1llvm-exegesis - LLVM Machine Instruction Benchmark
2==================================================
3
4.. program:: llvm-exegesis
5
6SYNOPSIS
7--------
8
9:program:`llvm-exegesis` [*options*]
10
11DESCRIPTION
12-----------
13
14:program:`llvm-exegesis` is a benchmarking tool that uses information available
15in LLVM to measure host machine instruction characteristics like latency,
16throughput, or port decomposition.
17
18Given an LLVM opcode name and a benchmarking mode, :program:`llvm-exegesis`
19generates a code snippet that makes execution as serial (resp. as parallel) as
20possible so that we can measure the latency (resp. inverse throughput/uop decomposition)
21of the instruction.
22The code snippet is jitted and executed on the host subtarget. The time taken
23(resp. resource usage) is measured using hardware performance counters. The
24result is printed out as YAML to the standard output.
25
26The main goal of this tool is to automatically (in)validate the LLVM's TableDef
27scheduling models. To that end, we also provide analysis of the results.
28
29:program:`llvm-exegesis` can also benchmark arbitrary user-provided code
30snippets.
31
32EXAMPLE 1: benchmarking instructions
33------------------------------------
34
35Assume you have an X86-64 machine. To measure the latency of a single
36instruction, run:
37
38.. code-block:: bash
39
40    $ llvm-exegesis -mode=latency -opcode-name=ADD64rr
41
42Measuring the uop decomposition or inverse throughput of an instruction works similarly:
43
44.. code-block:: bash
45
46    $ llvm-exegesis -mode=uops -opcode-name=ADD64rr
47    $ llvm-exegesis -mode=inverse_throughput -opcode-name=ADD64rr
48
49
50The output is a YAML document (the default is to write to stdout, but you can
51redirect the output to a file using `-benchmarks-file`):
52
53.. code-block:: none
54
55  ---
56  key:
57    opcode_name:     ADD64rr
58    mode:            latency
59    config:          ''
60  cpu_name:        haswell
61  llvm_triple:     x86_64-unknown-linux-gnu
62  num_repetitions: 10000
63  measurements:
64    - { key: latency, value: 1.0058, debug_string: '' }
65  error:           ''
66  info:            'explicit self cycles, selecting one aliasing configuration.
67  Snippet:
68  ADD64rr R8, R8, R10
69  '
70  ...
71
72To measure the latency of all instructions for the host architecture, run:
73
74.. code-block:: bash
75
76  #!/bin/bash
77  readonly INSTRUCTIONS=$(($(grep INSTRUCTION_LIST_END build/lib/Target/X86/X86GenInstrInfo.inc | cut -f2 -d=) - 1))
78  for INSTRUCTION in $(seq 1 ${INSTRUCTIONS});
79  do
80    ./build/bin/llvm-exegesis -mode=latency -opcode-index=${INSTRUCTION} | sed -n '/---/,$p'
81  done
82
83FIXME: Provide an :program:`llvm-exegesis` option to test all instructions.
84
85
86EXAMPLE 2: benchmarking a custom code snippet
87---------------------------------------------
88
89To measure the latency/uops of a custom piece of code, you can specify the
90`snippets-file` option (`-` reads from standard input).
91
92.. code-block:: bash
93
94    $ echo "vzeroupper" | llvm-exegesis -mode=uops -snippets-file=-
95
96Real-life code snippets typically depend on registers or memory.
97:program:`llvm-exegesis` checks the liveliness of registers (i.e. any register
98use has a corresponding def or is a "live in"). If your code depends on the
99value of some registers, you have two options:
100
101- Mark the register as requiring a definition. :program:`llvm-exegesis` will
102  automatically assign a value to the register. This can be done using the
103  directive `LLVM-EXEGESIS-DEFREG <reg name> <hex_value>`, where `<hex_value>`
104  is a bit pattern used to fill `<reg_name>`. If `<hex_value>` is smaller than
105  the register width, it will be sign-extended.
