xref: /openbsd/gnu/llvm/llvm/docs/XRayExample.rst (revision d415bd75)
1===================
2Debugging with XRay
3===================
4
5This document shows an example of how you would go about analyzing applications
6built with XRay instrumentation. Here we will attempt to debug ``llc``
7compiling some sample LLVM IR generated by Clang.
8
9.. contents::
10  :local:
11
12Building with XRay
13------------------
14
15To debug an application with XRay instrumentation, we need to build it with a
16Clang that supports the ``-fxray-instrument`` option. See `XRay <XRay.html>`_
17for more technical details of how XRay works for background information.
18
19In our example, we need to add ``-fxray-instrument`` to the list of flags
20passed to Clang when building a binary. Note that we need to link with Clang as
21well to get the XRay runtime linked in appropriately. For building ``llc`` with
22XRay, we do something similar below for our LLVM build:
23
24::
25
26  $ mkdir -p llvm-build && cd llvm-build
27  # Assume that the LLVM sources are at ../llvm
28  $ cmake -GNinja ../llvm -DCMAKE_BUILD_TYPE=Release \
29      -DCMAKE_C_FLAGS_RELEASE="-fxray-instrument" -DCMAKE_CXX_FLAGS="-fxray-instrument" \
30  # Once this finishes, we should build llc
31  $ ninja llc
32
33
34To verify that we have an XRay instrumented binary, we can use ``objdump`` to
35look for the ``xray_instr_map`` section.
36
37::
38
39  $ objdump -h -j xray_instr_map ./bin/llc
40  ./bin/llc:     file format elf64-x86-64
41
42  Sections:
43  Idx Name          Size      VMA               LMA               File off  Algn
44   14 xray_instr_map 00002fc0  00000000041516c6  00000000041516c6  03d516c6  2**0
45                    CONTENTS, ALLOC, LOAD, READONLY, DATA
46
47Getting Traces
48--------------
49
50By default, XRay does not write out the trace files or patch the application
51before main starts. If we run ``llc`` it should work like a normally built
52binary. If we want to get a full trace of the application's operations (of the
53functions we do end up instrumenting with XRay) then we need to enable XRay
54at application start. To do this, XRay checks the ``XRAY_OPTIONS`` environment
55variable.
56
57::
58
59  # The following doesn't create an XRay trace by default.
60  $ ./bin/llc input.ll
61
62  # We need to set the XRAY_OPTIONS to enable some features.
63  $ XRAY_OPTIONS="patch_premain=true xray_mode=xray-basic verbosity=1" ./bin/llc input.ll
64  ==69819==XRay: Log file in 'xray-log.llc.m35qPB'
65
66At this point we now have an XRay trace we can start analysing.
67
68The ``llvm-xray`` Tool
69----------------------
70
71Having a trace then allows us to do basic accounting of the functions that were
72instrumented, and how much time we're spending in parts of the code. To make
73sense of this data, we use the ``llvm-xray`` tool which has a few subcommands
74to help us understand our trace.
