilt will be. There is no equivalent to a ``make touch'' command, to mark files as permanently up-to-date. So any risk that is incurred by build pruning is mitigated. For release quality work, obviously, we recommend that you do not use build pruning (it's perfectly OK to use during integration, however, for checking compilation, etc. Just be sure to do an unconstrained build before committing the integration).
T.B.S.
Cons provides a very simple mechanism for overriding aspects of a build.
The essence is that you write an override file containing one or more
Override commands, and you specify this on the command line, when you run
cons:
cons -o over export
will build the export directory, with all derived files subject to the overrides present in the
file over. If you leave out the -o option, then everything necessary to remove all overrides will be rebuilt.
The override file can contain two types of overrides. The first is incoming
environment variables. These are normally accessible by the Construct
file from the %ENV hash variable. These can trivially be overridden in the override file by
setting the appropriate elements of %ENV (these could also be overridden in the user's environment, of course).
The second type of override is accomplished with the Override command, which looks like this:
Override <regexp>, <var1> => <value1>, <var2> => <value2>, ...;
The regular expression regexp is matched against every derived file that is a candidate for the build. If the derived file matches, then the variable/value pairs are used to override the values in the construction environment associated with the derived file.
Let's suppose that we have a construction environment like this:
$CONS = new cons(
COPT => '',
CDBG => '-g',
CFLAGS => '%COPT %CDBG',
);
Then if we have an override file over containing this command:
Override '\.o$', COPT => '-O', CDBG => '';
then any cons invocation with -o over that creates .o files via this environment will cause them to be compiled with -O and no -g. The override could, of course, be restricted to a single directory by the
appropriate selection of a regular expression.
Here's the original version of the Hello, World! program, built with this environment. Note that Cons rebuilds the appropriate pieces when the override is applied or removed:
% cons hello cc -g -c hello.c -o hello.o cc -o hello hello.o % cons -o over hello cc -O -c hello.c -o hello.o cc -o hello hello.o % cons -o over hello cons: "hello" is up-to-date. % cons hello cc -g -c hello.c -o hello.o cc -o hello hello.o
It's important that the Override command only be used for temporary, on-the-fly overrides necessary for
development because the overrides are not platform independent and because
they rely too much on intimate knowledge of the workings of the scripts.
For temporary use, however, they are exactly what you want.
Note that it is still useful to provide, say, the ability to create a fully
optimized version of a system for production use--from the Construct and
Conscript files. This way you can tailor the optimized system to the platform. Where
optimizer trade-offs need to be made (particular files may not be compiled
with full optimization, for example), then these can be recorded for
posterity (and reproducibility) directly in the scripts.
We have mentioned, and used, the concept of a construction environment, many times in the preceding pages. Now it's time to make this a little more concrete. With the following statement:
$env = new cons();
a reference to a new, default construction environment is created. This contains a number of construction variables and some methods. At the present writing, the default list of construction variables is defined as follows:
CC => 'cc',
CFLAGS => '',
CCCOM => '%CC %CFLAGS %_IFLAGS -c %< -o %>',
CPPPATH => '',
LINK => '%CC',
LINKCOM => '%LINK %LDFLAGS -o %> %< %_LDIRS %LIBS',
LIBPATH => '',
LIBS => '',
AR => 'ar',
ARCOM => "%AR %ARFLAGS %> %<\n%RANLIB %>",
ARFLAGS => 'r',
RANLIB => 'ranlib',
AS => 'as',
ASFLAGS => '',
ASCOM => '%AS %ASFLAGS %< -o %>',
LD => 'ld',
LDFLAGS => '',
SUFLIB => '.a',
SUFOBJ => '.o',
ENV => { PATH => '/bin:/usr/bin' },
These variables are used by the various methods associated with the
environment, in particular any method that ultimately invokes an external
command will substitute these variables into the final command, as
appropriate. For example, the Objects method takes a number of source files and arranges to derive, if necessary,
the corresponding object files. For example:
Objects $env 'foo.c', 'bar.c';
This will arrange to produce, if necessary, foo.o and bar.o. The command invoked is simply %CCOM, which expands through substitution, to the appropriate external command
required to build each object. We will explore the substitution rules
further under the Command method, below.
The construction variables are also used for other purposes. For example,
CPPPATH is used to specify a colon-separated path of include directories. These are
intended to be passed to the C preprocessor and are also used by the C-file
scanning machinery to determine the dependencies involved in a C
Compilation. Variables beginning with underscore, are created by various
methods, and should normally be considered ``internal'' variables. For
example, when a method is called which calls for the creation of an object
from a C source, the variable _IFLAGS is created: this corresponds to the -I switches required by the C compiler to represent the directories specified
by CPPPATH.
