|quackerd 07d1bc7f6f||1 month ago|
|arch||5 months ago|
|inc||5 months ago|
|kern||5 months ago|
|mk||11 months ago|
|scripts||5 months ago|
|test||6 months ago|
|.drone.yml||1 month ago|
|.gitattributes||2 years ago|
|.gitignore||6 months ago|
|CMakeLists.txt||5 months ago|
|Dockerfile||1 month ago|
|LICENSE||1 month ago|
|README.md||1 month ago|
|bochsrc||5 months ago|
Fedora: dnf install -y cmake make clang nasm xorriso lld grub2-pc-modules grub2-tools-extra
Load the iso with QEMU/your favorite simulator.
I would like my kernel code to be explicit so that I can reason about performance, memory allocation/deallocation. That mostly means when I look at a statement I know exactly what it does. The philosophy overlaps with Go's design quite a lot: https://commandcenter.blogspot.com/2012/06/less-is-exponentially-more.html.
Using fully-featured C++ is overly complicated for kernels and I'm dubious of OOP in general. Especially while enforcing “modern” C++, sometimes I find myself struggling more with the language itself than getting work done. I have had lengthy thoughts myself regarding C++ but decided to drop it. Here are some of the pros and cons I came up with.
C++ is stronger typed than C. Simply compiling the kernel itself with a C++ compiler provides more type safety than C.
They are compile time casts so no runtime overhead. They provide a bit better type safety than C style casts. The only two casts we would need are probably const_cast and reinterpret_cast.
For type safety for data structures. Linux's list.h isn't type safe. FreeBSD's queue.h tries to mimic templates with macros, which is less elegant than just using template. Update: this might not be true after trying to implement type safe in-place linked list in C++. I feel like template is just as inelegant for that.
Oh boy how I wish C standard would include namespace, if it weren't for backward compaibility and stable ABI.
But rust did better?
This list explains SOME of the banned features that might seem useful.
All data structures should be POD structs. Basically it means a struct without user-defined constructors and destructors. The reason is 1. encapsulation is pretty useless in the kernel. 2. constructors and destructors are implicitly called when declaring objects and when objects go out of scope, which violates explicitness. 3. Everything becomes easy to reason about. Now I don't need to worry about “Object obj;” implicitly invoking the allocator, acquiring a lock and querying a SQL database.
RAII which relies on ctors/dtors sure is nice but IMO Golang's defer is a much better approach being expressive as well as functionally “similar”. C has the extension “cleanup” supported by GCC and Clang which does the same thing as defer. I'll investigate the latter.
Member methods should just be functions that take the struct as a parameter. It's what C++ does internally anyway. Adding member methods also obfuscates the PODness of structs, ergo banned.
Inheritance is banned except for describing interfaces like file descriptors and other possible places that might benefit from it. Even then it shouldn't be abused - e.g. the kernel does NOT need a generic list interface with a billion implementations.
Just give it an expressive enough name. Function overloading is obfuscation.
Think about what “f();” could mean in C++ and the code executed by “a + b;". Need I say more?
I don't like mixing references with pointers. I don't find reference offering much more than raw pointers.
Totally useless for bond.