breadcrumbs: PNaCl >
This page is intended for developers making changes to the PNaCl tools. If you just want to build PNaCl-based applications, see the Native Client developer documentation. If you just want to build the toolchain, e.g. to package PNaCl or Chromium with NaCl support, see the shorter document on building.
In addition to what is needed to [build nacl], several packages may need to be installed on your system. For Ubuntu systems, e.g.:
sudo apt-get install git cmake texinfo flex bison ccache g++ g++-multilib gcc-arm-linux-gnueabihf g++-arm-linux-gnueabihf
test -e /usr/include/asm || sudo ln -s /usr/include/asm-generic /usr/include/asm
To build the PNaCl toolchain, you will need to check out the entire native client project with git. After that:
cd $NACL_DIR/native_client
Once you have the native client project checked out, you can check out the PNaCl sources.
Pre-built binaries for PNaCl are downloaded from the DEPS file by gclient, and are located in one of these directories, depending upon your architecture and operating system.
toolchain/linux_x86/pnacl_newlib_raw toolchain/mac_x86/pnacl_newlib_raw toolchain/win_x86/pnacl_newlib_raw
If you want to use these binaries outside of Scons, you must build and install the SDK libraries. This installs the libraries such as libnacl from the NaCl repo that go into the SDK and are needed to link programs without any special flags. (Note that this does not install the Pepper libraries, as those come from the Chromium repository):
pnacl/build.sh sdk
At this point, you probably want to test if the pre-built toolchain works. Let's start by creating a “hello world” program:
#include <stdio.h> int main(int argc, const char *argv[]) { printf("Hello World!\n"); return 0; }
Add the PNaCl toolchain to your path for convenience, if desired:
PATH=${PATH}:toolchain/linux_x86/pnacl_newlib/bin # Adjust the name for your system
Now compile hello.c:
pnacl-clang hello.c -o hello.nonfinal.pexe pnacl-finalize hello.nonfinal.pexe -o hello.pexe
(NOTE: If you get an error about missing crt1.o or -lnacl, you must run the SDK install step. See the installation instructions in the previous section.)
You have now produced a “pexe”, a PNaCl bitcode application. To run this pexe, it must be first translated to native code. Let's translate it to X86-32, X86-64 and ARM.
pnacl-translate -arch x86-32 hello.pexe -o hello_x86_32.nexe pnacl-translate -arch x86-64 hello.pexe -o hello_x86_64.nexe pnacl-translate -arch arm hello.pexe -o hello_arm.nexe
Now you have 3 native executables, one for each architecture. To run them inside of sel_ldr
, use the helper script, native_client/run.py
:
./run.py hello_x86_32.nexe ./run.py hello_x86_64.nexe ./run.py hello_arm.nexe # By default it will look for qemu to run the ARM nexe.
To use the PNaCl toolchain with the native client scons tests, just specify “bitcode=1” on the command-line.
Some sample command lines:
./scons platform=arm bitcode=1 # build everything ./scons platform=arm bitcode=1 smoke_tests # run smoke_tests ./scons platform=x86-32 bitcode=1 # build everything ./scons platform=x86-32 bitcode=1 smoke_tests # run smoke_tests ./scons platform=x86-64 bitcode=1 # build everything ./scons platform=x86-64 bitcode=1 smoke_tests # run smoke_tests ./scons -j32 -k --verbose ... # parallelize, don't abort on first error, verbose output ./scons platform=... bitcode=1 use_sz=1 ... # run using pnacl-sz instead of pnacl-llc
The scons tests are more conveniently wrapped in a testing script:
pnacl/test.sh test-x86-32 # Runs the smoke tests as above, plus nonpexe_tests pnacl/test.sh test-x86-64 pnacl/test.sh test-arm pnacl/test.sh test-x86-32-sbtc # Runs the same tests, but with the sandboxed translator pnacl/test.sh test-x86-64-sbtc pnacl/test.sh test-arm-sbtc
These tests can be run directly by running ‘make check’ in the build directory (pnacl/build/llvm_x86_64). There are currently a few known failures which are not marked as XFAIL in the usual LLVM way, so instead the regression tests can also be run in the following way, which will ignore the known failures. (This only works after you have built the PNaCl LLVM and have the build directory available).
pnacl/scripts/llvm-test.py --llvm-regression
The LLVM regression tests run quickly and should be run before every LLVM commit, along with the scons tests.