75
76One of the things we can do is to get an accounting of the functions that have
77been instrumented. We can see an example accounting with ``llvm-xray account``:
78
79::
80
81  $ llvm-xray account xray-log.llc.m35qPB --top=10 --sort=sum --sortorder=dsc --instr_map=./bin/llc
82  Functions with latencies: 29
83     funcid      count [      min,       med,       90p,       99p,       max]       sum  function
84        187        360 [ 0.000000,  0.000001,  0.000014,  0.000032,  0.000075]  0.001596  LLLexer.cpp:446:0: llvm::LLLexer::LexIdentifier()
85         85        130 [ 0.000000,  0.000000,  0.000018,  0.000023,  0.000156]  0.000799  X86ISelDAGToDAG.cpp:1984:0: (anonymous namespace)::X86DAGToDAGISel::Select(llvm::SDNode*)
86        138        130 [ 0.000000,  0.000000,  0.000017,  0.000155,  0.000155]  0.000774  SelectionDAGISel.cpp:2963:0: llvm::SelectionDAGISel::SelectCodeCommon(llvm::SDNode*, unsigned char const*, unsigned int)
87        188        103 [ 0.000000,  0.000000,  0.000003,  0.000123,  0.000214]  0.000737  LLParser.cpp:2692:0: llvm::LLParser::ParseValID(llvm::ValID&, llvm::LLParser::PerFunctionState*)
88         88          1 [ 0.000562,  0.000562,  0.000562,  0.000562,  0.000562]  0.000562  X86ISelLowering.cpp:83:0: llvm::X86TargetLowering::X86TargetLowering(llvm::X86TargetMachine const&, llvm::X86Subtarget const&)
89        125        102 [ 0.000001,  0.000003,  0.000010,  0.000017,  0.000049]  0.000471  Verifier.cpp:3714:0: (anonymous namespace)::Verifier::visitInstruction(llvm::Instruction&)
90         90          8 [ 0.000023,  0.000035,  0.000106,  0.000106,  0.000106]  0.000342  X86ISelLowering.cpp:3363:0: llvm::X86TargetLowering::LowerCall(llvm::TargetLowering::CallLoweringInfo&, llvm::SmallVectorImpl<llvm::SDValue>&) const
91        124         32 [ 0.000003,  0.000007,  0.000016,  0.000041,  0.000041]  0.000310  Verifier.cpp:1967:0: (anonymous namespace)::Verifier::visitFunction(llvm::Function const&)
92        123          1 [ 0.000302,  0.000302,  0.000302,  0.000302,  0.000302]  0.000302  LLVMContextImpl.cpp:54:0: llvm::LLVMContextImpl::~LLVMContextImpl()
93        139         46 [ 0.000000,  0.000002,  0.000006,  0.000008,  0.000019]  0.000138  TargetLowering.cpp:506:0: llvm::TargetLowering::SimplifyDemandedBits(llvm::SDValue, llvm::APInt const&, llvm::APInt&, llvm::APInt&, llvm::TargetLowering::TargetLoweringOpt&, unsigned int, bool) const
94
95This shows us that for our input file, ``llc`` spent the most cumulative time
96in the lexer (a total of 1 millisecond). If we wanted for example to work with
97this data in a spreadsheet, we can output the results as CSV using the
98``-format=csv`` option to the command for further analysis.
99
100If we want to get a textual representation of the raw trace we can use the
101``llvm-xray convert`` tool to get YAML output. The first few lines of that
102output for an example trace would look like the following:
103
104::
105
106  $ llvm-xray convert -f yaml --symbolize --instr_map=./bin/llc xray-log.llc.m35qPB
107  ---
108  header:
109    version:         1
110    type:            0
111    constant-tsc:    true
112    nonstop-tsc:     true
113    cycle-frequency: 2601000000
114  records:
115    - { type: 0, func-id: 110, function: __cxx_global_var_init.8, cpu: 37, thread: 69819, kind: function-enter, tsc: 5434426023268520 }
116    - { type: 0, func-id: 110, function: __cxx_global_var_init.8, cpu: 37, thread: 69819, kind: function-exit, tsc: 5434426023523052 }
117    - { type: 0, func-id: 164, function: __cxx_global_var_init, cpu: 37, thread: 69819, kind: function-enter, tsc: 5434426029925386 }
118    - { type: 0, func-id: 164, function: __cxx_global_var_init, cpu: 37, thread: 69819, kind: function-exit, tsc: 5434426030031128 }