Note that, for any particular environment, the value of a variable is set
once, and then never reset (to change a variable, you must create a new
environment. Methods are provided for copying existing environments for
this purpose). Some internal variables, such as _IFLAGS are created on demand, but once set, they remain fixed for the life of the
environment.
Another variable, ENV, is used to determine the system environment during the execution of an
external command. By default, the only environment variable that is set is PATH, which is the execution path for a UNIX command. For the utmost
reproducibility, you should really arrange to set your own execution path,
in your top-level Construct file (or perhaps by importing an appropriate construction package with the
Perl use
command). The default variables are intended to get you off the ground.
The list of default construction methods includes the following:
The new method is a Perl object constructor. That is, it is not invoked via a
reference to an existing construction environment reference, but, rather statically, using the name of the Perl package where the constructor is defined. The method is invoked like this:
$env = new cons(<overrides>);
The environment you get back is blessed into the package cons, which means that it will have associated with it the default methods
described below. Individual construction variables can be overridden by
providing name/value pairs in an override list. Note that to override any
command environment variable (i.e. anything under ENV), you will have to override all of them. You can get around this
difficulty by using the copy method on an existing construction environment.
The clone method creates a clone of an existing construction environment, and can be
called as in the following example:
$env2 = $env1->clone(<overrides>);
You can provide overrides in the usual manner to create a different environment from the original. If you just want a new name for the same environment (which may be helpful when exporting environments to existing components), you can just use simple assignment.
The copy method extracts the externally defined construction variables from an
environment and returns them as a list of name/value pairs. Overrides can
also be provided, in which case, the overridden values will be returned, as
appropriate. The returned list can be assigned to a hash, as shown in the
prototype, below, but it can also be manipulated in other ways:
%env = $env1->copy(<overrides>);
The value of ENV, which is itself a hash, is also copied to a new hash, so this may be
changed without fear of affecting the original environment. So, for
example, if you really want to override just the
PATH variable in the default environment, you could do the following:
%cons = new cons()->copy();
$cons{ENV}{PATH} = "<your path here>";
$cons = new cons(%cons);
This will leave anything else that might be in the default execution environment undisturbed.
The Install method arranges for the specified files to be installed in the specified
directory. The installation is optimized: the file is not copied if it can
be linked. If this is not the desired behavior, you will need to use a
different method to install the file. It is called as follows:
Install $env <directory>, <names>;
Note that, while the files to be installed may be arbitrarily named, only
the last component of each name is used for the installed target name. So,
for example, if you arrange to install foo/bar in baz, this will create a file bar in directory baz (not foo/bar).
The InstallAs method arranges for the specified source file(s) to be
installed as the specified target file(s). Multiple files
should be specified as a file list. The installation is optimized: the file
is not copied if it can be linked. If this is not the desired behavior, you
will need to use a different method to install the file. It is called as
follows:
InstallAs works in two ways:
Single File Install:
InstallAs $env TgtFile, SrcFile;
Multiple File Install:
InstallAs $env ['tgt1', 'tgt2'], ['src1', 'src2'];
Or, even as:
@srcs = qw(src1 src2 src3); @tgts = qw(tgt1 tgt2 tgt3); InstallAs $env [@tgts], [@srcs];
Both the target and the sources lists should be of the same length.
The Command method is a catchall method which can be used to arrange for any external
command to be called to update the target. For this command, a target file
and list of inputs is provided. In addition a construction command line, or
lines, is provided as a string (this string may have multiple commands
embedded within it, separated by new lines). Command is called as follows:
Command $env <target>, <inputs>, <construction command>;
The target is made dependent upon the list of input files specified, and the inputs must be built successfully or Cons will not attempt to build the target.
Within the construction command, any variable from the construction
environment may be introduced by prefixing the name of the construction
variable with %. This is recursive: the command is expanded until no more substitutions
can be made. If a construction variable is not defined in the environment,
then the null string will be substituted.
There are several pseudo variables which will also be expanded:
The target file name (in a multi-target command, this is always the first target mentioned).
Same as %>.
%2, ..., %9 These refer to the first through ninth input file,
respectively.
