SchOS/docs/plan/finch.md
Nexoniarz d7ad3d09f7 docs: Finch builds itself; what is left before main switches
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015f73FesGZGYPkmDvUcj6QX
2026-10-25 19:49:21 +01:00

4.7 KiB

Finch, for SchOS's own programs

Two things, and only one of them is later.

  • Bringing Finch into SchOS (below) is later — behind 1.0.md's freeze like everything else here that is not named there.
  • Finishing Finch itself, and making it fully self-hosting, is work on the Finch repository and does not touch this tree. It can proceed now; see the gap.

Finch is the author's own language: https://pandor.pl/git/nexoniarz/Finch (GitHub mirror: Nexoniarz/Finch). The plan is to write SchOS's own applications and tools in it, in place of C, and so leave behind what C does badly: silent narrowing, signed/unsigned mixing, undefined behaviour on an out-of-range index, a forgotten free.

What it is today (version 2.4.0)

Measured from the repository, not from its README alone:

  • LLVM front end, C-like syntax, no semicolons, no headers, no preprocessor. It compiles through LLVM's -O2 pipeline to native code; there is no garbage collector.
  • Values, not references: lists, text and structs copy on assignment and are freed at the end of their block. Out-of-range indexes, division by zero and null pointers stop with a runtime error rather than undefined behaviour.
  • Errors as values (int(s) or 0, try read_file(p)), maps, methods, modules.
  • C interop through libclang: import "foo.h" plus link "foo" calls any C library, structs by value and callbacks included. That is the way in to everything SchOS already has — musl, the /dev/* interfaces, Alpine's libraries.
  • Targets: Linux, macOS, Windows, on x86-64 and ARM64 (src/target.cpp; the System V, AAPCS64 and Microsoft x64 ABIs are in src/abi.cpp).
  • Tooling: a language server (finch lsp), VS Code and Kate support, 74 tests.
  • Self-hosting: boot/ is a Finch compiler written in Finch (about 3,100 lines, against about 5,900 for the C++ one in src/, language server included) that builds itself to a byte-identical result — tests/boot.sh.

What SchOS would need to do

  1. A schos target. finch picks a triple, a C compiler and a linker per target. SchOS runs unmodified Alpine binaries and Alpine's musl (see alpine.md), so the first target is the existing x86_64-unknown-linux-musl, and a Finch program built on the host should run on SchOS before anything in Finch is made SchOS-specific. Test that before writing a line of target code.
  2. The compiler on SchOS. Needs LLVM 21 and libclang, both of which Alpine packages. That is a question of apk and memory, not of SchOS.
  3. boot/ as the SchOS-hosted compiler. The C++ compiler needs LLVM's C++ libraries; the one in boot/ is the candidate for being the compiler that ships in the image, once it emits through LLVM's C API or a smaller back end.
  4. Bindings for SchOS's own interfaces (/dev/pci, /dev/netdev, schinit's job protocol, the framebuffer) — generated by the C importer from the headers in include/uapi/, so they are not hand-written and cannot drift from the kernel.

Where self-hosting stands

The Finch compiler written in Finch (boot/) now builds itself: stage 2 and stage 3 are byte-identical (58,660 lines of LLVM IR), and it passes all 22 of the repository's run tests (the C++ one passes 77/77). The work is on the finish-finch branch of github.com/Nexoniarz/Finch; before-bootstrap there is the all-C++ compiler, kept as the backup. main is untouched.

What the self-hosted compiler has: sized numbers, methods, defer, errors as values, maps, for k, v in, C import through libclang (written in Finch, in boot/cimport.fch), C structs by value and addr(fn) callbacks (System V x86-64), run, build, ir, version, -O0, -l, -I.

What it does not have yet, and the C++ compiler does:

-g debug info text IR needs the DWARF metadata written by hand
--target (Windows, ARM64, macOS) their struct-passing rules (abi.cpp) and linkers
lsp, the editor language server lsp.cpp, 580 lines
export fn, build --shared on the author's local embed branch

main becomes the Finch-written compiler when those are in, so that nothing a user does today stops working; until then it is the seed that builds the first stage. The new language features (enums, match, generics, a standard library) come before that switch, as simple as the rest of the language.

Rules

  • Inference, not an instruction: the kernel stays in C. Finch is for applications and tools; the user said "our apps or stuff like that".
  • No new capability starts until the register of unfinished work is empty; the SchOS half of this document is a plan, not an exception to that.