This document is work-in progress. Don't expect much from it yet.

The Boost.Build code is structured in four different components: "kernel", "util", "build" and "tools". The first two are relatively uninteresting, so we'll focus on the remaining pair. The "build" component provides classes necessary to declare targets, determine which properties should be used for their building, and for creating the dependency graph. The "tools" component provides user-visible functionality. It mostly allows to declare specific kind of main targets, and declare avaiable tools, which are then used when creating the dependency graph.

The build layer has just four main parts -- metatargets (abstract targets), virtual targets, generators and properties. Metatargets (see the "targets.jam" module) represent all the user-defined entities which can be built. The "meta" prefix signify that they don't really corrspond to files -- depending of build request, they can produce different set of files. Metatargets are created when Jamfiles are loaded. Each metagarget has a generate method which is given a property set and produces virtual targets for the passed properties. Virtual targets (see the "virtual-targets.jam" module) correspond to the atomic things which can be updated -- most typically files. Properties are just (name, value) pairs, specified by the user and describing how the targets should be built. Properties are stored using the property-set class. Generators are the objects which encapsulate tools -- they can take a list of source virtual targets and produce new virtual targets from them. The build process includes those steps: Top-level code calls the generate method of a metatarget with some properties. The metatarget combines the requested properties with requirements and passes the result, together with the list of sources, to the generators.construct function A generator appropriate for the build properties is selected and its run method is called. The method returns a list of virtual targets The targets are returned to the top level code. They are converted into bjam targets (via virtual-target.actualize) and passed to bjam for building.

There are several classes derived from "abstract-target". The "main-target" class represents top-level main target, the "project-target" acts like container for all main targets, and "basic-target" class is a base class for all further target types. Since each main target can have several alternatives, all top-level target objects are just containers, referring to "real" main target classes. The type is that container is "main-target". For example, given: alias a ; lib a : a.cpp : <toolset>gcc ; we would have one-top level instance of "main-target-class", which will contain one instance of "alias-target-class" and one instance of "lib-target-class". The "generate" method of "main-target" decides which of the alternative should be used, and call "generate" on the corresponding instance. Each alternative is a instance of a class derived from "basic-target". The "basic-target.generate" does several things that are always should be done: Determines what properties should be used for building the target. This includes looking at requested properties, requirements, and usage requirements of all sources. Builds all sources Computes the usage requirements which should be passes back. For the real work of constructing virtual target, a new method "construct" is called. The "construct" method can be implemented in any way by classes derived from "basic-target", but one specific derived class plays the central role -- "typed-target". That class holds the desired type of file to be produces, and calls the generators modules to do the job. This means that a specific metatarget subclass may avoid using generators at all. However, this is deprecated and we're trying to eliminate all such subsclasses at the moment. Note that the build/targets.jam file contains an UML diagram which might help.

Virtual targets correspond to the atomic things which can be updated. Each virtual target can be assigned an updating action -- instance of the action class. The action class, in turn, contains a list of source targets, properties, and a name of bjam action block which should be executed. We try hard to never create equal instances of the virtual-target class. Each code which creates virtual targets passes them though the virtual-target.register function, which detects if a target with the same name, sources, and properties was created. In that case, existing target is returned. When all virtual targets are produced, they are "actualized". This means that the real file names are computed, and the commands that should be run are generated. This is done by the virtual-target.actualize method and the action.actualize methods. The first is conceptually simple, while the second need additional explanation. The commands in bjam are generated in two-stage process. First, a rule with the appropriate name (for example "gcc.compile") is called and is given the names of targets. The rule sets some variables, like "OPTIONS". After that, the command string is taken, and variable are substitutes, so use of OPTIONS inside the command string become the real compile options. Boost.Build added a third stage to simplify things. It's now possible to automatically convert properties to appropriate assignments to variables. For example, <debug-symbols>on would add "-g" to the OPTIONS variable, without requiring to manually add this logic to gcc.compile. This functionality is part of the "toolset" module. Note that the build/virtual-targets.jam file contains an UML diagram which might help.

Above, we noted that metatargets are built with a set of properties. That set is represented with the property-set class. An important point is that handling of property sets can get very expensive. For that reason, we make sure that for each set of (name, value) pairs only one property-set instance is created. The property-set uses extensive caching for all operation, so most work is avoided. The property-set.create is the factory function which should be used to create instances of the property-set class.