The full set of inputs. If any of these have been used anywhere else in the
current command line (via %1, %2, etc.), then those will be deleted from the list provided by %<. Consider the following command found in a
Conscript file in the directory test:
Command $env 'tgt', qw(foo bar baz), qq(
echo %< -i %1 > %>
echo %< -i %2 >> %>
echo %< -i %3 >> %>
);
If tgt needed to be updated, then this would result in the execution of the
following commands, assuming that no remapping has been established for
directory test:
echo test/bar test/baz -i test/foo > test/tgt echo test/foo test/baz -i test/bar >> test/tgt echo test/foo test/bar -i test/baz >> test/tgt
Any of the above pseudo variables may be followed immediately by :d or
:f, to indicate the directory or file associated with the name. Continuing
with the above example, %<:f would expand to foo bar
baz, and %:d> would expand to test.
After substitution occurs, strings of white space are converted into single blanks, and leading and trailing white space is eliminated. It is therefore not possible to introduce variable length white space in strings passed into a command, without resorting to some sort of shell quoting.
If a multi-line command string is provided, the commands are executed sequentially. If any of the commands fails, then none of the rest are executed, and the target is not marked as updated, i.e. a new signature is not stored for the target.
Normally, if all the commands succeed, and return a zero status (or whatever platform-specific indication of success is required), then a new signature is stored for the target. If a command erroneously reports success even after a failure, then Cons will assume that the target file created by that command is accurate and up-to-date.
The first word of each command string, after expansion, is assumed to be an
executable command looked up on the PATH environment variable (which is, in turn, specified by the ENV construction variable). If this command is found on the path, then the
target will depend upon it: the command will therefore be automatically
built, as necessary. It's possible to write multi-part commands to some
shells, separated by semi-colons. Only the first command word will be
depended upon, however, so if you write your command strings this way, you
must either explicitly set up a dependency (with the
Depends method), or be sure that the command you are using is a system command
which is expected to be available. If it isn't available, you will, of
course, get an error.
If there are shell meta characters anywhere in the expanded command line,
such as <, >, quotes, or semi-colon, then the command will actually be executed by
invoking a shell. This means that a command such as:
cd foo
alone will typically fail, since there is no command cd on the path. But the command string:
cd $<:d; tar cf $>:f $<:f
when expanded will still contain the shell meta character semi-colon, and a
shell will be invoked to interpret the command. Since cd is interpreted by this sub-shell, the command will execute as expected.
To specify a command with multiple targets, you can specify a reference to a list of targets. In Perl, a list reference can be created by enclosing a list in square brackets. Hence the following command:
Command $env ['foo.h', 'foo.c'], 'foo.template', q(
gen %1
);
could be used in a case where the command gen creates two files, both
foo.h and foo.c.
The Objects method arranges to create the object files that correspond to the specified
source files. It is invoked as shown below:
@files = Objects $env <source or object files>;
Under Unix, source files ending in .s and .c are currently supported, and will be compiled into a name of the same file
ending in .o. By default, all files are created by invoking the external command which
results from expanding the CCOM construction variable, with %< and
%> set to the source and object files, respectively (see the Command
method for expansion details) . The variable CPPPATH is also used when scanning source files for dependencies. This is a colon
separated list of pathnames, and is also used to create the construction
variable _IFLAGS,
which will contain the appropriate list of -I options for the compilation. Any relative pathnames in CPPPATH is interpreted relative to the directory in which the associated
construction environment was created (absolute and top-relative names may
also be used). This variable is used by
CCOM. The behavior of this command can be modified by changing any of the
variables which are interpolated into CCOM, such as CC, CFLAGS, and, indirectly, CPPPATH. It's also possible to replace the value of
CCOM, itself. As a convenience, this file returns the list of object filenames.
The Program method arranges to link the specified program with the specified object
files. It is invoked in the following manner:
Program $env <program name>, <source or object files>;
Source files may be specified in place of objects files--the Objects
method will be invoked to arrange the conversion of all the files into
object files, and hence all the observations about the Objects method, above, apply to this method also. The actual linking of the program
will be handled by an external command which results from expanding the LINKCOM
construction variable, with %< set to the object files to be linked (in the order presented), and %> set to the target (see the Command method for expansion details). The user may set additional variables in the
construction environment, including LINK, to define which program to use for linking, LIBPATH, a colon-separated list of library search paths, for use with library
specifications of the form -llib, and LIBS, specifying the list of libraries to link against (in either -llib form or just as pathnames. Relative pathnames in both LIBPATH and LIBS are interpreted relative to the directory in which the associated
construction environment created (absolute and top-relative names may also
be used). Cons automatically sets up dependencies on any libraries
mentioned in LIBS: those libraries will be built before the command is linked.
