Customizing LLVMC: Reference Manual

Contents

Written by Mikhail Glushenkov

Introduction

LLVMC is a generic compiler driver, designed to be customizable and extensible. It plays the same role for LLVM as the gcc program does for GCC - LLVMC's job is essentially to transform a set of input files into a set of targets depending on configuration rules and user options. What makes LLVMC different is that these transformation rules are completely customizable - in fact, LLVMC knows nothing about the specifics of transformation (even the command-line options are mostly not hard-coded) and regards the transformation structure as an abstract graph. The structure of this graph is completely determined by plugins, which can be either statically or dynamically linked. This makes it possible to easily adapt LLVMC for other purposes - for example, as a build tool for game resources.

Because LLVMC employs TableGen as its configuration language, you need to be familiar with it to customize LLVMC.

Compiling with LLVMC

LLVMC tries hard to be as compatible with gcc as possible, although there are some small differences. Most of the time, however, you shouldn't be able to notice them:

$ # This works as expected:
$ llvmc -O3 -Wall hello.cpp
$ ./a.out
hello

One nice feature of LLVMC is that one doesn't have to distinguish between different compilers for different languages (think g++ vs. gcc) - the right toolchain is chosen automatically based on input language names (which are, in turn, determined from file extensions). If you want to force files ending with ".c" to compile as C++, use the -x option, just like you would do it with gcc:

$ # hello.c is really a C++ file
$ llvmc -x c++ hello.c
$ ./a.out
hello

On the other hand, when using LLVMC as a linker to combine several C++ object files you should provide the --linker option since it's impossible for LLVMC to choose the right linker in that case:

$ llvmc -c hello.cpp
$ llvmc hello.o
[A lot of link-time errors skipped]
$ llvmc --linker=c++ hello.o
$ ./a.out
hello

By default, LLVMC uses llvm-gcc to compile the source code. It is also possible to choose the clang compiler with the -clang option.

Predefined options

LLVMC has some built-in options that can't be overridden in the configuration libraries:

Compiling LLVMC plugins

It's easiest to start working on your own LLVMC plugin by copying the skeleton project which lives under $LLVMC_DIR/plugins/Simple:

$ cd $LLVMC_DIR/plugins
$ cp -r Simple MyPlugin
$ cd MyPlugin
$ ls
Makefile PluginMain.cpp Simple.td

As you can see, our basic plugin consists of only two files (not counting the build script). Simple.td contains TableGen description of the compilation graph; its format is documented in the following sections. PluginMain.cpp is just a helper file used to compile the auto-generated C++ code produced from TableGen source. It can also contain hook definitions (see below).

The first thing that you should do is to change the LLVMC_PLUGIN variable in the Makefile to avoid conflicts (since this variable is used to name the resulting library):

LLVMC_PLUGIN=MyPlugin

It is also a good idea to rename Simple.td to something less generic:

$ mv Simple.td MyPlugin.td

To build your plugin as a dynamic library, just cd to its source directory and run make. The resulting file will be called plugin_llvmc_$(LLVMC_PLUGIN).$(DLL_EXTENSION) (in our case, plugin_llvmc_MyPlugin.so). This library can be then loaded in with the -load option. Example:

$ cd $LLVMC_DIR/plugins/Simple
$ make
$ llvmc -load $LLVM_DIR/Release/lib/plugin_llvmc_Simple.so

Compiling standalone LLVMC-based drivers

By default, the llvmc executable consists of a driver core plus several statically linked plugins (Base and Clang at the moment). You can produce a standalone LLVMC-based driver executable by linking the core with your own plugins. The recommended way to do this is by starting with the provided Skeleton example ($LLVMC_DIR/example/Skeleton):

$ cd $LLVMC_DIR/example/
$ cp -r Skeleton mydriver
$ cd mydriver
$ vim Makefile
[...]
$ make

If you're compiling LLVM with different source and object directories, then you must perform the following additional steps before running make:

# LLVMC_SRC_DIR = $LLVM_SRC_DIR/tools/llvmc/
# LLVMC_OBJ_DIR = $LLVM_OBJ_DIR/tools/llvmc/
$ cp $LLVMC_SRC_DIR/example/mydriver/Makefile \
  $LLVMC_OBJ_DIR/example/mydriver/
$ cd $LLVMC_OBJ_DIR/example/mydriver
$ make

Another way to do the same thing is by using the following command:

$ cd $LLVMC_DIR
$ make LLVMC_BUILTIN_PLUGINS=MyPlugin LLVMC_BASED_DRIVER_NAME=mydriver

This works with both srcdir == objdir and srcdir != objdir, but assumes that the plugin source directory was placed under $LLVMC_DIR/plugins.