106- Mark the register as a "live in". :program:`llvm-exegesis` will benchmark
107  using whatever value was in this registers on entry. This can be done using
108  the directive `LLVM-EXEGESIS-LIVEIN <reg name>`.
109
110For example, the following code snippet depends on the values of XMM1 (which
111will be set by the tool) and the memory buffer passed in RDI (live in).
112
113.. code-block:: none
114
115  # LLVM-EXEGESIS-LIVEIN RDI
116  # LLVM-EXEGESIS-DEFREG XMM1 42
117  vmulps	(%rdi), %xmm1, %xmm2
118  vhaddps	%xmm2, %xmm2, %xmm3
119  addq $0x10, %rdi
120
121
122EXAMPLE 3: analysis
123-------------------
124
125Assuming you have a set of benchmarked instructions (either latency or uops) as
126YAML in file `/tmp/benchmarks.yaml`, you can analyze the results using the
127following command:
128
129.. code-block:: bash
130
131    $ llvm-exegesis -mode=analysis \
132  -benchmarks-file=/tmp/benchmarks.yaml \
133  -analysis-clusters-output-file=/tmp/clusters.csv \
134  -analysis-inconsistencies-output-file=/tmp/inconsistencies.html
135
136This will group the instructions into clusters with the same performance
137characteristics. The clusters will be written out to `/tmp/clusters.csv` in the
138following format:
139
140.. code-block:: none
141
142  cluster_id,opcode_name,config,sched_class
143  ...
144  2,ADD32ri8_DB,,WriteALU,1.00
145  2,ADD32ri_DB,,WriteALU,1.01
146  2,ADD32rr,,WriteALU,1.01
147  2,ADD32rr_DB,,WriteALU,1.00
148  2,ADD32rr_REV,,WriteALU,1.00
149  2,ADD64i32,,WriteALU,1.01
150  2,ADD64ri32,,WriteALU,1.01
151  2,MOVSX64rr32,,BSWAP32r_BSWAP64r_MOVSX64rr32,1.00
152  2,VPADDQYrr,,VPADDBYrr_VPADDDYrr_VPADDQYrr_VPADDWYrr_VPSUBBYrr_VPSUBDYrr_VPSUBQYrr_VPSUBWYrr,1.02
153  2,VPSUBQYrr,,VPADDBYrr_VPADDDYrr_VPADDQYrr_VPADDWYrr_VPSUBBYrr_VPSUBDYrr_VPSUBQYrr_VPSUBWYrr,1.01
154  2,ADD64ri8,,WriteALU,1.00
155  2,SETBr,,WriteSETCC,1.01
156  ...
157
158:program:`llvm-exegesis` will also analyze the clusters to point out
159inconsistencies in the scheduling information. The output is an html file. For
160example, `/tmp/inconsistencies.html` will contain messages like the following :
161
162.. image:: llvm-exegesis-analysis.png
163  :align: center
164
165Note that the scheduling class names will be resolved only when
166:program:`llvm-exegesis` is compiled in debug mode, else only the class id will
167be shown. This does not invalidate any of the analysis results though.
168
169OPTIONS
170-------
171
172.. option:: -help
173
174 Print a summary of command line options.
175
176.. option:: -opcode-index=<LLVM opcode index>
177
178 Specify the opcode to measure, by index. Specifying `-1` will result
179 in measuring every existing opcode. See example 1 for details.
180 Either `opcode-index`, `opcode-name` or `snippets-file` must be set.
181
182.. option:: -opcode-name=<opcode name 1>,<opcode name 2>,...
183
184 Specify the opcode to measure, by name. Several opcodes can be specified as
185 a comma-separated list. See example 1 for details.
186 Either `opcode-index`, `opcode-name` or `snippets-file` must be set.
187
188.. option:: -snippets-file=<filename>
189
190 Specify the custom code snippet to measure. See example 2 for details.
191 Either `opcode-index`, `opcode-name` or `snippets-file` must be set.
192
193.. option:: -mode=[latency|uops|inverse_throughput|analysis]
194
195 Specify the run mode. Note that if you pick `analysis` mode, you also need
196 to specify at least one of the `-analysis-clusters-output-file=` and
197 `-analysis-inconsistencies-output-file=`.
