/binaryen

Optimizer and compiler/toolchain library for WebAssembly

Primary LanguageWebAssemblyApache License 2.0Apache-2.0

CI

Binaryen

Binaryen is a compiler and toolchain infrastructure library for WebAssembly, written in C++. It aims to make compiling to WebAssembly easy, fast, and effective:

  • Easy: Binaryen has a simple C API in a single header, and can also be used from JavaScript. It accepts input in WebAssembly-like form but also accepts a general control flow graph for compilers that prefer that.

  • Fast: Binaryen's internal IR uses compact data structures and is designed for completely parallel codegen and optimization, using all available CPU cores. Binaryen's IR also compiles down to WebAssembly extremely easily and quickly because it is essentially a subset of WebAssembly.

  • Effective: Binaryen's optimizer has many passes (see an overview later down) that can improve code size and speed. These optimizations aim to make Binaryen powerful enough to be used as a compiler backend by itself. One specific area of focus is on WebAssembly-specific optimizations (that general-purpose compilers might not do), which you can think of as wasm minification, similar to minification for JavaScript, CSS, etc., all of which are language-specific.

Toolchains using Binaryen as a component (typically running wasm-opt) include:

For more on how some of those work, see the toolchain architecture parts of the V8 WasmGC porting blogpost.

Compilers using Binaryen as a library include:

  • AssemblyScript which compiles a variant of TypeScript to WebAssembly
  • wasm2js which compiles WebAssembly to JS
  • Asterius which compiles Haskell to WebAssembly
  • Grain which compiles Grain to WebAssembly

Binaryen also provides a set of toolchain utilities that can

  • Parse and emit WebAssembly. In particular this lets you load WebAssembly, optimize it using Binaryen, and re-emit it, thus implementing a wasm-to-wasm optimizer in a single command.
  • Interpret WebAssembly as well as run the WebAssembly spec tests.
  • Integrate with Emscripten in order to provide a complete compiler toolchain from C and C++ to WebAssembly.
  • Polyfill WebAssembly by running it in the interpreter compiled to JavaScript, if the browser does not yet have native support (useful for testing).

Consult the contributing instructions if you're interested in participating.

Binaryen IR

Binaryen's internal IR is designed to be

  • Flexible and fast for optimization.
  • As close as possible to WebAssembly so it is simple and fast to convert it to and from WebAssembly.

There are a few differences between Binaryen IR and the WebAssembly language:

