LLVM-PROFDATA(1)

📦
version 20.1.8 · 8 versions
, in fedora 43, fedora 44

💖 LLVM-PROFDATA(1) — the llvm-profdata glow-up 💖

okay so this page lives in the LLVM manual section, and it's literally about one little tool that does a LOT. Grab a snack, bestie. 🍓

✨ NAME ✨

llvm-profdata — Profile data tool 🎀

🌸 SYNOPSIS 🌸

llvm-profdata command [args...]

🦋 DESCRIPTION 🦋

The llvm-profdata tool is a small utility for working with profile data files. that's it, that's the whole vibe — small tool, big energy. 💅

👑 COMMANDS 👑

okay so here's the menu, pick your fighter:

  • merge
  • show
  • overlap
  • order

🔥 MERGE 🔥

🧚 SYNOPSIS 🧚

llvm-profdata merge [options] [filename...]

💜 DESCRIPTION 💜

llvm-profdata merge takes several profile data files generated by PGO instrumentation and merges them together into a single indexed profile data file. it's giving "squad assemble" fr fr.

By default profile data is merged without modification. what does that mean bestie? it means the relative importance of each input file is proportional to the number of samples or counts it contains. so in general, the input from a longer training run will be interpreted as relatively more important than a shorter run. depends on the nature of the training runs though, so it may be useful to adjust the weight given to each input file by using the -weighted-input option. 👀

Also important: Profiles passed in via -weighted-input, -input-files, or via positional arguments are processed once for each time they are seen.

⭐ OPTIONS ⭐

  • --help — Print a summary of command line options. the "what do I even do here" button. 🙈
  • --output=<output>, -o — Specify the output file name. Output cannot be - as the resulting indexed profile data can't be written to standard output. so no pipe dreams, sorry babe. 💀
  • --weighted-input=<weight,filename> — Specify an input file name along with a weight. The profile counts of the supplied filename will be scaled (multiplied) by the supplied weight, where weight is a decimal integer >= 1. Input files specified without using this option are assigned a default weight of 1. Examples are shown below. 🫶
  • --input-files=<path>, -f — Specify a file which contains a list of files to merge. The entries in this file are newline-separated. Lines starting with '#' are skipped. Entries may be of the form <filename> or <weight>,<filename>. ✨
  • --remapping-file=<path>, -r — Specify a file which contains a remapping from symbol names in the input profile to the symbol names that should be used in the output profile. The file should consist of lines of the form <input-symbol> <output-symbol>. Blank lines and lines starting with # are skipped.

The llvm-cxxmap tool can be used to generate the symbol remapping file. handy!! 🎀

  • --instr (default) — Specify that the input profile is an instrumentation-based profile.
  • --sample — Specify that the input profile is a sample-based profile.

The format of the generated file can be generated in one of three ways:

  • --binary (default) — Emit the profile using a binary encoding. For instrumentation-based profile the output format is the indexed binary format.
  • --extbinary — Emit the profile using an extensible binary encoding. This option can only be used with sample-based profile. The extensible binary encoding can be more compact with compression enabled and can be loaded faster than the default binary encoding. 🏎️
  • --text — Emit the profile in text mode. This option can also be used with both sample-based and instrumentation-based profile. When this option is used the profile will be dumped in the text format that is parsable by the profile reader.
  • --gcc — Emit the profile using GCC's gcov format (Not yet supported). 😭 someday bestie, someday.
  • --sparse[=true|false] — Do not emit function records with 0 execution count. Can only be used in conjunction with -instr. Defaults to false, since it can inhibit compiler optimization during PGO. so like, lowkey leave it alone unless you know. ⚠️
  • --num-threads=<N>, -j — Use N threads to perform profile merging. When N=0, llvm-profdata auto-detects an appropriate number of threads to use. This is the default. multitasking queen energy. 👑
  • --failure-mode=[any|all] — Set the failure mode. There are two options: 'any' causes the merge command to fail if any profiles are invalid, and 'all' causes the merge command to fail only if all profiles are invalid. If 'all' is set, information from any invalid profiles is excluded from the final merged product. The default failure mode is 'any'. 💅
  • --prof-sym-list=<path> — Specify a file which contains a list of symbols to generate profile symbol list in the profile. This option can only be used with sample-based profile in extbinary format. The entries in this file are newline-separated.
  • --compress-all-sections=[true|false] — Compress all sections when writing the profile. This option can only be used with sample-based profile in extbinary format.
  • --use-md5=[true|false] — Use MD5 to represent string in name table when writing the profile. This option can only be used with sample-based profile in extbinary format.
  • --gen-partial-profile=[true|false] — Mark the profile to be a partial profile which only provides partial profile coverage for the optimized target. This option can only be used with sample-based profile in extbinary format.
  • --split-layout=[true|false] — Split the profile data section to two with one containing sample profiles with inlined functions and the other not. This option can only be used with sample-based profile in extbinary format.
  • --convert-sample-profile-layout=[nest|flat] — Convert the merged profile into a profile with a new layout. Supported layout are nest (Nested profile, the input should be CS flat profile) and flat (Profile with nested inlinees flattened out).
  • --supplement-instr-with-sample=<file> — Supplement an instrumentation profile with sample profile. The sample profile is the input of the flag. Output will be in instrumentation format (only works with -instr).
  • --zero-counter-threshold=<float> — For the function which is cold in instr profile but hot in sample profile, if the ratio of the number of zero counters divided by the total number of counters is above the threshold, the profile of the function will be regarded as being harmful for performance and will be dropped. bye felicia 🗑️
  • --instr-prof-cold-threshold=<int> — User specified cold threshold for instr profile which will override the cold threshold got from profile summary.
  • --suppl-min-size-threshold=<int> — If the size of a function is smaller than the threshold, assume it can be inlined by PGO early inliner and it will not be adjusted based on sample profile.
  • --debug-info=<path> — Specify the executable or .dSYM that contains debug info for the raw profile. When --debug-info-correlate or --profile-correlate=debug-info was used for instrumentation, use this option to correlate the raw profile.
  • --binary-file=<path> — Specify the executable that contains profile data and profile name sections for the raw profile. When -profile-correlate=binary was used for instrumentation, use this option to correlate the raw profile.
  • --debuginfod — Use debuginfod to find the associated executables that contain profile data and name sections for the raw profiles to correlate them. When -profile-correlate=binary was used for instrumentation, this option can be used for correlation.
  • --debug-file-directory=<dir> — Use provided local directories to search for executables that contain profile data and name sections for the raw profiles to correlate them. When -profile-correlate=binary was used for instrumentation, this option can be used for correlation.
  • --correlate=<kind> — Specify the correlation kind (debug_info or binary) to use when -debuginfod or -debug-file-directory=<dir> option is provided.
  • --temporal-profile-trace-reservoir-size — The maximum number of temporal profile traces to be stored in the output profile. If more traces are added, we will use reservoir sampling to select which traces to keep. Note that changing this value between different merge invocations on the same indexed profile could result in sample bias. The default value is 100. ⚠️
  • --temporal-profile-max-trace-length — The maximum number of functions in a single temporal profile trace. Longer traces will be truncated. The default value is 1000.
  • --function=<string> — Only keep functions matching the regex in the output, all others are erased from the profile. ✂️
  • --no-function=<string> — Remove functions matching the regex from the profile. If both --function and --no-function are specified and a function matches both, it is removed. no takebacks, it's gone. 💀

🌈 EXAMPLES 🌈

💫 Basic Usage 💫

Merge three profiles:

llvm-profdata merge foo.profdata bar.profdata baz.profdata -output merged.profdata

💖 Weighted Input 💖

The input file foo.profdata is especially important, multiply its counts by 10:

llvm-profdata merge --weighted-input=10,foo.profdata bar.profdata baz.profdata --output merged.profdata

Exactly equivalent to the previous invocation (explicit form; useful for programmatic invocation):

llvm-profdata merge --weighted-input=10,foo.profdata --weighted-input=1,bar.profdata --weighted-input=1,baz.profdata --output merged.profdata

👀 SHOW 👀

SYNOPSIS

llvm-profdata show [options] [filename]

DESCRIPTION

okay bestie, llvm-profdata show is the nosy one 💅 — you hand it a profile data file and it spills all the tea about the profile counters, both for the file itself and for any function(s) you specifically asked about.

And if you leave out filename, or you just pass -? It reads its input from standard input. shy king behavior honestly 🌸

