Ekam Build System
Switch branches/tags
Nothing to show
Clone or download
kentonv Use regular qemu for cross-tests, not qemu-static.
TBH I don't know why I ever used qemu-static.
Latest commit aa3ee6f Aug 8, 2018

README.md

Ekam Build System

Ekam ("make" backwards) is a build system which automatically figures out what to build and how to build it purely based on the source code. No separate "makefile" is needed.

Ekam works by exploration. For example, when it encounters a file ending in ".cpp", it tries to compile the file, intercepting system calls to find out its dependencies (e.g. included headers). If some of these are missing, Ekam continues to explore until it finds headers matching them. When Ekam builds an object file and discovers that it contains a "main" symbol, it tries to link it, searching for other object files to satisfy all symbol references therein. When Ekam sees a test, it runs the test, again intercepting system calls to dynamically discover what inputs the test needs (which may not have been built yet). Ekam can be extended to understand new file types by writing simple shell scripts telling it what to do with them.

Thus Ekam is, in fact, Make in reverse. Make starts by reading a Makefile, sees what executables it wants to build, then from there figures out what source files need to be compiled to link into them, then compiles them. Ekam starts by looking for source files to compile, then determines what executables it can build from them, and, in the end, might output a Makefile describing what it did.

Ekam is a work in progress.

Building Ekam

Warning

Ekam is an experimental project that is not ready for wide use.

That said, I (Kenton) have successfully used Ekam as my primary build system throughout the development of Cap'n Proto and Sandstorm.

Supported Platforms

At this time, Ekam only runs on Linux. It requires GCC 4.8+ or Clang 3.3+, as it uses C++11 features.

In the past, Ekam worked on FreeBSD and Max OSX, but the code to support that atrophied and was eventually deleted. Ekam uses a lot of OS-specific hacks and so is unlikely to work on any platform which is not explicitly supported.

We'd like to see other platforms supported, but Ekam's primary user and maintainer right now (Sandstorm.io) is itself highly Linux-specific, so there is not much pressure. (Let us know if you want to help.)

Bootstrapping the build

Download and compile Ekam like so:

git clone https://github.com/sandstorm-io/ekam.git
cd ekam
make

If successful, Ekam should have built itself, with the output binary at "bin/ekam".

Yes, we use make in order bootstrap Ekam, mostly just because it's slightly nicer than a shell script.

Compiling Ekam with Ekam

Compiling Ekam requires the GCC flag -std=gnu++0x to enable C++11 features, but currently there is no way for the code itself to specify compiler flags that it requires. You can only specify them via environment variable. So, to build Ekam with Ekam, type this command at the top of the Ekam repository:

CXXFLAGS=-std=gnu++0x ekam -j4

The -j4 tells Ekam to run up to four tasks at once. You may want to adjust this number depending on how many CPU cores you have.

Note that Ekam looks for a directory called src within the current directory, and scans it for source code. The Ekam source repository is already set up with such a src subdirectory containing the Ekam code. You could, however, place the entire Ekam repository inside some other directory called src, and then run Ekam from the directory above that, and it will still find the code. The Protocol Buffers instructions below will take advantage of this to create a directory tree containing both Ekam and protobufs.

Ekam places its output in siblings of src called tmp (for intermediate files), bin (for output binaries), lib (for output libraries, although currently Ekam doesn't support building libraries), etc. These are intended to model Unix directory tree conventions.

Continuous Building

If you invoke Ekam with the -c option, it will watch the source tree for changes and rebuild derived files as needed. In this way, you can simply leave Ekam running while you work on your code, and get information about errors almost immediately on saving.

Note that continuous building is the only way to do incremental builds with Ekam -- any time you run a new Ekam process, it always starts from scratch. I generally just leave Ekam running in a console window 24/7.

IDE plugins and other external clients

Ekam can, while running, export a network interface which allows other programs to query the state of the build, including receiving the task tree and error logs.

Ekam Server

If you pass the -n flag to Ekam, it will listen for connections on a port and stream build status updates to anyone who connects. Invoke like:

ekam -n :41315

Ekam Client

ekam-client is a very simple program that prints an Ekam build status stream to the console exactly as Ekam itself does. Currently ekam-client doesn't actually know how to create a network connection but instead reads the stream from standard input, so you can invoke it like this:

nc localhost 41315 | ekam-client

ekam-client is mostly just a tech demo, since it displays the same info that is already visible in the console where Ekam itself is running.

Qt Creator Plugin

The qtcreator directory in the Ekam repository contains source code for a Qt Creator plugin. Qt Creator is an open source C++ IDE that, despite its name, is quite nice even for non-Qt code.