  • Tree structure
    • Binaryen IR is a tree, i.e., it has hierarchical structure, for convenience of optimization. This differs from the WebAssembly binary format which is a stack machine.
    • Consequently Binaryen's text format allows only s-expressions. WebAssembly's official text format is primarily a linear instruction list (with s-expression extensions). Binaryen can't read the linear style, but it can read a wasm text file if it contains only s-expressions.
    • Binaryen uses Stack IR to optimize "stacky" code (that can't be represented in structured form).
    • When stacky code must be represented in Binaryen IR, such as with multivalue instructions and blocks, it is represented with tuple types that do not exist in the WebAssembly language. In addition to multivalue instructions, locals and globals can also have tuple types in Binaryen IR but not in WebAssembly. Experiments show that better support for multivalue could enable useful but small code size savings of 1-3%, so it has not been worth changing the core IR structure to support it better.
    • Block input values (currently only supported in catch blocks in the exception handling feature) are represented as pop subexpressions.
  • Types and unreachable code
    • WebAssembly limits block/if/loop types to none and the concrete value types (i32, i64, f32, f64). Binaryen IR has an unreachable type, and it allows block/if/loop to take it, allowing local transforms that don't need to know the global context. As a result, Binaryen's default text output is not necessarily valid wasm text. (To get valid wasm text, you can do --generate-stack-ir --print-stack-ir, which prints Stack IR, this is guaranteed to be valid for wasm parsers.)
    • Binaryen ignores unreachable code when reading WebAssembly binaries. That means that if you read a wasm file with unreachable code, that code will be discarded as if it were optimized out (often this is what you want anyhow, and optimized programs have no unreachable code anyway, but if you write an unoptimized file and then read it, it may look different). The reason for this behavior is that unreachable code in WebAssembly has corner cases that are tricky to handle in Binaryen IR (it can be very unstructured, and Binaryen IR is more structured than WebAssembly as noted earlier). Note that Binaryen does support unreachable code in .wat text files, since as we saw Binaryen only supports s-expressions there, which are structured.
  • Blocks
    • Binaryen IR has only one node that contains a variable-length list of operands: the block. WebAssembly on the other hand allows lists in loops, if arms, and the top level of a function. Binaryen's IR has a single operand for all non-block nodes; this operand may of course be a block. The motivation for this property is that many passes need special code for iterating on lists, so having a single IR node with a list simplifies them.
    • As in wasm, blocks and loops may have names. Branch targets in the IR are resolved by name (as opposed to nesting depth). This has 2 consequences:
      • Blocks without names may not be branch targets.
      • Names are required to be unique. (Reading .wat files with duplicate names is supported; the names are modified when the IR is constructed).
    • As an optimization, a block that is the child of a loop (or if arm, or function toplevel) and which has no branches targeting it will not be emitted when generating wasm. Instead its list of operands will be directly used in the containing node. Such a block is sometimes called an "implicit block".
  • Reference Types
  • The wasm text and binary formats require that a function whose address is taken by ref.func must be either in the table, or declared via an (elem declare func $..). Binaryen will emit that data when necessary, but it does not represent it in IR. That is, IR can be worked on without needing to think about declaring function references.
  • Binaryen IR allows non-nullable locals in the form that the wasm spec does, (which was historically nicknamed "1a"), in which a local.get must be structurally dominated by a local.set in order to validate (that ensures we do not read the default value of null). Despite being aligned with the wasm spec, there are some minor details that you may notice:
    • A nameless Block in Binaryen IR does not interfere with validation. Nameless blocks are never emitted into the binary format (we just emit their contents), so we ignore them for purposes of non-nullable locals. As a result, if you read wasm text emitted by Binaryen then you may see what seems to be code that should not validate per the spec (and may not validate in wasm text parsers), but that difference will not exist in the binary format (binaries emitted by Binaryen will always work everywhere, aside for bugs of course).
    • The Binaryen pass runner will automatically fix up validation after each pass (finding things that do not validate and fixing them up, usually by demoting a local to be nullable). As a result you do not need to worry much about this when writing Binaryen passes. For more details see the requiresNonNullableLocalFixups() hook in pass.h and the LocalStructuralDominance class.
  • Binaryen IR uses the most refined types possible for references, specifically:
    • The IR type of a ref.func is always a specific function type, and not plain funcref. It is also non-nullable.
    • Non-nullable types are also used for the type that try_table sends on branches (if we branch, a null is never sent), that is, it sends (ref exn) and not (ref null exn). In both cases if GC is not enabled then we emit the less-refined type in the binary. When reading a binary, the more refined types will be applied as we build the IR.
  • br_if output types are more refined in Binaryen IR: they have the type of the value, when a value flows in. In the wasm spec the type is that of the branch target, which may be less refined. Using the more refined type here ensures that we optimize in the best way possible, using all the type information, but it does mean that some roundtripping operations may look a little different. In particular, when we emit a br_if whose type is more refined in Binaryen IR then we emit a cast right after it, so that the output has the right type in the wasm spec. That may cause a few bytes of extra size in rare cases (we avoid this overhead in the common case where the br_if value is unused).
  • Strings
    • Binaryen allows string views (stringview_wtf16 etc.) to be cast using ref.cast. This simplifies the IR, as it allows ref.cast to always be used in all places (and it is lowered to ref.as_non_null where possible in the optimizer). The stringref spec does not seem to allow this though, and to fix that the binary writer will replace ref.cast that casts a string view to a non-nullable type to ref.as_non_null. A ref.cast of a string view that is a no-op is skipped entirely.

As a result, you might notice that round-trip conversions (wasm => Binaryen IR => wasm) change code a little in some corner cases.

