www/features.html

   1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
   2           "http://www.w3.org/TR/html4/strict.dtd">
   3 <html>
   4 <head>
   5   <META http-equiv="Content-Type" content="text/html; charset=ISO-8859-1" />
   6   <title>Clang - Features and Goals</title>
   7   <link type="text/css" rel="stylesheet" href="menu.css" />
   8   <link type="text/css" rel="stylesheet" href="content.css" />
   9   <style type="text/css">
  10 </style>
  11 </head>
  12 <body>
  13
  14 <!--#include virtual="menu.html.incl"-->
  15
  16 <div id="content">
  17
  18 <!--*************************************************************************-->
  19 <h1>Clang - Features and Goals</h1>
  20 <!--*************************************************************************-->
  21
  22 <p>
  23 This page describes the <a href="index.html#goals">features and goals</a> of
  24 Clang in more detail and gives a more broad explanation about what we mean.
  25 These features are:
  26 </p>
  27
  28 <p>End-User Features:</p>
  29
  30 <ul>
  31 <li><a href="#performance">Fast compiles and low memory use</a></li>
  32 <li><a href="#expressivediags">Expressive diagnostics</a></li>
  33 <li><a href="#gcccompat">GCC compatibility</a></li>
  34 </ul>
  35
  36 <p>Utility and Applications:</p>
  37
  38 <ul>
  39 <li><a href="#libraryarch">Library based architecture</a></li>
  40 <li><a href="#diverseclients">Support diverse clients</a></li>
  41 <li><a href="#ideintegration">Integration with IDEs</a></li>
  42 <li><a href="#license">Use the LLVM 'BSD' License</a></li>
  43 </ul>
  44
  45 <p>Internal Design and Implementation:</p>
  46
  47 <ul>
  48 <li><a href="#real">A real-world, production quality compiler</a></li>
  49 <li><a href="#simplecode">A simple and hackable code base</a></li>
  50 <li><a href="#unifiedparser">A single unified parser for C, Objective C, C++,
  51     and Objective C++</a></li>
  52 <li><a href="#conformance">Conformance with C/C++/ObjC and their
  53     variants</a></li>
  54 </ul>
  55
  56 <!--*************************************************************************-->
  57 <h2><a name="enduser">End-User Features</a></h2>
  58 <!--*************************************************************************-->
  59
  60
  61 <!--=======================================================================-->
  62 <h3><a name="performance">Fast compiles and Low Memory Use</a></h3>
  63 <!--=======================================================================-->
  64
  65 <p>A major focus of our work on clang is to make it fast, light and scalable.
  66 The library-based architecture of clang makes it straight-forward to time and
  67 profile the cost of each layer of the stack, and the driver has a number of
  68 options for performance analysis.</p>
  69
  70 <p>While there is still much that can be done, we find that the clang front-end
  71 is significantly quicker than gcc and uses less memory  For example, when
  72 compiling "Carbon.h" on Mac OS/X, we see that clang is 2.5x faster than GCC:</p>
  73
  74 <img class="img_slide" src="feature-compile1.png" width="400" height="300" />
  75
  76 <p>Carbon.h is a monster: it transitively includes 558 files, 12.3M of code,
  77 declares 10000 functions, has 2000 struct definitions, 8000 fields, 20000 enum
  78 constants, etc (see slide 25+ of the <a href="clang_video-07-25-2007.html">clang
  79 talk</a> for more information). It is also #include'd into almost every C file
  80 in a GUI app on the Mac, so its compile time is very important.</p>
  81
  82 <p>From the slide above, you can see that we can measure the time to preprocess
  83 the file independently from the time to parse it, and independently from the
  84 time to build the ASTs for the code.  GCC doesn't provide a way to measure the
  85 parser without AST building (it only provides -fsyntax-only).  In our
  86 measurements, we find that clang's preprocessor is consistently 40% faster than
  87 GCCs, and the parser + AST builder is ~4x faster than GCC's.  If you have
  88 sources that do not depend as heavily on the preprocessor (or if you
  89 use Precompiled Headers) you may see a much bigger speedup from clang.
