blob: a7d2da37711e9b5a53cba85a5a49089c8100ca5c [file] [log] [blame]
<title>Clang Driver Manual</title>
<link type="text/css" rel="stylesheet" href="../menu.css" />
<link type="text/css" rel="stylesheet" href="../content.css" />
<style type="text/css">
td {
vertical-align: top;
<!--#include virtual="../menu.html.incl"-->
<div id="content">
<h1>Driver Design &amp; Internals</h1>
<li><a href="#intro">Introduction</a></li>
<li><a href="#features">Features and Goals</a></li>
<li><a href="#gcccompat">GCC Compatibility</a></li>
<li><a href="#components">Flexible</a></li>
<li><a href="#performance">Low Overhead</a></li>
<li><a href="#simple">Simple</a></li>
<li><a href="#design">Design</a></li>
<li><a href="#int_intro">Internals Introduction</a></li>
<li><a href="#int_overview">Design Overview</a></li>
<li><a href="#int_notes">Additional Notes</a></li>
<li><a href="#int_compilation">The Compilation Object</a></li>
<li><a href="#int_unified_parsing">Unified Parsing &amp; Pipelining</a></li>
<li><a href="#int_toolchain_translation">ToolChain Argument Translation</a></li>
<li><a href="#int_unused_warnings">Unused Argument Warnings</a></li>
<li><a href="#int_gcc_concepts">Relation to GCC Driver Concepts</a></li>
<!-- ======================================================================= -->
<h2 id="intro">Introduction</h2>
<!-- ======================================================================= -->
<p>This document describes the Clang driver. The purpose of this
document is to describe both the motivation and design goals
for the driver, as well as details of the internal
<!-- ======================================================================= -->
<h2 id="features">Features and Goals</h2>
<!-- ======================================================================= -->
<p>The Clang driver is intended to be a production quality
compiler driver providing access to the Clang compiler and
tools, with a command line interface which is compatible with
the gcc driver.</p>
<p>Although the driver is part of and driven by the Clang
project, it is logically a separate tool which shares many of
the same goals as Clang:</p>
<li><a href="#gcccompat">GCC Compatibility</a></li>
<li><a href="#components">Flexible</a></li>
<li><a href="#performance">Low Overhead</a></li>
<li><a href="#simple">Simple</a></li>
<h3 id="gcccompat">GCC Compatibility</h3>
<p>The number one goal of the driver is to ease the adoption of
Clang by allowing users to drop Clang into a build system
which was designed to call GCC. Although this makes the driver
much more complicated than might otherwise be necessary, we
decided that being very compatible with the gcc command line
interface was worth it in order to allow users to quickly test
clang on their projects.</p>
<h3 id="components">Flexible</h3>
<p>The driver was designed to be flexible and easily accomodate
new uses as we grow the clang and LLVM infrastructure. As one
example, the driver can easily support the introduction of
tools which have an integrated assembler; something we hope to
add to LLVM in the future.</p>
<p>Similarly, most of the driver functionality is kept in a
library which can be used to build other tools which want to
implement or accept a gcc like interface. </p>
<h3 id="performance">Low Overhead</h3>
<p>The driver should have as little overhead as possible. In
practice, we found that the gcc driver by itself incurred a
small but meaningful overhead when compiling many small
files. The driver doesn't do much work compared to a
compilation, but we have tried to keep it as efficient as
possible by following a few simple principles:</p>
<li>Avoid memory allocation and string copying when
<li>Don't parse arguments more than once.</li>
<li>Provide a few simple interfaces for efficiently searching
<h3 id="simple">Simple</h3>
<p>Finally, the driver was designed to be "as simple as
possible", given the other goals. Notably, trying to be
completely compatible with the gcc driver adds a significant
amount of complexity. However, the design of the driver
attempts to mitigate this complexity by dividing the process
into a number of independent stages instead of a single
monolithic task.</p>
<!-- ======================================================================= -->
<h2 id="design">Internal Design and Implementation</h2>
<!-- ======================================================================= -->
<li><a href="#int_intro">Internals Introduction</a></li>
<li><a href="#int_overview">Design Overview</a></li>
<li><a href="#int_notes">Additional Notes</a></li>
<li><a href="#int_gcc_concepts">Relation to GCC Driver Concepts</a></li>
<h3><a name="int_intro">Internals Introduction</a></h3>