  • When optimizing Binaryen uses an additional IR, Stack IR (see src/wasm-stack.h). Stack IR allows a bunch of optimizations that are tailored for the stack machine form of WebAssembly's binary format (but Stack IR is less efficient for general optimizations than the main Binaryen IR). If you have a wasm file that has been particularly well-optimized, a simple round-trip conversion (just read and write, without optimization) may cause more noticeable differences, as Binaryen fits it into Binaryen IR's more structured format. If you also optimize during the round-trip conversion then Stack IR opts will be run and the final wasm will be better optimized.

Notes when working with Binaryen IR:

  • As mentioned above, Binaryen IR has a tree structure. As a result, each expression should have exactly one parent - you should not "reuse" a node by having it appear more than once in the tree. The motivation for this limitation is that when we optimize we modify nodes, so if they appear more than once in the tree, a change in one place can appear in another incorrectly.
  • For similar reasons, nodes should not appear in more than one functions.

Intrinsics

Binaryen intrinsic functions look like calls to imports, e.g.,

(import "binaryen-intrinsics" "foo" (func $foo))

Implementing them that way allows them to be read and written by other tools, and it avoids confusing errors on a binary format error that could happen in those tools if we had a custom binary format extension.

An intrinsic method may be optimized away by the optimizer. If it is not, it must be lowered before shipping the wasm, as otherwise it will look like a call to an import that does not exist (and VMs will show an error on not having a proper value for that import). That final lowering is not done automatically. A user of intrinsics must run the pass for that explicitly, because the tools do not know when the user intends to finish optimizing, as the user may have a pipeline of multiple optimization steps, or may be doing local experimentation, or fuzzing/reducing, etc. Only the user knows when the final optimization happens before the wasm is "final" and ready to be shipped. Note that, in general, some additional optimizations may be possible after the final lowering, and so a useful pattern is to optimize once normally with intrinsics, then lower them away, then optimize after that, e.g.:

wasm-opt input.wasm -o output.wasm -O --intrinsic-lowering -O

Each intrinsic defines its semantics, which includes what the optimizer is allowed to do with it and what the final lowering will turn it to. See intrinsics.h for the detailed definitions. A quick summary appears here:

  • call.without.effects: Similar to a call_ref in that it receives parameters, and a reference to a function to call, and calls that function with those parameters, except that the optimizer can assume the call has no side effects, and may be able to optimize it out (if it does not have a result that is used, generally).

Tools

This repository contains code that builds the following tools in bin/ (see the building instructions):

  • wasm-opt: Loads WebAssembly and runs Binaryen IR passes on it.
  • wasm-as: Assembles WebAssembly in text format (currently S-Expression format) into binary format (going through Binaryen IR).
  • wasm-dis: Un-assembles WebAssembly in binary format into text format (going through Binaryen IR).
  • wasm2js: A WebAssembly-to-JS compiler. This is used by Emscripten to generate JavaScript as an alternative to WebAssembly.
  • wasm-reduce: A testcase reducer for WebAssembly files. Given a wasm file that is interesting for some reason (say, it crashes a specific VM), wasm-reduce can find a smaller wasm file that has the same property, which is often easier to debug. See the docs for more details.
  • wasm-shell: A shell that can load and interpret WebAssembly code. It can also run the spec test suite.
  • wasm-emscripten-finalize: Takes a wasm binary produced by llvm+lld and performs emscripten-specific passes over it.
  • wasm-ctor-eval: A tool that can execute functions (or parts of functions) at compile time.
  • wasm-merge: Merges multiple wasm files into a single file, connecting corresponding imports to exports as it does so. Like a bundler for JS, but for wasm.
  • wasm-metadce: A tool to remove parts of Wasm files in a flexible way that depends on how the module is used.
  • binaryen.js: A standalone JavaScript library that exposes Binaryen methods for creating and optimizing Wasm modules. For builds, see binaryen.js on npm (or download it directly from GitHub or unpkg). Minimal requirements: Node.js v15.8 or Chrome v75 or Firefox v78.

