2 <title>LLVM: bugpoint tool</title>
6 <center><h1>LLVM: <tt>bugpoint</tt> tool</h1></center>
13 <tt>bugpoint [options] [input LLVM ll/bc files] [LLVM passes] --args <program arguments>...</tt>
15 <img src="img/Debugging.gif" width=444 height=314 align=right>
18 The <tt>bugpoint</tt> tool narrows down the source of
19 problems in LLVM tools and passes. It can be used to debug three types of
20 failures: optimizer crashes, miscompilations by optimizers, or bad native
21 code generation (including problems in the static and JIT compilers). It aims
22 to reduce large test cases to small, useful ones. For example,
23 if <tt><a href="CommandGuide/gccas.html">gccas</a></tt> crashes while optimizing a file, it
24 will identify the optimization (or combination of optimizations) that causes the
25 crash, and reduce the file down to a small example which triggers the crash.<p>
27 <a name="designphilosophy">
28 <h4>Design Philosophy</h4>
30 <tt>bugpoint</tt> is designed to be a useful tool without requiring any
31 hooks into the LLVM infrastructure at all. It works with any and all LLVM
32 passes and code generators, and does not need to "know" how they work. Because
33 of this, it may appear to do stupid things or miss obvious
34 simplifications. <tt>bugpoint</tt> is also designed to trade off programmer
35 time for computer time in the compiler-debugging process; consequently, it may
36 take a long period of (unattended) time to reduce a test case, but we feel it
37 is still worth it. Note that <tt>bugpoint</tt> is generally very quick unless
38 debugging a miscompilation where each test of the program (which requires
39 executing it) takes a long time.<p>
41 <a name="automaticdebuggerselection">
42 <h4>Automatic Debugger Selection</h4>
44 <tt>bugpoint</tt> reads each <tt>.bc</tt> or <tt>.ll</tt> file
45 specified on the command line and links them together into a single module,
46 called the test program. If any LLVM passes are
47 specified on the command line, it runs these passes on the test program. If
48 any of the passes crash, or if they produce malformed output (which causes the
50 <tt>bugpoint</tt> starts the <a href="#crashdebug">crash debugger</a>.<p>
52 Otherwise, if the <a href="#opt_output"><tt>-output</tt></a> option was not
53 specified, <tt>bugpoint</tt> runs the test program with the C backend (which is
54 assumed to generate good code) to generate a reference output. Once
55 <tt>bugpoint</tt> has a reference output for the test program, it tries
56 executing it with the <a href="#opt_run-">selected</a> code generator. If the
57 selected code generator crashes, <tt>bugpoint</tt> starts the <a
58 href="#crashdebug">crash debugger</a> on the code generator. Otherwise, if the
59 resulting output differs from the reference output, it assumes the difference
60 resulted from a code generator failure, and starts the <a
61 href="#codegendebug">code generator debugger</a>.<p>
63 Finally, if the output of the selected code generator matches the reference
64 output, <tt>bugpoint</tt> runs the test program after all of the LLVM passes
65 have been applied to it. If its output differs from the reference output, it
66 assumes the difference resulted from a failure in one of the LLVM passes, and
67 enters the <a href="#miscompilationdebug">miscompilation
68 debugger</a>. Otherwise, there is no problem <tt>bugpoint</tt> can debug.<p>
71 <h4>Crash debugger</h4>
73 If an optimizer or code generator crashes, <tt>bugpoint</tt> will try as hard as
74 it can to reduce the list of passes (for optimizer crashes) and the size of the
75 test program. First, <tt>bugpoint</tt> figures out which combination of
76 optimizer passes triggers the bug. This is useful when debugging a problem
77 exposed by <tt>gccas</tt>, for example, because it runs over 38 passes.<p>
79 Next, <tt>bugpoint</tt> tries removing functions from the test program, to
80 reduce its size. Usually it is able to reduce a test program to a single
81 function, when debugging intraprocedural optimizations. Once the number of
