Skip to content
/ Cnx Public

A Proof-of-Concept for a Modern Standard Library for GNU C 11+

License

Notifications You must be signed in to change notification settings

braxtons12/Cnx

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Cnx

CMake

Cnx is a GNU C (11+) library providing type-safe collections and ergonomic features typical in higher-level languages, to C. It aims to be a proof-of-concept view of what a more modern standard (or at least widely used base) library could look like in the future if it only had to support C23, with much more modern abstractions and ergonomics over the standard C library. It is currently under active development and has not yet hit a stable release point.

Some features of Cnx include:

  • Type-Safe collections implemented (mostly) as manually instantiated templates. Currently implemented collections include CnxString, CnxVector(T), and CnxArray(T, N)
  • Error handling facilities similar to Rust and boost::outcome with equivalent semantics and similar API to Rust, via CnxError and CnxResult(T)
  • Optional value type, CnxOption(T) based on Rust's std::option::Option with equivalent semantics and similar API to Rust
  • Facilities for defining, implementing, and using polymorphic interfaces via Traits
  • Iterators, Ranges, and foreach loops (equivalent to C++'s for(elem : collection))
  • Human-readable (no more having to remember which character combination corresonds to which type in printf), type-safe string formatting and formatted I/O via the CnxFormat Trait
  • More intuitive, more performant file API with CnxFile, along with simple file system interaction API with CnxPath
Why not standard C?

Cnx makes heavy use of __auto_type and some of its features are only possible with it. It also uses statement-expressions for several features and __attribute__(cleanup) for scoped destruction. It could be refactored to be Standard C in C23 or a subsequent standard, if C gets auto, typeof, __VA_OPT__, lamdas, and defer.

Can I use this in production?

Can you? Sure! Should you? Probably not yet. Testing right now is minimal and only enough to ensure a majority of functionality works in the general case, and is definitely not thorough enough yet to ensure reliability in production.

Documentation

You can view the documentation here

Getting Started

Cnx uses CMake, and incorporating it into your project is easy!

First, set up your CMake project. In CMakeLists.txt:

FetchContent_Declare(Cnx
        GIT_REPOSITORY "https://github.com/braxtons12/Cnx"
        GIT_TAG "0.1.0"
        )

FetchContent_MakeAvailable(Cnx)

### Setup your target......

target_link_libraries(your_target Cnx)

Then, include your desired headers, either the main header, Cnx/Cnx.h, for everything, or individual ones for granular imports.

Example

#include <Cnx/CnxDef.h>
#define VECTOR_INCLUDE_DEFAULT_INSTANTIATIONS TRUE
#define RANGE_INCLUDE_DEFAULT_INSTANTIATIONS TRUE
#include <Cnx/Cnx.h>

void transform(i32* restrict elem) {
	*elem *= 3;
}

i32 main(i32 argc, const_cstring* argv) {

	// vector is already instantiated for builtins like `i32` and some provided types like `CnxString`,
	// so, we can just use it directly here
	let_mut vec = cnx_vector_new_with_capacity(i32, 10);

	// insert 10 elements, 0 through 9, into the vector
	ranged_for(i, 0, 10) {
		cnx_vector_push_back(vec, i);
	}

	// print information about the vector (size, capacity, whether it is currently in
	// small-size-optimization mode) to `stdout`, followed by a newline.
	println("{}", as_format_t(CnxVector(i32), vec));

	// print each element, followed by a newline, to `stdout`
	// prints 0 through 9
	foreach(elem, vec) {
		println("{}", elem);	
	}

	// transform the elements in the vector with the above-defined `transform` function
	// and returns a view of the vector as a `CnxRange(i32)` (that we ignore)
	cnx_transform(i32, vec, transform);

	// print each element, followed by a newline, to `stdout`
	// prints multiples of 3 from 0 through 27
	foreach(elem, vec) {
		println("{}", elem);	
	}
}

Collections

Cnx collections provide type-safety, iterators, the ability to use user-provided default and copy constructors, and destructors, for stored types, and, where applicable, are allocator aware. For more details on collections, see the documentation on the specific collection.

Formatting

String formatting uses a very simple syntax. In a format string, {} specifies that an argument should be formatted and placed in that location. Format specifiers can be placed inside the braces to specify the way the argument should be formatted. Currently supported specifiers are limited:

  • {}: The default format
  • {d}: Format the associated argument as a decimal (base-10) number. This is the default for integral types. Only applicable to signed, unsigned, and floating point numeric types
  • {dyy}: Format a floating point number as a decimal (base-10) number, with yy digits after the decimal point
  • {e}: Format a floating point number as scientific notation with the default number of significant figures (3) after the decimal point. This is the default for floating point types.
  • {eyy}: Format a floating point number as scientific notation with yy number of significant figures after the decimal point.
  • {x}: Format an integral number as lower-case hexadecimal
  • {X}: Format an integral number as upper-case hexadecimal

Providing a specifier that is invalid for the associated argument will result in a runtime assert in debug builds and unspecified behavior in release builds (for builtin and provided types).

For more details on string formatting, see the documentation for the CnxFormat Trait, cnx_format(format_string, ...), and the [CnxFormat](@ref format) module.

Performance

I know, you're probably thinking "So, something that nice can't be fast right? What's the cost?" And the answer is: not necessarily anything. Generally, performance is faster than traditional standard C functionality, even with the increased ergonomics and composability gained from using the library's features.