The Library method arranges to create the specified library from the specified object
files. It is invoked as follows:
Library $env <library name>, <source or object files>;
Source files may be specified in place of objects files--the Objects
method will be invoked to arrange the conversion of all the files into
object files, and hence all the observations about the Objects method, above, apply to this method also. The actual creation of the
library will be handled by an external command which results from expanding
the ARCOM
construction variable, with %< set to the library members (in the order presented), and %> to the library to be created (see the Command
method for expansion details). The user may set variables in the
construction environment which will affect the operation of the command.
These include AR, the archive program to use, ARFLAGS, which can be used to modify the flags given to the program specified by AR, and
RANLIB, the name of a archive index generation program, if needed (if the
particular need does not require the latter functionality, then ARCOM
must be redefined to not reference RANLIB).
The Library method allows the same library to be specified in multiple method
invocations. All of the contributing objects from all the invocations
(which may be from different directories) are combined and generated by a
single archive command. Note, however, that if you prune a build so that
only part of a library is specified, then only that part of the library
will be generated (the rest will disappear!).
The Module method is a combination of the Program and Command
methods. Rather than generating an executable program directly, this
command allows you to specify your own command to actually generate a
module. The method is invoked as follows:
Module $env <module name>, <source or object files>, <construction command>;
This command is useful in instances where you wish to create, for example, dynamically loaded modules, or statically linked code libraries.
The Depends method allows you to specify additional dependencies for a target. It is
invoked as follows:
Depends $env <target>, <dependencies>;
This may be occasionally useful, especially in cases where no scanner exists (or is writable) for particular types of files. Normally, dependencies are calculated automatically from a combination of the explicit dependencies set up by the method invocation or by scanning source files.
There are several ways of extending Cons, which vary in degree of
difficulty. The simplest method is to define your own construction
environment, based on the default environment, but modified to reflect your
particular needs. This will often suffice for C-based applications. You can
use the new constructor, and the clone and copy methods to create hybrid environments. These changes can be entirely
transparent to the underlying Conscript files.
For slightly more demanding changes, you may wish to add new methods to the
cons package. Here's an example of a very simple extension,
InstallScript, which installs a tcl script in a requested location, but edits the script
first to reflect a platform-dependent path that needs to be installed in
the script:
# cons::InstallScript - Create a platform dependent version of a shell # script by replacing string ``#!your-path-here'' with platform specific # path $BIN_DIR.
sub cons::InstallScript {
my($env, $dst, $src) = shift;
Command $env $dst, $src, qq(
sed s+your-path-here+$BIN_DIR+ %< > %>
chmod oug+x %>
);
}
Notice that this method is defined directly in the cons package (by prefixing the name with cons::). A change made in this manner will be globally visible to all
environments, and could be called as in the following example:
InstallScript $env "$BIN/foo", "foo.tcl";
For a small improvement in generality, the BINDIR variable could be passed in as an argument or taken from the construction
environment--as
%BINDIR.
Instead of adding the method to the cons name space, you could define a new package which inherits existing methods
from the cons package and overrides or adds others. This can be done using Perl's
inheritance mechanisms.
The following example defines a new package cons::switch which overrides
the standard Library method. The overridden method builds linked library modules, rather than
library archives. A new constructor is provided. Environments created with
this Constructor will have the new library method; others won't.
package cons::switch;
BEGIN {@ISA = `cons'}
sub new {
shift;
bless new cons(@_);
}
sub Library {
my($env) = shift;
my($lib) = shift;
my(@objs) = Objects $env @_;
Command $env $lib, @objs, q(
%LD -r %LDFLAGS %< -o %>
);
}
This functionality could be invoked as in the following example:
$env = new cons::switch(@overrides); ... Library $env 'lib.o', 'foo.c', 'bar.c';
The cons command is always invoked from the root of the build tree. A
Construct file must exist in that directory. If the -f argument is used, then an alternate Construct file may be used (and, possibly, an alternate root, since cons will cd to Construct file's containing directory). The command is invoked as follows:
cons <arguments> -- <construct-args>
where arguments can be any of the following, in any order:
Build the specified target. If target is a directory, then recursively build everything within that directory.
Limit the Conscript files considered to just those that match pattern, which is a Perl regular expression. Multiple + arguments are accepted.
Sets name to value val in the ARG hash passed to the top-level
Construct file.
Show command that would have been executed, when retrieving from cache. No indication that the file has been retrieved is given; this is useful for generating build logs that can be compared with real build logs.
Disable all caching. Do not retrieve from cache nor flush to cache.
Build dependencies in random order. This is useful when building multiple similar trees with caching enabled.
Synchronize existing build targets that are found to be up-to-date with cache. This is useful if caching has been disabled with -cc or just recently enabled with UseCache.