  90 </p>
  91
  92 <p>Compile time performance is important, but when using clang as an API, often
  93 memory use is even moreso: the less memory the code takes the more code you can
  94 fit into memory at a time (useful for whole program analysis tools, for
  95 example).</p>
  96
  97 <img class="img_slide" src="feature-memory1.png" width="400" height="300" />
  98
  99 <p>Here we see a huge advantage of clang: its ASTs take <b>5x less memory</b>
 100 than GCC's syntax trees, despite the fact that clang's ASTs capture far more
 101 source-level information than GCC's trees do.  This feat is accomplished through
 102 the use of carefully designed APIs and efficient representations.</p>
 103
 104 <p>In addition to being efficient when pitted head-to-head against GCC in batch
 105 mode, clang is built with a <a href="#libraryarch">library based
 106 architecture</a> that makes it relatively easy to adapt it and build new tools
 107 with it.  This means that it is often possible to apply out-of-the-box thinking
 108 and novel techniques to improve compilation in various ways.</p>
 109
 110 <img class="img_slide" src="feature-compile2.png" width="400" height="300" />
 111
 112 <p>This slide shows how the clang preprocessor can be used to make "distcc"
 113 parallelization <b>3x</b> more scalable than when using the GCC preprocessor.
 114 "distcc" quickly bottlenecks on the preprocessor running on the central driver
 115 machine, so a fast preprocessor is very useful.  Comparing the first two bars
 116 of each group shows how a ~40% faster preprocessor can reduce preprocessing time
 117 of these large C++ apps by about 40% (shocking!).</p>
 118
 119 <p>The third bar on the slide is the interesting part: it shows how trivial
 120 caching of file system accesses across invocations of the preprocessor allows
 121 clang to reduce time spent in the kernel by 10x, making distcc over 3x more
 122 scalable.  This is obviously just one simple hack, doing more interesting things
 123 (like caching tokens across preprocessed files) would yield another substantial
 124 speedup.</p>
 125
 126 <p>The clean framework-based design of clang means that many things are possible
 127 that would be very difficult in other systems, for example incremental
 128 compilation, multithreading, intelligent caching, etc.  We are only starting
 129 to tap the full potential of the clang design.</p>
 130
 131
 132 <!--=======================================================================-->
 133 <h3><a name="expressivediags">Expressive Diagnostics</a></h3>
 134 <!--=======================================================================-->
 135
 136 <p>In addition to being fast and functional, we aim to make Clang extremely user
 137 friendly.  As far as a command-line compiler goes, this basically boils down to
 138 making the diagnostics (error and warning messages) generated by the compiler
 139 be as useful as possible.  There are several ways that we do this, but the
 140 most important are pinpointing exactly what is wrong in the program,
 141 highlighting related information so that it is easy to understand at a glance,
 142 and making the wording as clear as possible.</p>
 143
 144 <p>Here is one simple example that illustrates the difference between a typical
 145 GCC and Clang diagnostic:</p>
 146
 147 <pre>
 148   $ <b>gcc-4.2 -fsyntax-only t.c</b>
 149   t.c:7: error: invalid operands to binary + (have 'int' and 'struct A')
 150   $ <b>clang -fsyntax-only t.c</b>
 151   t.c:7:39: error: invalid operands to binary expression ('int' and 'struct A')
 152   <font color="darkgreen">  return y + func(y ? ((SomeA.X + 40) + SomeA) / 42 + SomeA.X : SomeA.X);</font>
 153   <font color="blue">                       ~~~~~~~~~~~~~~ ^ ~~~~~</font>
 154 </pre>
 155
 156 <p>Here you can see that you don't even need to see the original source code to
 157 understand what is wrong based on the Clang error: Because clang prints a
 158 caret, you know exactly <em>which</em> plus it is complaining about.  The range
 159 information highlights the left and right side of the plus which makes it
 160 immediately obvious what the compiler is talking about, which is very useful for
 161 cases involving precedence issues and many other situations.</p>
 162
 163 <p>Clang diagnostics are very polished and have many features.  For more
 164 information and examples, please see the <a href="diagnostics.html">Expressive
 165 Diagnostics</a> page.</p>
 166
 167 <!--=======================================================================-->
 168 <h3><a name="gcccompat">GCC Compatibility</a></h3>
 169 <!--=======================================================================-->
 170
 171 <p>GCC is currently the defacto-standard open source compiler today, and it
 172 routinely compiles a huge volume of code.  GCC supports a huge number of
 173 extensions and features (many of which are undocumented) and a lot of
 174 code and header files depend on these features in order to build.</p>
 175
 176 <p>While it would be nice to be able to ignore these extensions and focus on
 177 implementing the language standards to the letter, pragmatics force us to
 178 support the GCC extensions that see the most use.  Many users just want their
 179 code to compile, they don't care to argue about whether it is pedantically C99
 180 or not.</p>
 181
 182 <p>As mentioned above, all
 183 extensions are explicitly recognized as such and marked with extension
 184 diagnostics, which can be mapped to warnings, errors, or just ignored.
