1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
5 <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
6 <title>LLVM bugpoint tool: design and usage</title>
7 <link rel="stylesheet" href="_static/llvm.css" type="text/css">
11 LLVM bugpoint tool: design and usage
15 <li><a href="#desc">Description</a></li>
16 <li><a href="#design">Design Philosophy</a>
18 <li><a href="#autoselect">Automatic Debugger Selection</a></li>
19 <li><a href="#crashdebug">Crash debugger</a></li>
20 <li><a href="#codegendebug">Code generator debugger</a></li>
21 <li><a href="#miscompilationdebug">Miscompilation debugger</a></li>
23 <li><a href="#advice">Advice for using <tt>bugpoint</tt></a></li>
24 <li><a href="#notEnough">What to do when <tt>bugpoint</tt> isn't enough</a></li>
27 <div class="doc_author">
28 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a></p>
31 <!-- *********************************************************************** -->
33 <a name="desc">Description</a>
35 <!-- *********************************************************************** -->
39 <p><tt>bugpoint</tt> narrows down the source of problems in LLVM tools and
40 passes. It can be used to debug three types of failures: optimizer crashes,
41 miscompilations by optimizers, or bad native code generation (including problems
42 in the static and JIT compilers). It aims to reduce large test cases to small,
43 useful ones. For example, if <tt>opt</tt> crashes while optimizing a
44 file, it will identify the optimization (or combination of optimizations) that
45 causes the crash, and reduce the file down to a small example which triggers the
48 <p>For detailed case scenarios, such as debugging <tt>opt</tt>, or one of the
49 LLVM code generators, see <a href="HowToSubmitABug.html">How To Submit a Bug
50 Report document</a>.</p>
54 <!-- *********************************************************************** -->
56 <a name="design">Design Philosophy</a>
58 <!-- *********************************************************************** -->
62 <p><tt>bugpoint</tt> is designed to be a useful tool without requiring any
63 hooks into the LLVM infrastructure at all. It works with any and all LLVM
64 passes and code generators, and does not need to "know" how they work. Because
65 of this, it may appear to do stupid things or miss obvious
66 simplifications. <tt>bugpoint</tt> is also designed to trade off programmer
67 time for computer time in the compiler-debugging process; consequently, it may
68 take a long period of (unattended) time to reduce a test case, but we feel it
69 is still worth it. Note that <tt>bugpoint</tt> is generally very quick unless
70 debugging a miscompilation where each test of the program (which requires
71 executing it) takes a long time.</p>
73 <!-- ======================================================================= -->
75 <a name="autoselect">Automatic Debugger Selection</a>
80 <p><tt>bugpoint</tt> reads each <tt>.bc</tt> or <tt>.ll</tt> file specified on
81 the command line and links them together into a single module, called the test
82 program. If any LLVM passes are specified on the command line, it runs these
83 passes on the test program. If any of the passes crash, or if they produce
84 malformed output (which causes the verifier to abort), <tt>bugpoint</tt> starts
85 the <a href="#crashdebug">crash debugger</a>.</p>
87 <p>Otherwise, if the <tt>-output</tt> option was not specified,
88 <tt>bugpoint</tt> runs the test program with the C backend (which is assumed to
89 generate good code) to generate a reference output. Once <tt>bugpoint</tt> has
90 a reference output for the test program, it tries executing it with the
91 selected code generator. If the selected code generator crashes,
92 <tt>bugpoint</tt> starts the <a href="#crashdebug">crash debugger</a> on the
93 code generator. Otherwise, if the resulting output differs from the reference
94 output, it assumes the difference resulted from a code generator failure, and
95 starts the <a href="#codegendebug">code generator debugger</a>.</p>
97 <p>Finally, if the output of the selected code generator matches the reference
98 output, <tt>bugpoint</tt> runs the test program after all of the LLVM passes
99 have been applied to it. If its output differs from the reference output, it
100 assumes the difference resulted from a failure in one of the LLVM passes, and
101 enters the <a href="#miscompilationdebug">miscompilation debugger</a>.
