wasm-opt over Wado output

wasm-opt reads a core module, not a component. jco splits a Wado component and hands back its core module, so the pipeline runs through jco. What wasm-opt then removes shows what Wado's own codegen could remove.

Setup

binaryen is not in the mise registry, and mise's github: backend gets a 403 from the GitHub API in a cloud session. Download the release tarball instead:

curl -sSL https://github.com/WebAssembly/binaryen/releases/download/version_133/binaryen-version_133-x86_64-linux.tar.gz | tar xz -C scratchpad

Pipeline

B=scratchpad/binaryen-version_133/bin
S=scratchpad/wasm-opt
FEATURES=(--mvp-features --enable-sign-ext --enable-mutable-globals
  --enable-nontrapping-float-to-int --enable-simd --enable-relaxed-simd
  --enable-bulk-memory --enable-bulk-memory-opt --enable-call-indirect-overlong
  --enable-exception-handling --enable-tail-call --enable-reference-types
  --enable-multivalue --enable-gc --enable-extended-const --enable-multimemory)
mkdir -p $S
wado compile -Os -f no-wide-arithmetic prog.wado -o $S/prog.wasm
node scripts/jco/transpile-released.mjs $S/prog.wasm $S/jco   # → $S/jco/prog.core.wasm
$B/wasm-opt "${FEATURES[@]}" $S/jco/prog.core.wasm -o $S/none.wasm
$B/wasm-opt "${FEATURES[@]}" -O3 $S/jco/prog.core.wasm -o $S/O3.wasm

• Make -Os the baseline. It is -O2 with the name section stripped and assert messages dropped. An -O2 baseline counts about 9KB of names that wasm-opt drops by default.
• -f no-wide-arithmetic is required, since the module is run on V8 (see the jco skill).
• Name the features explicitly, as an array because zsh does not word-split a string. Never pass -all: it lets wasm-opt emit proposals V8 cannot load, compact imports among them.
• The run with no pass (none.wasm) re-encodes the module and nothing more, so it separates encoding waste from optimization.
• Add -g to keep function names for a diff. Measure size without it.

Reading the difference

wado-dev-tools stands in for the wasm-tools CLI (scripts/dev-bin.sh wado-dev-tools <command> <file.wasm>).

• objdump gives the size and count of each section.
• print --offsets puts each instruction's byte offset beside it. Use it for encoding questions: an instruction takes the bytes up to the next offset.
• print gives WAT to diff. A whole-module diff is too large to read, so compare per-function instruction counts and opcode histograms with a small Node script instead.
• -O2 --skip-pass=NAME measures what one pass is worth: the bytes -O2 grows by without it. Passes overlap, so these do not add up to the whole gap. Running a pass alone misleads for the -optimizing variants, which rerun the whole function pipeline on what they touch.

To run the result, copy the jco output directory and replace its <name>.core.wasm. Then run it as the jco skill describes. Keep a median of at least 7 runs per variant, since the spread on a cloud VM is ±15%.

Findings: zlib benchmark (binaryen 133, 2026-10)

Sizes of the zlib benchmark's core module:

Variant	Bytes
Wado -Os	76,476
wasm-opt, no pass	76,943
wasm-opt -O2	70,073
wasm-opt -O3	67,979
wasm-opt -Oz	67,853

Re-encoding alone makes this module 467 bytes larger, so binaryen's encoding finds nothing to save in Wado's.

In 2026-09, on Node 26, decompression ran faster under wasm-opt -O2 and -Os: the median rose 10–40% across two sessions. Compression, and everything under -O3, stayed within the noise.

What each -O2 pass is worth, by --skip-pass:

Pass	Bytes	What it does to Wado's output
coalesce-locals	2,230	Mostly cleans up after inlining: 337 run alone
inlining-optimizing	1,799	Takes 114 functions to 53
optimize-instructions	1,164	Peepholes, ref.as_non_null among them
remove-unused-brs	1,062	Turns block/br_if/return into if/else
precompute-propagate	782	Constant propagation
dae-optimizing	758	Drops unused and always-constant parameters
heap2local	664	Turns non-escaping structs into locals
vacuum, merge-blocks	824	Removes the blocks the passes above leave behind

What wasm-opt removes, as open codegen targets:

• Local declarations: the emitter wrote one entry per local, 2,942 entries and 7.1KB in all. Grouped by type, they take 346 entries and 1.0KB. wir_optimize/local_layout.rs groups them, and puts the most used first.
• Duplicate function types: 136 types, of which -O2 keeps 69. Most of the rest go with the functions it inlines.
• ref.as_non_null: -O2 takes the count from 388 to 16. Codegen added most of them, after reads of nullable global slots. A global promoted to a non-null constant now has a non-null slot (wir_optimize/const_global.rs). The object of a GC access, which traps on null by itself, is not narrowed (wir_optimize/cleanup.rs). The -Os build keeps 17.
• Set-then-get: a local.set X followed at once by local.get X occurred 1,191 times. wasm-opt turns it into a local.tee or removes it. The last WIR pass fuses each pair into a tee (fuse_remaining_local_tees), and copy propagation removes a copy of a tee'd local. The -Os build keeps 39 pairs.
• Set sinking: simplify-locals moves a local.set into the one read of it, so the value stays on the stack. -O2 removes 565 local.set and 725 local.get, and adds 188 local.tee.
• Local coalescing: wasm-opt merges locals whose live ranges never overlap, as a register allocator would. wir_optimize/local_coalesce.rs does the same per Wasm type. On the -Os build it takes 2,752 locals to 1,020.
• Control-flow shape: -O2 takes block from 229 to 57, br_if from 191 to 54, and return from 184 to 44, while if grows from 1,278 to 1,385.
• Single-caller inlining: -O2 inlines functions with one caller whatever their size, inflate_fast among them.
• Identical functions: the -Os build has none. The -O2 build keeps 9 copies of the outlined $cold0 bounds-check path of List<T>::index_value and index_assign, one per element type, and i32's Display::fmt equals its Inspect::inspect. --duplicate-function-elimination saves 4,722 bytes there. Two param_spec clones whose bindings differ can fold to one body (fmt_float_special$spec0 and $spec1 in json_twitter).
• Empty functions: a module whose globals all become constants keeps an empty $initialize_module. wir_optimize/empty_work.rs removes the calls to it and the once-flag that guarded them, and DCE drops both.