 185 </p>
 186
 187
 188 <!--*************************************************************************-->
 189 <h2><a name="applications">Utility and Applications</a></h2>
 190 <!--*************************************************************************-->
 191
 192 <!--=======================================================================-->
 193 <h3><a name="libraryarch">Library Based Architecture</a></h3>
 194 <!--=======================================================================-->
 195
 196 <p>A major design concept for clang is its use of a library-based
 197 architecture.  In this design, various parts of the front-end can be cleanly
 198 divided into separate libraries which can then be mixed up for different needs
 199 and uses.  In addition, the library-based approach encourages good interfaces
 200 and makes it easier for new developers to get involved (because they only need
 201 to understand small pieces of the big picture).</p>
 202
 203 <blockquote>
 204 "The world needs better compiler tools, tools which are built as libraries.
 205 This design point allows reuse of the tools in new and novel ways. However,
 206 building the tools as libraries isn't enough: they must have clean APIs, be as
 207 decoupled from each other as possible, and be easy to modify/extend. This
 208 requires clean layering, decent design, and keeping the libraries independent of
 209 any specific client."</blockquote>
 210
 211 <p>
 212 Currently, clang is divided into the following libraries and tool:
 213 </p>
 214
 215 <ul>
 216 <li><b>libsupport</b> - Basic support library, from LLVM.</li>
 217 <li><b>libsystem</b> - System abstraction library, from LLVM.</li>
 218 <li><b>libbasic</b> - Diagnostics, SourceLocations, SourceBuffer abstraction,
 219     file system caching for input source files.</li>
 220 <li><b>libast</b> - Provides classes to represent the C AST, the C type system,
 221     builtin functions, and various helpers for analyzing and manipulating the
 222     AST (visitors, pretty printers, etc).</li>
 223 <li><b>liblex</b> - Lexing and preprocessing, identifier hash table, pragma
 224     handling, tokens, and macro expansion.</li>
 225 <li><b>libparse</b> - Parsing. This library invokes coarse-grained 'Actions'
 226     provided by the client (e.g. libsema builds ASTs) but knows nothing about
 227     ASTs or other client-specific data structures.</li>
 228 <li><b>libsema</b> - Semantic Analysis.  This provides a set of parser actions
 229     to build a standardized AST for programs.</li>
 230 <li><b>libcodegen</b> - Lower the AST to LLVM IR for optimization &amp; code
 231     generation.</li>
 232 <li><b>librewrite</b> - Editing of text buffers (important for code rewriting
 233     transformation, like refactoring).</li>
 234 <li><b>libanalysis</b> - Static analysis support.</li>
 235 <li><b><a href="docs/libIndex.html">libindex</a></b> - Cross-translation-unit infrastructure and indexing support.</li>
 236 <li><b>clang</b> - A driver program, client of the libraries at various
 237     levels.</li>
 238 </ul>
 239
 240 <p>As an example of the power of this library based design....  If you wanted to
 241 build a preprocessor, you would take the Basic and Lexer libraries. If you want
 242 an indexer, you would take the previous two and add the Parser library and
 243 some actions for indexing. If you want a refactoring, static analysis, or
 244 source-to-source compiler tool, you would then add the AST building and
 245 semantic analyzer libraries.</p>
 246
 247 <p>For more information about the low-level implementation details of the
 248 various clang libraries, please see the <a href="docs/InternalsManual.html">
 249 clang Internals Manual</a>.</p>
 250
 251 <!--=======================================================================-->
 252 <h3><a name="diverseclients">Support Diverse Clients</a></h3>
 253 <!--=======================================================================-->
 254
 255 <p>Clang is designed and built with many grand plans for how we can use it.  The
 256 driving force is the fact that we use C and C++ daily, and have to suffer due to
 257 a lack of good tools available for it.  We believe that the C and C++ tools
