Assaf Arkin wrote:
On Mon, Jul 28, 2008 at 11:23 PM, Ittay Dror <[EMAIL PROTECTED]> wrote:
I merged the other email (ordering) and comments. My comments inline
Assaf Arkin wrote:
On Mon, Jul 28, 2008 at 2:42 AM, Ittay Dror <[EMAIL PROTECTED]> wrote:
Hi,
I'm working on adding C++ support to buildr. I already have a prototype
that
builds libraries and executables in Linux. I'd like to share some of the
difficulties I had and request changes to buildr to accommodate C++ more
easily. (Right now, I've created parallel route to that of building
Java-like code)
compile
========
overview
--------------------
the compile method in project returns a CompileTask that is generic and
uses
a Compiler instance to do the actual compilation. In C++, compilation is
also dependency based (.o => .cpp, sometimes precompiling headers). Also,
the same code can produce several results (static and shared libraries,
oj
files with debug, profiling, preprocessor defines turned on and off). [1]
there is the 'build' task, which is used as a stub to attach dependencies
to.
suggestion
---------------------
* there should be an array of compile tasks (as in packages)
* #compile should delegate the call to a factory method which returns a
task
(again, as in packages)
Yes. And I know a few people just waiting for the change to compile
multiple things in the same project, so here's another reason for
adding this feature.
But I have to warn you, it's not as simple as it looks, I took a stab
at it before and deciding to downscale support to one compiler per
project. It's worth doing because a lot of languages would benefit
from it, but that's also what makes it tricky. I think it would be
easier to get C support working without it first, and separately work
on this feature and then improve C support using it.
How about this: classify compile commands with symbolic names. like
compile('java') or compile('c++:shared') ? on bootstrap, the different
extensions can create compile tasks based on directory structure (so the
Java extension can see that the directory [:source, :main, :java] exists and
create compile('java') with some default values.
All compile tasks are prerequisits of 'build'
Then 'package :jar' can create a package that depends on compile('java'),
compile('groovy') or whatever makes sense to put in a jar, as long as the
compile task exist of course (not to create them if they don't) (BTW, I have
some issues with the lack of command-query separation, normally when using a
query method, I wouldn't want a task to be created if it doesn't exist)
Rake::Task.task_defined? will tell if if a task is defined without
creating it. Rake::Task[] (same as calling task) would find you the
task, creating it if necessary by looking at the rules, existing files
or creating a generic task.
I want to avoid discussing the issues with
compile('java')/compile('groovy') here. It's a big issue that belongs
in its own thread and affects more than just C/C++. I'm just pointing
out that it looks as easy as adding a language flag to compile, but
when you get down to look at all the details involved, it's a pretty
damn big change.
And separately, see comments below, it will not replace the generic
compile task but add more tasks for compile to orchestrate.
* generic pre-requisites (like 'resources') should either be tacked on
'build' (relying on order of prerequisites), or the compile task can be
defined to be a composite (that is, from the outside it is a single task,
but it can use other tasks to accomplish its job).
compile already is: resources is a prerequisite for compile, some
other tasks (e.g. byte code enhancing) are tacked on to compile by
enhancing it.
yes, but the compilation of the java family of languages is one task
(calling javac), while compiling c++ is several tasks: task per obj file and
task per link. so there's a chain of tasks already. having a generic method
receive a task from the factory method and make it depend on 'resources'
won't do, since the lower level tasks should be the ones that depend.
I don't see why the existing compile task can't orchestrate all the
smaller compile tasks. It already orchestrates several tasks,
compiling a project will compile all its sub-projects, dependencies,
resources, etc. Think of it as the compile stage of the build, more
than just running the compiler. In fact all top-level projects have a
compile task, but many don't have anything to compile, just use it to
orchestate compilation of all their child projects.
If you let compile orchestrate smaller tasks, you can get the Rake
dependency mechanism working for you to handle individual object
files, compiling only that which is necessary, but also get the Buildr
dependency mechanism orchestrating the different steps of the build
and dependencies between projects.
can you give an example of how a task can orchestrate other tasks? also,
as far as i could tell, the 'compile' method always create a
CompileTask. i can't use it as is because it expects some compiler which
i can't give it because i want to use tasks and also, i can't add
dependencies to it because it depends directly on tasks like 'resources'
which the prerequisites should depend on.