102 Otherwise, there is no problem <tt>bugpoint</tt> can debug.</p>
106 <!-- ======================================================================= -->
108 <a name="crashdebug">Crash debugger</a>
113 <p>If an optimizer or code generator crashes, <tt>bugpoint</tt> will try as hard
114 as it can to reduce the list of passes (for optimizer crashes) and the size of
115 the test program. First, <tt>bugpoint</tt> figures out which combination of
116 optimizer passes triggers the bug. This is useful when debugging a problem
117 exposed by <tt>opt</tt>, for example, because it runs over 38 passes.</p>
119 <p>Next, <tt>bugpoint</tt> tries removing functions from the test program, to
120 reduce its size. Usually it is able to reduce a test program to a single
121 function, when debugging intraprocedural optimizations. Once the number of
122 functions has been reduced, it attempts to delete various edges in the control
123 flow graph, to reduce the size of the function as much as possible. Finally,
124 <tt>bugpoint</tt> deletes any individual LLVM instructions whose absence does
125 not eliminate the failure. At the end, <tt>bugpoint</tt> should tell you what
126 passes crash, give you a bitcode file, and give you instructions on how to
127 reproduce the failure with <tt>opt</tt> or <tt>llc</tt>.</p>
131 <!-- ======================================================================= -->
133 <a name="codegendebug">Code generator debugger</a>
138 <p>The code generator debugger attempts to narrow down the amount of code that
139 is being miscompiled by the selected code generator. To do this, it takes the
140 test program and partitions it into two pieces: one piece which it compiles
141 with the C backend (into a shared object), and one piece which it runs with
142 either the JIT or the static LLC compiler. It uses several techniques to
143 reduce the amount of code pushed through the LLVM code generator, to reduce the
144 potential scope of the problem. After it is finished, it emits two bitcode
145 files (called "test" [to be compiled with the code generator] and "safe" [to be
146 compiled with the C backend], respectively), and instructions for reproducing
147 the problem. The code generator debugger assumes that the C backend produces
152 <!-- ======================================================================= -->
154 <a name="miscompilationdebug">Miscompilation debugger</a>
159 <p>The miscompilation debugger works similarly to the code generator debugger.
160 It works by splitting the test program into two pieces, running the
161 optimizations specified on one piece, linking the two pieces back together, and
162 then executing the result. It attempts to narrow down the list of passes to
163 the one (or few) which are causing the miscompilation, then reduce the portion
164 of the test program which is being miscompiled. The miscompilation debugger
165 assumes that the selected code generator is working properly.</p>
171 <!-- *********************************************************************** -->
173 <a name="advice">Advice for using bugpoint</a>
175 <!-- *********************************************************************** -->
179 <tt>bugpoint</tt> can be a remarkably useful tool, but it sometimes works in
180 non-obvious ways. Here are some hints and tips:<p>
183 <li>In the code generator and miscompilation debuggers, <tt>bugpoint</tt> only
184 works with programs that have deterministic output. Thus, if the program
185 outputs <tt>argv[0]</tt>, the date, time, or any other "random" data,
186 <tt>bugpoint</tt> may misinterpret differences in these data, when output,
187 as the result of a miscompilation. Programs should be temporarily modified
188 to disable outputs that are likely to vary from run to run.
190 <li>In the code generator and miscompilation debuggers, debugging will go
191 faster if you manually modify the program or its inputs to reduce the
192 runtime, but still exhibit the problem.
194 <li><tt>bugpoint</tt> is extremely useful when working on a new optimization:
195 it helps track down regressions quickly. To avoid having to relink
196 <tt>bugpoint</tt> every time you change your optimization however, have
197 <tt>bugpoint</tt> dynamically load your optimization with the
198 <tt>-load</tt> option.
200 <li><p><tt>bugpoint</tt> can generate a lot of output and run for a long period
201 of time. It is often useful to capture the output of the program to file.
202 For example, in the C shell, you can run:</p>
204 <div class="doc_code">
205 <p><tt>bugpoint ... |& tee bugpoint.log</tt></p>
208 <p>to get a copy of <tt>bugpoint</tt>'s output in the file
209 <tt>bugpoint.log</tt>, as well as on your terminal.</p>
211 <li><tt>bugpoint</tt> cannot debug problems with the LLVM linker. If
212 <tt>bugpoint</tt> crashes before you see its "All input ok" message,
213 you might try <tt>llvm-link -v</tt> on the same set of input files. If
214 that also crashes, you may be experiencing a linker bug.
