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