At the risk of spending a lot of time on the obvious (i have a feeling
we're talking about different things):
say a project has 2 cpp files A.cpp and B.cpp, with matching headers,
and no other headers, which compile to shared and static libraries. my
dependency tree is:
compile:cpp ----- libsomething.so --- A.o --- A.cpp
\ \ / \ A.h
\ X
\ / \ B.o --- B.cpp
\- libsomething.a/-----/ \ B.h
these should be rake tasks for two reasons: timestamp checking and the
fact that two artifacts rely on the same set of objects. also linking
and compiling are two different commands and finally, if i call the
compiler twice, it will do the work twice (that is, it doesn't have any
internal mechanism that tells it there's no need to recreate the obj
files or libraries).
note that all of this tree needs to rely on the 'resources' task, since
some headers may be generated. so 'resources' need to run before all the
timestamp checking and compilation is done.
of course the factory method can create just one task that does all the rest
in its action (compile obj files and link), but i do want to use tasks for
the following reasons:
1. it makes the logic more like make, which will assist acceptance
2. it can use mechanisms in unix compilers to help make. specifically, most
(if not all) unix compilers have an option to spit out dependencies of the
source files on headers.
3. it reuses timestamp checking code in rake (and if ever rake implements
checksum based recompilation)
4. if rake will implement a job execution engine (like -j in make), then
structuring compilation by tasks will allow it to parallelize the execution.
but, i think the solution is easy: similar to the 'build' "pseudo task", i
can create a 'compile:prepare' pseudo task that depends on 'resources' etc.
then, the factory method needs only to depend on 'compile:prepare' (the
logic is that another extension can then add other things to do before
compile without needing to change the compile extensions)
We had compile:prepare in the past which invokes resources and ...
well, that's about it. It turns out that just having compile and
doing everything else as prerequisite is good enough.
package & artifacts
=========
overview
---------------
buildr has a cool concept that all dependencies (in 'compile.with') are
converted to tasks that are then simple rake dependencies. However, the
conversion is not generic enough. to compile C++ code against a
dependency
one needs 2 paths: a folder containing headers and another containing
libraries. To put this in a repository, these need to be packaged into
one
file. To use after pulling from the repository, one needs to unpack. So a
task representing a repository artifact is in fact an unzip task, that
depends on the 'Artifact' task to pull the package from a remote
repository.
Let's take Java for example, let's say we have a task that depends on
the contents of another WAR. Specifically the classes (in
WEB-INF/classes) and libraries (WEB-INF/lib). A generic unzipping
artifact won't help much, you'll get the root path which is useless.
You need the classes path for one, and each file in the lib (pointing
to the directory itself does nothing interesting). It won't work with
EAR either, when you unzip those, you end up with a WAR which you need
to unzip again.
But this hypothetical task that uses WAR could be smarter. It
understands the semantics of the packages it uses, and all these
packages follow a common convention, so it only needs to unpack the
portions of the WAR it cares about, it knows how to construct the
relevant paths, one to class and one to every JAR inside the lib
directory.
I think the same analogy applies to C packages. If by convention you
always use include and lib, you can unpack only the portion of the
package you need, find the relevant paths and use them appropriately.
(note: not sure i'm following you here. )
Artifacts by themselves are a generic mechanism for getting packages
into the local repository. Their only responsibility if the artifact
and its metadata, so a task representing a repository artifact would
only know how to download it.
You can have a separate task that knows how to extract an artifact
task and use it instead, that way you get the unpacking you need, but
not all downloaded artifacts have to be unpacked.
yes, this is what i'm currently doing, as i explained below.
but what i want is for me to be able to do that by integrating with the
existing 'artifacts' task. right now it will only return Artifact
objects. I'd like to have a more elegant solution than just to run over
them and create my own objects, which i think will be more tricky with
transitive dependencies (where transitivity may come from my artifacts,
e.g. the project's artifacts)
my current implementation creates classes that have methods to retrieve the
include paths, the library paths and the library names. I don't use the task
name, since it is useless (as you mentioned). so I have an
ExtractedRepoArtifact FileTask class that implements these methods by
relying on the structure of the package ('include' and 'lib' directories),
it depends on the Artifact class and its action is to extract the artifact.