119    - { type: 0, func-id: 142, function: '(anonymous namespace)::CommandLineParser::ParseCommandLineOptions(int, char const* const*, llvm::StringRef, llvm::raw_ostream*)', cpu: 37, thread: 69819, kind: function-enter, tsc: 5434426046951388 }
120    - { type: 0, func-id: 142, function: '(anonymous namespace)::CommandLineParser::ParseCommandLineOptions(int, char const* const*, llvm::StringRef, llvm::raw_ostream*)', cpu: 37, thread: 69819, kind: function-exit, tsc: 5434426047282020 }
121    - { type: 0, func-id: 187, function: 'llvm::LLLexer::LexIdentifier()', cpu: 37, thread: 69819, kind: function-enter, tsc: 5434426047857332 }
122    - { type: 0, func-id: 187, function: 'llvm::LLLexer::LexIdentifier()', cpu: 37, thread: 69819, kind: function-exit, tsc: 5434426047984152 }
123    - { type: 0, func-id: 187, function: 'llvm::LLLexer::LexIdentifier()', cpu: 37, thread: 69819, kind: function-enter, tsc: 5434426048036584 }
124    - { type: 0, func-id: 187, function: 'llvm::LLLexer::LexIdentifier()', cpu: 37, thread: 69819, kind: function-exit, tsc: 5434426048042292 }
125    - { type: 0, func-id: 187, function: 'llvm::LLLexer::LexIdentifier()', cpu: 37, thread: 69819, kind: function-enter, tsc: 5434426048055056 }
126    - { type: 0, func-id: 187, function: 'llvm::LLLexer::LexIdentifier()', cpu: 37, thread: 69819, kind: function-exit, tsc: 5434426048067316 }
127
128Controlling Fidelity
129--------------------
130
131So far in our examples, we haven't been getting full coverage of the functions
132we have in the binary. To get that, we need to modify the compiler flags so
133that we can instrument more (if not all) the functions we have in the binary.
134We have two options for doing that, and we explore both of these below.
135
136Instruction Threshold
137`````````````````````
138
139The first "blunt" way of doing this is by setting the minimum threshold for
140function bodies to 1. We can do that with the
141``-fxray-instruction-threshold=N`` flag when building our binary. We rebuild
142``llc`` with this option and observe the results:
143
144::
145
146  $ rm CMakeCache.txt
147  $ cmake -GNinja ../llvm -DCMAKE_BUILD_TYPE=Release \
148      -DCMAKE_C_FLAGS_RELEASE="-fxray-instrument -fxray-instruction-threshold=1" \
149      -DCMAKE_CXX_FLAGS="-fxray-instrument -fxray-instruction-threshold=1"
150  $ ninja llc
151  $ XRAY_OPTIONS="patch_premain=true" ./bin/llc input.ll
152  ==69819==XRay: Log file in 'xray-log.llc.5rqxkU'
153
154  $ llvm-xray account xray-log.llc.5rqxkU --top=10 --sort=sum --sortorder=dsc --instr_map=./bin/llc
155  Functions with latencies: 36652
156   funcid      count [      min,       med,       90p,       99p,       max]       sum  function
157       75          1 [ 0.672368,  0.672368,  0.672368,  0.672368,  0.672368]  0.672368  llc.cpp:271:0: main
158       78          1 [ 0.626455,  0.626455,  0.626455,  0.626455,  0.626455]  0.626455  llc.cpp:381:0: compileModule(char**, llvm::LLVMContext&)
159   139617          1 [ 0.472618,  0.472618,  0.472618,  0.472618,  0.472618]  0.472618  LegacyPassManager.cpp:1723:0: llvm::legacy::PassManager::run(llvm::Module&)
160   139610          1 [ 0.472618,  0.472618,  0.472618,  0.472618,  0.472618]  0.472618  LegacyPassManager.cpp:1681:0: llvm::legacy::PassManagerImpl::run(llvm::Module&)
161   139612          1 [ 0.470948,  0.470948,  0.470948,  0.470948,  0.470948]  0.470948  LegacyPassManager.cpp:1564:0: (anonymous namespace)::MPPassManager::runOnModule(llvm::Module&)
162   139607          2 [ 0.147345,  0.315994,  0.315994,  0.315994,  0.315994]  0.463340  LegacyPassManager.cpp:1530:0: llvm::FPPassManager::runOnModule(llvm::Module&)
