SysString

This header-only library provides a C++ string class template optimized for interoperability with external native string types. It is immutable, Unicode-first, and exposes convenient operations similar to Python or ECMAScript strings. It uses a separate builder class template to construct strings. It provides fast concatenation via the + operator that does not allocate temporary strings. The library exposes bidirectional UTF-8/UTF-16/UTF-32 and grapheme cluster views of sys_string_t as well as of other C++ ranges of characters.

What does it mean?

• Interoperability with external native string types means that SysString makes conversions to and from such types as efficient as possible and, ideally, zero-cost operations. Native string types are things like:

  • NSString * or CFStringRef on macOS/iOS
  • Java String on Android
  • Python str in Python extensions
  • JavaScript String on Emscripten/WebAssembly
  • const wchar_t *, HSTRING or BSTR on Windows
  • const char * on Linux

  sys_string_t<Storage> and sys_string_builder_t<Storage> are parametrized on Storage, which defines what kind of native string type to use internally and interoperate with. Different Storage implementations are provided for all the external types above.

  For example, the storage for Apple's platforms uses NSString * internally, allowing zero-cost conversions between C++ and native sides.

  On Android and Emscripten/WebAssembly, no-op conversions from C++ to native strings are technically impossible. The storage implementations for these platforms still make conversions as cheap as possible (avoiding UTF conversions, for example).

  The library also provides sys_string/sys_string_builder typedefs that use the "default" storage type on each platform (you can change which one it is via compilation options). Regardless of which storage is the default, you can always directly use other specializations in your code if necessary.

• Immutability. String instances cannot be modified. To do modifications, you use a separate "builder" class. This is similar to how many other languages do it and results in improved performance and elimination of a whole class of errors.

• Unicode-first. Instances of sys_string_t always store Unicode characters in either UTF-8, UTF-16 or UTF-32, depending on their storage. Iteration can be done in all of these encodings, and all operations (case conversion, case-insensitive comparisons, trimming) are specified as actions on a sequence of Unicode codepoints using Unicode algorithms. The library provides common Unicode string operations such as case conversion, normalization, and grapheme cluster iteration.

• Operations similar to Python or ECMAScript strings. You can do things like rtrim, split, join, starts_with, etc. on sys_string_t in a way proven to be natural and productive in those languages.

• Concatenation does not allocate temporaries. You can safely do things like result = s1 + s2 + s3. It will result in one memory allocation and 3 calls to memcpy to copy the contents of s1, s2 and s3 into the final result. Not 2 allocations and 5 copies like in other languages or with std::string.

• Bidirectional UTF-8/UTF-16/UTF-32 views. You can view sys_string_t as a sequence of UTF-8/16/32 characters and iterate forward or backward equally efficiently. Consider trying to find the last instance of Unicode whitespace in UTF-8 data. Doing it as fast as finding the first instance is non-trivial.
  The views also work on any C++ input ranges (C array, std::array, std::vector, std::string, or even std::ranges::istream_view) of characters (char, char8_t, char16_t, char32_t and wchar_t on platforms where it is Unicode). Thus you can iterate in UTF-8 over std::vector<char16_t> etc.

• Bidirectional grapheme cluster views. Similarly, you can also further view any of the UTF-8/UTF-16/UTF-32 views of sys_string_t as a sequence of grapheme clusters and iterate over them forward or backward equally efficiently. Consider the task of erasing the last user-perceived "character" from a string. To do so correctly, you need to erase the last grapheme cluster. Doing it and doing it fast is very non-trivial. This functionality also works on any C++ range of characters (but requires a forward range).

Why bother? Doesn't std::string work well?

An std::string storing UTF-8 (or std::wstring storing UTF-16 on Windows) works very well for some scenarios but fails miserably for others. The sys_string class is an attempt to create something that works well in situations where std::string would be a bad choice.

Specifically, std::basic_string is an STL container of a character type that owns its memory and controls it via a user-supplied allocator. These design choices make it very fast for direct character access but create the following problems:

• They rule out (efficient) reference-counted implementations, which means that when you copy an std::string instance, it must copy its content. Some of the penalty for that is alleviated by modern small string optimization, but this is, at best, a band-aid. There are workarounds, of course, such as using std::shared_ptr<std::string> "when it matters", but they result in even more complexity for something that is quite fundamental to any data processing.

• They foreclose any ability to efficiently interchange data with some other string type. It becomes problematic if your code needs to frequently ping-pong data between C++ and your OS string abstraction. Consider Apple's platforms (macOS, iOS). Applications written for these platforms often have to extensively interoperate with code that requires usage of the NSString * native string type. If you have to ping-pong string data a lot and/or store the same string data on both sides, using std::string will mean a large performance and memory penalty.

