1LLVM-EXEGESIS(1) LLVM LLVM-EXEGESIS(1)
2
6 llvm-exegesis - LLVM Machine Instruction Benchmark
7
9 llvm-exegesis [options]
10
12 llvm-exegesis is a benchmarking tool that uses information available in
13 LLVM to measure host machine instruction characteristics like latency,
14 throughput, or port decomposition.
15 Given an LLVM opcode name and a benchmarking mode, llvm-exegesis gener‐
17 ates a code snippet that makes execution as serial (resp. as parallel)
18 as possible so that we can measure the latency (resp. inverse through‐
19 put/uop decomposition) of the instruction. The code snippet is jitted
20 and executed on the host subtarget. The time taken (resp. resource us‐
21 age) is measured using hardware performance counters. The result is
22 printed out as YAML to the standard output.
23 The main goal of this tool is to automatically (in)validate the LLVM's
25 TableDef scheduling models. To that end, we also provide analysis of
26 the results.
27 llvm-exegesis can also benchmark arbitrary user-provided code snippets.
29
31 Assume you have an X86-64 machine. To measure the latency of a single
32 instruction, run:
33 $ llvm-exegesis -mode=latency -opcode-name=ADD64rr
35 Measuring the uop decomposition or inverse throughput of an instruction
37 works similarly:
38 $ llvm-exegesis -mode=uops -opcode-name=ADD64rr
40 $ llvm-exegesis -mode=inverse_throughput -opcode-name=ADD64rr
41 The output is a YAML document (the default is to write to stdout, but
43 you can redirect the output to a file using -benchmarks-file):
44 ---
46 key:
47 opcode_name: ADD64rr
48 mode: latency
49 config: ''
50 cpu_name: haswell
51 llvm_triple: x86_64-unknown-linux-gnu
52 num_repetitions: 10000
53 measurements:
54 - { key: latency, value: 1.0058, debug_string: '' }
55 error: ''
56 info: 'explicit self cycles, selecting one aliasing configuration.
57 Snippet:
58 ADD64rr R8, R8, R10
59 '
60 ...
61 To measure the latency of all instructions for the host architecture,
63 run:
64 #!/bin/bash
66 readonly INSTRUCTIONS=$(($(grep INSTRUCTION_LIST_END build/lib/Target/X86/X86GenInstrInfo.inc | cut -f2 -d=) - 1))
67 for INSTRUCTION in $(seq 1 ${INSTRUCTIONS});
68 do
69 ./build/bin/llvm-exegesis -mode=latency -opcode-index=${INSTRUCTION} | sed -n '/---/,$p'
70 done
71 FIXME: Provide an llvm-exegesis option to test all instructions.
73
75 To measure the latency/uops of a custom piece of code, you can specify
76 the snippets-file option (- reads from standard input).
77 $ echo "vzeroupper" | llvm-exegesis -mode=uops -snippets-file=-
79 Real-life code snippets typically depend on registers or memory.
81 llvm-exegesis checks the liveliness of registers (i.e. any register use
82 has a corresponding def or is a "live in"). If your code depends on the
83 value of some registers, you have two options:
84 • Mark the register as requiring a definition. llvm-exegesis will auto‐
86 matically assign a value to the register. This can be done using the
87 directive LLVM-EXEGESIS-DEFREG <reg name> <hex_value>, where
88 <hex_value> is a bit pattern used to fill <reg_name>. If <hex_value>
89 is smaller than the register width, it will be sign-extended.
90 • Mark the register as a "live in". llvm-exegesis will benchmark using
92 whatever value was in this registers on entry. This can be done using
93 the directive LLVM-EXEGESIS-LIVEIN <reg name>.
94 For example, the following code snippet depends on the values of XMM1
96 (which will be set by the tool) and the memory buffer passed in RDI
97 (live in).
98 # LLVM-EXEGESIS-LIVEIN RDI
100 # LLVM-EXEGESIS-DEFREG XMM1 42
101 vmulps (%rdi), %xmm1, %xmm2
102 vhaddps %xmm2, %xmm2, %xmm3
103 addq $0x10, %rdi
104
106 Assuming you have a set of benchmarked instructions (either latency or
107 uops) as YAML in file /tmp/benchmarks.yaml, you can analyze the results
108 using the following command:
109 $ llvm-exegesis -mode=analysis \
111 -benchmarks-file=/tmp/benchmarks.yaml \
112 -analysis-clusters-output-file=/tmp/clusters.csv \
113 -analysis-inconsistencies-output-file=/tmp/inconsistencies.html
114 This will group the instructions into clusters with the same perfor‐
116 mance characteristics. The clusters will be written out to /tmp/clus‐
117 ters.csv in the following format:
118 cluster_id,opcode_name,config,sched_class
120 ...
