.\" Man page generated from reStructuredText. . . .nr rst2man-indent-level 0 . .de1 rstReportMargin \\$1 \\n[an-margin] level \\n[rst2man-indent-level] level margin: \\n[rst2man-indent\\n[rst2man-indent-level]] - \\n[rst2man-indent0] \\n[rst2man-indent1] \\n[rst2man-indent2] .. .de1 INDENT .\" .rstReportMargin pre: . RS \\$1 . nr rst2man-indent\\n[rst2man-indent-level] \\n[an-margin] . nr rst2man-indent-level +1 .\" .rstReportMargin post: .. .de UNINDENT . RE .\" indent \\n[an-margin] .\" old: \\n[rst2man-indent\\n[rst2man-indent-level]] .nr rst2man-indent-level -1 .\" new: \\n[rst2man-indent\\n[rst2man-indent-level]] .in \\n[rst2man-indent\\n[rst2man-indent-level]]u .. .TH "LLVM-EXEGESIS" "1" "2024-02-05" "15" "LLVM" .SH NAME llvm-exegesis \- LLVM Machine Instruction Benchmark .SH SYNOPSIS .sp \fBllvm\-exegesis\fP [\fIoptions\fP] .SH DESCRIPTION .sp \fBllvm\-exegesis\fP is a benchmarking tool that uses information available in LLVM to measure host machine instruction characteristics like latency, throughput, or port decomposition. .sp Given an LLVM opcode name and a benchmarking mode, \fBllvm\-exegesis\fP generates a code snippet that makes execution as serial (resp. as parallel) as possible so that we can measure the latency (resp. inverse throughput/uop decomposition) of the instruction. The code snippet is jitted and, unless requested not to, executed on the host subtarget. The time taken (resp. resource usage) is measured using hardware performance counters. The result is printed out as YAML to the standard output. .sp The main goal of this tool is to automatically (in)validate the LLVM’s TableDef scheduling models. To that end, we also provide analysis of the results. .sp \fBllvm\-exegesis\fP can also benchmark arbitrary user\-provided code snippets. .SH SUPPORTED PLATFORMS .sp \fBllvm\-exegesis\fP currently only supports X86 (64\-bit only), ARM (AArch64 only), MIPS, and PowerPC (PowerPC64LE only) on Linux for benchmarking. Not all benchmarking functionality is guaranteed to work on every platform. \fBllvm\-exegesis\fP also has a separate analysis mode that is supported on every platform on which LLVM is. .SH SNIPPET ANNOTATIONS .sp \fBllvm\-exegesis\fP supports benchmarking arbitrary snippets of assembly. However, benchmarking these snippets often requires some setup so that they can execute properly. \fBllvm\-exegesis\fP has two annotations and some additional utilities to help with setup so that snippets can be benchmarked properly. .INDENT 0.0 .IP \(bu 2 \fILLVM\-EXEGESIS\-DEFREG \fP \- Adding this annotation to the text assembly snippet to be benchmarked marks the register as requiring a definition. A value will automatically be provided unless a second parameter, a hex value, is passed in. This is done with the \fILLVM\-EXEGESIS\-DEFREG \fP format. \fI\fP is a bit pattern used to fill the register. If it is a value smaller than the register, it is sign extended to match the size of the register. .IP \(bu 2 \fILLVM\-EXEGESIS\-LIVEIN \fP \- This annotation allows specifying registers that should keep their value upon starting the benchmark. Values can be passed through registers from the benchmarking setup in some cases. The registers and the values assigned to them that can be utilized in the benchmarking script with a \fILLVM\-EXEGESIS\-LIVEIN\fP are as follows: .INDENT 2.0 .IP \(bu 2 Scratch memory register \- The specific register that this value is put in is platform dependent (e.g., it is the RDI register on X86 Linux). Setting this register as a live in ensures that a pointer to a block of memory (1MB) is placed within this register that can be used by the snippet. .UNINDENT .IP \(bu 2 \fILLVM\-EXEGESIS\-MEM\-DEF \fP \- This annotation allows specifying memory definitions that can later be mapped into the execution process of a snippet with the \fILLVM\-EXEGESIS\-MEM\-MAP\fP annotation. Each value is named using the \fI\fP argument so that it can be referenced later within a map annotation. The size is specified in bytes the the value is taken in hexadecimal. If the size of the value is less than the specified size, the value will be repeated until it fills the entire section of memory. Using this annotation requires using the subprocess execution mode. .IP \(bu 2 \fILLVM\-EXEGESIS\-MEM\-MAP