Enable dependency debugging.
Use the specified file instead of Construct (but first change to containing directory of file).
Show a help message local to the current build if one such is defined, and exit.
Read override file file.
Keep going as far as possible after errors.
Show construction products in specified trees. No build is attempted.
Show construction products and associated actions. No build is attempted.
Show products and where they are defined. No build is attempted.
Don't be verbose about Installing and Removing targets.
Write all filenames considered into file.
Remove construction products associated with <targets>. No build is attempted.
Search for files in repos. Multiple -R repos directories are searched in the order specified.
Show cons version and continue processing.
Show cons version and exit.
Show a help message similar to this one, and exit.
And construct-args can be any arguments that you wish to process in the Construct file. Note that there should be a -- separating the arguments to cons and the arguments that you wish to process in the Construct file.
Processing of construct-args can be done by any standard package like Getopt or its variants, or any user defined package. cons will pass in the construct-args as @ARGV and will not attempt to interpret anything after the --.
cons -R /usr/local/repository -d os=solaris +driver -- -c test -f DEBUG
would pass the following to cons
-R /usr/local/repository -d os=solaris +driver
and the following, to the top level Construct file as @ARGV
-c test -f DEBUG
Note that cons -r . is equivalent to a full recursive make clean, but requires no support in the Construct file or any Conscript files. This is most useful if you are compiling files into source
directories (if you separate the build/export directories, then you can just remove the directories).
The options -p, -pa, and -pw are extremely useful for use as an aid in reading scripts or debugging
them. If you want to know what script installs export/include/foo.h, for example, just type:
cons -pw export/include/foo.h
QuickScan allows simple target-independent scanners to be set up for source files. Only one QuickScan scanner may be associated with any given source file and environment.
QuickScan is invoked as follows:
QuickScan CONSENV CODEREF, FILENAME [, PATH]
The subroutine referenced by CODEREF is expected to return a list of
filenames included directly by FILE. These filenames will, in turn, be
scanned. The optional PATH argument supplies a lookup path for finding
FILENAME and/or files returned by the user-supplied subroutine. The
subroutine is called once for each line in the file, with $_
set to the current line. If the subroutine needs to look at additional
lines, or, for that matter, the entire file, then it may read them itself,
from the filehandle SCAN. It may also terminate the loop, if it knows that
no further include information is available, by closing the filehandle.
Whether or not a lookup path is provided, QuickScan first tries to lookup the file relative to the current directory (for the top-level file supplied directly to QuickScan), or from the directory containing the file which referenced the file. This is not very general, but seems good enough--especially if you have the luxury of writing your own utilities and can control the use of the search path in a standard way. Finally, the search path is, currently, colon separated. This may not make the NT camp happy.
Here's a real example, taken from a Construct file here:
sub cons::SMFgen { my($env, @tables) = @_; foreach
$t (@tables) { $env->QuickScan(sub { /\b\S*?\.smf\b/g },
``$t.smf'', $env->{SMF_INCLUDE_PATH}); $env->Command(
[``$t.smdb.cc'',``$t.smdb.h'',``$t.snmp.cc'',``$t.ami.cc'',
``$t.http.cc''], ``$t.smf'', q( smfgen %(
%SMF_INCLUDE_OPT %) %< ) ); } }
[NOTE that the form $env->QuickScan ... and $env->Command ... should not be necessary, but, for some reason, is required for this particular invocation. This appears to be a bug in Perl or a misunderstanding on my part; this invocation style does not always appear to be necessary.]
This finds all names of the form <name>.smf in the file. It will return the names even if they're found within comments, but that's OK (the mechanism is forgiving of extra files; they're just ignored on the assumption that the missing file will be noticed when the program, in this example, smfgen, is actually invoked).
A scanner is only invoked for a given source file if it is needed by some target in the tree. It is only ever invoked once for a given source file.
Here is another way to build the same scanner. This one uses an explicit code reference, and also (unecessarily, in this case) reads the whole file itself:
sub myscan { my(@includes); do { push(@includes,
/\b\S*?\.smf\b/g); } while <SCAN>; @includes }
Note that the order of the loop is reversed, with the loop test at the end. This is because the first line is already read for you. This scanner can be attached to a source file by:
QuickScan $env \myscan, "$_.smf";
Cons is maintained by the user community. To subscribe, send mail to cons-discuss-request@eng.fore.com with body subscribe.
Please report any bugs/fixes/suggestions through the cons-discuss@eng.fore.com mailing list.
The Official cons site is: http://www.dsmit.com/cons/