Draft
I propose that we modularise the build system for NaCl so that it has the following properties, based on what I outlined in a [blog post] (http://lackingrhoticity.blogspot.com/2009/12/modular-vs-monolithic-build-systems.html):
In practice, this means that each module should: 1. Locate its dependencies through environment variables and explicitly-passed arguments. Modules should not reach outside their source/build directories to find dependencies using relative pathnames. (However, for practical reasons, using some system-installed dependencies, i.e. programs in /usr, is OK.) 1. Support --prefix
(or an equivalent). 1. Support build dirs that are separate from source dirs (not essential, but preferable). 1. Support the DESTDIR
argument to make install
or an equivalent. (If DESTDIR
is missing, we can work around this using --prefix
, but it's a pain.)
We contrast modular build systems to monolithic build systems. The main characteristic of a monolithic build is that any part of the build can reference any other part via relative filenames.
There are three areas where we might wish to apply this: * The toolchain: binutils, gcc (pre-gcc and the full gcc), newlib, libnacl, pthread library, etc. This is currently built by tools/Makefile, which invokes Scons. There is a separate script for building the ARM toolchain. * The trusted codebase: sel_ldr, ncval, NPAPI plugin. Built by Scons. * Package collection (native_client_sdk in SVN). Built by shell scripts. (The toolchain and package collection are both sometimes referred to as “the SDK”, but I'm using different terms here in order to distinguish them.)
In the long term, the package collection is the most important part to modularise, because we expect it to become large and take a long time to build. However, the package collection is less of a priority in the short term because fewer people are working on it. Modularising the NaCl Scons build is less of a priority because it builds relatively quickly. Therefore I am focusing initially on modularising the toolchain build.
The toolchain build takes a relatively long time, and covering multiple architectures makes this worse. Currently our Buildbot-based continuous integration system rebuilds the toolchain in a cycle that is separate from the main build. The trybots don't rebuild the toolchain because that would take too long for most changes; the trybots cannot tell what parts of the toolchain need to be rebuilt. This means the trybots are not very useful for testing toolchain changes. This situation could be improved if the toolchain could be rebuilt incrementally.
I have written a prototype replacement for tools/Makefile.
git clone git://github.com/mseaborn/nacl-modular-build.git
This is now in the NaCl tree, under [tools/modular-build] (http://code.google.com/p/nativeclient/source/browse/trunk/src/native_client/tools/modular-build).
nacl-newlib expects to get NaCl header files from its source tree. This means tools/Makefile must copy the header files into the newlib source tree. This violates the modular build model. It is better if we can separate source directories from build directories, so that source directories are not modified during the build.
--with-headers=path
option similar to what glibc provides.nacl-newlib installs 32-bit libraries into lib/32, and tools/Makefile works around this to make them visible in lib32.
NaCl's Scons build uses “naclsdk_mode=custom:...” as both an input tree and an output tree for nacl_extra_sdk. build.py tries to use this argument to specify the output tree only (it passes the path of an empty directory). We could tell scons how to find nacl-gcc by setting PATH instead. Scons resets PATH by default, but we disable this by passing USE_ENVIRON=1 (an option that was added for this purpose).
Is there a way to tell Scons to use a build directory other than scons-out
? We need a way to ensure that Scons-built files are rebuilt whenever nacl-gcc is rebuilt.
The toolchain patches now require -p2
to apply. (This is a minor point, but it is a divergence from the usual practice.)
Missing features in the modular build widely varies by source. Some features can be easily added by implementing Makefile-like functionality that is just not there yet. Some other features might require an in-depth rewrite of important modular-build concepts. The list is not sorted by importance.
The generic part of the build should be better documented. Take AddAutoconfModule
as example. It accepts arguments through **kwargs
that are not easy to follow by reading sources. Examples of parameters with somewhat cryptic names: explicitly_passed_deps
, use_install_root
. The directory structure of the build must be documented too. What is the concept of ‘install-prefix’ and how is it different from ‘install’?
The build should support building from sources prepared by a human as opposed to be checking out sources from known HTTP locations automatically. I.e. it should be possible to fetch all sources somehow, then modular-build them. In the current implementation one has to make a full build first to fetch all sources, then remove everything except out/source
, replace with my own sources, then rebuild. This could be solved by having all source targets grouped under a single target.
Document incremental rebuild options. Currently it is hard to see a result of changing something in gcc sources by rebuilding the whole toolchain incrementally. Should an out/something
directory be removed or should build.py
be invoked with some parameters to invoke an incremental rebuild? Ideally it should allow asking modules if they can incrementally rebuild by themselves. This does not always have to be clean in terms of the final correctness. Sometimes a developer wants to rebuild just one component (or only those components that changed) and install the results to the destination plus maybe rebuild (some) dependent components. In this situation a developer knowingly sacrifices correctness in favor of faster build.
The build should print all parameters of the scripts invoked by system() (make, configure). It is significantly harder to debug problems if all invocations are not logged properly.
The build should log file operations that it performs. Currently all ad hoc operations performed are not logged which makes it hard to understand sources of many problems. For example, treemappers
that manipulate file trees should record their file operations in the log. Preferably into something like ‘cp -r A/ B’, not something like ‘cp A/1 A/2 A/3 A/4/4 B’. Once run on buildbots this should make it easier to debug problems.
Parallel build option analogous to -j in GNU Make.
Reduce the amount of copies of the same file. This is probably unfixable. For example, there are 3 places where sys/nacl_imc_api.h is being copied during the build: out/headers
, out/headers/nacl/abi
, out/input-prefix/glibc_toolchain/nacl/include
(the tools/Makefile build has only one replica). This makes it harder for a novice user to follow where this file comes from and whether it changes on the way.
The build should improve the readability of it's rules. It shold look rather as a static datafile description. Currently it is a complex program build.py
that has loops, conditionals os.path.join
, os.path.abspath
trickery, etc.