To build the Qt Creator plugin, you'll want to download the Qt Creator source code release and build it from scratch. Then, go into the qtcreator directory in the Ekam repo and do the following:

export QTC_SOURCE=<path to qt creator source directory>
qmake
make

This should build the plugin and install it directly into your Qt Creator build.

Now you can start Qt Creator and choose the "Ekam Actions" view in the navigation frame. The plugin connects to a local Ekam instance running on port 41315; so, start ekam like this:

ekam -c -n :41315

The plugin will also add error markers to you source code, visible in the issue tracker. Double-clicking on a failed rule in the tree view will navigate the issues view to the first message from that action, which you can in turn use to navigate to the source location of the error.

Using Ekam in your own code

Preferred project layout

An Ekam project directory must contain a directory called src which contains all source code. Ekam will ignore everything other than this src directory.

Ekam will create sibling directories called tmp, bin, and lib; you probably shouldn't have any files in those directories to start, because it's nice to be able to rm -rf them to clean.

Import the rule files

In order for Ekam to be able to build a project, it needs to find its rule files, which are under src/ekam/rules in Ekam's own source code. You will want to symlink this directory into your project's source tree somewhere. It does not have to be anywhere in particular, but src/ekam-rules is probably a good bet.

Compiler flags

You can set the following environment variables to control how Ekam compiles your code:

  • CXX: Sets the C++ compiler, e.g. CXX=clang++.
  • CXXFLAGS: Sets C++ compilation flags, e.g. CXXFLAGS=-std=c++11 -O2 -Wall.
  • LIBS: Sets linker flags, e.g. LIBS=-lsodium -lz

Building binaries

When Ekam finds or compiles a .o file that defines the symbol main, it will attempt to link the file with other .o files defining all needed symbols (transitively) in order to produce a binary. The binary takes the name of the .o with the extension removed, e.g. foo.o produces a binary foo.

The binary is initially just dropped into the tmp directory, which mirrors the src directory. For example, src/foo/bar.c++ will be compiled to src/foo/bar.o, which will link (if it defines main) into src/foo/bar. If you'd like for the binary to be output to the bin directory, you must declare a manifest file with the extension .ekam-manifest. This file is a simple text file where each line names a file and its output directory. For example:

bar bin

This says: "Once you've built bar (within this directory), copy it into bin."

You can also choose to rename the file when installing:

bar bin/baz

This says: "Once you've built bar (within this directory), copy it to bin/baz."

Building libraries

Currently, Ekam does not support building libraries. This seems complicated to support since there's no automated way for Ekam to decide what makes a good library. You'd need to declare some set of modules representing the public interface, which is a bit sad.

For now, Ekam is suitable for projects where the final output is a binary. A "library" in Ekam is just a directory that contains some code. It does not build to a .a or .so file; instead, the objects may be directly linked into binaries found in other directories as needed.

Handling Magic Singleton Registries

There is a common pattern in C++ code in which a particular file's symbols are not referenced from other files, but instead the file registers itself in some global registry at program startup which makes its functionality discoverable to the rest of the program.

By default, this pattern does not work with Ekam, because Ekam has no way to tell which of these magic self-registering files should be linked into which binaries, since their symbols are not referenced from other .o files.

To solve this problem, don't. This design pattern sucks and you should not use it. Rewrite your code so that its functionality is not dependent on what objects were specified on the link line. For example, have your main() function explicitly call registerFoo(), registerBar(), etc., for all the magic modules you want registered, ideally passing a registry object to each function so you don't need a singleton registry.

See Singletons Considered Harmful for extended discussion.

(We make a special exception for test frameworks, described below.)

Debugging

If you want gdb to be able to find your code when debugging Ekam-built binaries, you should set up a couple symlinks as follows:

mkdir ekam-provider
ln -s ../src ekam-provider/canonical
ln -s ../src ekam-provider/c++header

To understand the reason for these symlinks, see the explanation of intercept.so later in this document.

Tests

If Ekam builds a binary which ends in -test or _test, it will run that binary and report pass/fail depending on the exit status. Passing tests are marked in green in the output (whereas regular successful build actions are blue) in order to help you develop a pavlovian attachment to writing tests and seeing them pass.

Ekam has additional built-in support for two test frameworks: Google Test and KJ tests. When Ekam compiles a source file which declares test cases using one of these frameworks (e.g. using Google Test's TEST or TEST_F macros, or KJ's KJ_TEST macro) but without a main function, it will automatically link against the respective framework's test runner (which supplies main) in order to produce a test binary. Note that the detection of test declarations is based on linker symbols, not on scanning the source code, so don't worry if you've declared your own wrapper macros.