Write me!

NOTE: THIS SECTION IS NOT EXPECTED TO BE READ! There are two user-visible kinds of targets in Boost.Build. First are "abstract" — they correspond to things declared by user, for example, projects and executable files. The primary thing about abstract target is that it's possible to request them to be build with a particular values of some properties. Each combination of properties may possible yield different set of real file, so abstract target do not have a direct correspondence with files. File targets, on the contary, are associated with concrete files. Dependency graphs for abstract targets with specific properties are constructed from file targets. User has no was to create file targets, however it can specify rules that detect file type for sources, and also rules for transforming between file targets of different types. That information is used in constructing dependency graph, as desribed in the "next section". [ link? ] Note:File targets are not the same as targets in Jam sense; the latter are created from file targets at the latest possible moment. Note:"File target" is a proposed name for what we call virtual targets. It it more understandable by users, but has one problem: virtual targets can potentially be "phony", and not correspond to any file.

Dependency scanning is the process of finding implicit dependencies, like "#include" statements in C++. The requirements for right dependency scanning mechanism are: Support for different scanning algorithms. C++ and XML have quite different syntax for includes and rules for looking up included files. Ability to scan the same file several times. For example, single C++ file can be compiled with different include paths. Proper detection of dependencies on generated files. Proper detection of dependencies from generated file.

Different scanning algorithm are encapsulated by objects called "scanners". Please see the documentation for "scanner" module for more details.

As said above, it's possible to compile a C++ file twice, with different include paths. Therefore, include dependencies for those compilations can be different. The problem is that bjam does not allow several scans of the same target. The solution in Boost.Build is straigtforward. When a virtual target is converted to bjam target (via virtual-target.actualize method), we specify the scanner object to be used. The actualize method will create different bjam targets for different scanners. All targets with specific scanner are made dependent on target without scanner, which target is always created. This is done in case the target is updated. The updating action will be associated with target without scanner, but if sources for that action are touched, all targets — with scanner and without should be considered outdated. For example, assume that "a.cpp" is compiled by two compilers with different include path. It's also copied into some install location. In turn, it's produced from "a.verbatim". The dependency graph will look like: a.o (<toolset>gcc) <--(compile)-- a.cpp (scanner1) ----+ a.o (<toolset>msvc) <--(compile)-- a.cpp (scanner2) ----| a.cpp (installed copy) <--(copy) ----------------------- a.cpp (no scanner) ^ | a.verbose --------------------------------+

This requirement breaks down to the following ones. If when compiling "a.cpp" there's include of "a.h", the "dir" directory is in include path, and a target called "a.h" will be generated to "dir", then bjam should discover the include, and create "a.h" before compiling "a.cpp". Since almost always Boost.Build generates targets to a "bin" directory, it should be supported as well. I.e. in the scanario above, Jamfile in "dir" might create a main target, which generates "a.h". The file will be generated to "dir/bin" directory, but we still have to recornize the dependency. The first requirement means that when determining what "a.h" means, when found in "a.cpp", we have to iterate over all directories in include paths, checking for each one: If there's file "a.h" in that directory, or If there's a target called "a.h", which will be generated to that directory. Classic Jam has built-in facilities for point (1) above, but that's not enough. It's hard to implement the right semantic without builtin support. For example, we could try to check if there's targer called "a.h" somewhere in dependency graph, and add a dependency to it. The problem is that without search in include path, the semantic may be incorrect. For example, one can have an action which generated some "dummy" header, for system which don't have the native one. Naturally, we don't want to depend on that generated header on platforms where native one is included. There are two design choices for builtin support. Suppose we have files a.cpp and b.cpp, and each one includes header.h, generated by some action. Dependency graph created by classic jam would look like: a.cpp -----> <scanner1>header.h [search path: d1, d2, d3] <d2>header.h --------> header.y [generated in d2] b.cpp -----> <scanner2>header.h [ search path: d1, d2, d4] In this case, Jam thinks all header.h target are not realated. The right dependency graph might be: a.cpp ---- \ \ >----> <d2>header.h --------> header.y / [generated in d2] / b.cpp ---- or a.cpp -----> <scanner1>header.h [search path: d1, d2, d3] | (includes) V <d2>header.h --------> header.y [generated in d2] ^ (includes) | b.cpp -----> <scanner2>header.h [ search path: d1, d2, d4] The first alternative was used for some time. The problem however is: what include paths should be used when scanning header.h? The second alternative was suggested by Matt Armstrong. It has similiar effect: add targets which depend on <scanner1>header.h will also depend on <d2>header.h. But now we have two different target with two different scanners, and those targets can be scanned independently. The problem of first alternative is avoided, so the second alternative is implemented now. The second sub-requirements is that targets generated to "bin" directory are handled as well. Boost.Build implements semi-automatic approach. When compiling C++ files the process is: The main target to which compiled file belongs is found. All other main targets that the found one depends on are found. Those include main target which are used as sources, or present as values of "dependency" features. All directories where files belonging to those main target will be generated are added to the include path. After this is done, dependencies are found by the approach explained previously. Note that if a target uses generated headers from other main target, that main target should be explicitly specified as dependency property. It would be better to lift this requirement, but it seems not very problematic in practice. For target types other than C++, adding of include paths must be implemented anew.