Sometimes, you will want a 'bare-bones' version of LLVMC that has no built-in plugins. It can be compiled with the following command:

$ cd $LLVMC_DIR
$ make LLVMC_BUILTIN_PLUGINS=""

Customizing LLVMC: the compilation graph

Each TableGen configuration file should include the common definitions:

include "llvm/CompilerDriver/Common.td"

Internally, LLVMC stores information about possible source transformations in form of a graph. Nodes in this graph represent tools, and edges between two nodes represent a transformation path. A special "root" node is used to mark entry points for the transformations. LLVMC also assigns a weight to each edge (more on this later) to choose between several alternative edges.

The definition of the compilation graph (see file plugins/Base/Base.td for an example) is just a list of edges:

def CompilationGraph : CompilationGraph<[
    Edge<"root", "llvm_gcc_c">,
    Edge<"root", "llvm_gcc_assembler">,
    ...

    Edge<"llvm_gcc_c", "llc">,
    Edge<"llvm_gcc_cpp", "llc">,
    ...

    OptionalEdge<"llvm_gcc_c", "opt", (case (switch_on "opt"),
                                      (inc_weight))>,
    OptionalEdge<"llvm_gcc_cpp", "opt", (case (switch_on "opt"),
                                              (inc_weight))>,
    ...

    OptionalEdge<"llvm_gcc_assembler", "llvm_gcc_cpp_linker",
        (case (input_languages_contain "c++"), (inc_weight),
              (or (parameter_equals "linker", "g++"),
                  (parameter_equals "linker", "c++")), (inc_weight))>,
    ...

    ]>;

As you can see, the edges can be either default or optional, where optional edges are differentiated by an additional case expression used to calculate the weight of this edge. Notice also that we refer to tools via their names (as strings). This makes it possible to add edges to an existing compilation graph in plugins without having to know about all tool definitions used in the graph.

The default edges are assigned a weight of 1, and optional edges get a weight of 0 + 2*N where N is the number of tests that evaluated to true in the case expression. It is also possible to provide an integer parameter to inc_weight and dec_weight - in this case, the weight is increased (or decreased) by the provided value instead of the default 2. It is also possible to change the default weight of an optional edge by using the default clause of the case construct.

When passing an input file through the graph, LLVMC picks the edge with the maximum weight. To avoid ambiguity, there should be only one default edge between two nodes (with the exception of the root node, which gets a special treatment - there you are allowed to specify one default edge per language).

When multiple plugins are loaded, their compilation graphs are merged together. Since multiple edges that have the same end nodes are not allowed (i.e. the graph is not a multigraph), an edge defined in several plugins will be replaced by the definition from the plugin that was loaded last. Plugin load order can be controlled by using the plugin priority feature described above.

To get a visual representation of the compilation graph (useful for debugging), run llvmc --view-graph. You will need dot and gsview installed for this to work properly.

Describing options

Command-line options that the plugin supports are defined by using an OptionList:

def Options : OptionList<[
(switch_option "E", (help "Help string")),
(alias_option "quiet", "q")
...
]>;

As you can see, the option list is just a list of DAGs, where each DAG is an option description consisting of the option name and some properties. A plugin can define more than one option list (they are all merged together in the end), which can be handy if one wants to separate option groups syntactically.

External options

Sometimes, when linking several plugins together, one plugin needs to access options defined in some other plugin. Because of the way options are implemented, such options must be marked as extern. This is what the extern option property is for. Example:

...
(switch_option "E", (extern))
...

If an external option has additional attributes besides 'extern', they are ignored. See also the section on plugin priorities.

Conditional evaluation

The 'case' construct is the main means by which programmability is achieved in LLVMC. It can be used to calculate edge weights, program actions and modify the shell commands to be executed. The 'case' expression is designed after the similarly-named construct in functional languages and takes the form (case (test_1), statement_1, (test_2), statement_2, ... (test_N), statement_N). The statements are evaluated only if the corresponding tests evaluate to true.

Examples:

// Edge weight calculation

// Increases edge weight by 5 if "-A" is provided on the
// command-line, and by 5 more if "-B" is also provided.
(case
    (switch_on "A"), (inc_weight 5),
    (switch_on "B"), (inc_weight 5))


// Tool command line specification

// Evaluates to "cmdline1" if the option "-A" is provided on the
// command line; to "cmdline2" if "-B" is provided;
// otherwise to "cmdline3".