 258 ecosystem has been significantly limited by how difficult it is to parse and
 259 represent the source code for these languages, and we aim to rectify this
 260 problem in clang.</p>
 261
 262 <p>The problem with this goal is that different clients have very different
 263 requirements.  Consider code generation, for example: a simple front-end that
 264 parses for code generation must analyze the code for validity and emit code
 265 in some intermediate form to pass off to a optimizer or backend.  Because
 266 validity analysis and code generation can largely be done on the fly, there is
 267 not hard requirement that the front-end actually build up a full AST for all
 268 the expressions and statements in the code.  TCC and GCC are examples of
 269 compilers that either build no real AST (in the former case) or build a stripped
 270 down and simplified AST (in the later case) because they focus primarily on
 271 codegen.</p>
 272
 273 <p>On the opposite side of the spectrum, some clients (like refactoring) want
 274 highly detailed information about the original source code and want a complete
 275 AST to describe it with.  Refactoring wants to have information about macro
 276 expansions, the location of every paren expression '(((x)))' vs 'x', full
 277 position information, and much more.  Further, refactoring wants to look
 278 <em>across the whole program</em> to ensure that it is making transformations
 279 that are safe.  Making this efficient and getting this right requires a
 280 significant amount of engineering and algorithmic work that simply are
 281 unnecessary for a simple static compiler.</p>
 282
 283 <p>The beauty of the clang approach is that it does not restrict how you use it.
 284 In particular, it is possible to use the clang preprocessor and parser to build
 285 an extremely quick and light-weight on-the-fly code generator (similar to TCC)
 286 that does not build an AST at all.   As an intermediate step, clang supports
 287 using the current AST generation and semantic analysis code and having a code
 288 generation client free the AST for each function after code generation. Finally,
 289 clang provides support for building and retaining fully-fledged ASTs, and even
 290 supports writing them out to disk.</p>
 291
 292 <p>Designing the libraries with clean and simple APIs allows these high-level
 293 policy decisions to be determined in the client, instead of forcing "one true
 294 way" in the implementation of any of these libraries.  Getting this right is
 295 hard, and we don't always get it right the first time, but we fix any problems
 296 when we realize we made a mistake.</p>
 297
 298 <!--=======================================================================-->
 299 <h3><a name="ideintegration">Integration with IDEs</h3>
 300 <!--=======================================================================-->
 301
 302 <p>
 303 We believe that Integrated Development Environments (IDE's) are a great way
 304 to pull together various pieces of the development puzzle, and aim to make clang
 305 work well in such an environment.  The chief advantage of an IDE is that they
 306 typically have visibility across your entire project and are long-lived
 307 processes, whereas stand-alone compiler tools are typically invoked on each
 308 individual file in the project, and thus have limited scope.</p>
 309
 310 <p>There are many implications of this difference, but a significant one has to
 311 do with efficiency and caching: sharing an address space across different files
 312 in a project, means that you can use intelligent caching and other techniques to
 313 dramatically reduce analysis/compilation time.</p>
 314
 315 <p>A further difference between IDEs and batch compiler is that they often
 316 impose very different requirements on the front-end: they depend on high
 317 performance in order to provide a "snappy" experience, and thus really want
 318 techniques like "incremental compilation", "fuzzy parsing", etc.  Finally, IDEs
 319 often have very different requirements than code generation, often requiring
 320 information that a codegen-only frontend can throw away.  Clang is
 321 specifically designed and built to capture this information.