<p>In order to satisfy the stated goals, the driver was designed
to completely subsume the functionality of the gcc executable;
that is, the driver should not need to delegate to gcc to
perform subtasks. On Darwin, this implies that the Clang
driver also subsumes the gcc driver-driver, which is used to
implement support for building universal images (binaries and
object files). This also implies that the driver should be
able to call the language specific compilers (e.g. cc1)
directly, which means that it must have enough information to
forward command line arguments to child processes
<h3><a name="int_overview">Design Overview</a></h3>
<p>The diagram below shows the significant components of the
driver architecture and how they relate to one another. The
orange components represent concrete data structures built by
the driver, the green components indicate conceptually
distinct stages which manipulate these data structures, and
the blue components are important helper classes. </p>
<a href="DriverArchitecture.png" alt="Driver Architecture Diagram">
<img width=400 src="DriverArchitecture.png">
<h3><a name="int_stages">Driver Stages</a></h3>
<p>The driver functionality is conceptually divided into five stages:</p>
<b>Parse: Option Parsing</b>
<p>The command line argument strings are decomposed into
arguments (<tt>Arg</tt> instances). The driver expects to
understand all available options, although there is some
facility for just passing certain classes of options
through (like <tt>-Wl,</tt>).</p>
<p>Each argument corresponds to exactly one
abstract <tt>Option</tt> definition, which describes how
the option is parsed along with some additional
metadata. The Arg instances themselves are lightweight and
merely contain enough information for clients to determine
which option they correspond to and their values (if they
have additional parameters).</p>
<p>For example, a command line like "-Ifoo -I foo" would
parse to two Arg instances (a JoinedArg and a SeparateArg
instance), but each would refer to the same Option.</p>
<p>Options are lazily created in order to avoid populating
all Option classes when the driver is loaded. Most of the
driver code only needs to deal with options by their
unique ID (e.g., <tt>options::OPT_I</tt>),</p>
<p>Arg instances themselves do not generally store the
values of parameters. In many cases, this would
simply result in creating unnecessary string
copies. Instead, Arg instances are always embedded inside
an ArgList structure, which contains the original vector
of argument strings. Each Arg itself only needs to contain
an index into this vector instead of storing its values
<p>The clang driver can dump the results of this
stage using the <tt>-ccc-print-options</tt> flag (which
must preceed any actual command line arguments). For
$ <b>clang -ccc-print-options -Xarch_i386 -fomit-frame-pointer -Wa,-fast -Ifoo -I foo t.c</b>
Option 0 - Name: "-Xarch_", Values: {"i386", "-fomit-frame-pointer"}
Option 1 - Name: "-Wa,", Values: {"-fast"}
Option 2 - Name: "-I", Values: {"foo"}
Option 3 - Name: "-I", Values: {"foo"}
Option 4 - Name: "&lt;input&gt;", Values: {"t.c"}
<p>After this stage is complete the command line should be
broken down into well defined option objects with their
appropriate parameters. Subsequent stages should rarely,
if ever, need to do any string processing.</p>
<b>Pipeline: Compilation Job Construction</b>
<p>Once the arguments are parsed, the tree of subprocess
jobs needed for the desired compilation sequence are
constructed. This involves determining the input files and
their types, what work is to be done on them (preprocess,
compile, assemble, link, etc.), and constructing a list of
Action instances for each task. The result is a list of
one or more top-level actions, each of which generally
corresponds to a single output (for example, an object or
linked executable).</p>
<p>The majority of Actions correspond to actual tasks,
however there are two special Actions. The first is
InputAction, which simply serves to adapt an input
argument for use as an input to other Actions. The second
is BindArchAction, which conceptually alters the
architecture to be used for all of its input Actions.</p>
<p>The clang driver can dump the results of this
stage using the <tt>-ccc-print-phases</tt> flag. For
$ <b>clang -ccc-print-phases -x c t.c -x assembler t.s</b>
0: input, "t.c", c
1: preprocessor, {0}, cpp-output
2: compiler, {1}, assembler
3: assembler, {2}, object
4: input, "t.s", assembler
5: assembler, {4}, object
6: linker, {3, 5}, image
<p>Here the driver is constructing seven distinct actions,
four to compile the "t.c" input into an object file, two to
assemble the "t.s" input, and one to link them together.</p>