(case
    (switch_on "A"), "cmdline1",
    (switch_on "B"), "cmdline2",
    (default), "cmdline3")

Note the slight difference in 'case' expression handling in contexts of edge weights and command line specification - in the second example the value of the "B" switch is never checked when switch "A" is enabled, and the whole expression always evaluates to "cmdline1" in that case.

Case expressions can also be nested, i.e. the following is legal:

(case (switch_on "E"), (case (switch_on "o"), ..., (default), ...)
      (default), ...)

You should, however, try to avoid doing that because it hurts readability. It is usually better to split tool descriptions and/or use TableGen inheritance instead.

Writing a tool description

As was said earlier, nodes in the compilation graph represent tools, which are described separately. A tool definition looks like this (taken from the include/llvm/CompilerDriver/Tools.td file):

def llvm_gcc_cpp : Tool<[
    (in_language "c++"),
    (out_language "llvm-assembler"),
    (output_suffix "bc"),
    (cmd_line "llvm-g++ -c $INFILE -o $OUTFILE -emit-llvm"),
    (sink)
    ]>;

This defines a new tool called llvm_gcc_cpp, which is an alias for llvm-g++. As you can see, a tool definition is just a list of properties; most of them should be self-explanatory. The sink property means that this tool should be passed all command-line options that aren't mentioned in the option list.

The complete list of all currently implemented tool properties follows.

Actions

A tool often needs to react to command-line options, and this is precisely what the actions property is for. The next example illustrates this feature:

def llvm_gcc_linker : Tool<[
    (in_language "object-code"),
    (out_language "executable"),
    (output_suffix "out"),
    (cmd_line "llvm-gcc $INFILE -o $OUTFILE"),
    (join),
    (actions (case (not_empty "L"), (forward "L"),
                   (not_empty "l"), (forward "l"),
                   (not_empty "dummy"),
                             [(append_cmd "-dummy1"), (append_cmd "-dummy2")])
    ]>;

The actions tool property is implemented on top of the omnipresent case expression. It associates one or more different actions with given conditions - in the example, the actions are forward, which forwards a given option unchanged, and append_cmd, which appends a given string to the tool execution command. Multiple actions can be associated with a single condition by using a list of actions (used in the example to append some dummy options). The same case construct can also be used in the cmd_line property to modify the tool command line.

The "join" property used in the example means that this tool behaves like a linker.

The list of all possible actions follows.

  • Possible actions:

    • append_cmd - Append a string to the tool invocation command. Example: (case (switch_on "pthread"), (append_cmd "-lpthread")).
    • error - Exit with error. Example: (error "Mixing -c and -S is not allowed!").
    • warning - Print a warning. Example: (warning "Specifying both -O1 and -O2 is meaningless!").
    • forward - Forward the option unchanged. Example: (forward "Wall").
    • forward_as - Change the option's name, but forward the argument unchanged. Example: (forward_as "O0", "--disable-optimization").
    • forward_value - Forward only option's value. Cannot be used with switch options (since they don't have values), but works fine with lists. Example: (forward_value "Wa,").
    • forward_transformed_value - As above, but applies a hook to the option's value before forwarding (see below). When forward_transformed_value is applied to a list option, the hook must have signature std::string hooks::HookName (const std::vector<std::string>&). Example: (forward_transformed_value "m", "ConvertToMAttr").
    • output_suffix - Modify the output suffix of this tool. Example: (output_suffix "i").
    • stop_compilation - Stop compilation after this tool processes its input. Used without arguments. Example: (stop_compilation).

Language map

If you are adding support for a new language to LLVMC, you'll need to modify the language map, which defines mappings from file extensions to language names. It is used to choose the proper toolchain(s) for a given input file set. Language map definition looks like this:

def LanguageMap : LanguageMap<
    [LangToSuffixes<"c++", ["cc", "cp", "cxx", "cpp", "CPP", "c++", "C"]>,
     LangToSuffixes<"c", ["c"]>,
     ...
    ]>;

For example, without those definitions the following command wouldn't work:

$ llvmc hello.cpp
llvmc: Unknown suffix: cpp

The language map entries should be added only for tools that are linked with the root node. Since tools are not allowed to have multiple output languages, for nodes "inside" the graph the input and output languages should match. This is enforced at compile-time.