82 functions has been reduced, it attempts to delete various edges in the control
83 flow graph, to reduce the size of the function as much as possible. Finally,
84 <tt>bugpoint</tt> deletes any individual LLVM instructions whose absence does
85 not eliminate the failure. At the end, <tt>bugpoint</tt> should tell you what
86 passes crash, give you a bytecode file, and give you instructions on how to
87 reproduce the failure with <tt><a href="CommandGuide/opt.html">opt</a></tt>, <tt><a
88 href="CommandGuide/analyze.html">analyze</a></tt>, or <tt><a href="CommandGuide/llc.html">llc</a></tt>.<p>
90 <a name="codegendebug">
91 <h4>Code generator debugger</h4>
93 <p>The code generator debugger attempts to narrow down the amount of code that
94 is being miscompiled by the <a href="#opt_run-">selected</a> code generator. To
95 do this, it takes the test program and partitions it into two pieces: one piece
96 which it compiles with the C backend (into a shared object), and one piece which
97 it runs with either the JIT or the static LLC compiler. It uses several
98 techniques to reduce the amount of code pushed through the LLVM code generator,
99 to reduce the potential scope of the problem. After it is finished, it emits
100 two bytecode files (called "test" [to be compiled with the code generator] and
101 "safe" [to be compiled with the C backend], respectively), and instructions for
102 reproducing the problem. The code generator debugger assumes that the C backend
103 produces good code.</p>
105 <a name="miscompilationdebug">
106 <h4>Miscompilation debugger</h4>
108 The miscompilation debugger works similarly to the code generator
109 debugger. It works by splitting the test program into two pieces, running the
110 optimizations specified on one piece, linking the two pieces back together,
111 and then executing the result.
112 It attempts to narrow down the list of passes to the one (or few) which are
113 causing the miscompilation, then reduce the portion of the test program which is
114 being miscompiled. The miscompilation debugger assumes that the selected
115 code generator is working properly.<p>
117 <a name="bugpoint notes">
118 <h4>Advice for using <tt>bugpoint</tt></h4>
120 <tt>bugpoint</tt> can be a remarkably useful tool, but it sometimes works in
121 non-obvious ways. Here are some hints and tips:<p>
124 <li>In the code generator and miscompilation debuggers, <tt>bugpoint</tt> only
125 works with programs that have deterministic output. Thus, if the program
126 outputs <tt>argv[0]</tt>, the date, time, or any other "random" data, <tt>bugpoint</tt> may
127 misinterpret differences in these data, when output, as the result of a
128 miscompilation. Programs should be temporarily modified to disable
129 outputs that are likely to vary from run to run.
131 <li>In the code generator and miscompilation debuggers, debugging will go
132 faster if you manually modify the program or its inputs to reduce the
133 runtime, but still exhibit the problem.
135 <li><tt>bugpoint</tt> is extremely useful when working on a new optimization:
136 it helps track down regressions quickly. To avoid having to relink
137 <tt>bugpoint</tt> every time you change your optimization however, have
138 <tt>bugpoint</tt> dynamically load your optimization with the <a
139 href="#opt_load"><tt>-load</tt></a> option.
141 <li><tt>bugpoint</tt> can generate a lot of output and run for a long period of
142 time. It is often useful to capture the output of the program to file. For
143 example, in the C shell, you can type:<br>
144 <tt>bugpoint ..... |& tee bugpoint.log</tt>
145 <br>to get a copy of <tt>bugpoint</tt>'s output in the file
146 <tt>bugpoint.log</tt>, as well as on your terminal.
148 <li><tt>bugpoint</tt> cannot debug problems with the LLVM linker. If
149 <tt>bugpoint</tt> crashes before you see its "All input ok" message,
150 you might try <tt>llvm-link -v</tt> on the same set of input files. If
151 that also crashes, you may be experiencing a linker bug.
