check-symbols

Command-line utility for checking ELF symbols.

It lists the dynamic symbols exported by a shared library, checks them against a list of expected symbols, or compares them with the symbols an executable or another library imports. Symbol names are demangled and normalised before they are printed or compared.

Requirements

• Linux, or another platform that uses ELF binaries
• GNU nm (part of binutils) on the PATH
• Python 3.9 or higher

Installation

pip3 install check-symbols

Usage

check-symbols is a command-line application:

usage: check-symbols [-h] [--version] [--check <path>] [--diff <path>] [--include <regex> [<regex> ...]]
                     [--exclude <regex> [<regex> ...]] [--abi-symbols]
                     [<path>]

Command-line utility for checking ELF symbols.

positional arguments:
  <path>                path to the shared library, or a directory containing it (default: discover automatically)

options:
  -h, --help            show this help message and exit
  --version             print the version and exit
  --check <path>        path to a text file containing a list of expected symbols (one per line). use a dash ('-') to
                        read from stdin.
  --diff <path>         compiled executable or shared library expected to consume the symbols
  --include <regex> [<regex> ...]
                        one or more regular expressions to select the listed/checked symbols.
  --exclude <regex> [<regex> ...]
                        one or more regular expressions to exclude from the listed/checked symbols.
  --abi-symbols         include compiler-generated ABI symbols such as vtables, typeinfo, thunks and guard variables

v0.2.0 - github.com/marzer/check-symbols

With neither --check nor --diff, the exported symbols are printed one per line. This output is valid input for --check, so check-symbols libfoo.so > symbols.txt makes a first check file.

When the path is a directory, the first .so file in it (in sorted order) is used. Without a path, the current directory is searched.

--check compares the exported symbols with a check file. --diff compares them with the undefined symbols of an executable or another shared library. Every import that the library does not export counts as missing, including imports from other libraries such as libc. A diff is therefore usually narrowed with --include or --exclude. Both modes can be used in one run.

--include keeps the symbols that match every pattern, and --exclude drops the symbols that match any pattern. They apply to both sides of a check or diff. Both options take one or more values, so give the library path before them, or end the patterns with --.

Check files

A check file holds one symbol per line, in the form that list mode prints. Blank lines are ignored. Comments start with #, ; or // and run to the end of the line. A marker inside braces, square brackets or parentheses, such as the # in {lambda(int)#1} or the ; in [u8; 16], does not start a comment.

Lines in a check file are normalised in the same way as the library's symbols, so ns::foo(int) const matches ns::foo.

Without --abi-symbols, lines for compiler-generated ABI symbols are ignored and never reported as missing. This includes the forms that earlier versions printed, such as X::$vtable. To check these symbols, use --abi-symbols.

Normalisation

Symbols are read with nm -PBDCg, which prints the demangled dynamic symbols, and are then normalised:

• A library exports the symbols of type B and D (data), R (read-only data), T (code), V and W (weak definitions), u (unique globals) and i (indirect functions). Types are matched case-sensitively. U, v and w are imports, and only U counts on the consumer side of --diff. A, the version definitions, is dropped, as are linker-generated symbols such as _edata and __bss_start.
• Compiler-generated ABI symbols are dropped: vtables, VTTs, typeinfo, thunks, guard variables, TLS init and wrapper functions, transaction clones, reference temporaries, template parameter objects and module initialisers. Where these appear depends on LTO and the toolchain. --abi-symbols keeps them.
• Version tags are removed: adler32@@ZLIB_1.2.0 becomes adler32.
• Parameter lists and cv- and ref-qualifiers are removed, so all overloads of a function share one name: ns::vec::get(int) const becomes ns::vec::get.
• nm prints return types only for function templates. Everything before the first space outside brackets is removed, so int ns::vec::get<int>(int) becomes ns::vec::get<int> and unsigned int ns::vec::size<int>() becomes int ns::vec::size<int>.
• Operators keep their names: ns::vec::operator new, ns::vec::operator(), ns::operator<<. Conversion operators keep their target type: ns::vec::operator bool.
• Lambdas and other local entities keep the name and parameter list of the enclosing function: ns::foo()::{lambda(int)#1}::operator().
• Common standard library types are shortened, and the [abi:cxx11] tag is removed (other ABI tags are kept): std::basic_string<char, std::char_traits<char>, std::allocator<char> > becomes std::string.
• With --abi-symbols, vtables, typeinfo objects, typeinfo names and VTTs become X::$vtable, X::$typeinfo, X::$typeinfo_name and X::$VTT.
• The list is sorted, and duplicates are removed.

Exit codes

Values 1 and 2 are bit flags. When --check and --diff are used together, their results are combined, so a missing symbol in either mode sets bit 1.

On invalid command-line arguments, argparse exits with 2. A script cannot tell this apart from bit 2 by the exit code alone.

Contributing

See CONTRIBUTING.md.

Changelog

v0.3.0 - 2026-10-15

• fixed --diff counting the consumer's own unique globals (u) as imports
• fixed some function templates being listed as X::$typeinfo, X::$vtable or X::$VTT
• ignored compiler-generated ABI symbols by default
• added symbol types R, V, W, u and i
• added --abi-symbols

v0.2.0 - 2026-10-15

• fixed versioned symbols (foo@@VER) keeping a trailing @, which also stopped their parameter lists being removed
• fixed soname version tags (foo@@libfoo-1.0.so.0) not being removed
• fixed operator new, operator delete and conversion operators being listed as new, delete and the target type
• fixed operator() names changing when read back from a check file
• fixed conversion operators to function and member function pointer types changing when read back from a check file
• fixed typeinfo name for X being normalised to name for X::$typeinfo (it is now X::$typeinfo_name)
• fixed names containing # (e.g. lambdas) being cut short; comments are now only stripped from check files, and never inside braces, square brackets or parentheses
• fixed a crash on check-file lines with an unmatched )
• fixed a missing nm and invalid --include/--exclude patterns being reported as internal bugs
• fixed directory discovery picking an arbitrary .so file when there were several (it now takes the first in sorted order)
• improved speed on libraries with many symbols
• dropped support for Python 3.8

v0.1.0 - 2024-06-14

• First public release 🎉