\fP \- This annotation allows for mapping previously defined memory definitions into the execution context of a process. The value name refers to a previously defined memory definition and the address is a decimal number that specifies the address the memory definition should start at. Note that a single memory definition can be mapped multiple times. Using this annotation requires the subprocess execution mode. .UNINDENT .SH EXAMPLE 1: BENCHMARKING INSTRUCTIONS .sp Assume you have an X86\-64 machine. To measure the latency of a single instruction, run: .INDENT 0.0 .INDENT 3.5 .sp .EX $ llvm\-exegesis \-mode=latency \-opcode\-name=ADD64rr .EE .UNINDENT .UNINDENT .sp Measuring the uop decomposition or inverse throughput of an instruction works similarly: .INDENT 0.0 .INDENT 3.5 .sp .EX $ llvm\-exegesis \-mode=uops \-opcode\-name=ADD64rr $ llvm\-exegesis \-mode=inverse_throughput \-opcode\-name=ADD64rr .EE .UNINDENT .UNINDENT .sp The output is a YAML document (the default is to write to stdout, but you can redirect the output to a file using \fI\-benchmarks\-file\fP): .INDENT 0.0 .INDENT 3.5 .sp .EX \-\-\- key: opcode_name: ADD64rr mode: latency config: \(aq\(aq cpu_name: haswell llvm_triple: x86_64\-unknown\-linux\-gnu num_repetitions: 10000 measurements: \- { key: latency, value: 1.0058, debug_string: \(aq\(aq } error: \(aq\(aq info: \(aqexplicit self cycles, selecting one aliasing configuration. Snippet: ADD64rr R8, R8, R10 \(aq \&... .EE .UNINDENT .UNINDENT .sp To measure the latency of all instructions for the host architecture, run: .INDENT 0.0 .INDENT 3.5 .sp .EX $ llvm\-exegesis \-mode=latency \-opcode\-index=\-1 .EE .UNINDENT .UNINDENT .SH EXAMPLE 2: BENCHMARKING A CUSTOM CODE SNIPPET .sp To measure the latency/uops of a custom piece of code, you can specify the \fIsnippets\-file\fP option (\fI\-\fP reads from standard input). .INDENT 0.0 .INDENT 3.5 .sp .EX $ echo \(dqvzeroupper\(dq | llvm\-exegesis \-mode=uops \-snippets\-file=\- .EE .UNINDENT .UNINDENT .sp Real\-life code snippets typically depend on registers or memory. \fBllvm\-exegesis\fP checks the liveliness of registers (i.e. any register use has a corresponding def or is a “live in”). If your code depends on the value of some registers, you need to use snippet annotations to ensure setup is performed properly. .sp For example, the following code snippet depends on the values of XMM1 (which will be set by the tool) and the memory buffer passed in RDI (live in). .INDENT 0.0 .INDENT 3.5 .sp .EX # LLVM\-EXEGESIS\-LIVEIN RDI # LLVM\-EXEGESIS\-DEFREG XMM1 42 vmulps (%rdi), %xmm1, %xmm2 vhaddps %xmm2, %xmm2, %xmm3 addq $0x10, %rdi .EE .UNINDENT .UNINDENT .SH EXAMPLE 3: BENCHMARKING WITH MEMORY ANNOTATIONS .sp Some snippets require memory setup in specific places to execute without crashing. Setting up memory can be accomplished with the \fILLVM\-EXEGESIS\-MEM\-DEF\fP and \fILLVM\-EXEGESIS\-MEM\-MAP\fP annotations. To execute the following snippet: .INDENT 0.0 .INDENT 3.5 .sp .EX movq $8192, %rax movq (%rax), %rdi .EE .UNINDENT .UNINDENT .sp We need to have at least eight bytes of memory allocated starting \fI0x2000\fP\&. We can create the necessary execution environment with the following annotations added to the snippet: .INDENT 0.0 .INDENT 3.5 .sp .EX # LLVM\-EXEGESIS\-MEM\-DEF test1 4096 2147483647 # LLVM\-EXEGESIS\-MEM\-MAP test1 8192 movq $8192, %rax movq (%rax), %rdi .EE .UNINDENT .UNINDENT .SH EXAMPLE 4: ANALYSIS .sp Assuming you have a set of benchmarked instructions (either latency or uops) as YAML in file \fI/tmp/benchmarks.yaml\fP, you can analyze the results using the following command: .INDENT 0.0 .INDENT 3.5 .sp .EX $ llvm\-exegesis \-mode=analysis \e \-benchmarks\-file=/tmp/benchmarks.yaml \e \-analysis\-clusters\-output\-file=/tmp/clusters.csv \e \-analysis\-inconsistencies\-output\-file=/tmp/inconsistencies.html .EE .UNINDENT .UNINDENT .sp This will group the instructions into clusters with the same performance characteristics. The clusters will be written out to \fI/tmp/clusters.csv\fP in the following format: .INDENT 0.0 .INDENT 3.5 .sp .EX cluster_id,opcode_name,config,sched_class \&... 2,ADD32ri8_DB,,WriteALU,1.00 2,ADD32ri_DB,,WriteALU,1.01 2,ADD32rr,,WriteALU,1.01 2,ADD32rr_DB,,WriteALU,1.00 2,ADD32rr_REV,,WriteALU,1.00 2,ADD64i32,,WriteALU,1.01 2,ADD64ri32,,WriteALU,1.01 2,MOVSX64rr32,,BSWAP32r_BSWAP64r_MOVSX64rr32,1.00 2,VPADDQYrr,,VPADDBYrr_VPADDDYrr_VPADDQYrr_VPADDWYrr_VPSUBBYrr_VPSUBDYrr_VPSUBQYrr_VPSUBWYrr,1.02 2,VPSUBQYrr,,VPADDBYrr_VPADDDYrr_VPADDQYrr_VPADDWYrr_VPSUBBYrr_VPSUBDYrr_VPSUBQYrr_VPSUBWYrr,1.01 2,ADD64ri8,,WriteALU,1.00 2,SETBr,,WriteSETCC,1.01 \&... .EE .UNINDENT .UNINDENT .sp \fBllvm\-exegesis\fP will also analyze the clusters to point out inconsistencies in the scheduling information. The output is an html file. For example, \fI/tmp/inconsistencies.html\fP will contain messages like the following : [image] .sp Note that the scheduling class names will be resolved only when \fBllvm\-exegesis\fP is compiled in debug mode, else only the class id will be shown. This does not invalidate any of the analysis results though. .SH OPTIONS .INDENT 0.0 .TP .B \-help Print a summary of command line options. .UNINDENT .INDENT 0.0 .TP .B \-opcode\-index= Specify the opcode to measure, by index. Specifying \fI\-1\fP will result in measuring every existing opcode. See example 1 for details. Either \fIopcode\-index\fP, \fIopcode\-name\fP or \fIsnippets\-file\fP must be set. .UNINDENT .INDENT 0.0 .TP .B \-opcode\-name=,,... Specify the opcode to measure, by name. Several opcodes can be specified as a comma\-separated list. See example 1 for details. Either \fIopcode\-index\fP, \fIopcode\-name\fP or \fIsnippets\-file\fP must be set. .UNINDENT .INDENT 0.0 .TP .B \-snippets\-file= Specify the custom code snippet to measure. See example 2 for details. Either \fIopcode\-index\fP, \fIopcode\-name\fP or \fIsnippets\-file\fP must be set. .UNINDENT .INDENT 0.0 .TP .B \-mode=[latency|uops|inverse_throughput|analysis] Specify the run mode. Note that some modes have additional requirements and options. .sp \fIlatency\fP mode can be make use of either RDTSC or LBR. \fIlatency[LBR]\fP is only available on X86 (at least \fISkylake\fP). To run in \fIlatency\fP mode, a positive value must be specified for \fIx86\-lbr\-sample\-period\fP and \fI–repetition\-mode=loop\fP\&. .sp In \fIanalysis\fP mode, you also need to specify at least one of the \fI\-analysis\-clusters\-output\-file=\fP and \fI\-analysis\-inconsistencies\-output\-file=\fP\&. .UNINDENT .INDENT 0.0 .TP .B \-\-benchmark\-phase=[prepare\-snippet|prepare\-and\-assemble\-snippet|assemble\-measured\-code|measure] By default, when \fI\-mode=\fP is specified, the generated snippet will be executed and measured, and that requires that we are running on the hardware for which the snippet was generated, and that supports performance measurements. However, it is possible to stop at some stage before measuring. Choices are: * \fBprepare\-snippet\fP: Only generate the minimal instruction sequence. * \fBprepare\-and\-assemble\-snippet\fP: Same as \fBprepare\-snippet\fP, but also dumps an excerpt of the sequence (hex