163   139605         21 [ 0.000002,  0.000002,  0.102593,  0.213336,  0.213336]  0.463331  LegacyPassManager.cpp:1491:0: llvm::FPPassManager::runOnFunction(llvm::Function&)
164   139563      26096 [ 0.000002,  0.000002,  0.000037,  0.000063,  0.000215]  0.225708  LegacyPassManager.cpp:1083:0: llvm::PMDataManager::findAnalysisPass(void const*, bool)
165   108055        188 [ 0.000002,  0.000120,  0.001375,  0.004523,  0.062624]  0.159279  MachineFunctionPass.cpp:38:0: llvm::MachineFunctionPass::runOnFunction(llvm::Function&)
166    62635         22 [ 0.000041,  0.000046,  0.000050,  0.126744,  0.126744]  0.127715  X86TargetMachine.cpp:242:0: llvm::X86TargetMachine::getSubtargetImpl(llvm::Function const&) const
167
168
169Instrumentation Attributes
170``````````````````````````
171
172The other way is to use configuration files for selecting which functions
173should always be instrumented by the compiler. This gives us a way of ensuring
174that certain functions are either always or never instrumented by not having to
175add the attribute to the source.
176
177To use this feature, you can define one file for the functions to always
178instrument, and another for functions to never instrument. The format of these
179files are exactly the same as the SanitizerLists files that control similar
180things for the sanitizer implementations. For example:
181
182::
183
184  # xray-attr-list.txt
185  # always instrument functions that match the following filters:
186  [always]
187  fun:main
188
189  # never instrument functions that match the following filters:
190  [never]
191  fun:__cxx_*
192
193Given the file above we can re-build by providing it to the
194``-fxray-attr-list=`` flag to clang. You can have multiple files, each defining
195different sets of attribute sets, to be combined into a single list by clang.
196
197The XRay stack tool
198-------------------
199
200Given a trace, and optionally an instrumentation map, the ``llvm-xray stack``
201command can be used to analyze a call stack graph constructed from the function
202call timeline.
203
204The way to use the command is to output the top stacks by call count and time spent.
205
206::
207
208  $ llvm-xray stack xray-log.llc.5rqxkU --instr_map=./bin/llc
209
210  Unique Stacks: 3069
211  Top 10 Stacks by leaf sum:
212
213  Sum: 9633790
214  lvl   function                                                            count              sum
215  #0    main                                                                    1         58421550
216  #1    compileModule(char**, llvm::LLVMContext&)                               1         51440360
217  #2    llvm::legacy::PassManagerImpl::run(llvm::Module&)                       1         40535375
218  #3    llvm::FPPassManager::runOnModule(llvm::Module&)                         2         39337525
219  #4    llvm::FPPassManager::runOnFunction(llvm::Function&)                     6         39331465
220  #5    llvm::PMDataManager::verifyPreservedAnalysis(llvm::Pass*)             399         16628590
221  #6    llvm::PMTopLevelManager::findAnalysisPass(void const*)               4584         15155600
222  #7    llvm::PMDataManager::findAnalysisPass(void const*, bool)            32088          9633790
223
224  ..etc..
225
226In the default mode, identical stacks on different threads are independently
227aggregated. In a multithreaded program, you may end up having identical call
228stacks fill your list of top calls.
229
230To address this, you may specify the ``--aggregate-threads`` or
231``--per-thread-stacks`` flags. ``--per-thread-stacks`` treats the thread id as an
232implicit root in each call stack tree, while ``--aggregate-threads`` combines
233identical stacks from all threads.