The following tests take longer and are generally not run by the developers on every commit, but they should be run for large changes, and are run on the toolchain testing bots on every commit.
To run the gcc torture test suite against PNaCl, do the following:
tools/toolchain_tester/torture_test.py pnacl x86-64 --concurrency=12
Other options are documented in torture_test.py and tools/toolchain_tester/toolchain_tester.py
To run the LLVM test suite against PNaCl, do
pnacl/scripts/llvm-test.py --testsuite-all --arch x86-64
This is equivalent to
pnacl/scripts/llvm-test.py --testsuite-prereq --arch x86-64 pnacl/scripts/llvm-test.py --testsuite-clean pnacl/scripts/llvm-test.py --testsuite-configure pnacl/scripts/llvm-test.py --testsuite-run --arch x86-64 pnacl/scripts/llvm-test.py --testsuite-report --arch x86-64
The test suite needs to be cleaned and re-configured each time you switch architectures, but the results files persist after cleaning, so testsuite-report will give the result from the last run for that architecture. Because there is no good way to filter the set of tests that are actually run and there are a lot of known failures (e.g. tests that have not been ported to use newlib), this test suite takes a long time to run.
You need to set LD_LIBRARY_PATH before running the test suite, otherwise the native build of each test, whose output the PNaCl build is validated against, will fail to run. It's also helpful to set PNACL_CONCURRENCY.
LD_LIBRARY_PATH=$NACL_DIR/native_client/toolchain/linux_x86/pnacl_newlib/lib PNACL_CONCURRENCY=12 pnacl/scripts/llvm-test.py --testsuite-all --arch x86-64
If you have access to the spec2k benchmark suite you can use the harness in tests/spec2k/ to run some more extensive tests.
If SPEC_DIR is the directory containing the benchmark code/data, run:
pnacl/test.sh test-spec <SPEC_DIR> SetupPnaclX8664Opt
test.sh is a convenience script and clobbers old results, etc. The actual steps are:
pushd tests/spec2k ./run_all.sh CleanBenchmarks ./run_all.sh PopulateFromSpecHarness <SPEC_DIR> ./run_all.sh TimedBuildAndRunBenchmarks SETUP popd
where SETUP is for example “SetupPnaclArmOpt”. See “run_all.sh” for a full list of setups.
These tests depend on having built PNaCl's branch of LLVM from source (uses some of the headers, etc.). See “checking out PNaCl sources, building and testing” below.
In general look at the Makefile.standalone file for various targets and options. E.g., you can compile with DEBUG=1 while debugging, so that variables aren‘t optimized-out and functions aren’t so inlined, etc.
cd toolchain_build/src/subzero make -j32 -f Makefile.standalone check
If you are using Chromium's build of clang to build PNaCl (the default) and you have never done so, you need to install it.
From the native_client/ directory:
../tools/clang/scripts/update.py
From now on, when you run ‘gclient sync’ clang will also be updated if needed.
Now you can build the PNaCl toolchain:
toolchain_build/toolchain_build_pnacl.py --verbose --sync --clobber --install toolchain/linux_x86/pnacl_newlib_raw rm -rf toolchain/linux_x86/pnacl_newlib ln -s pnacl_newlib_raw toolchain/linux_x86/pnacl_newlib # syncs the sources and builds the developer toolchain # Add the --build-sbtc flag to also build the sandboxed translator
If everything built correctly, run the scons tests as described above.
Most tasks related to the toolchains are focused a few scripts.
pnacl/driver/
toolchain_build/toolchain_build_pnacl.py
This script checks out and builds the complete PNaCl toolchain. Run with the -h flag for a full list of flags and build targets. Here are the most important flags:
optional arguments: --build-sbtc Build the sandboxed translators ... Usage: toolchain_build_pnacl.py [options] [targets...] ... Options: -h, --help show this help message and exit -v, --verbose Produce more output. -c, --clobber Clobber working directories before building. -y, --sync Run source target commands [i.e. sync the targets' sources before building them] -i, --ignore-dependencies Ignore target dependencies and build only the specified target. --install=INSTALL After building, copy contents of build packages to the specified directory
The build.sh script is now only used to build the sandboxed translator. It has 2 relevant arguments:
However, it is called from toolchain_build_pnacl.py for the bots, and it expects to use the toolchain built by the translator_compiler target of toolchain_build_pnacl.py. Use the --build-sbtc flag of toolchain_build_pnacl.py and build everything (or just the sandboxed_translators target and its dependencies) for that.
