rimathia/fmt-code-size-experiments
experiments for using fmt with a custom string library with the goal of minimal code size overhead for the consumers of the custom string library
fmt-code-size-experiments
Experiments for using fmt with a custom string library with the goal of minimal code size overhead for the consumers of the custom string library and reasonable performance. The underlying assumption is that there are many executables running which use the custom string library, therefore having some code in the shared custom string library instead of in every executable saves quite a bit of code size.
A dummy custom string is used here which is just a std::vector<char>.
The Dockerfile serves to determine code sizes on linux using gcc 9, it can be run right from the root directory: docker build . --tag=fmt-code-size-experiments. It omits the microbenchmarks.
The baseline for using fmt with a custom string class is the library usefmt_directpushback which is only a header. The other usefmt_ libraries all try to get the code into a shared library. For the fmt family of functions the code size overhead when using fmt_directpushback is about 1.6K, but for the printf-style functions it is about 45K.
usefmt variant |
description |
|---|---|
| usefmt_originalfmt | just a using declaration for the functions in the fmt library, creates std::string instead of the custom string, for comparison |
| usefmt_directpushback | directly call push_back on the custom string in a header |
| usefmt_erasedpushback | use the type-erased fmt::vformat_to and fmt::basic_printf_context which get a std::back_inserter of the custom string, the actual formatting functions end up in the shared library |
| usefmt_erasedbuffer | similar to usefmt_erasedpushback but use the fmt functions targetting a temporary fmt::memory_buffer, then construct the custom string from the memory_buffer |
| usefmt_singlealloc | use fmt::format_to_n to a fixed-size buffer first, if the fixed capacity is not sufficient then reserve sufficient capacity in the custom string before formatting into the custom string |
| usefmt_bufxyz | use a simple custom-made replacement for fmt::memory_buffer with configurable initial capacity to check the dependene on the initial capacity |
performance
The summary is: usefmt_erasedbuffer is the way to go. Although this always results in a copy from the internal buffer to the custom string using a fixed-size buffer on the stack (as in usefmt_singlealloc) apparently isn't worth it.
format a small string into a string of size n
The performance characteristics of fmt::format and fmt::printf are rather different for reasons I don't yet understand.
fmt::format
fmt::printf