<p>A rather different compilation pipeline is shown here; in
this example there are two top level actions to compile
the input files into two separate object files, where each
object file is built using <tt>lipo</tt> to merge results
built for two separate architectures.</p>
$ <b>clang -ccc-print-phases -c -arch i386 -arch x86_64 t0.c t1.c</b>
0: input, "t0.c", c
1: preprocessor, {0}, cpp-output
2: compiler, {1}, assembler
3: assembler, {2}, object
4: bind-arch, "i386", {3}, object
5: bind-arch, "x86_64", {3}, object
6: lipo, {4, 5}, object
7: input, "t1.c", c
8: preprocessor, {7}, cpp-output
9: compiler, {8}, assembler
10: assembler, {9}, object
11: bind-arch, "i386", {10}, object
12: bind-arch, "x86_64", {10}, object
13: lipo, {11, 12}, object
<p>After this stage is complete the compilation process is
divided into a simple set of actions which need to be
performed to produce intermediate or final outputs (in
some cases, like <tt>-fsyntax-only</tt>, there is no
"real" final output). Phases are well known compilation
steps, such as "preprocess", "compile", "assemble",
"link", etc.</p>
<b>Bind: Tool &amp; Filename Selection</b>
<p>This stage (in conjunction with the Translate stage)
turns the tree of Actions into a list of actual subprocess
to run. Conceptually, the driver performs a top down
matching to assign Action(s) to Tools. The ToolChain is
responsible for selecting the tool to perform a particular
action; once selected the driver interacts with the tool
to see if it can match additional actions (for example, by
having an integrated preprocessor).
<p>Once Tools have been selected for all actions, the driver
determines how the tools should be connected (for example,
using an inprocess module, pipes, temporary files, or user
provided filenames). If an output file is required, the
driver also computes the appropriate file name (the suffix
and file location depend on the input types and options
such as <tt>-save-temps</tt>).
<p>The driver interacts with a ToolChain to perform the Tool
bindings. Each ToolChain contains information about all
the tools needed for compilation for a particular
architecture, platform, and operating system. A single
driver invocation may query multiple ToolChains during one
compilation in order to interact with tools for separate
<p>The results of this stage are not computed directly, but
the driver can print the results via
the <tt>-ccc-print-bindings</tt> option. For example:</p>
$ <b>clang -ccc-print-bindings -arch i386 -arch ppc t0.c</b>
# "i386-apple-darwin9" - "clang", inputs: ["t0.c"], output: "/tmp/cc-Sn4RKF.s"
# "i386-apple-darwin9" - "darwin::Assemble", inputs: ["/tmp/cc-Sn4RKF.s"], output: "/tmp/cc-gvSnbS.o"
# "i386-apple-darwin9" - "darwin::Link", inputs: ["/tmp/cc-gvSnbS.o"], output: "/tmp/cc-jgHQxi.out"
# "ppc-apple-darwin9" - "gcc::Compile", inputs: ["t0.c"], output: "/tmp/cc-Q0bTox.s"
# "ppc-apple-darwin9" - "gcc::Assemble", inputs: ["/tmp/cc-Q0bTox.s"], output: "/tmp/cc-WCdicw.o"
# "ppc-apple-darwin9" - "gcc::Link", inputs: ["/tmp/cc-WCdicw.o"], output: "/tmp/cc-HHBEBh.out"
# "i386-apple-darwin9" - "darwin::Lipo", inputs: ["/tmp/cc-jgHQxi.out", "/tmp/cc-HHBEBh.out"], output: "a.out"
<p>This shows the tool chain, tool, inputs and outputs which
have been bound for this compilation sequence. Here clang
is being used to compile t0.c on the i386 architecture and
darwin specific versions of the tools are being used to
assemble and link the result, but generic gcc versions of
the tools are being used on PowerPC.</p>
<b>Translate: Tool Specific Argument Translation</b>
<p>Once a Tool has been selected to perform a particular
Action, the Tool must construct concrete Jobs which will be
executed during compilation. The main work is in translating
from the gcc style command line options to whatever options
the subprocess expects.</p>
<p>Some tools, such as the assembler, only interact with a
handful of arguments and just determine the path of the
executable to call and pass on their input and output
arguments. Others, like the compiler or the linker, may
translate a large number of arguments in addition.</p>
<p>The ArgList class provides a number of simple helper
methods to assist with translating arguments; for example,
to pass on only the last of arguments corresponding to some
option, or all arguments for an option.</p>
<p>The result of this stage is a list of Jobs (executable
paths and argument strings) to execute.</p>
<p>Finally, the compilation pipeline is executed. This is
mostly straightforward, although there is some interaction
with options
like <tt>-pipe</tt>, <tt>-pass-exit-codes</tt>
and <tt>-time</tt>.</p>
<h3><a name="int_notes">Additional Notes</a></h3>