When given a project dependency, i return the build task which implements
the artifact methods mentioned above by returning the
[:source,:main,:include] and [:target, Platform.id, :lib] paths. It also
allows the user to add include paths (e.g., for generated files) which are
then both used for compilation and returned by the artifact methods.
furthermore, when building against another project, there is no need to
pack
and unpack in the repository. one can simply use the artifacts produced
in
the 'build' phase of the other project.
Yes. Right now it points to the package, which gets invoked and so
packs everything, whether you need the packing or not. You don't,
however, have to unpack it, if you know the packaging type you can be
smarter and go directly to the source.
but i don't want to pack if there's no use for it. speed is critical in this
project, since there's no eclipse to constantly compile code for you, so
developers need to run the build after each change. having it pack
unnecessarily wasts time.
One step at a time. I would worry if we can't do that at all, but if
it's just optimization, we can get to the more problematic issues
first.
finally, in C++ in many cases you rely on a system library.
in all cases the resulting dependency is two-fold: on a include dir paths
and on a library paths. note that these do not necessarily reside under a
shared folder. for example, a dependency on another project may depend on
two include folders: one just a folder in the source tree, the other of
generated files in the target directory
suggestion
-------------------
While usage of Buildr.artifacts is only as a utility method, so one can
easily write his own implementation and use that, I think it will be nice
to
be able to get some reuse.
* when given a project, use it as is (not 'spec.packages'), or allow it
to
return its artifacts ('spec.artifacts').
Yes. Except we're missing that whole dependency later (that's
something 1.4 will add). Ideally the project would have dependency
lists it can populates (at least compile and runtime), and other
projects can get these dependency lists and pick what they want. So
the compile dependency list would be the place to put headers and
libraries, without having to package them. We don't have that right
now.
this is the purpose for the 'spec.artifacts' suggestion (that is, an
'artifacts' method in Project). maybe need to classify them similarly to my
suggestion for 'compile', so the Buildr.artifacts method receives a
'classifier' argument, whose value can be, for example, 'java' and calls
'spec.artifacts(classifier)'. are we on the same page here?
I'm looking at each of your use cases and trying to identify in my mind:
a) What you can do right now to make it happen.
b) What, if we added another feature, we should accommodate for.
c) What new feature we would need for this.
I'm starting with a) because you can get it working right now, it may
not be elegant and not work as fast, but we can get that out of the
way so we can focus about doing the rest. There are some things we're
planning on changing anyway, so I'm also trying to see if future
changes would address the elegant/fast use cases, I can tell you what
I have in mind, but no code yet to make it happen. And then identify
anything not addressed by current plans and decide how to support that
directly.
i got it working now. but i'm doing several code paths in parallel. i
have a 'make' method instead of 'compile'. the reason are both because i
need to create several tasks, not a 'compiler' object (and i want to
create them before rake's execution starts) , and because i need to
create different implementations per platform.
Right now, project.packages is good enough for what you need. It's an
array of tasks, you can throw any task you want in there and the
dependent project would pick on it. You don't have to throw ZIP files
in there, you can add a header file or a directory of header files, or
a task that knows it's a directly of header files.
It's inelegant because project.packages is intent to be the list of
things that get installed and released, so it's an "off the label" use
for that part of the API. But, it will work, and if you just add
things to the end of project.packages, they won't get installed or
released. So project.packages is that same as project.artifacts, just
with a different name.
or i can implement my own 'artifacts' method, which is what i did
because i need different artifact objects than what Buildr.artifacts
returns.
Separately, we need (and planning and working on) a smarter dependency
management, which you can populate and anything referencing the
project can access. It won't be called artifacts but dependencies, it
will do a lot more, and it will be more elegant and documented for
specific use cases like this.
* if a symbol, recursively call on the spec from the namespace
* if a struct, recursively call
* otherwise, classify the artifact and call a factory method to create
it.
classification can be by packaging (e.g. jar). but actually, i don't have
a
very good idea here. note that for c++, there need to be a way of
defining
an artifact to look in the system for include files and libraries (maybe
something like 'openssl:system'? - version and group ids are
meaningless).
* the factory method can create different artifacts. for c++ there would
be
RepositoryArtifact (downloads and unpacks), ProjectArtifact (short
circuit
to the project's target and source directories) and SystemArtifact.
I think that the use of artifact namespaces can help here as it allows to
create a more verbose syntax for declaring artifacts, while still
allowing
the user to create shorter names for them. (as an example in C++ it will
allow me to add to the artifact the list of flags to use when
compiling/linking with it, assuming they're not inherent to the artifact,
e.g. turn debug on). The factory method receives the artifact definition
(which can actually be defined by each plugin) and decides what to do
with
it.
