[pbaam]

I guess he means that in order to achieve incremental compilation they need to write their own code generation and linker for every architecture and format. This is needed because incremental compilation here doesn't just mean storing object files in a cache directory (which has always worked that way). They also want to cache every analyzed function and declaration. So they have to serialize compiler state to a file. But after analysis is done, LLVM will start code generation from the beginning (which is the time expensive thing, even in debug builds)

[jvillasante]

Yes, but isn't that an implementation detail? Shouldn't they prioritize getting to 1.0 (the language itself) and then work in implementation details like that? I mean, It's a monumental task to write compiler and linker from scratch!

[pbaam]

Well, if your compilations turn to be submilisecond it's not an implementation detail :) *. As of now it is only supported for x86_64 Linux (only ELF format) and it has some bugs; incremental compilation is in its very early stages. Andrew talked about it in the 2024 roadmap video[1] why they are digging so low on the multiplatform toolchain (for besides incremental compilation):

- Fast build times (closely related to IC, but LLVM gives very slow development iterations, even for the compiler development)

- Language innovations: besides IC, async/await is a feature Andrew determined to not be feasable to implement with LLVM's coroutines. Async will likely not make it into 1.0, as noted in the 0.13 release notes. It is not discarted yet but neither is it the priority.

- There are architectures that don't work very well on LLVM: SPARC and RISC-V are the ones I remember

My personal point is that a language that is meant to compete with C cannot have a hard dependency on a C++ project. That, and that it's appealing to have an alternative to LLVM whenever you want to do some JIT but don't want to bring a heavy dependency

[1] https://www.youtube.com/watch?v=5eL_LcxwwHg

* There is also the `--watch` flag for `zig build` which re-runs every step (IC helps) everytime a file is saved.

[edit: formatting]*

[geodel]

> Shouldn't they prioritize getting to 1.0 (the language itself)

Nope. Different languages have different priorities and different USPs. For Zig sub-second compilation / incremental compilation, cross compiling toolchain are flagship features. Without those there is no point in releasing 1.0.