<h4 id="int_compilation">The Compilation Object</h4>
<p>The driver constructs a Compilation object for each set of
command line arguments. The Driver itself is intended to be
invariant during construction of a Compilation; an IDE should be
able to construct a single long lived driver instance to use
for an entire build, for example.</p>
<p>The Compilation object holds information that is particular
to each compilation sequence. For example, the list of used
temporary files (which must be removed once compilation is
finished) and result files (which should be removed if
compilation files).</p>
<h4 id="int_unified_parsing">Unified Parsing &amp; Pipelining</h4>
<p>Parsing and pipelining both occur without reference to a
Compilation instance. This is by design; the driver expects that
both of these phases are platform neutral, with a few very well
defined exceptions such as whether the platform uses a driver
<h4 id="int_toolchain_translation">ToolChain Argument Translation</h4>
<p>In order to match gcc very closely, the clang driver
currently allows tool chains to perform their own translation of
the argument list (into a new ArgList data structure). Although
this allows the clang driver to match gcc easily, it also makes
the driver operation much harder to understand (since the Tools
stop seeing some arguments the user provided, and see new ones
<p>For example, on Darwin <tt>-gfull</tt> gets translated into two
separate arguments, <tt>-g</tt>
and <tt>-fno-eliminate-unused-debug-symbols</tt>. Trying to write Tool
logic to do something with <tt>-gfull</tt> will not work, because Tool
argument translation is done after the arguments have been
<p>A long term goal is to remove this tool chain specific
translation, and instead force each tool to change its own logic
to do the right thing on the untranslated original arguments.</p>
<h4 id="int_unused_warnings">Unused Argument Warnings</h4>
<p>The driver operates by parsing all arguments but giving Tools
the opportunity to choose which arguments to pass on. One
downside of this infrastructure is that if the user misspells
some option, or is confused about which options to use, some
command line arguments the user really cared about may go
unused. This problem is particularly important when using
clang as a compiler, since the clang compiler does not support
anywhere near all the options that gcc does, and we want to make
sure users know which ones are being used.</p>
<p>To support this, the driver maintains a bit associated with
each argument of whether it has been used (at all) during the
compilation. This bit usually doesn't need to be set by hand,
as the key ArgList accessors will set it automatically.</p>
<p>When a compilation is successful (there are no errors), the
driver checks the bit and emits an "unused argument" warning for
any arguments which were never accessed. This is conservative
(the argument may not have been used to do what the user wanted)
but still catches the most obvious cases.</p>
<h3><a name="int_gcc_concepts">Relation to GCC Driver Concepts</a></h3>
<p>For those familiar with the gcc driver, this section provides
a brief overview of how things from the gcc driver map to the
clang driver.</p>
<b>Driver Driver</b>
<p>The driver driver is fully integrated into the clang
driver. The driver simply constructs additional Actions to
bind the architecture during the <i>Pipeline</i>
phase. The tool chain specific argument translation is
responsible for handling <tt>-Xarch_</tt>.</p>
<p>The one caveat is that this approach
requires <tt>-Xarch_</tt> not be used to alter the
compilation itself (for example, one cannot
provide <tt>-S</tt> as an <tt>-Xarch_</tt> argument). The
driver attempts to reject such invocations, and overall
there isn't a good reason to abuse <tt>-Xarch_</tt> to
that end in practice.</p>
<p>The upside is that the clang driver is more efficient and
does little extra work to support universal builds. It also
provides better error reporting and UI consistency.</p>
<p>The clang driver has no direct correspondant for
"specs". The majority of the functionality that is
embedded in specs is in the Tool specific argument
translation routines. The parts of specs which control the
compilation pipeline are generally part of
the <ii>Pipeline</ii> stage.</p>
<p>The gcc driver has no direct understanding of tool
chains. Each gcc binary roughly corresponds to the
information which is embedded inside a single
<p>The clang driver is intended to be portable and support
complex compilation environments. All platform and tool
chain specific code should be protected behind either
abstract or well defined interfaces (such as whether the
platform supports use as a driver driver).</p>