encoded). * \fBassemble\-measured\-code\fP: Same as \fBprepare\-and\-assemble\-snippet\fP\&. but also creates the full sequence that can be dumped to a file using \fB\-\-dump\-object\-to\-disk\fP\&. * \fBmeasure\fP: Same as \fBassemble\-measured\-code\fP, but also runs the measurement. .UNINDENT .INDENT 0.0 .TP .B \-x86\-lbr\-sample\-period= Specify the LBR sampling period \- how many branches before we take a sample. When a positive value is specified for this option and when the mode is \fIlatency\fP, we will use LBRs for measuring. On choosing the “right” sampling period, a small value is preferred, but throttling could occur if the sampling is too frequent. A prime number should be used to avoid consistently skipping certain blocks. .UNINDENT .INDENT 0.0 .TP .B \-x86\-disable\-upper\-sse\-registers Using the upper xmm registers (xmm8\-xmm15) forces a longer instruction encoding which may put greater pressure on the frontend fetch and decode stages, potentially reducing the rate that instructions are dispatched to the backend, particularly on older hardware. Comparing baseline results with this mode enabled can help determine the effects of the frontend and can be used to improve latency and throughput estimates. .UNINDENT .INDENT 0.0 .TP .B \-repetition\-mode=[duplicate|loop|min] Specify the repetition mode. \fIduplicate\fP will create a large, straight line basic block with \fInum\-repetitions\fP instructions (repeating the snippet \fInum\-repetitions\fP/\fIsnippet size\fP times). \fIloop\fP will, optionally, duplicate the snippet until the loop body contains at least \fIloop\-body\-size\fP instructions, and then wrap the result in a loop which will execute \fInum\-repetitions\fP instructions (thus, again, repeating the snippet \fInum\-repetitions\fP/\fIsnippet size\fP times). The \fIloop\fP mode, especially with loop unrolling tends to better hide the effects of the CPU frontend on architectures that cache decoded instructions, but consumes a register for counting iterations. If performing an analysis over many opcodes, it may be best to instead use the \fImin\fP mode, which will run each other mode, and produce the minimal measured result. .UNINDENT .INDENT 0.0 .TP .B \-num\-repetitions= Specify the target number of executed instructions. Note that the actual repetition count of the snippet will be \fInum\-repetitions\fP/\fIsnippet size\fP\&. Higher values lead to more accurate measurements but lengthen the benchmark. .UNINDENT .INDENT 0.0 .TP .B \-loop\-body\-size= Only effective for \fI\-repetition\-mode=[loop|min]\fP\&. Instead of looping over the snippet directly, first duplicate it so that the loop body contains at least this many instructions. This potentially results in loop body being cached in the CPU Op Cache / Loop Cache, which allows to which may have higher throughput than the CPU decoders. .UNINDENT .INDENT 0.0 .TP .B \-max\-configs\-per\-opcode= Specify the maximum configurations that can be generated for each opcode. By default this is \fI1\fP, meaning that we assume that a single measurement is enough to characterize an opcode. This might not be true of all instructions: for example, the performance characteristics of the LEA instruction on X86 depends on the value of assigned registers and immediates. Setting a value of \fI\-max\-configs\-per\-opcode\fP larger than \fI1\fP allows \fIllvm\-exegesis\fP to explore more configurations to discover if some register or immediate assignments lead to different performance characteristics. .UNINDENT .INDENT 0.0 .TP .B \-benchmarks\-file= File to read (\fIanalysis\fP mode) or write (\fIlatency\fP/\fIuops\fP/\fIinverse_throughput\fP modes) benchmark results. “\-” uses stdin/stdout. .UNINDENT .INDENT 0.0 .TP .B \-analysis\-clusters\-output\-file= If provided, write the analysis clusters as CSV to this file. “\-” prints