3 ====================================
4 LLVM bugpoint tool: design and usage
5 ====================================
13 ``bugpoint`` narrows down the source of problems in LLVM tools and passes. It
14 can be used to debug three types of failures: optimizer crashes, miscompilations
15 by optimizers, or bad native code generation (including problems in the static
16 and JIT compilers). It aims to reduce large test cases to small, useful ones.
17 For example, if ``opt`` crashes while optimizing a file, it will identify the
18 optimization (or combination of optimizations) that causes the crash, and reduce
19 the file down to a small example which triggers the crash.
21 For detailed case scenarios, such as debugging ``opt``, or one of the LLVM code
22 generators, see `How To Submit a Bug Report document <HowToSubmitABug.html>`_.
27 ``bugpoint`` is designed to be a useful tool without requiring any hooks into
28 the LLVM infrastructure at all. It works with any and all LLVM passes and code
29 generators, and does not need to "know" how they work. Because of this, it may
30 appear to do stupid things or miss obvious simplifications. ``bugpoint`` is
31 also designed to trade off programmer time for computer time in the
32 compiler-debugging process; consequently, it may take a long period of
33 (unattended) time to reduce a test case, but we feel it is still worth it. Note
34 that ``bugpoint`` is generally very quick unless debugging a miscompilation
35 where each test of the program (which requires executing it) takes a long time.
37 Automatic Debugger Selection
38 ----------------------------
40 ``bugpoint`` reads each ``.bc`` or ``.ll`` file specified on the command line
41 and links them together into a single module, called the test program. If any
42 LLVM passes are specified on the command line, it runs these passes on the test
43 program. If any of the passes crash, or if they produce malformed output (which
44 causes the verifier to abort), ``bugpoint`` starts the `crash debugger`_.
46 Otherwise, if the ``-output`` option was not specified, ``bugpoint`` runs the
47 test program with the "safe" backend (which is assumed to generate good code) to
48 generate a reference output. Once ``bugpoint`` has a reference output for the
49 test program, it tries executing it with the selected code generator. If the
50 selected code generator crashes, ``bugpoint`` starts the `crash debugger`_ on
51 the code generator. Otherwise, if the resulting output differs from the
52 reference output, it assumes the difference resulted from a code generator
53 failure, and starts the `code generator debugger`_.
55 Finally, if the output of the selected code generator matches the reference
56 output, ``bugpoint`` runs the test program after all of the LLVM passes have
57 been applied to it. If its output differs from the reference output, it assumes
58 the difference resulted from a failure in one of the LLVM passes, and enters the
59 `miscompilation debugger`_. Otherwise, there is no problem ``bugpoint`` can
67 If an optimizer or code generator crashes, ``bugpoint`` will try as hard as it
68 can to reduce the list of passes (for optimizer crashes) and the size of the
69 test program. First, ``bugpoint`` figures out which combination of optimizer
70 passes triggers the bug. This is useful when debugging a problem exposed by
71 ``opt``, for example, because it runs over 38 passes.
73 Next, ``bugpoint`` tries removing functions from the test program, to reduce its
74 size. Usually it is able to reduce a test program to a single function, when
75 debugging intraprocedural optimizations. Once the number of functions has been
76 reduced, it attempts to delete various edges in the control flow graph, to
77 reduce the size of the function as much as possible. Finally, ``bugpoint``
78 deletes any individual LLVM instructions whose absence does not eliminate the
79 failure. At the end, ``bugpoint`` should tell you what passes crash, give you a
80 bitcode file, and give you instructions on how to reproduce the failure with
83 .. _code generator debugger:
85 Code generator debugger
86 -----------------------
88 The code generator debugger attempts to narrow down the amount of code that is
89 being miscompiled by the selected code generator. To do this, it takes the test
90 program and partitions it into two pieces: one piece which it compiles with the
91 "safe" backend (into a shared object), and one piece which it runs with either
92 the JIT or the static LLC compiler. It uses several techniques to reduce the
93 amount of code pushed through the LLVM code generator, to reduce the potential
94 scope of the problem. After it is finished, it emits two bitcode files (called
95 "test" [to be compiled with the code generator] and "safe" [to be compiled with
96 the "safe" backend], respectively), and instructions for reproducing the
97 problem. The code generator debugger assumes that the "safe" backend produces
100 .. _miscompilation debugger:
102 Miscompilation debugger
103 -----------------------
105 The miscompilation debugger works similarly to the code generator debugger. It
106 works by splitting the test program into two pieces, running the optimizations
107 specified on one piece, linking the two pieces back together, and then executing
108 the result. It attempts to narrow down the list of passes to the one (or few)
109 which are causing the miscompilation, then reduce the portion of the test
110 program which is being miscompiled. The miscompilation debugger assumes that
111 the selected code generator is working properly.