153 <li>If your program is <b>supposed</b> to crash, <tt>bugpoint</tt> will be
154 confused. One way to deal with this is to cause bugpoint to ignore the exit
155 code from your program, by giving it the <tt>-check-exit-code=false</tt>
163 <li><tt>-additional-so <library></tt><br>
164 Load <tt><library></tt> into the test program whenever it is run.
165 This is useful if you are debugging programs which depend on non-LLVM
166 libraries (such as the X or curses libraries) to run.<p>
168 <li><tt>-args <program args></tt><br>
169 Pass all arguments specified after <tt>-args</tt> to the
170 test program whenever it runs. Note that if any of
171 the <tt><program args></tt> start with a '-', you should use:
173 <tt>bugpoint <bugpoint args> -args -- <program args></tt>
175 The "<tt>--</tt>" right after the <tt>-args</tt> option tells
176 <tt>bugpoint</tt> to consider any options starting with <tt>-</tt> to be
177 part of the <tt>-args</tt> option, not as options to <tt>bugpoint</tt>
180 <li><tt>-tool-args <tool args></tt><br>
181 Pass all arguments specified after <tt>-tool-args</tt> to the
182 LLVM tool under test (llc, lli, etc.) whenever it runs.
183 You should use this option in the following way:
185 <tt>bugpoint <bugpoint args> -tool-args -- <tool args></tt>
187 The "<tt>--</tt>" right after the <tt>-tool-args</tt> option tells
188 <tt>bugpoint</tt> to consider any options starting with <tt>-</tt> to be
189 part of the <tt>-tool-args</tt> option, not as options to
190 <tt>bugpoint</tt> itself. (See <tt>-args</tt>, above.)<p>
192 <li><tt>-check-exit-code={true,false}</tt><br>
193 Assume a non-zero exit code or core dump from the test program is
194 a failure. Defaults to true.<p>
196 <li><tt>-disable-{dce,simplifycfg}</tt><br>
197 Do not run the specified passes to clean up and reduce the size of the
198 test program. By default, <tt>bugpoint</tt> uses these passes internally
199 when attempting to reduce test programs. If you're trying to find
200 a bug in one of these passes, <tt>bugpoint</tt> may crash.<p>
202 <li> <tt>-help</tt><br>
203 Print a summary of command line options.<p>
205 <a name="opt_input"><li><tt>-input <filename></tt><br>
206 Open <tt><filename></tt> and redirect the standard input of the
207 test program, whenever it runs, to come from that file.
210 <a name="opt_load"><li> <tt>-load <plugin></tt><br>
211 Load the dynamic object <tt><plugin></tt> into <tt>bugpoint</tt>
212 itself. This object should register new
213 optimization passes. Once loaded, the object will add new command line
214 options to enable various optimizations. To see the new complete list
215 of optimizations, use the -help and -load options together:
217 <tt>bugpoint -load <plugin> -help</tt>
220 <a name="opt_output"><li><tt>-output <filename></tt><br>
221 Whenever the test program produces output on its standard output
222 stream, it should match the contents of <tt><filename></tt>
223 (the "reference output"). If you do not use this option,
224 <tt>bugpoint</tt> will attempt to generate a reference output by
225 compiling the program with the C backend and running it.<p>
227 <li><tt>-profile-info-file <filename></tt><br>
228 Profile file loaded by -profile-loader.<p>
230 <a name="opt_run-"><li><tt>-run-{int,jit,llc,cbe}</tt><br>
231 Whenever the test program is compiled, <tt>bugpoint</tt> should generate
232 code for it using the specified code generator. These options allow
233 you to choose the interpreter, the JIT compiler, the static native
234 code compiler, or the C backend, respectively.<p>
239 If <tt>bugpoint</tt> succeeds in finding a problem, it will exit with 0.
240 Otherwise, if an error occurs, it will exit with a non-zero value.
243 Maintained by the <a href="http://llvm.cs.uiuc.edu">LLVM Team</a>.