1.4 will have a better dependency mechanism, and one thing I looked at
is associating meta-data with each dependency. So perhaps that would
address things like compiling/linking flags.
ordering
=========
overview
-------------------
to support jni, one needs to first compile java classes, then run javah
to
generate headers and then compile c code that implements these headers.
so
the javah task should be able to specify it depends on the java compile
task. this can't be by depending on all compile tasks of course or on
'build'.
Alternatively:
compile do |task|
javah task.target
end
This will run javah each time the compiler runs.
but running each time is what i want to avoid. not only do i want to avoid
the invocation of 'javah', but when invoked it will change the timestamp of
the generated headers and so many source files will get recompiled.
Rake separates invocation from execution. Invoking a task tells it to
invoke its prerequisites, then use those to decide if it needs
executing, and if so execute. Whether you put javah at the end of
compile, or a prerequisite to build, it will get invoked and it should
be smart enough to decide whether there's any work to be done.
i think i'm missing something here. in the code snippet above, didn't
you add an action to 'compile' and in that action call the javah
command? to me it looks like at the end of compile javah is run.
But there is a significant difference between the two. If you add it
to compile, it gets invoked during compilation -- and compilation
implies there's a change to the source code which might lead to change
in the header files -- and that happens as often as is necessary. If
you put is as prerequisite to build, it only happens when the build
task runs. If you run rake task, which doesn't run the build task,
you may end up testing the wrong header files.
there should be a rule to the effect of:
jni_headers_dir => [classes] do |task|
javah classes # with whatefer flags to put generated headers in
jni_headers_dir
touch jni_headers_dir
end
so if the classes are newer than the directory (and only then) javah
runs. if i run it every time it will generate headers, changing the
timestamp, which will cause all dependent cpp classes to recompile which
will take a lot of time.
note that compiling a C/C++ source file is a much slower process than
compiling java.
suggestion
-------------------
when creating a compile task (whose name can be, as in the case of c++,
the
result library name - to allow for dependency checking), also create a
"for
ordering only" task with a symbolic name (e.g., 'java:compile') which
depends on the actual task. then other tasks can depend on that task
And yes, you'll still need that if you want to run the C compiler
after the Java compiler, so I think the right thing to do would have
separate compile tasks.
I hope all this makes sense, and I'm looking forward to comments. I
intend
to share the code once I'm finished.
Unfortunately, the last time I wrote C code was over tens years ago,
so my rustiness is showing. I'm sure I missed some points because of
that.
I hope I cleared things. I think it is worth investing in C/C++ as it is a
space where there's still no solutions (that i know of) that handle module
dependency.
Definitely.
To make sure it is clear, I'm not asking for the buildr team to implement
C/C++ building, I intend to do that, and have already made a demo of it
working, but I do want to ask for the infrastructure in buildr to make it
easier, since currently it looks like a "stepson".
In addition, two things we should look at.
First, find out a good intersection between C/C++ and other languages.
There may be some changes that are only necessary for C/C++, but
hopefully most of these can be shared across languages, that way we
get better features all around.
Second, make sure we exhausted all our options before making a change.
If there's another way of doing something, even stop-gap measure
while we cook up a better feature all around, then we have less
changes to worry about.
It's an exercise we did before with Groovy and Scala (earlier versions
were married to Java) and it worked out pretty well. We started by
not making any changes in Buildr to accommodate it, instead using a
separate task specifically for compiling Scala code that relied on
some hacks and inelegant code to actually work. Then took the time to
build multi-lingual support out of that.
i'm already past that. i have ~20 modules compiling, with transitive
dependencies on other modules and on third party modules.
so i'm now at a stage where i want better integration with buildr.
Assaf
Ittay
Assaf
Thank you,
Ittay
Notes:
[1] I don't consider linking a library as packaging. First, the obj files
are not used by themselves as in other languages. Second, packaging is
required to manage dependencies, because in order for project P to be
built
against dependency D, D needs to contain both headers and libraries -
this
is the package.
--
--
Ittay Dror <[EMAIL PROTECTED]>
--
--
Ittay Dror <[EMAIL PROTECTED]>
--
--
Ittay Dror <[EMAIL PROTECTED]>
