• The rust version can just be cargo run'd.

• The C++ version can be built with make remem, it requires re2, a flake is available if useful.

• Both take a single regex through stdin (beware newlines and shell escaping / evaluation if using echo), compile it, then runs it against nonsense input to try and estimate capture memory overhead, the two values printed to stdout are the allocations (minus deallocations) for compilation, and the allocations (minus deallocations) for capture.

• make will build both and run them against the regexen sample via the run script (which requires Python), printing the values for each then the regex on stdout in an ascii table.

  Various variants of each gets run:

  • re2 utf8 mode
  • re2 latin1 mode, not entirely clear what it does but it impacts memory use so seemed fair
  • regex unicode + string
  • regex ascii + string -- this is the only mode which can fail to compile a regex, in which case the entries are -
  • regex ascii + bytes

  regex unicode + bytes is a valid mode for the binary, but it seems to have the same memory behaviour as unicode + string so was not included in the run script.

The Rust version uses a custom allocator to measure allocations, the C++ version overrides new and delete and stashes the requested size in the allocation to know how much allocations decrease it.