In order for Google Test or KJ test integration to work, the respective test framework's code must be in your source tree as a dependency (see below).

Note that tests are run with intercept.so injected, which has implications if your test does any filesystem access. See the explanation of intercept.so later in this document.

Dependencies

If your project depends on other projects, and you want to build those other projects as part of your own build (rather than require the user to install the libraries on their system), you should follow the pattern that Ekam itself does.

The basic idea is to have a directory called deps into which you download the repositories of your external dependencies. Then, under src, place symlinks that deep-link into deps and pull out the code you want.

For example, to depend on Cap'n Proto, you would do something like:

mkdir deps
git clone https://github.com/sandstorm-io/capnproto.git deps/capnproto
ln -s ../deps/capnproto/c++/src/{kj,capnp} src

Note that deps should not be checked into your repository. Instead, you should provide a script that people run after cloning your repo that downloads dependencies. For Ekam's own build, we handle this in the Makefile, so you can look at that for an example. We may eventually add functionality to Ekam to handle dependencies so that you no longer need a Makefile for this.

Now, here's the fun part: If you work on a lot of different projects, instead of nesting dependencies as described above, you can symlink deps to .., such that all your git clones are siblings:

rm -rf deps
ln -s .. deps

You can do this with Ekam, for instance, if you happen to have Cap'n Proto cloned as a sibling to Ekam.

Non-Ekam-clean Dependencies

If your dependency is not organized for Ekam, you might need to link to its top-level directory rather than a source subdirectory. For instance, Google Test separates its source code into src and include directories, so if you only link in its src you'll miss the headers. Instead, link in the whole repo like:

ln -s ../../gtest src

When Ekam sees a directory named include, it adds that directory to the include path for all other code it compiles. Thus the Google Test headers will end up includable by your code.

Note that projects imported this way will usually have a bunch of errors reported since they are not Ekam-clean. However, often (such as in the case of Google Test) the important parts will compile well enough to use.

Per-directory compile options

To specify some additional compiler options that should apply within a particular directory (and all children, recursively), you may create a file called compile.ekam-flags and populate it with simple variable assignments in shell syntax. For example:

# Define FOO to 1 when compiling code in this directory.
CXXFLAGS=$CXXFLAGS -DFOO=1

.ekam-flags files are actually executed using /bin/sh just before invoking the compiler. When multiple flags files are in-scope, the flags file in the outermost directory runs first.

Cross-compiling

Ekam currently supports cross-compiling to multiple target architectures at once, by listing additional (non-host) architectures in the CROSS_TARGETS environment variable:

CROSS_TARGETS="aarch64-linux-gnu" ekam

Notes:

  • This is designed to work e.g. with the crossbuild-essential-* Debian packages.
  • Ekam always compiles for the host architecture in addition to these targets. This is to support building tools like the Cap'n Proto compiler and then immediately using them in the same build.
  • Ekam assumes the inter-object dependencies are the same on all targets. This tends to mean that it work well for targetting alternate CPU architectures, but not as well for targeting other operating systems.
  • You may specify target-specific CXXFLAS and LIBS like CXXFLAGS_aarch64_linux_gnu and LIBS_aarch64_linux_gnu. If present, these completely replace the default CXXFLAGS and LIBS.
  • If any unit tests are built, Ekam will try to use qemu to run them.

Custom Rules

You may teach Ekam how to handle a new type of file -- or introduce a one-off build rule not triggered by any file -- by creating a rule file. A rule file is any executable with the extension .ekam-rule. Often, they are shell scripts, but this is not a requirement. Rule files can themselves be the output of other rules, so you could compile a C++ program that acts as a rule.

Ekam executes custom rules and then communicates with them on stdin/stdout using a simple line-based text protocol described below. The rule may also write error logs intended for the user to stderr.

When Ekam encounters an executable with the .ekam-rule extension, it first runs it with no arguments in order to "learn" it. At that time, the program may tell ekam what kinds of files it should trigger on using the trigger command (below). When Ekam encounters a trigger file, it will execute the rule again with the trigger file's canonical name as the argument. Alternatively, if you just want to implement a one-off action, then the rule may simply perform that action at "learn" time without registering any triggers.

Canonical file names

A file's "canonical" name is the name without the src/ or tmp/ prefix. Thus canonical names do not distinguish between source files and build outputs. No two files can have the same canonical name -- it is an error for a build action to output a file which exists in the src directory, or for two actions to produce the same file under tmp.

You can ask Ekam to map a canonical name to a physical disk name using the findInput command.

Tags

Ekam rules are triggered by tags. A tag is a simple string, conventionally of the format type:value. Tags are applied to files. For example, a .o file declaring the symbol main might be given the tag c++symbol:main. Rules may assign new tags to any file (source or output).