to stdout. By default, this analysis is not run. .UNINDENT .INDENT 0.0 .TP .B \-analysis\-inconsistencies\-output\-file= If non\-empty, write inconsistencies found during analysis to this file. \fI\-\fP prints to stdout. By default, this analysis is not run. .UNINDENT .INDENT 0.0 .TP .B \-analysis\-filter=[all|reg\-only|mem\-only] By default, all benchmark results are analysed, but sometimes it may be useful to only look at those that to not involve memory, or vice versa. This option allows to either keep all benchmarks, or filter out (ignore) either all the ones that do involve memory (involve instructions that may read or write to memory), or the opposite, to only keep such benchmarks. .UNINDENT .INDENT 0.0 .TP .B \-analysis\-clustering=[dbscan,naive] Specify the clustering algorithm to use. By default DBSCAN will be used. Naive clustering algorithm is better for doing further work on the \fI\-analysis\-inconsistencies\-output\-file=\fP output, it will create one cluster per opcode, and check that the cluster is stable (all points are neighbours). .UNINDENT .INDENT 0.0 .TP .B \-analysis\-numpoints= Specify the numPoints parameters to be used for DBSCAN clustering (\fIanalysis\fP mode, DBSCAN only). .UNINDENT .INDENT 0.0 .TP .B \-analysis\-clustering\-epsilon= Specify the epsilon parameter used for clustering of benchmark points (\fIanalysis\fP mode). .UNINDENT .INDENT 0.0 .TP .B \-analysis\-inconsistency\-epsilon= Specify the epsilon parameter used for detection of when the cluster is different from the LLVM schedule profile values (\fIanalysis\fP mode). .UNINDENT .INDENT 0.0 .TP .B \-analysis\-display\-unstable\-clusters If there is more than one benchmark for an opcode, said benchmarks may end up not being clustered into the same cluster if the measured performance characteristics are different. by default all such opcodes are filtered out. This flag will instead show only such unstable opcodes. .UNINDENT .INDENT 0.0 .TP .B \-ignore\-invalid\-sched\-class=false If set, ignore instructions that do not have a sched class (class idx = 0). .UNINDENT .INDENT 0.0 .TP .B \-mtriple= Target triple. See \fI\-version\fP for available targets. .UNINDENT .INDENT 0.0 .TP .B \-mcpu= If set, measure the cpu characteristics using the counters for this CPU. This is useful when creating new sched models (the host CPU is unknown to LLVM). (\fI\-mcpu=help\fP for details) .UNINDENT .INDENT 0.0 .TP .B \-\-analysis\-override\-benchmark\-triple\-and\-cpu By default, llvm\-exegesis will analyze the benchmarks for the triple/CPU they were measured for, but if you want to analyze them for some other combination (specified via \fI\-mtriple\fP/\fI\-mcpu\fP), you can pass this flag. .UNINDENT .INDENT 0.0 .TP .B \-\-dump\-object\-to\-disk=true If set, llvm\-exegesis will dump the generated code to a temporary file to enable code inspection. Disabled by default. .UNINDENT .INDENT 0.0 .TP .B \-\-use\-dummy\-perf\-counters If set, llvm\-exegesis will not read any real performance counters and return a dummy value instead. This can be used to ensure a snippet doesn’t crash when hardware performance counters are unavailable and for debugging \fBllvm\-exegesis\fP itself. .UNINDENT .INDENT 0.0 .TP .B \-\-execution\-mode=[inprocess,subprocess] This option specifies what execution mode to use. The \fIinprocess\fP execution mode is the default. The \fIsubprocess\fP execution mode allows for additional features such as memory annotations but is currently restricted to X86\-64 on Linux. .UNINDENT .SH EXIT STATUS .sp \fBllvm\-exegesis\fP returns 0 on success. Otherwise, an error message is printed to standard error, and the tool returns a non 0 value. .SH AUTHOR Maintained by the LLVM Team (https://llvm.org/). .SH COPYRIGHT 2003-2024, LLVM Project .\" Generated by docutils manpage writer. .