216 <li><tt>bugpoint</tt> is useful for proactively finding bugs in LLVM.
217 Invoking <tt>bugpoint</tt> with the <tt>-find-bugs</tt> option will cause
218 the list of specified optimizations to be randomized and applied to the
219 program. This process will repeat until a bug is found or the user
220 kills <tt>bugpoint</tt>.
224 <!-- *********************************************************************** -->
226 <a name="notEnough">What to do when bugpoint isn't enough</a>
228 <!-- *********************************************************************** -->
232 <p>Sometimes, <tt>bugpoint</tt> is not enough. In particular, InstCombine and
233 TargetLowering both have visitor structured code with lots of potential
234 transformations. If the process of using bugpoint has left you with
235 still too much code to figure out and the problem seems
236 to be in instcombine, the following steps may help. These same techniques
237 are useful with TargetLowering as well.</p>
239 <p>Turn on <tt>-debug-only=instcombine</tt> and see which transformations
240 within instcombine are firing by selecting out lines with "<tt>IC</tt>"
243 <p>At this point, you have a decision to make. Is the number
244 of transformations small enough to step through them using a debugger?
245 If so, then try that.</p>
247 <p>If there are too many transformations, then a source modification
248 approach may be helpful.
249 In this approach, you can modify the source code of instcombine
250 to disable just those transformations that are being performed on your
251 test input and perform a binary search over the set of transformations.
252 One set of places to modify are the "<tt>visit*</tt>" methods of
253 <tt>InstCombiner</tt> (<I>e.g.</I> <tt>visitICmpInst</tt>) by adding a
254 "<tt>return false</tt>" as the first line of the method.</p>
256 <p>If that still doesn't remove enough, then change the caller of
257 <tt>InstCombiner::DoOneIteration</tt>, <tt>InstCombiner::runOnFunction</tt>
258 to limit the number of iterations.</p>
260 <p>You may also find it useful to use "<tt>-stats</tt>" now to see what parts
261 of instcombine are firing. This can guide where to put additional reporting
264 <p>At this point, if the amount of transformations is still too large, then
265 inserting code to limit whether or not to execute the body of the code
266 in the visit function can be helpful. Add a static counter which is
267 incremented on every invocation of the function. Then add code which
268 simply returns false on desired ranges. For example:</p>
270 <div class="doc_code">
271 <p><tt>static int calledCount = 0;</tt></p>
272 <p><tt>calledCount++;</tt></p>
273 <p><tt>DEBUG(if (calledCount < 212) return false);</tt></p>
274 <p><tt>DEBUG(if (calledCount > 217) return false);</tt></p>
275 <p><tt>DEBUG(if (calledCount == 213) return false);</tt></p>
276 <p><tt>DEBUG(if (calledCount == 214) return false);</tt></p>
277 <p><tt>DEBUG(if (calledCount == 215) return false);</tt></p>
278 <p><tt>DEBUG(if (calledCount == 216) return false);</tt></p>
279 <p><tt>DEBUG(dbgs() << "visitXOR calledCount: " << calledCount
280 << "\n");</tt></p>
281 <p><tt>DEBUG(dbgs() << "I: "; I->dump());</tt></p>
284 <p>could be added to <tt>visitXOR</tt> to limit <tt>visitXor</tt> to being
285 applied only to calls 212 and 217. This is from an actual test case and raises
286 an important point---a simple binary search may not be sufficient, as
287 transformations that interact may require isolating more than one call.
288 In TargetLowering, use <tt>return SDNode();</tt> instead of
289 <tt>return false;</tt>.</p>
291 <p>Now that that the number of transformations is down to a manageable
292 number, try examining the output to see if you can figure out which
293 transformations are being done. If that can be figured out, then
294 do the usual debugging. If which code corresponds to the transformation
295 being performed isn't obvious, set a breakpoint after the call count
296 based disabling and step through the code. Alternatively, you can use
297 "printf" style debugging to report waypoints.</p>
301 <!-- *********************************************************************** -->
305 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
306 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
307 <a href="http://validator.w3.org/check/referer"><img
308 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
310 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
311 <a href="http://llvm.org/">LLVM Compiler Infrastructure</a><br>
312 Last modified: $Date$