Optimization

Optimization is opt-in: -O runs rasc’s own passes, -O2 adds a pass through Binaryen’s wasm-opt. The default build stays fast and reproducible.

-O: rasc’s own passes

• Inliner — splices callee bodies at call sites for functions of at most 32 instructions or marked @inline; two rounds, recursion excluded. This matters more than it sounds: generic stdlib code produces deep dispatch chains that generic wasm optimizers cannot untangle from outside.
• Allocation sinking — a post-inline pass that scalarizes non-escaping new into locals, deleting the allocation, its header writes, and its shadow-stack spills. A sunk object never reaches any GC. Naive allocate-per-frame code measured 12.3× faster on the WAMR interpreter — the tier un-optimized user content actually runs on — and 2.8× on V8. --no-alloc-sink turns it off for bisecting.
• Bounds-check elimination — the canonical for (let i = 0; i < arr.length; i++) read loop routes element reads through unchecked accessors. Stores keep the checked path. unchecked(expr) is honored under -O; debug builds stay fully checked.

-O2: the binaryen post-pass

-O2 is -O plus a wasm-opt round trip (-O3 -fimfs=100 -ifwl -g), discovering the binary via RASC_WASM_OPT or PATH. The post-pass is best-effort: builds without binaryen keep working. It passes --enable-multivalue only when a signature actually carries several results, so byte-identity is preserved for modules that never use it.

Know the shape of the cost: on the Box2D port the full pipeline is 32 ms debug, 41 ms at -O, and about 1 s at -O2 — wasm-opt dominates. And binaryen alone recovers little from naive codegen; rasc’s own inlining is the prerequisite that opens the module up.