Option preprocessor

It is sometimes useful to run error-checking code before processing the compilation graph. For example, if optimization options "-O1" and "-O2" are implemented as switches, we might want to output a warning if the user invokes the driver with both of these options enabled.

The OptionPreprocessor feature is reserved specially for these occasions. Example (adapted from the built-in Base plugin):

def Preprocess : OptionPreprocessor<
(case (and (switch_on "O3"), (any_switch_on ["O0", "O1", "O2"])),
           [(unset_option ["O0", "O1", "O2"]),
            (warning "Multiple -O options specified, defaulted to -O3.")],
      (and (switch_on "O2"), (any_switch_on ["O0", "O1"])),
           (unset_option ["O0", "O1"]),
      (and (switch_on "O1"), (switch_on "O0")),
           (unset_option "O0"))
>;

Here, OptionPreprocessor is used to unset all spurious optimization options (so that they are not forwarded to the compiler).

OptionPreprocessor is basically a single big case expression, which is evaluated only once right after the plugin is loaded. The only allowed actions in OptionPreprocessor are error, warning and a special action unset_option, which, as the name suggests, unsets a given option. For convenience, unset_option also works on lists.

More advanced topics

Hooks and environment variables

Normally, LLVMC executes programs from the system PATH. Sometimes, this is not sufficient: for example, we may want to specify tool paths or names in the configuration file. This can be easily achieved via the hooks mechanism. To write your own hooks, just add their definitions to the PluginMain.cpp or drop a .cpp file into the your plugin directory. Hooks should live in the hooks namespace and have the signature std::string hooks::MyHookName ([const char* Arg0 [ const char* Arg2 [, ...]]]). They can be used from the cmd_line tool property:

(cmd_line "$CALL(MyHook)/path/to/file -o $CALL(AnotherHook)")

To pass arguments to hooks, use the following syntax:

(cmd_line "$CALL(MyHook, 'Arg1', 'Arg2', 'Arg # 3')/path/to/file -o1 -o2")

It is also possible to use environment variables in the same manner:

(cmd_line "$ENV(VAR1)/path/to/file -o $ENV(VAR2)")

To change the command line string based on user-provided options use the case expression (documented above):

(cmd_line
  (case
    (switch_on "E"),
       "llvm-g++ -E -x c $INFILE -o $OUTFILE",
    (default),
       "llvm-g++ -c -x c $INFILE -o $OUTFILE -emit-llvm"))

How plugins are loaded

It is possible for LLVMC plugins to depend on each other. For example, one can create edges between nodes defined in some other plugin. To make this work, however, that plugin should be loaded first. To achieve this, the concept of plugin priority was introduced. By default, every plugin has priority zero; to specify the priority explicitly, put the following line in your plugin's TableGen file:

def Priority : PluginPriority<$PRIORITY_VALUE>;
# Where PRIORITY_VALUE is some integer > 0

Plugins are loaded in order of their (increasing) priority, starting with 0. Therefore, the plugin with the highest priority value will be loaded last.

Debugging

When writing LLVMC plugins, it can be useful to get a visual view of the resulting compilation graph. This can be achieved via the command line option --view-graph. This command assumes that Graphviz and Ghostview are installed. There is also a --write-graph option that creates a Graphviz source file (compilation-graph.dot) in the current directory.

Another useful llvmc option is --check-graph. It checks the compilation graph for common errors like mismatched output/input language names, multiple default edges and cycles. These checks can't be performed at compile-time because the plugins can load code dynamically. When invoked with --check-graph, llvmc doesn't perform any compilation tasks and returns the number of encountered errors as its status code.

Conditioning on the executable name

For now, the executable name (the value passed to the driver in argv[0]) is accessible only in the C++ code (i.e. hooks). Use the following code:

namespace llvmc {
extern const char* ProgramName;
}

namespace hooks {

std::string MyHook() {
//...
if (strcmp(ProgramName, "mydriver") == 0) {
   //...

}

} // end namespace hooks

In general, you're encouraged not to make the behaviour dependent on the executable file name, and use command-line switches instead. See for example how the Base plugin behaves when it needs to choose the correct linker options (think g++ vs. gcc).


Valid CSS Valid XHTML 1.0 Transitional Mikhail Glushenkov
LLVM Compiler Infrastructure
Last modified: $Date: 2008-12-11 11:34:48 -0600 (Thu, 11 Dec 2008) $