This document describes the requirements, design, and configuration of the LLVM compiler driver, llvmc. The compiler driver knows about LLVM's tool set and can be configured to know about a variety of compilers for source languages. It uses this knowledge to execute the tools necessary to accomplish general compilation, optimization, and linking tasks. The main purpose of llvmc is to provide a simple and consistent interface to all compilation tasks. This reduces the burden on the end user who can just learn to use llvmc instead of the entire LLVM tool set and all the source language compilers compatible with LLVM.

The llvmc tool is a configurable compiler driver. As such, it isn't a compiler, optimizer, or a linker itself but it drives (invokes) other software that perform those tasks. If you are familiar with the GNU Compiler Collection's gcc tool, llvmc is very similar.

The following introductory sections will help you understand why this tool is necessary and what it does.

llvmc was invented to make compilation of user programs with LLVM-based tools easier. To accomplish this, llvmc strives to:

• Be the single point of access to most of the LLVM tool set.
• Hide the complexities of the LLVM tools through a single interface.
• Provide a consistent interface for compiling all languages.

Additionally, llvmc makes it easier to write a compiler for use with LLVM, because it:

• Makes integration of existing non-LLVM tools simple.
• Extends the capabilities of minimal compiler tools by optimizing their output.
• Reduces the number of interfaces a compiler writer must know about before a working compiler can be completed (essentially only the VMCore interfaces need to be understood).
• Supports source language translator invocation via both dynamically loadable shared objects and invocation of an executable.

At a high level, llvmc operation is very simple. The basic action taken by llvmc is to simply invoke some tool or set of tools to fill the user's request for compilation. Every execution of llvmctakes the following sequence of steps:

Collect Command Line Options

The command line options provide the marching orders to llvmc on what actions it should perform. This is the request the user is making of llvmc and it is interpreted first. See the llvmc manual page for details on the options.

Read Configuration Files

Based on the options and the suffixes of the filenames presented, a set of configuration files are read to configure the actions llvmc will take. Configuration files are provided by either LLVM or the compiler tools that llvmc invokes. These files determine what actions llvmc will take in response to the user's request. See the section on configuration for more details.

Determine Phases To Execute

Based on the command line options and configuration files, llvmc determines the compilation phases that must be executed by the user's request. This is the primary work of llvmc.

Determine Actions To Execute

Each phase to be executed can result in the invocation of one or more actions. An action is either a whole program or a function in a dynamically linked shared library. In this step, llvmc determines the sequence of actions that must be executed. Actions will always be executed in a deterministic order.

Execute Actions

The actions necessary to support the user's original request are executed sequentially and deterministically. All actions result in either the invocation of a whole program to perform the action or the loading of a dynamically linkable shared library and invocation of a standard interface function within that library.

Termination

If any action fails (returns a non-zero result code), llvmc also fails and returns the result code from the failing action. If everything succeeds, llvmc will return a zero result code.

llvmc's operation must be simple, regular and predictable. Developers need to be able to rely on it to take a consistent approach to compilation. For example, the invocation:

must produce exactly the same results as:

    llvmc -O2 x.c -o x.o
    llvmc -O2 y.c -o y.o
    llvmc -O2 z.c -o z.o
    llvmc -O2 x.o y.o z.o -o xyz

To accomplish this, llvmc uses a very simple goal oriented procedure to do its work. The overall goal is to produce a functioning executable. To accomplish this, llvmc always attempts to execute a series of compilation phases in the same sequence. However, the user's options to llvmc can cause the sequence of phases to start in the middle or finish early.

llvmc breaks every compilation task into the following five distinct phases:

Preprocessing

Not all languages support preprocessing; but for those that do, this phase can be invoked. This phase is for languages that provide combining, filtering, or otherwise altering with the source language input before the translator parses it. Although C and C++ are the most common users of this phase, other languages may provide their own preprocessor (whether its the C pre-processor or not).

Translation

The translation phase converts the source language input into something that LLVM can interpret and use for downstream phases. The translation is essentially from "non-LLVM form" to "LLVM form".

Optimization

Once an LLVM Module has been obtained from the translation phase, the program enters the optimization phase. This phase attempts to optimize all of the input provided on the command line according to the options provided.

Linking

The inputs are combined to form a complete program.

The following table shows the inputs, outputs, and command line options applicable to each phase.

An action, with regard to llvmc is a basic operation that it takes in order to fulfill the user's request. Each phase of compilation will invoke zero or more actions in order to accomplish that phase.