Ekam has some default rules that assign tags meant for broad use:

  • canonical:<filename>: Each file receives this tag, where <filename> is its canonical name.
  • filetype:<extension>: Each regular file that has a file type extension receives this tag, e.g. filetype:.c++.
  • directory:*: Each directory receives this tag.

Commands

A rule may perform the following commands by writing them to standard output.

  • trigger <tag>: Used during the learning phase to tell Ekam that the rule should be executed on any file tagged with <tag>.
  • verb <text>: Use during the learning phase to tell Ekam the rule's "verb", which is what is displayed to the user when the rule later runs. This should be a simple, descriptive word. For instance, for a C++ compile action, the verb is compile.
  • silent: Use during the learning phase to indicate that when this command later runs, it should not be reported to the user unless it fails. Use this to reduce noise caused by very simple commands that perform trivial actions.
  • findInput <file>: Obtains the canonical name of the given file. Ekam will reply by writing one line to the rule's standard input containing the full disk path of the file (e.g. including src/ or tmp/). Ekam will remember that the build action depended on this file, so if the file changes, the action will be re-run. If no match was found, Ekam will return a blank line.
  • findProvider <tag>: Find a file tagged with <tag>. If there are multiple matches, Ekam heuristically chooses the "preferred" one, which generally means the one closest in the directory tree to the file which triggered the rule. The path is returned as with findInput. Also as with findInput, the file is considered a dependency of the action. Ekam will re-run this action if the file changes or if the file Ekam chose to match <tag> changes.
  • findModifiers <name>: Search for the file <name> in the trigger file's directory and every parent up to the source root. For each place that it is found (in order starting from the greatest ancestor), return the full disk path and mark it as an input. After returning all results, return a blank line to indicate the end of the list. This command is intended for finding "modifier" files which specify options that should apply within a particular directory. For instance, compile.ekam-flags is implemented this way.
  • noteInput <external-file>: Tells Ekam that the action depends on <external-file>, which is a path outside of the project's source tree. For instance, /usr/include/stdlib.h. Currently Ekam ignores this, but in theory it could watch these files and re-run the action if they change.
  • newOutput <canonical-name>: Create a new output file with the given canonical name. Ekam replies by writing the on-disk path where the file should be created to the rule's standard input.
  • provide <filename> <tag>: Tag <filename> (a canonical name) with <tag>. The file must be a known input our output of this rule; i.e. it must have been the subeject of a previous call to findInput, findProvider, or newOutput.
  • install <filename> <location>: Take the canonical filename <filename> and copy it to <location>, where <location> should start with bin/, lib/, etc.
  • passed: Indicate that this action ran a test, and the test passed.

intercept.so

Sometimes, it's hard to know what a build tool's exact inputs and outputs will be ahead of time. For instance, a C++ compiler run will need to input all of the header files #includeed by the source file. There's no reasonable way to know what these might be in advance, much less look up the locations of files to satisfy each.

To solve this, Ekam implements a library which can be injected into a tool via LD_PRELOAD in order to intercept filesystem calls, automatically issues the proper Ekam commands to register inputs and outputs, and then map the call to the real disk path.

The interceptor library is implemented in src/ekam/rules/intercept.c and built by src/ekam/rules/intercept.ekam-rule. You may request it by requesting the tag special:ekam-interceptor, which maps to the compiled intercept.so, which you may then inject into an arbitrary command using LD_PRELOAD. See src/ekam/rules/compile.ekam-rule to see how this is done with the C++ compiler.

When filesystem calls are being intercepted, an attempt to open a regular filename will be heuristically mapped to the closest file whose canonical name contains the full requested name as a suffix. For example, when processing src/foo/bar/baz, if the tool tries to open qux/corge, this could map to src/foo/bar/qux/corge or src/qux/corge or src/grault/qux/corge, but not src/grault/corge nor src/qux/corge/grault nor src/corge.

If you want to open a file by tag, the special virtual path /ekam-provider/<tag-type>/<tag-value> can be used. For example, since the include.ekam-rule rule tags all C++ header files with c++header:<include-path>, when we invoke the compiler using the inteceptor, we pass the flag -I/ekam-provider/c++header.

If you want to open a file purely by its whole canonical path (not using the heuristic that finds nearby files), you may do so by opening /ekam-provider/canonical/<canonical-name>, since as described above every file gets tagged with canonical:<canonical-name>.

Get Involved

Have a question about Ekam, or want to contribute? Talk to us on the Ekam discussion group.

Ekam is currently developed as part of the Sandstorm.io project and is primarily used by Sandstorm. If you like Ekam, consider getting involved with Sandstorm.