OPTIONS

  • --all-functions — spill the deets for every single function. no crumbs left, periodt ✨
  • --binary-ids — print the embedded binary ids that are chillin' in the profile.
  • --counts — print the actual counter values for the functions being displayed.
  • --show-format=<text|json|yaml> — emit the output in whichever format you picked, if the profile type you gave it supports that format. no promises otherwise 🫶
  • --function=<string> — print details for a function when that function's name contains the given string. substring vibes, not exact match.
  • --help — print a summary of the command line options.
  • --output=<output>, -o — pick the output file name. If output is -, or you just don't specify it at all, the output goes to standard output.
  • --instr (this one's the default) — declares that your input profile is an instrumentation-based profile.
  • --text — tells the profile dumper to show profile counts in the text format of the instrumentation-based profile data representation. By default the profile info gets dumped in a more human readable form (also text, bestie, still text) with annotations. so this is the raw text mode 🌈
  • --topn=<n> — tells the profile dumper to show the top n functions with the hottest basic blocks in the summary section. By default, no topn functions get dumped at all.
  • --sample — declares that your input profile is a sample-based profile.
  • --memop-sizes — show the profiled sizes of the memory intrinsic calls, for the functions being shown.
  • --value-cutoff=<n> — only show the functions whose max count values are greater than or equal to n. By default, the value-cutoff is set to 0.
  • --list-below-cutoff — only output the names of functions whose max count value is below the cutoff value. (yes, the opposite of the one above, iconic 🎀)
  • --profile-version — print the profile version.
  • --showcs — only show context sensitive profile counts. The default is to filter out all context sensitive profile counts.
  • --show-prof-sym-list=[true|false] — show the profile symbol list if one exists in the profile. Heads up: this option is only meaningful for sample-based profiles in extbinary format.
  • --show-sec-info-only=[true|false] — show basic info about each section in the profile. Again: only meaningful for sample-based profiles in extbinary format.
  • --debug-info=<path> — specify the executable or .dSYM that contains the debug info for the raw profile. When --debug-info-correlate or --profile-correlate=debug-info was used for instrumentation, use this option to show the correlated functions from the raw profile 🔥
  • --covered — show only the functions that actually got executed, aka the ones with non-zero counts. the real ones 💜

🫂 OVERLAP 🫂

SYNOPSIS

llvm-profdata overlap [options] [base profile file] [test profile file]

DESCRIPTION

llvm-profdata overlap takes two profile data files and shows you the overlap of the counter distribution — both between the whole files and between any of the specific functions you name.

okay so how is overlap even defined here? buckle up bestie, math time 😭

Suppose the base profile file has these counts: {c1_1, c1_2, ..., c1_n, c1_u_1, c2_u_2, ..., c2_u_s}, and the test profile file has {c2_1, c2_2, ..., c2_n, c2_v_1, c2_v_2, ..., c2_v_t}.

Here c{1|2}_i (for i = 1 .. n) are the matched counters, and c1_u_i (for i = 1 .. s) and c2_v_i (for i = 1 .. v) are the unmatched counters — meaning counters that only exist in the base profile file and only exist in the test profile file, respectively.

Then:

sum_1 = c1_1 + c1_2 + ... + c1_n + c1_u_1 + c2_u_2 + ... + c2_u_s
sum_2 = c2_1 + c2_2 + ... + c2_n + c2_v_1 + c2_v_2 + ... + c2_v_t

overlap = min(c1_1/sum_1, c2_1/sum_2) + min(c1_2/sum_1, c2_2/sum_2) + ...
          + min(c1_n/sum_1, c2_n/sum_2)

The resulting overlap distribution is a percentage, ranging from 0.0% to 100.0% — where 0.0% means there's literally no overlap and 100.0% means a perfect overlap. soulmates fr fr 💖

Example time: if the base profile file has counts of {400, 600}, and the test profile file has matched counts of {60000, 40000}, then the overlap is 80%. slay.

OPTIONS

  • --function=<string> — print details for a function when that function's name contains the given string.
  • --help — print a summary of the command line options.
  • --output=<output>, -o — pick the output file name. If output is -, or it isn't specified, the output goes to standard output.
  • --value-cutoff=<n> — show only those functions whose max count values are greater than or equal to n. By default, the value-cutoff is set to the max of unsigned long long — so basically nothing gets filtered unless you say so 🌈
  • --cs — only show overlap for the context sensitive profile counts. The default is to show the non-context sensitive profile counts.

🔢 ORDER 🔢

SYNOPSIS

llvm-profdata order [options] [filename]

DESCRIPTION

llvm-profdata order takes the temporal profiling traces from a profile and finds a function order that reduces the number of page faults for those traces. You can hand this output straight to lld — via --symbol-ordering-file= for ELF, or -order-file for Mach-O.

And if the traces found in the profile are actually representative of the real world, then this order should improve startup performance. main character energy for your binaries, honestly 👑

OPTIONS

  • --help — print a summary of the command line options.
  • --output=<output>, -o — pick the output file name. If output is -, or it isn't specified, the output goes to standard output.

🚪 EXIT STATUS 🚪

llvm-profdata returns 1 if the command is omitted or is invalid, if it can't read input files, or if there's a mismatch between their data. so if you see a 1, something's up 💀

✍️ AUTHOR ✍️

Maintained by the LLVM Team (https://llvm.org/).

2003-2025, LLVM Project

and that's a wrap bestie 🫶 this lil manual is from the LLVM Project, version 20, dated 2025-09-03, and it's page LLVM-PROFDATA(1) of the manual — go forth and profile responsibly 💅✨

📦 where this page comes from