Actions come in two forms:

• Invokable Executables
• Functions in a shared library

This section of the document describes the configuration files used by llvmc. Configuration information is relatively static for a given release of LLVM and a compiler tool. However, the details may change from release to release of either. Users are encouraged to simply use the various options of the llvmc command and ignore the configuration of the tool. These configuration files are for compiler writers and LLVM developers. Those wishing to simply use llvmc don't need to understand this section but it may be instructive on how the tool works.

llvmc is highly configurable both on the command line and in configuration files. The options it understands are generic, consistent and simple by design. Furthermore, the llvmc options apply to the compilation of any LLVM enabled programming language. To be enabled as a supported source language compiler, a compiler writer must provide a configuration file that tells llvmc how to invoke the compiler and what its capabilities are. The purpose of the configuration files then is to allow compiler writers to specify to llvmc how the compiler should be invoked. Users may but are not advised to alter the compiler's llvmc configuration.

Because llvmc just invokes other programs, it must deal with the available command line options for those programs regardless of whether they were written for LLVM or not. Furthermore, not all compiler tools will have the same capabilities. Some compiler tools will simply generate LLVM assembly code, others will be able to generate fully optimized bitcode. In general, llvmc doesn't make any assumptions about the capabilities or command line options of a sub-tool. It simply uses the details found in the configuration files and leaves it to the compiler writer to specify the configuration correctly.

This approach means that new compiler tools can be up and working very quickly. As a first cut, a tool can simply compile its source to raw (unoptimized) bitcode or LLVM assembly and llvmc can be configured to pick up the slack (translate LLVM assembly to bitcode, optimize the bitcode, generate native assembly, link, etc.). In fact, the compiler tools need not use any LLVM libraries, and it could be written in any language (instead of C++). The configuration data will allow the full range of optimization, assembly, and linking capabilities that LLVM provides to be added to these kinds of tools. Enabling the rapid development of front-ends is one of the primary goals of llvmc.

As a compiler tool matures, it may utilize the LLVM libraries and tools to more efficiently produce optimized bitcode directly in a single compilation and optimization program. In these cases, multiple tools would not be needed and the configuration data for the compiler would change.

Configuring llvmc to the needs and capabilities of a source language compiler is relatively straight-forward. A compiler writer must provide a definition of what to do for each of the five compilation phases for each of the optimization levels. The specification consists simply of prototypical command lines into which llvmc can substitute command line arguments and file names. Note that any given phase can be completely blank if the source language's compiler combines multiple phases into a single program. For example, quite often pre-processing, translation, and optimization are combined into a single program. The specification for such a compiler would have blank entries for pre-processing and translation but a full command line for optimization.

Each configuration file provides the details for a single source language that is to be compiled. This configuration information tells llvmc how to invoke the language's pre-processor, translator, optimizer, assembler and linker. Note that a given source language needn't provide all these tools as many of them exist in llvm currently.

llvmc always looks for files of a specific name. It uses the first file with the name its looking for by searching directories in the following order:

1. Any directory specified by the -config-dir option will be checked first.
2. If the environment variable LLVM_CONFIG_DIR is set, and it contains the name of a valid directory, that directory will be searched next.
3. If the user's home directory (typically /home/user contains a sub-directory named .llvm and that directory contains a sub-directory named etc then that directory will be tried next.
4. If the LLVM installation directory (typically /usr/local/llvm contains a sub-directory named etc then that directory will be tried last.
5. A standard "system" directory will be searched next. This is typically /etc/llvm on UNIX™ and C:\WINNT on Microsoft Windows™.
6. If the configuration file sought still can't be found, llvmc will print an error message and exit.

The first file found in this search will be used. Other files with the same name will be ignored even if they exist in one of the subsequent search locations.

In the directories searched, each configuration file is given a specific name to foster faster lookup (so llvmc doesn't have to do directory searches). The name of a given language specific configuration file is simply the same as the suffix used to identify files containing source in that language. For example, a configuration file for C++ source might be named cpp, C, or cxx. For languages that support multiple file suffixes, multiple (probably identical) files (or symbolic links) will need to be provided.

Which configuration files are read depends on the command line options and the suffixes of the file names provided on llvmc's command line. Note that the -x LANGUAGE option alters the language that llvmc uses for the subsequent files on the command line. Only the configuration files actually needed to complete llvmc's task are read. Other language specific files will be ignored.