This script runs toolchain tests. It deletes old testing artifacts before running. The main invocation is test-all.
This directory contains the compiler drivers, which are in Python. If you change them, run e.g. “toolchain_build/toolchain_build_pnacl.py --install toolchain/linux_x86/pnacl_newlib driver_i686_linux” to install them.
Git Repositories
Most of the code required to build the PNaCl toolchain lives in git repositories hosted on chromium.googlesource.com. The build.sh script will automatically check out and update these repositories. The checked-out code is placed in toolchain_build/src.
Relevant Directories
toolchain_build/src holds all the git checkouts managed by toolchain_build_pnacl.py toolchain/${os}_x86/pnacl_newlib root of PNaCl toolchain installed by package_version.py toolchain_build/out holds build directories for the pnacl toolchain toolchain_build/out/toolchain_build.log holds the logs of the builds
If you've made changes to PNaCl toolchain, either to the driver or git repositories, you can rebuild PNaCl by doing:
toolchain_build/toolchain_build_pnacl.py -v --install toolchain/linux_x86/pnacl_newlib_raw # Incrementally build and reinstall pnacl/test.sh test-all # Run the scons tests
To send out reviews for the git repos, you must have depot tools. Code reviews are sent using the git cl script in depot_tools.
If you are going to commit, you will need to have a credential in your ~/.netrc
file:
The first time you set up code review you'll need to let git cl
do some setup:
**```none cd toolchain_build/src/llvm git cl config
Just press Enter at the prompts. To do local development, create a new branch: **```none git checkout -b mybranch origin/master ```** **The last argument to git checkout (the start point) is important because if your local branch does not track the remote master branch, then `git cl land` will not work)** Edit your code, and commit as many times as you want: **```none git commit -a -m “comment describing changes” ```** **Then use git-cl to upload to the code review site:** **```none git cl upload ```** You'll need to authenticate to codereview.chromium.org also. Most likely this will be your ldap email. This will upload the review, and give you an issue number, e.g. http://codereview.chromium.org/1557002. Go to the issue webpage, and edit the change summary, add reviewers, then hit send mail. Command line options from gcl such as reviewers ("-r") and send mail ("--send_mail") are also available with git cl. Git cl associates each branch with a review issue ID. So you if you `git cl upload` again from the same branch, it will upload a new patch set to the same issue. When you get the LGTM, you can push the changes. ```none git cl land
If there have been changes on the master branch since you forked your local branch, git cl may complain, in which case you should rebase it against the new HEAD before you push:
git fetch origin git rebase origin/master
At this point the official toolchain builder is still not incorporating your changes as it builds off of specific commit IDs in the external repo.
To fix this modify native_client/pnacl/COMPONENT_REVISIONS to point at your new commit ID which look like
llvm=efa7b3d5c85f0242787eb091dadedc2727215df4
You can then make a standard native_client change list and test the CL with trybots (CLs affecting pnacl/ will end up selecting pnacl trybots). Upload, trybot, get an LGTM, then commit. You can also test locally by cleaning the toolchain directory, rebuilding, and testing as described earlier.
Or, you can automatically update the file using a script: pnacl/deps_update.py
Finally, to allow users to download the new revision built by the toolchain builder, look in the toolchain_revisions/pnacl_newlib.json file. The current revision is in a line near the end:
"revision": 13392,
Update this to an svn revision number that is new enough to include your COMPONENT_REVISIONS change by running:
build/package_version/package_version.py setrevision --revision-package pnacl_newlib --revision 12345
Now create a CL and test it with some try jobs (e.g., git cl try my_cl_name
). NOTE: You will need to wait for a toolchain to be built (see the waterfall) before this change actually works (downloads a toolchain tarball). Also see build/update_pnacl_tool_revisions.py to automate this process.
Building Validators
In the case that you want to run the native client validators on your nexes, build the validators like so:
./scons targetplatform=arm sdl=none arm-ncval-core ./scons platform=x86-32 sdl=none ncval ./scons platform=x86-64 sdl=none ncval
The validators will be placed in scons-out/{host}-{platform}/staging/.