We have two benchmarks to showcase this. The first pits println against printf from the standard library. The second pits the Cnx file API against the standard library's fprintf. For the code used for each benchmark and more detailed results (Std. Dev., Median, individual runs), see the project/code in the "benchmark" subfolder.

The implementation of println makes heavy use of most of the functionality presented in the example (except for CnxRange(T)), as well as the Result(T) type, which means it's using many of the facilities currently provided by the library. Meanwhile, the file API uses everything println does, in addition to providing a safer and more ergonomic API for file manipulation than the standard library, while also taking better advantage of features like buffering to further improve performance. So, lets start by taking a look at the benchmark comparing the relative speed of println to printf.

This benchmark consisted of printing out N strings to stdout, each consisting of:

  • the Mth multiple of 1024, unsigned
  • the Mth multiple of negative 1024
  • the Mth multiple of negative 1024.1024
  • a CnxString pre-initialized to "This is a string"

where M is in [0, N)

It was run 10 times for each N, with the average of the 10 runs taken. The benchmark was performed with builds from both Clang 14.0.0 and GCC 11.2.0, and with both the default system allocator and jemalloc.
All benchmarks were run on an Intel Core i7-8750H with 16GB RAM running EndeavorOS (Arch Linux) with the Zen Kernel 5.17.5-zen1-1-zen.
Clang builds were compiled with "-flto -Ofast -ffast-math -DNDEBUG".
GCC builds were compiled with "-flto -ffat-lto-objects -Ofast -DNDEBUG" All numbers are relative performance compared to printf

While these numbers are Linux specific, in general you can expect comparable numbers on Windows using native clang, and even better numbers using MinGW, but benchmark on your specific platform for platform specific numbers.

println vs printf Results

N Clang + System Allocator Clang + jemalloc GCC + System Allocator GCC + jemalloc
1 1.6316 1.9758 2.9815 2.1675
10 1.8762 1.4501 1.7318 1.6026
100 1.2305 1.2312 1.0868 1.0535
1000 1.1330 1.1666 1.0652 1.0045
10000 1.3016 1.0991 1.0266 1.0610
100000 1.1154 1.1374 1.0509 1.0517
average 1.3814 1.3434 1.4905 1.3235

As you can see, on average you can expect around a 38% performance boost compared to printf. For infrequent I/O you can expect up to a 200% performance boost, and for high frequency I/O you can expect around a 10% performance boost. If you dig into the detailed benchmark results, you'll see that in general you can expect it to be fairly allocator insensitive with clang (somewhat less so with GCC), and that building with clang and using the default system allocator is the most consistently fast option to use.

So not only does using Cnx for string formatting and I/O give greatly improved ergonomics and composability over traditional methods, you also get a performance increase too!

Now for the benchmark pitting the file API against fprintf. This is nearly identical to the previous benchmark as fas as setup and methodologies go. The only difference is that the strings are printed to two separate files, one for the file API and one for fprintf, each using the API being benchmarked for performing the string formatting and I/O.

Cnx File API vs fprintf Results

N Clang + System Allocator Clang + jemalloc GCC + System Allocator GCC + jemalloc
1 5.2710 6.6535 3.6145 5.0965
10 3.6245 3.3975 3.2620 2.3175
100 2.4345 2.7620 1.6815 1.6745
1000 1.5145 1.5195 1.1050 1.1800
10000 1.3485 1.3350 0.9645 1.0255
100000 1.3335 1.3115 1.0365 1.0005
average 2.5878 2.8298 1.9440 2.0491

So, on average you can expect around a 100% performance boost compared to fprintf. For infrequent I/O you can expect at least a 250% performance boost, or even higher, and for high frequency I/O you can expect around a 17% performance boost. You can see that, particularly at very sparse workloads, taking better advantage of buffering the way Cnx's file API does leads to significant performance improvements over the standard functionality. If you dig into the detailed benchmark results, you'll see that allocator choice plays a bigger role here than with println, in particular at very sparse workloads. Unforntunately, GCC begins to sputter out as the workload increases, and in some cases actually manages to perform worse than the standard library functions. However, the gcc + system allocator pairing does the most consistent performance (lowest average standard deviation and smallest abs(averaged median - averaged average)). That said, it's also the slowest pairing, and the clang + system allocator pairing comes in at a close second in terms of consistency while still being the second fastest overall. Lastly, you'll see the clang + jemalloc pairing is the most consistently fast option to use (highest average and highest averaged median).

This was the performance for formatting builtin types (u32, i32, f32, cstring. CnxString's CnxFormat implementation simply forwards itself, so it's about as expensive as passing a cstring to printf). That said, it would be reasonable to expect this performance to extend to printing for custom types as well.

Testing

Tests are set up as a separate "Cnx-Test" target in the CMake project.
To run the tests, simply configure and build the test target, then run the resulting "Cnx-Test" executable.
Please feel free to submit new tests!

Inside the Cnx main directory:

cmake -B build -G "Ninja"
cmake --build build --target Cnx-Test
./build/Cnx-Test

Contributing

Feel free to submit issues, pull requests, etc.!
When contributing code, please adhead to the project .clang-tidy, follow the project .clang-format (except in judicious cases of macros ruining things), prefer let or let_mut over explicit typing (where possible), and prefer simplicity and correctness over performance by default.

License

Cnx uses the MIT license.

Special Thanks

Special thanks should be given to the Rust team and C++ standardization committee, and contributors of each, for creating great programming languages that have inspired a lot of the functionality provided by this library. Thanks should also go to TotallyNotChase for inspiring the design of the Trait system and Iterators with c-iterators and to Hirrolot for also inspiring the Trait system with interface99.