ud-arch-codec

The shared spine every univdreams arch backend speaks. One trait, one open registry; the compile and decompile pipelines dispatch through it so every arch sees the same plumbing without duplicating it.

What's in the crate

ArchCodec — the trait. Methods cover the encoder surface the rest of the pipeline needs (assemble, jump, call, conditional jump, switch dispatch, move, arith, return, size queries) with Unsupported defaults so arches only implement what they model.
ArchError — soft (Unsupported) vs hard (Assemble, OutOfRange, UnknownArch) errors. The pipeline treats Unsupported as "fall back to pinned bytes."
EncodeHints — per-call interpretation hints (today just wide, x86's short-vs-rel32 toggle).
SwitchSpec — shared dispatch descriptor (today only "msvc-jmp-table" is recognised).
registry::register / for_arch — open registry of (arch_name, e_machine) -> Box<dyn ArchCodec> factories.

Adding a new arch

Create the arch crate (crates/ud-arch-<name>/).
Add ud-arch-codec to its Cargo.toml.
Implement ArchCodec for your codec type. Start with the four "always-supported" methods (assemble_one, encode_jump, encode_call, encode_cond_jump) and the three size queries — return Unsupported from everything else until you need it.

Expose a pub fn register() that submits your factory:

pub fn register() {
    ud_arch_codec::register(factory);
}

fn factory(
    arch_name: Option<&str>,
    e_machine: Option<u64>,
) -> Option<Box<dyn ArchCodec>> {
    if arch_name == Some("riscv64") { Some(Box::new(RiscV64Codec)) }
    else { None }
}

Call your_crate::register() from ud_translate::register_all_arches (the workspace's single entry point for arch registration). The CLI and wasm playground both invoke that automatically.

That's the contract. The lower path and decompile-side byte-drop pass start using your codec as soon as for_arch resolves it from a parsed @module block.

Stmt → trait method map

`Stmt::` variant	Codec method asked at lower time	Falls back to
`Asm { text }` (empty bytes)	`assemble_one(text, ip)` (+ desymbolize)	Hard error
`Goto { target, wide }`	`encode_jump(source, target, hints)`	Hard error
`IfGoto { cond_code, target, wide }`	`encode_cond_jump_with_code(cond_code, …)`	Hard error
`Call { direct_target }` (when set)	`encode_call(source, target, hints)`	Hard error
`Switch { dispatch, … }`	`encode_switch_dispatch(spec)`	Hard error
`IfBlock` cond/tail (empty)	`encode_cond_jump(text, …)` / `encode_jump(…)`	Hard error
`WhileBlock` entry/tail (empty)	same	Hard error
`Move { dst, src }` (empty bytes)	`encode_move(dst, src)`	Hard error
`Return { value }` (empty bytes)	`encode_return(value)`	Hard error

The decompile-side byte-drop pass mirrors each of these: it only clears bytes when the codec re-encodes them to match the original. The byte-identity guard is unconditional.

Layering note

ud-arch-codec deliberately doesn't depend on ud-ast — ud-ast already depends on ud-arch-x86 (for the canonical-emit derivation of head_bytes), so making ud-arch-codec depend on ud-ast would close a cycle. The registry takes raw (arch_name, e_machine) pairs that ud-translate extracts from a parsed module.

If you need richer Stmt-aware encoder types here later, do it through a side crate that depends on both — not by broadening ud-arch-codec.

ud-arch-codec 0.2.0

ud-arch-codec

What's in the crate

Adding a new arch

Stmt → trait method map

Layering note