ADR-0017: Linear-backend bump/arena allocator; one fixed GC strategy, not a pluggable one

Status: Accepted

Context

The linear-memory backend (src/codegen-linear/, the alternative to the WasmGC backend per ADR-0003 and the codegen-axes split) owns its own allocation — there is no host GC underneath it. We had to decide both which allocation strategy it uses and whether to leave that strategy pluggable.

The field consensus for AOT source→Wasm compilers, captured as recommendation R10 in docs/architecture/compiler-design-lessons.md, is twofold:

1. Do not build a pluggable / interchangeable GC. Supporting tracing and reference-counting as swappable strategies is documented as not viable — reference counting cannot collect cycles, so a real system bolts tracing on anyway and ends up maintaining both. The GC subsystem is the most-redesigned part of every compiler in this space; over-abstracting it is a known trap.
2. A bump/arena "allocate-and-never-free" mode is a near-free win for short-lived programs. Most conformance- and CLI-style WASI programs allocate and then exit; for them a tracing collector is pure overhead and code size.

Decision

• The linear backend's allocator is a bump/arena: __malloc advances a single __heap_ptr global (8-byte aligned) and frees nothing. Reclamation happens implicitly when the Wasm instance is dropped — the allocate-and-exit model. This is the default and only mode (#1856).
• __malloc grows linear memory on demand (memory.grow) when a request would exceed the current pages, so the arena is usable for non-trivial programs rather than silently overflowing the initial 64 KiB page.
• An opt-in allocator: "arena-reset" (--allocator arena-reset) adds __arena_reset() (O(1) rewind of the whole arena) and __arena_used() for embedders that reuse one instance across many short-lived tasks. Off by default so the common case pays nothing.
• We do not ship a pluggable GC abstraction. If reclamation within a single long-lived run is ever genuinely needed, we will commit to one fixed strategy (tracing, or RC-with-a-cycle-collector) — chosen and recorded then, not abstracted now. This need is deferred: no current standalone / WASI workload requires intra-run reclamation, and the WasmGC backend already covers long-lived, cyclic-graph workloads by delegating to the host GC.

Consequences

• Smallest-binary, fastest path for standalone/WASI programs; the bump runtime is ~135 bytes and adds no per-allocation metadata. The arena-reset exports cost ~86 bytes and only when requested.
• No cycle-collection or fragmentation handling on the linear backend — by design. Workloads that need that target the WasmGC backend instead.
• The "one fixed strategy, decided later" stance keeps the door open without paying the abstraction tax up front. Revisit only when a concrete workload demonstrates the need.