 322 </p>
 323
 324
 325 <!--=======================================================================-->
 326 <h3><a name="license">Use the LLVM 'BSD' License</a></h3>
 327 <!--=======================================================================-->
 328
 329 <p>We actively intend for clang (and LLVM as a whole) to be used for
 330 commercial projects, and the BSD license is the simplest way to allow this.  We
 331 feel that the license encourages contributors to pick up the source and work
 332 with it, and believe that those individuals and organizations will contribute
 333 back their work if they do not want to have to maintain a fork forever (which is
 334 time consuming and expensive when merges are involved).  Further, nobody makes
 335 money on compilers these days, but many people need them to get bigger goals
 336 accomplished: it makes sense for everyone to work together.</p>
 337
 338 <p>For more information about the LLVM/clang license, please see the <a
 339 href="http://llvm.org/docs/DeveloperPolicy.html#license">LLVM License
 340 Description</a> for more information.</p>
 341
 342
 343
 344 <!--*************************************************************************-->
 345 <h2><a name="design">Internal Design and Implementation</a></h2>
 346 <!--*************************************************************************-->
 347
 348 <!--=======================================================================-->
 349 <h3><a name="real">A real-world, production quality compiler</a></h3>
 350 <!--=======================================================================-->
 351
 352 <p>
 353 Clang is designed and built by experienced compiler developers who
 354 are increasingly frustrated with the problems that <a
 355 href="comparison.html">existing open source compilers</a> have.  Clang is
 356 carefully and thoughtfully designed and built to provide the foundation of a
 357 whole new generation of C/C++/Objective C development tools, and we intend for
 358 it to be production quality.</p>
 359
 360 <p>Being a production quality compiler means many things: it means being high
 361 performance, being solid and (relatively) bug free, and it means eventually
 362 being used and depended on by a broad range of people.  While we are still in
 363 the early development stages, we strongly believe that this will become a
 364 reality.</p>
 365
 366 <!--=======================================================================-->
 367 <h3><a name="simplecode">A simple and hackable code base</a></h3>
 368 <!--=======================================================================-->
 369
 370 <p>Our goal is to make it possible for anyone with a basic understanding
 371 of compilers and working knowledge of the C/C++/ObjC languages to understand and
 372 extend the clang source base.  A large part of this falls out of our decision to
 373 make the AST mirror the languages as closely as possible: you have your friendly
 374 if statement, for statement, parenthesis expression, structs, unions, etc, all
 375 represented in a simple and explicit way.</p>
 376
 377 <p>In addition to a simple design, we work to make the source base approachable
 378 by commenting it well, including citations of the language standards where
 379 appropriate, and designing the code for simplicity.  Beyond that, clang offers
 380 a set of AST dumpers, printers, and visualizers that make it easy to put code in
 381 and see how it is represented.</p>
 382
 383 <!--=======================================================================-->
 384 <h3><a name="unifiedparser">A single unified parser for C, Objective C, C++,
 385 and Objective C++</a></h3>
 386 <!--=======================================================================-->
 387
 388 <p>Clang is the "C Language Family Front-end", which means we intend to support
 389 the most popular members of the C family.  We are convinced that the right
 390 parsing technology for this class of languages is a hand-built recursive-descent
 391 parser.  Because it is plain C++ code, recursive descent makes it very easy for
 392 new developers to understand the code, it easily supports ad-hoc rules and other
 393 strange hacks required by C/C++, and makes it straight-forward to implement
 394 excellent diagnostics and error recovery.</p>
 395
 396 <p>We believe that implementing C/C++/ObjC in a single unified parser makes the
 397 end result easier to maintain and evolve than maintaining a separate C and C++
 398 parser which must be bugfixed and maintained independently of each other.</p>
 399
 400 <!--=======================================================================-->
 401 <h3><a name="conformance">Conformance with C/C++/ObjC and their
 402  variants</a></h3>
 403 <!--=======================================================================-->
 404
 405 <p>When you start work on implementing a language, you find out that there is a
 406 huge gap between how the language works and how most people understand it to
 407 work.  This gap is the difference between a normal programmer and a (scary?
 408 super-natural?) "language lawyer", who knows the ins and outs of the language
 409 and can grok standardese with ease.</p>
 410
 411 <p>In practice, being conformant with the languages means that we aim to support
 412 the full language, including the dark and dusty corners (like trigraphs,
 413 preprocessor arcana, C99 VLAs, etc).  Where we support extensions above and
 414 beyond what the standard officially allows, we make an effort to explicitly call
 415 this out in the code and emit warnings about it (which are disabled by default,
 416 but can optionally be mapped to either warnings or errors), allowing you to use
 417 clang in "strict" mode if you desire.</p>
 418
 419 <p>We also intend to support "dialects" of these languages, such as C89, K&amp;R
 420 C, C++'03, Objective-C 2, etc.</p>
 421
 422 </div>
 423 </body>
 424 </html>