# Project status

Champollion is a learning project, developed in the open at
[dorian-zhao/champollion](https://github.com/dorian-zhao/champollion). This
page is a snapshot; the repository's
[design brief](https://github.com/dorian-zhao/champollion/blob/main/docs/design.md)
is the source of truth, including the baseline every performance change is
reported against.

## What works

- **M5, done.** Multithreaded C and Go programs run under Direct mode. Alpine's
  busybox (pipelines, job control, subshells) and CPython 3.14 (subprocess,
  asyncio, multiprocessing) run unmodified against an Alpine riscv64 sysroot
  and match native output in the test suite.
- **M6 — VM mode (route A), done criterion met, not yet declared.** riscv64 container
  images run on the Mac through `rvcontainer`, and Alpine and Debian roots run
  `busybox`, `bash`, `coreutils`, `dpkg`, `apt`, `perl`, and CPython;
  `apt-get update` and package installation work as well.

## Performance (October 2026)

Measured on Apple silicon against native arm64, as the factor of the native
runtime (lower is better):

| Program | Factor of native |
| --- | --- |
| C | 1.5× |
| n-body | 1.37× |
| Lua | ~2.0× |
| SQLite | 1.87× |

Against `qemu-riscv64` in the same OrbStack VM: Lua 4.8× and SQLite 4.2×
faster.

## Known gaps

Direct mode (route B), in the order a program is likely to notice: dynamically
linked programs get no persistent code cache; realtime signals do not queue; of the
interval timers, only `ITIMER_REAL` works; startup costs about 0.5 ms more than
a static `hello` did before the host signal thread (3.4 ms against native
1.9 ms).

VM mode (route A): every `run` unpacks the image again; the Kata kernel is
6.12 while Containerization tests from 6.14.9; Linux builds linked by GNU `ld` run SQLite
about 12% slower than zig-linked ones, cause unknown.

## License

MIT or Apache-2.0, at your option.