Suppose file "a.cpp" includes "a.h" and both are generated by some action. Note that classic jam has two stages. In first stage dependency graph graph is build and actions which should be run are determined. In second stage the actions are executed. Initially, neither file exists, so the include is not found. As the result, jam might attempt to compile a.cpp before creating a.h, and compilation will fail. The solution in Boost.Jam is to perform additional dependency scans after targets are updated. This break separation between build stages in jam — which some people consider a good thing — but I'm not aware of any better solution. In order to understand the rest of this section, you better read some details about jam dependency scanning, available at this link. Whenever a target is updated, Boost.Jam rescans it for includes. Consider this graph, created before any actions are run. A -------> C ----> C.pro / B --/ C-includes ---> D Both A and B have dependency on C and C-includes (the latter dependency is not shown). Say during building we've tried to create A, then tried to create C and successfully created C. In that case, the set of includes in C might well have changed. We do not bother to detect precisely which includes were added or removed. Instead we create another internal node C-includes-2. Then we determine what actions should be run to update the target. In fact this mean that we perform logic of first stage while already executing stage. After actions for C-includes-2 are determined, we add C-includes-2 to the list of A's dependents, and stage 2 proceeds as usual. Unfortunately, we can't do the same with target B, since when it's not visited, C target does not know B depends on it. So, we add a flag to C which tells and it was rescanned. When visiting B target, the flag is notices and C-includes-2 will be added to the list of B's dependencies. Note also that internal nodes are sometimes updated too. Consider this dependency graph: a.o ---> a.cpp a.cpp-includes --> a.h (scanned) a.h-includes ------> a.h (generated) | | a.pro <-------------------------------------------+ Here, out handling of generated headers come into play. Say that a.h exists but is out of date with respect to "a.pro", then "a.h (generated)" and "a.h-includes" will be marking for updating, but "a.h (scanned)" won't be marked. We have to rescan "a.h" file after it's created, but since "a.h (generated)" has no scanner associated with it, it's only possible to rescan "a.h" after "a.h-includes" target was updated. Tbe above consideration lead to decision that we'll rescan a target whenever it's updated, no matter if this target is internal or not. The remainder of this document is not indended to be read at all. This will be rearranged in future.

As described above, file targets corresponds to files that Boost.Build manages. User's may be concerned about file targets in three ways: when declaring file target types, when declaring transformations between types, and when determining where file target will be placed. File targets can also be connected with actions, that determine how the target is created. Both file targets and actions are implemented in the virtual-target module.

A file target can be given a file, which determines what transformations can be applied to the file. The type.register rule declares new types. File type can also be assigned a scanner, which is used to find implicit dependencies. See "dependency scanning" [ link? ] below.

To distinguish targets build with different properties, they are put in different directories. Rules for determining target paths are given below: All targets are placed under directory corresponding to the project where they are defined. Each non free, non incidental property cause an additional element to be added to the target path. That element has the form <feature-name>-<feature-value> for ordinary features and <feature-value> for implicit ones. [Note about composite features]. If the set of free, non incidental properties is different from the set of free, non incidental properties for the project in which the main target that uses the target is defined, a part of the form main_target-<name> is added to the target path. Note:It would be nice to completely track free features also, but this appears to be complex and not extremely needed. For example, we might have these paths: debug/optimization-off debug/main-target-a