The syntax of the configuration files is very simple and somewhat compatible with Java's property files. Here are the syntax rules:

• The file encoding is ASCII.
• The file is line oriented. There should be one configuration definition per line. Lines are terminated by the newline (0x0A) and/or carriage return characters (0x0D)
• A backslash (\) before a newline causes the newline to be ignored. This is useful for line continuation of long definitions. A backslash anywhere else is recognized as a backslash.
• A configuration item consists of a name, an = and a value.
• A name consists of a sequence of identifiers separated by period.
• An identifier consists of specific keywords made up of only lower case and upper case letters (e.g. lang.name).
• Values come in four flavors: booleans, integers, commands and strings.
• Valid "false" boolean values are false False FALSE no No NO off Off and OFF.
• Valid "true" boolean values are true True TRUE yes Yes YES on On and ON.
• Integers are simply sequences of digits.
• Commands start with a program name and are followed by a sequence of words that are passed to that program as command line arguments. Program arguments that begin and end with the % sign will have their value substituted. Program names beginning with / are considered to be absolute. Otherwise the PATH will be applied to find the program to execute.
• Strings are composed of multiple sequences of characters from the character class [-A-Za-z0-9_:%+/\\|,] separated by white space.
• White space on a line is folded. Multiple blanks or tabs will be reduced to a single blank.
• White space before the configuration item's name is ignored.
• White space on either side of the = is ignored.
• White space in a string value is used to separate the individual components of the string value but otherwise ignored.
• Comments are introduced by the # character. Everything after a # and before the end of line is ignored.

The table below provides definitions of the allowed configuration items that may appear in a configuration file. Every item has a default value and does not need to appear in the configuration file. Missing items will have the default value. Each identifier may appear as all lower case, first letter capitalized or all upper case.

On any configuration item that ends in command, you must specify substitution tokens. Substitution tokens begin and end with a percent sign (%) and are replaced by the corresponding text. Any substitution token may be given on any command line but some are more useful than others. In particular each command should have both an %in% and an %out% substitution. The table below provides definitions of each of the allowed substitution tokens.

Since an example is always instructive, here's how the Stacker language configuration file looks.

# Stacker Configuration File For llvmc

##########################################################
# Language definitions
##########################################################
  lang.name=Stacker 
  lang.opt1=-simplifycfg -instcombine -mem2reg
  lang.opt2=-simplifycfg -instcombine -mem2reg -load-vn \
    -gcse -dse -scalarrepl -sccp 
  lang.opt3=-simplifycfg -instcombine -mem2reg -load-vn \
    -gcse -dse -scalarrepl -sccp -branch-combine -adce \
    -globaldce -inline -licm 
  lang.opt4=-simplifycfg -instcombine -mem2reg -load-vn \
    -gcse -dse -scalarrepl -sccp -ipconstprop \
    -branch-combine -adce -globaldce -inline -licm 
  lang.opt5=-simplifycfg -instcombine -mem2reg --load-vn \
    -gcse -dse scalarrepl -sccp -ipconstprop \
    -branch-combine -adce -globaldce -inline -licm \
    -block-placement

##########################################################
# Pre-processor definitions
##########################################################

  # Stacker doesn't have a preprocessor but the following
  # allows the -E option to be supported
  preprocessor.command=cp %in% %out%
  preprocessor.required=false

##########################################################
# Translator definitions
##########################################################

  # To compile stacker source, we just run the stacker
  # compiler with a default stack size of 2048 entries.
  translator.command=stkrc -s 2048 %in% -o %out% %time% \
    %stats% %force% %args%

  # stkrc doesn't preprocess but we set this to true so
  # that we don't run the cp command by default.
  translator.preprocesses=true

  # The translator is required to run.
  translator.required=true

  # stkrc doesn't handle the -On options
  translator.output=bitcode

##########################################################
# Optimizer definitions
##########################################################
  
  # For optimization, we use the LLVM "opt" program
  optimizer.command=opt %in% -o %out% %opt% %time% %stats% \
    %force% %args%

  optimizer.required = true

  # opt doesn't translate
  optimizer.translates = no

  # opt doesn't preprocess
  optimizer.preprocesses=no

  # opt produces bitcode
  optimizer.output = bc

##########################################################
# Assembler definitions
##########################################################
  assembler.command=llc %in% -o %out% %target% %time% %stats%

This document uses precise terms in reference to the various artifacts and concepts related to compilation. The terms used throughout this document are defined below.

assembly

A compilation phase in which LLVM bitcode or LLVM assembly code is assembled to a native code format (either target specific aseembly language or the platform's native object file format).

compiler

Refers to any program that can be invoked by llvmc to accomplish the work of one or more compilation phases.

driver

Refers to llvmc itself.

linking

A compilation phase in which LLVM bitcode files and (optionally) native system libraries are combined to form a complete executable program.

optimization

A compilation phase in which LLVM bitcode is optimized.

phase

Refers to any one of the five compilation phases that that llvmc supports. The five phases are: preprocessing, translation, optimization, assembly, linking.

source language

Any common programming language (e.g. C, C++, Java, Stacker, ML, FORTRAN). These languages are distinguished from any of the lower level languages (such as LLVM or native assembly), by the fact that a translation phase is required before LLVM can be applied.

tool

Refers to any program in the LLVM tool set.

translation

A compilation phase in which source language code is translated into either LLVM assembly language or LLVM bitcode.