• They make std::basic_string Unicode-hostile. By being oblivious to the difference between a "storage unit" and a "character", std::basic_string cannot really handle Unicode encodings where the two differ (such as UTF-8 or UTF-16). Yes, you can store data in these encodings in an std::basic_string, but you need to be extremely careful how you use it. What will erase(it) do if the iterator points to the middle of a 4-byte UTF-8 sequence?

Finally, and unrelatedly to the above, std::string lacks some simple things that are taken for granted these days by users of pretty much all other languages. There are no case-insensitive comparisons, no "trim" or "split", etc. It is possible to write those yourself, of course, but here the Unicode-unfriendliness raises its ugly head. To do any of these correctly, you need to be able to handle a string as a sequence of Unicode characters, and doing so with std::string is cumbersome.

Non-goals

The following requirements that other string classes often have are specifically non-goals of this library.

• Support the C++ allocators mechanism. Since sys_string_t is meant to interoperate with other string class/types, it necessarily has to use the same allocation mechanisms as those. Different allocation behavior can be accomplished via a custom Storage class.

• Have a zero-cost const char * c_str() method on all platforms. The goal of the library is to provide an efficient conversion to the native string types rather than specifically const char *. While the ability to obtain const char * is provided everywhere, it might involve additional memory allocations and other overhead. Of course, when the storage of sys_string_t is char, it can be obtained with 0 cost.

• Make sys_string_t an STL container. Conceptually, a string is not a container. You can view the contents of a string as a sequence of UTF-8 or UTF-16, or UTF-32 codepoints, and the library provides such views that function as STL ranges.

• Support non-Unicode "narrow" and "wide" character encodings. sys_string_t only understands Unicode. Conversions to/from non-Unicode encodings are a job for a different library. Most significantly, it means that any char * passed to this library's methods is required to be in UTF-8.

• Provide locale-dependent functionality. Properly supporting locales with Unicode is an important area, but it belongs to another library, not this one. This library is focused on locale-independent behavior that works the same everywhere. For example, the to_lower method implements the locale-independent part of the Unicode specification. (Final uppercase 'Σ' transforms to 'ς', but 'I' always transforms to 'i'.)

Performance

In general, sys_string_t aims to have the same performance of its operations as the best hand-crafted code that uses corresponding native string types on every platform. For example, on macOS, code using sys_string should be as fast as code manually using NSString */CFStringRef.

In particular, using sys_string_generic (which interoperates with char * and is the default sys_string on non-Apple Unix or if the SYS_STRING_USE_GENERIC macro is set to 1) should be as fast as or better than using std::string.

This consideration needs to be kept in mind when evaluating whether sys_string is a better choice for your application than std::string. To continue Apple's example, an std::string is generally faster for direct character access than NSString * and thus faster than the default sys_string on Apple platforms too. If your code rarely transfers data from NSString * to std::string and spends most of the time iterating over std::string characters, then using std::string might be the right choice. Of course, in such a case, using sys_string_generic might still be a better choice.

Another way to look at it is that sys_string_t sometimes trades micro-benchmarking performance of individual string operations for reduced copying, allocations, and memory pressure overall. Whether this is the right tradeoff for you depends on the specifics of your codebase.

Unicode tables

Some sys_string Unicode operations, such as case conversions, case-insensitive comparisons, and grapheme cluster iteration, require the use of Unicode lookup tables. Currently, the size of the tables is ~80kB, if you use all of these features. It will be less if you only use some of them and zero if you use none.

If your application already contains ICU, it is possible to use it instead for these features. This will save around 100kB in the final executable (data and supporting code, again if using all of the features). On the other hand, using ICU makes these operations somewhat slower.

On yet another hand, using ICU provides consistency w.r.t. the Unicode standard supported within the rest of your application. Using internal tables will use the Unicode version of this library (currently 17), which might be different from what your ICU version uses.

Of course, if you don't already use ICU, integrating it might be painful, and its total size (unless you use a system-provided version) is orders of magnitude larger than 100kB.

Which approach to prefer thus depends on the specifics of your application.

Compatibility

Starting from version 3, this library requires a C++20 compiler. In particular, the following C++20 features must be available:

• Ranges support in the standard library (__cpp_lib_ranges >= 201911)
• Three-way comparison (spaceship operator)
• char8_t type
• std::endian support in the standard library (__cpp_lib_endian >= 201907)

The library is known to work with at least:

• Xcode 15.4
• MSVC 17.6
• Clang 16
• GCC 12
• Emscripten 3.1.70

Version 2 of this library was the last version supporting C++17. It is known to work at least with:

• Xcode 13
• MSVC 16.9
• Clang 12.0.5
• GCC 9.3
• Emscripten 3.1.21

Usage

• Building/Integrating
• Usage