What's this?

A WebAssembly interpreter written in C.

Goals

• Learn the spec by implementing it
• Have a fun
• Correctness
• Clean code
• Small footprint
• Reasonable performance
• Portability
  • Core wasm: C11
  • WASI: POSIX-like
  • Clang/LLVM extentions are ok where an alternative implmentation in pure C is reasonably easy.

Non-goals

• Top-notch performance
• Stable API/ABI

What are implemented?

Note: As this runtime is relatively new, all proposals which had finished when I started this implementation are just included in the WebAssembly in the above table. It includes mutable-global, nontrapping-float-to-int-conversions, sign-extension-ops, multi-value, reference-types, bulk-memory-operations, simd.

Where can this run?

Use as a command

See toywasm command help message.

Usage:
    toywasm [OPTIONS] [--] <MODULE> [WASI-ARGS...]
Options:
    --allow-unresolved-functions
    --disable-jump-table
    --disable-localtype-cellidx
    --disable-resulttype-cellidx
    --dyld
    --dyld-bindnow
    --dyld-dlfcn
    --dyld-path LIBRARY_DIR
    --dyld-stack-size C_STACK_SIZE_FOR_PIE_IN_BYTES
    --invoke FUNCTION[ FUNCTION_ARGS...]
    --load MODULE_PATH
    --max-frames NUMBER_OF_FRAMES
    --max-memory MEMORY_LIMIT_IN_BYTES
    --max-stack-cells NUMBER_OF_CELLS
    --repl
    --repl-prompt STRING
    --print-build-options
    --print-stats
    --timeout TIMEOUT_MS
    --version
    --wasi
    --wasi-dir HOST_DIR[::GUEST_DIR]
    --wasi-env NAME=VAR
    --wasi-littlefs-dir LITTLEFS_IMAGE_PATH::LFS_DIR[::GUEST_DIR]
Examples:
    Run a wasi module
        toywasm --wasi module
    Load a module and invoke its function
        toywasm --load module --invoke "func arg1 arg2"

Use as a library

See example apps:

• Investigate module

• Load and execute WASI module

Toywasm provides cmake config files for its libraries. If your app is using cmake, you can use find_package to find toywasm libraries as it's done in the CMakeLists.txt of the above example apps.

Release binaries

Warning Toywasm version numbers are NOT a semver. It's just an increasing number, which doesn't imply anything about compatibilities or features.

• Release binaries are built with clang, with ThinLTO where it's available.

• For macOS, we ship a universal binary. (amd64 and arm64)

• We ship binaries with two configurations for each platforms/architectures. The assets with full- in the filename have most features enabled. Another one is with the default configuration.

• Each asset contains the toywasm cli command, static libraries, and cmake config files.

Build from source

• To build toywasm, we recommend to use clang for the following features, which some parts of toywasm codebase is kind of assuming to have. They are not required for the correctness though.

  • musttail attribute (LLVM 13 and later)

  • LTO

• For macOS and similar posix-like environment, you can build it with cmake in a usual way.

  % cmake -B build
  % cmake --build build

  For build-time options, see the ToywasmConfig.cmake.

  By default, it uses ThinLTO if available. You might want to use regular LTO instead to get a bit better optimization:

  % cmake -B build . -DUSE_IPO=OFF -DCMAKE_C_FLAGS=-flto=full
  % cmake --build build

  By default, it requires wabt and cmocka for tests. If you don't want to install them, you can disable tests.

  % cmake -B build -D BUILD_TESTING=OFF
  % cmake --build build

  Otherwise, after building it, you can run tests with:

  % cd build
  % ctest

• For wasm32-wasi and wasm32-wasi-threads targets, we provide a convenient script build-wasm32-wasi.sh to download wasi-sdk and build toywasm with it.

• For NuttX, it's probably easiest to use interpreters/toywasm in NuttX apps. There are a few sample defconfig with toywasm enabled, including sim:toywasm and esp32s3-devkit:toywasm.

How slow/fast is this?

• Computation benchmark

• Startup benchmark

Why is this slow?

• Unlike many of interpreters, toywasm aims to execute wasm bytecode directly where possible. That is, it doesn't "compile" wasm bytecode into intermediate code.

  Unfortunately, wasm bytecode is simply not efficient to execute that way. It's basically designed to be somehow "compiled" at the load time.

  Many of interpreters out there translate wasm bytecode to their internal bytecode for performance reasons. Wasm3 and WAMR "fast" interpreter work exactly that way. Even WAMR "classic" interpreter replaces some critical instructions in-place.

  While toywasm maps wasm modules read-only and never modifies them in-place, it still generates a few types of offline annotations on the bytecode to avoid being too slow. While they are smaller than a full translation, you might consider them a kind of translation. See Annotations for details.

• I don't like to use huge-switch statements or labels as values GNU C extension, which are well-known techniques to implement efficient interpreters.

Internals

• Annotations

• check_interrupt and restartable errors

References

• Wasm3

• WAMR "fast" interpreter

• labels as values GNU C extension