www/performance-2008-10-31.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 - Performance</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 - Performance</h1>
  20 <!--*************************************************************************-->
  21
  22 <p>This page tracks the compile time performance of Clang on two
  23 interesting benchmarks:
  24 <ul>
  25   <li><i>Sketch</i>: The Objective-C example application shipped on
  26     Mac OS X as part of Xcode. <i>Sketch</i> is indicative of a
  27     "typical" Objective-C app. The source itself has a relatively
  28     small amount of code (~7,500 lines of source code), but it relies
  29     on the extensive Cocoa APIs to build its functionality. Like many
  30     Objective-C applications, it includes
  31     <tt>Cocoa/Cocoa.h</tt> in all of its source files, which represents a
  32     significant stress test of the front-end's performance on lexing,
  33     preprocessing, parsing, and syntax analysis.</li>
  34   <li><i>176.gcc</i>: This is the gcc-2.7.2.2 code base as present in
  35   SPECINT 2000. In contrast to Sketch, <i>176.gcc</i> consists of a
  36   large amount of C source code (~220,000 lines) with few system
  37   dependencies. This stresses the back-end's performance on generating
  38   assembly code and debug information.</li>
  39 </ul>
  40 </p>
  41
  42 <!--*************************************************************************-->
  43 <h2><a name="enduser">Experiments</a></h2>
  44 <!--*************************************************************************-->
  45
  46 <p>Measurements are done by serially processing each file in the
  47 respective benchmark, using Clang, gcc, and llvm-gcc as compilers. In
  48 order to track the performance of various subsystems the timings have
  49 been broken down into separate stages where possible:
  50
  51 <ul>
  52   <li><tt>-Eonly</tt>: This option runs the preprocessor but does not
  53     perform any output. For gcc and llvm-gcc, the -MM option is used
  54     as a rough equivalent to this step.</li>
  55   <li><tt>-parse-noop</tt>: This option runs the parser on the input,
  56     but without semantic analysis or any output. gcc and llvm-gcc have
  57     no equivalent for this option.</li>
  58   <li><tt>-fsyntax-only</tt>: This option runs the parser with semantic
  59     analysis.</li>
  60   <li><tt>-emit-llvm -O0</tt>: For Clang and llvm-gcc, this option
  61     converts to the LLVM intermediate representation but doesn't
  62     generate native code.</li>
  63   <li><tt>-S -O0</tt>: Perform actual code generation to produce a
  64     native assembler file.</li>
  65   <li><tt>-S -O0 -g</tt>: This adds emission of debug information to
  66     the assembly output.</li>
  67 </ul>
  68 </p>
  69
  70 <p>This set of stages is chosen to be approximately additive, that is
  71 each subsequent stage simply adds some additional processing. The
  72 timings measure the delta of the given stage from the previous
  73 one. For example, the timings for <tt>-fsyntax-only</tt> below show
  74 the difference of running with <tt>-fsyntax-only</tt> versus running
  75 with <tt>-parse-noop</tt> (for clang) or <tt>-MM</tt> with gcc and
  76 llvm-gcc. This amounts to a fairly accurate measure of only the time
  77 to perform semantic analysis (and parsing, in the case of gcc and llvm-gcc).</p>
  78
  79 <p>These timings are chosen to break down the compilation process for
  80 clang as much as possible. The graphs below show these numbers
  81 combined so that it is easy to see how the time for a particular task
  82 is divided among various components. For example, <tt>-S -O0</tt>
  83 includes the time of <tt>-fsyntax-only</tt> and <tt>-emit-llvm -O0</tt>.</p>
  84
  85 <p>Note that we already know that the LLVM optimizers are substantially (30-40%)
  86 faster than the GCC optimizers at a given -O level, so we only focus on -O0
  87 compile time here.</p>
  88
  89 <!--*************************************************************************-->
  90 <h2><a name="enduser">Timing Results</a></h2>
  91 <!--*************************************************************************-->
  92
  93 <!--=======================================================================-->
  94 <h3><a name="2008-10-31">2008-10-31</a></h3>
  95 <!--=======================================================================-->
  96
  97 <center><h4>Sketch</h4></center>
  98 <img class="img_slide"
  99      src="timing-data/2008-10-31/sketch.png" alt="Sketch Timings"/>
 100
 101 <p>This shows Clang's substantial performance improvements in
 102 preprocessing and semantic analysis; over 90% faster on
 103 -fsyntax-only. As expected, time spent in code generation for this
 104 benchmark is relatively small. One caveat, Clang's debug information
 105 generation for Objective-C is very incomplete; this means the <tt>-S
 106 -O0 -g</tt> numbers are unfair since Clang is generating substantially
 107 less output.</p>
 108
 109 <p>This chart also shows the effect of using precompiled headers (PCH)
 110 on compiler time. gcc and llvm-gcc see a large performance improvement
 111 with PCH; about 4x in wall time. Unfortunately, Clang does not yet
 112 have an implementation of PCH-style optimizations, but we are actively
 113 working to address this.</p>
 114
 115 <center><h4>176.gcc</h4></center>
 116 <img class="img_slide"
 117      src="timing-data/2008-10-31/176.gcc.png" alt="176.gcc Timings"/>
 118
 119 <p>Unlike the <i>Sketch</i> timings, compilation of <i>176.gcc</i>
 120 involves a large amount of code generation. The time spent in Clang's
 121 LLVM IR generation and code generation is on par with gcc's code
 122 generation time but the improved parsing & semantic analysis
 123 performance means Clang still comes in at ~29% faster versus gcc
 124 on <tt>-S -O0 -g</tt> and ~20% faster versus llvm-gcc.</p>
 125
 126 <p>These numbers indicate that Clang still has room for improvement in
 127 several areas, notably our LLVM IR generation is significantly slower
 128 than that of llvm-gcc, and both Clang and llvm-gcc incur a
 129 significantly higher cost for adding debugging information compared to
 130 gcc.</p>
 131
 132 </div>
 133 </body>
 134 </html>