198
199.. option:: -repetition-mode=[duplicate|loop|min]
200
201 Specify the repetition mode. `duplicate` will create a large, straight line
202 basic block with `num-repetitions` copies of the snippet. `loop` will wrap
203 the snippet in a loop which will be run `num-repetitions` times. The `loop`
204 mode tends to better hide the effects of the CPU frontend on architectures
205 that cache decoded instructions, but consumes a register for counting
206 iterations. If performing an analysis over many opcodes, it may be best
207 to instead use the `min` mode, which will run each other mode, and produce
208 the minimal measured result.
209
210.. option:: -num-repetitions=<Number of repetitions>
211
212 Specify the number of repetitions of the asm snippet.
213 Higher values lead to more accurate measurements but lengthen the benchmark.
214
215.. option:: -max-configs-per-opcode=<value>
216
217 Specify the maximum configurations that can be generated for each opcode.
218 By default this is `1`, meaning that we assume that a single measurement is
219 enough to characterize an opcode. This might not be true of all instructions:
220 for example, the performance characteristics of the LEA instruction on X86
221 depends on the value of assigned registers and immediates. Setting a value of
222 `-max-configs-per-opcode` larger than `1` allows `llvm-exegesis` to explore
223 more configurations to discover if some register or immediate assignments
224 lead to different performance characteristics.
225
226
227.. option:: -benchmarks-file=</path/to/file>
228
229 File to read (`analysis` mode) or write (`latency`/`uops`/`inverse_throughput`
230 modes) benchmark results. "-" uses stdin/stdout.
231
232.. option:: -analysis-clusters-output-file=</path/to/file>
233
234 If provided, write the analysis clusters as CSV to this file. "-" prints to
235 stdout. By default, this analysis is not run.
236
237.. option:: -analysis-inconsistencies-output-file=</path/to/file>
238
239 If non-empty, write inconsistencies found during analysis to this file. `-`
240 prints to stdout. By default, this analysis is not run.
241
242.. option:: -analysis-clustering=[dbscan,naive]
243
244 Specify the clustering algorithm to use. By default DBSCAN will be used.
245 Naive clustering algorithm is better for doing further work on the
246 `-analysis-inconsistencies-output-file=` output, it will create one cluster
247 per opcode, and check that the cluster is stable (all points are neighbours).
248
249.. option:: -analysis-numpoints=<dbscan numPoints parameter>
250
251 Specify the numPoints parameters to be used for DBSCAN clustering
252 (`analysis` mode, DBSCAN only).
253
254.. option:: -analysis-clustering-epsilon=<dbscan epsilon parameter>
255
256 Specify the epsilon parameter used for clustering of benchmark points
257 (`analysis` mode).
258
259.. option:: -analysis-inconsistency-epsilon=<epsilon>
260
261 Specify the epsilon parameter used for detection of when the cluster
262 is different from the LLVM schedule profile values (`analysis` mode).
263
264.. option:: -analysis-display-unstable-clusters
265
266 If there is more than one benchmark for an opcode, said benchmarks may end up
267 not being clustered into the same cluster if the measured performance
268 characteristics are different. by default all such opcodes are filtered out.
269 This flag will instead show only such unstable opcodes.
270
271.. option:: -ignore-invalid-sched-class=false
272
273 If set, ignore instructions that do not have a sched class (class idx = 0).
274
275.. option:: -mcpu=<cpu name>
276
277 If set, measure the cpu characteristics using the counters for this CPU. This
278 is useful when creating new sched models (the host CPU is unknown to LLVM).
279
280.. option:: --dump-object-to-disk=true
281
282 By default, llvm-exegesis will dump the generated code to a temporary file to
283 enable code inspection. You may disable it to speed up the execution and save
284 disk space.
285
286EXIT STATUS
287-----------
288
289:program:`llvm-exegesis` returns 0 on success. Otherwise, an error message is
290printed to standard error, and the tool returns a non 0 value.
291