113 Advice for using bugpoint
114 =========================
116 ``bugpoint`` can be a remarkably useful tool, but it sometimes works in
117 non-obvious ways. Here are some hints and tips:
119 * In the code generator and miscompilation debuggers, ``bugpoint`` only works
120 with programs that have deterministic output. Thus, if the program outputs
121 ``argv[0]``, the date, time, or any other "random" data, ``bugpoint`` may
122 misinterpret differences in these data, when output, as the result of a
123 miscompilation. Programs should be temporarily modified to disable outputs
124 that are likely to vary from run to run.
126 * In the code generator and miscompilation debuggers, debugging will go faster
127 if you manually modify the program or its inputs to reduce the runtime, but
128 still exhibit the problem.
130 * ``bugpoint`` is extremely useful when working on a new optimization: it helps
131 track down regressions quickly. To avoid having to relink ``bugpoint`` every
132 time you change your optimization however, have ``bugpoint`` dynamically load
133 your optimization with the ``-load`` option.
135 * ``bugpoint`` can generate a lot of output and run for a long period of time.
136 It is often useful to capture the output of the program to file. For example,
137 in the C shell, you can run:
141 bugpoint ... |& tee bugpoint.log
143 to get a copy of ``bugpoint``'s output in the file ``bugpoint.log``, as well
146 * ``bugpoint`` cannot debug problems with the LLVM linker. If ``bugpoint``
147 crashes before you see its "All input ok" message, you might try ``llvm-link
148 -v`` on the same set of input files. If that also crashes, you may be
149 experiencing a linker bug.
151 * ``bugpoint`` is useful for proactively finding bugs in LLVM. Invoking
152 ``bugpoint`` with the ``-find-bugs`` option will cause the list of specified
153 optimizations to be randomized and applied to the program. This process will
154 repeat until a bug is found or the user kills ``bugpoint``.
156 What to do when bugpoint isn't enough
157 =====================================
159 Sometimes, ``bugpoint`` is not enough. In particular, InstCombine and
160 TargetLowering both have visitor structured code with lots of potential
161 transformations. If the process of using bugpoint has left you with still too
162 much code to figure out and the problem seems to be in instcombine, the
163 following steps may help. These same techniques are useful with TargetLowering
166 Turn on ``-debug-only=instcombine`` and see which transformations within
167 instcombine are firing by selecting out lines with "``IC``" in them.
169 At this point, you have a decision to make. Is the number of transformations
170 small enough to step through them using a debugger? If so, then try that.
172 If there are too many transformations, then a source modification approach may
173 be helpful. In this approach, you can modify the source code of instcombine to
174 disable just those transformations that are being performed on your test input
175 and perform a binary search over the set of transformations. One set of places
176 to modify are the "``visit*``" methods of ``InstCombiner`` (*e.g.*
177 ``visitICmpInst``) by adding a "``return false``" as the first line of the
180 If that still doesn't remove enough, then change the caller of
181 ``InstCombiner::DoOneIteration``, ``InstCombiner::runOnFunction`` to limit the
182 number of iterations.
184 You may also find it useful to use "``-stats``" now to see what parts of
185 instcombine are firing. This can guide where to put additional reporting code.
187 At this point, if the amount of transformations is still too large, then
188 inserting code to limit whether or not to execute the body of the code in the
189 visit function can be helpful. Add a static counter which is incremented on
190 every invocation of the function. Then add code which simply returns false on
191 desired ranges. For example:
196 static int calledCount = 0;
198 DEBUG(if (calledCount < 212) return false);
199 DEBUG(if (calledCount > 217) return false);
200 DEBUG(if (calledCount == 213) return false);
201 DEBUG(if (calledCount == 214) return false);
202 DEBUG(if (calledCount == 215) return false);
203 DEBUG(if (calledCount == 216) return false);
204 DEBUG(dbgs() << "visitXOR calledCount: " << calledCount << "\n");
205 DEBUG(dbgs() << "I: "; I->dump());
207 could be added to ``visitXOR`` to limit ``visitXor`` to being applied only to
208 calls 212 and 217. This is from an actual test case and raises an important
209 point---a simple binary search may not be sufficient, as transformations that
210 interact may require isolating more than one call. In TargetLowering, use
211 ``return SDNode();`` instead of ``return false;``.
213 Now that that the number of transformations is down to a manageable number, try
214 examining the output to see if you can figure out which transformations are
215 being done. If that can be figured out, then do the usual debugging. If which
216 code corresponds to the transformation being performed isn't obvious, set a
217 breakpoint after the call count based disabling and step through the code.
218 Alternatively, you can use "``printf``" style debugging to report waypoints.