Request: wasm SIMD (v128) support in Dwarfsm inline WAT parser

[mizchi]

Motivation

I'm building a SIMD abstraction module for MoonBit (mizchi/simd) that provides optimized array operations (sum, dot product, adler32, etc.). The native target achieves 13–33x speedup via C FFI, but the wasm target is stuck on scalar fallback because Dwarfsm doesn't fully support v128 SIMD instructions.

Related: #2844

Current Dwarfsm v128 Support Status

I investigated the Dwarfsm source (dwarfsm_parse.ml in moonbitlang/moonbit-compiler) and tested inline WAT with various v128 instructions:

Feature	Status	Notes
v128 as valtype (param/result)	✅ Works
v128.load / v128.store	✅ Works	Memory access
f32x4.add, f32x4.mul, f64x2.add, f64x2.mul	✅ Works	Float SIMD ops
v128.const	❌ Parse error	Constant initialization
i32x4.add, i32x4.mul, etc.	❌ Parse error	Integer SIMD ops
i32x4.extract_lane, i8x16.shuffle	❌ Parse error	Lane manipulation

Two Remaining Blockers

1. Missing instruction table entries in Dwarfsm

i32x4.*, i8x16.*, v128.const, and other integer SIMD instructions are not registered in dwarfsm_parse.ml. The float SIMD ops (f32x4.*, f64x2.*) already work, so adding the missing patterns to the instruction table should be straightforward.

2. #external type maps to externref in wasm

When defining #external type V128 in MoonBit, it compiles to externref in the wasm output. This causes a type mismatch with inline WAT that uses v128:

CompileError: type mismatch (externref vs v128)

A language-level v128 type or a way to control the wasm representation of #external type would be needed.

Context

• Repository with benchmarks and investigation: https://github.com/mizchi/simd
• The wasm_memory_ptr pattern (FixedArray ptr + 8 byte offset) for linear memory access has been verified and is ready for use once SIMD support lands.

Requested Changes

1. Add missing v128 instruction patterns to dwarfsm_parse.ml (integer SIMD ops + v128.const)
2. Provide a way to use v128 as a proper wasm value type (not externref) from MoonBit code

[mizchi]

One additional note on motivation: SIMD instructions are one of the primary areas where wasm can achieve a significant performance advantage over JavaScript in browser environments. Without v128 support, MoonBit's wasm output loses one of its strongest differentiators against JS for compute-heavy workloads (matrix operations, audio/image processing, checksums, etc.).

[Yu-zh]

Thanks for bringing this up. I have made similar attempt to support simd in MoonBit: https://github.com/Yu-zh/v128. I had to use a hacked compiler to make things work. So your requests are mostly reasonable to me.

Adding the missing v128 instructions should be simple. For the type v128, we still need to discuss about it within our team.