234
235Flame Graph Generation
236----------------------
237
238The ``llvm-xray stack`` tool may also be used to generate flamegraphs for
239visualizing your instrumented invocations. The tool does not generate the graphs
240themselves, but instead generates a format that can be used with Brendan Gregg's
241FlameGraph tool, currently available on `github
242<https://github.com/brendangregg/FlameGraph>`_.
243
244To generate output for a flamegraph, a few more options are necessary.
245
246- ``--all-stacks`` - Emits all of the stacks.
247- ``--stack-format`` - Choose the flamegraph output format 'flame'.
248- ``--aggregation-type`` - Choose the metric to graph.
249
250You may pipe the command output directly to the flamegraph tool to obtain an
251svg file.
252
253::
254
255  $ llvm-xray stack xray-log.llc.5rqxkU --instr_map=./bin/llc --stack-format=flame --aggregation-type=time --all-stacks | \
256  /path/to/FlameGraph/flamegraph.pl > flamegraph.svg
257
258If you open the svg in a browser, mouse events allow exploring the call stacks.
259
260Chrome Trace Viewer Visualization
261---------------------------------
262
263We can also generate a trace which can be loaded by the Chrome Trace Viewer
264from the same generated trace:
265
266::
267
268  $ llvm-xray convert --symbolize --instr_map=./bin/llc \
269    --output-format=trace_event xray-log.llc.5rqxkU \
270      | gzip > llc-trace.txt.gz
271
272From a Chrome browser, navigating to ``chrome:///tracing`` allows us to load
273the ``sample-trace.txt.gz`` file to visualize the execution trace.
274
275Further Exploration
276-------------------
277
278The ``llvm-xray`` tool has a few other subcommands that are in various stages
279of being developed. One interesting subcommand that can highlight a few
280interesting things is the ``graph`` subcommand. Given for example the following
281toy program that we build with XRay instrumentation, we can see how the
282generated graph may be a helpful indicator of where time is being spent for the
283application.
284
285.. code-block:: c++
286
287  // sample.cc
288  #include <iostream>
289  #include <thread>
290
291  [[clang::xray_always_instrument]] void f() {
292    std::cerr << '.';
293  }
294
295  [[clang::xray_always_instrument]] void g() {
296    for (int i = 0; i < 1 << 10; ++i) {
297      std::cerr << '-';
298    }
299  }
300
301  int main(int argc, char* argv[]) {
302    std::thread t1([] {
303      for (int i = 0; i < 1 << 10; ++i)
304        f();
305    });
306    std::thread t2([] {
307      g();
308    });
309    t1.join();
310    t2.join();
311    std::cerr << '\n';
312  }
313
314We then build the above with XRay instrumentation:
315
316::
317
318  $ clang++ -o sample -O3 sample.cc -std=c++11 -fxray-instrument -fxray-instruction-threshold=1
319  $ XRAY_OPTIONS="patch_premain=true xray_mode=xray-basic" ./sample
320
321We can then explore the graph rendering of the trace generated by this sample
322application. We assume you have the graphviz tools available in your system,
323including both ``unflatten`` and ``dot``. If you prefer rendering or exploring
324the graph using another tool, then that should be feasible as well. ``llvm-xray
325graph`` will create DOT format graphs which should be usable in most graph
326rendering applications. One example invocation of the ``llvm-xray graph``
327command should yield some interesting insights to the workings of C++
328applications:
329
330::
331
332  $ llvm-xray graph xray-log.sample.* -m sample --color-edges=sum --edge-label=sum \
333      | unflatten -f -l10 | dot -Tsvg -o sample.svg
334
335
336Next Steps
337----------
338
339If you have some interesting analyses you'd like to implement as part of the
340llvm-xray tool, please feel free to propose them on the llvm-dev@ mailing list.
341The following are some ideas to inspire you in getting involved and potentially
342making things better.
343
344  - Implement a query/filtering library that allows for finding patterns in the
345    XRay traces.
346  - Collecting function call stacks and how often they're encountered in the
347    XRay trace.
348