121 2,ADD32ri8_DB,,WriteALU,1.00
122 2,ADD32ri_DB,,WriteALU,1.01
123 2,ADD32rr,,WriteALU,1.01
124 2,ADD32rr_DB,,WriteALU,1.00
125 2,ADD32rr_REV,,WriteALU,1.00
126 2,ADD64i32,,WriteALU,1.01
127 2,ADD64ri32,,WriteALU,1.01
128 2,MOVSX64rr32,,BSWAP32r_BSWAP64r_MOVSX64rr32,1.00
129 2,VPADDQYrr,,VPADDBYrr_VPADDDYrr_VPADDQYrr_VPADDWYrr_VPSUBBYrr_VPSUBDYrr_VPSUBQYrr_VPSUBWYrr,1.02
130 2,VPSUBQYrr,,VPADDBYrr_VPADDDYrr_VPADDQYrr_VPADDWYrr_VPSUBBYrr_VPSUBDYrr_VPSUBQYrr_VPSUBWYrr,1.01
131 2,ADD64ri8,,WriteALU,1.00
132 2,SETBr,,WriteSETCC,1.01
133 ...
134 llvm-exegesis will also analyze the clusters to point out inconsisten‐
136 cies in the scheduling information. The output is an html file. For ex‐
137 ample, /tmp/inconsistencies.html will contain messages like the follow‐
138 ing : [image]
139 Note that the scheduling class names will be resolved only when
141 llvm-exegesis is compiled in debug mode, else only the class id will be
142 shown. This does not invalidate any of the analysis results though.
143
145 -help Print a summary of command line options.
146 -opcode-index=<LLVM opcode index>
148 Specify the opcode to measure, by index. See example 1 for de‐
149 tails. Either opcode-index, opcode-name or snippets-file must
150 be set.
151 -opcode-name=<opcode name 1>,<opcode name 2>,...
153 Specify the opcode to measure, by name. Several opcodes can be
154 specified as a comma-separated list. See example 1 for details.
155 Either opcode-index, opcode-name or snippets-file must be set.
156 -snippets-file=<filename>
158 Specify the custom code snippet to measure. See example 2
159 for details. Either opcode-index, opcode-name or snip‐
160 pets-file must be set.
161 -mode=[latency|uops|inverse_throughput|analysis]
163 Specify the run mode. Note that if you pick analysis mode, you
164 also need to specify at least one of the -analysis-clusters-out‐
165 put-file= and -analysis-inconsistencies-output-file=.
166 -num-repetitions=<Number of repetition>
168 Specify the number of repetitions of the asm snippet. Higher
169 values lead to more accurate measurements but lengthen the
170 benchmark.
171 -benchmarks-file=</path/to/file>
173 File to read (analysis mode) or write (latency/uops/in‐
174 verse_throughput modes) benchmark results. "-" uses stdin/std‐
175 out.
176 -analysis-clusters-output-file=</path/to/file>
178 If provided, write the analysis clusters as CSV to this file.
179 "-" prints to stdout. By default, this analysis is not run.
180 -analysis-inconsistencies-output-file=</path/to/file>
182 If non-empty, write inconsistencies found during analysis to
183 this file. - prints to stdout. By default, this analysis is not
184 run.
185 -analysis-clustering=[dbscan,naive]
187 Specify the clustering algorithm to use. By default DBSCAN will
188 be used. Naive clustering algorithm is better for doing further
189 work on the -analysis-inconsistencies-output-file= output, it
190 will create one cluster per opcode, and check that the cluster
191 is stable (all points are neighbours).
192 -analysis-numpoints=<dbscan numPoints parameter>
194 Specify the numPoints parameters to be used for DBSCAN cluster‐
195 ing (analysis mode, DBSCAN only).
196 -analysis-clustering-epsilon=<dbscan epsilon parameter>
198 Specify the epsilon parameter used for clustering of benchmark
199 points (analysis mode).
200 -analysis-inconsistency-epsilon=<epsilon>
202 Specify the epsilon parameter used for detection of when the
203 cluster is different from the LLVM schedule profile values
204 (analysis mode).
205 -analysis-display-unstable-clusters
207 If there is more than one benchmark for an opcode, said bench‐
208 marks may end up not being clustered into the same cluster if
209 the measured performance characteristics are different. by de‐
210 fault all such opcodes are filtered out. This flag will instead
211 show only such unstable opcodes.
212 -ignore-invalid-sched-class=false
214 If set, ignore instructions that do not have a sched class
215 (class idx = 0).
216 -mcpu=<cpu name>
218 If set, measure the cpu characteristics using the counters for
219 this CPU. This is useful when creating new sched models (the
220 host CPU is unknown to LLVM).
221 --dump-object-to-disk=true
223 By default, llvm-exegesis will dump the generated code to a tem‐
224 porary file to enable code inspection. You may disable it to
225 speed up the execution and save disk space.
226
228 llvm-exegesis returns 0 on success. Otherwise, an error message is
229 printed to standard error, and the tool returns a non 0 value.
230
232 Maintained by the LLVM Team (https://llvm.org/).
233
235 2003-2021, LLVM Project
2369 2021-07-22 LLVM-EXEGESIS(1)