All of the Binaryen tools are deterministic, that is, given the same inputs you should always get the same outputs. (If you see a case that behaves otherwise, please file an issue.)

Usage instructions for each are below.

Binaryen Optimizations

Binaryen contains a lot of optimization passes to make WebAssembly smaller and faster. You can run the Binaryen optimizer by using wasm-opt, but also they can be run while using other tools, like wasm2js and wasm-metadce.

  • The default optimization pipeline is set up by functions like addDefaultFunctionOptimizationPasses.
  • There are various pass options that you can set, to adjust the optimization and shrink levels, whether to ignore unlikely traps, inlining heuristics, fast-math, and so forth. See wasm-opt --help for how to set them and other details.

See each optimization pass for details of what it does, but here is a quick overview of some of the relevant ones:

  • CoalesceLocals - Key "register allocation" pass. Does a live range analysis and then reuses locals in order to minimize their number, as well as to remove copies between them.
  • CodeFolding - Avoids duplicate code by merging it (e.g. if two if arms have some shared instructions at their end).
  • CodePushing - "Pushes" code forward past branch operations, potentially allowing the code to not be run if the branch is taken.
  • DeadArgumentElimination - LTO pass to remove arguments to a function if it is always called with the same constants.
  • DeadCodeElimination
  • Directize - Turn an indirect call into a normal call, when the table index is constant.
  • DuplicateFunctionElimination - LTO pass.
  • Inlining - LTO pass.
  • LocalCSE - Simple local common subexpression elimination.
  • LoopInvariantCodeMotion
  • MemoryPacking - Key "optimize data segments" pass that combines segments, removes unneeded parts, etc.
  • MergeBlocks - Merge a block to an outer one where possible, reducing their number.
  • MergeLocals - When two locals have the same value in part of their overlap, pick in a way to help CoalesceLocals do better later (split off from CoalesceLocals to keep the latter simple).
  • MinifyImportsAndExports - Minifies them to "a", "b", etc.
  • OptimizeAddedConstants - Optimize a load/store with an added constant into a constant offset.
  • OptimizeInstructions - Key peephole optimization pass with a constantly increasing list of patterns.
  • PickLoadSigns - Adjust whether a load is signed or unsigned in order to avoid sign/unsign operations later.
  • Precompute - Calculates constant expressions at compile time, using the built-in interpreter (which is guaranteed to be able to handle any constant expression).
  • ReReloop - Transforms wasm structured control flow to a CFG and then goes back to structured form using the Relooper algorithm, which may find more optimal shapes.
  • RedundantSetElimination - Removes a local.set of a value that is already present in a local. (Overlaps with CoalesceLocals; this achieves the specific operation just mentioned without all the other work CoalesceLocals does, and therefore is useful in other places in the optimization pipeline.)
  • RemoveUnsedBrs - Key "minor control flow optimizations" pass, including jump threading and various transforms that can get rid of a br or br_table (like turning a block with a br in the middle into an if when possible).
  • RemoveUnusedModuleElements - "Global DCE", an LTO pass that removes imports, functions, globals, etc., when they are not used.
  • ReorderFunctions - Put more-called functions first, potentially allowing the LEB emitted to call them to be smaller (in a very large program).
  • ReorderLocals - Put more-used locals first, potentially allowing the LEB emitted to use them to be smaller (in a very large function). After the sorting, it also removes locals not used at all.
  • SimplifyGlobals - Optimizes globals in various ways, for example, coalescing them, removing mutability from a global never modified, applying a constant value from an immutable global, etc.
  • SimplifyLocals - Key "local.get/set/tee" optimization pass, doing things like replacing a set and a get with moving the set’s value to the get (and creating a tee) where possible. Also creates block/if/loop return values instead of using a local to pass the value.
  • Vacuum - Key "remove silly unneeded code" pass, doing things like removing an if arm that has no contents, a drop of a constant value with no side effects, a block with a single child, etc.

"LTO" in the above means an optimization is Link Time Optimization-like in that it works across multiple functions, but in a sense Binaryen is always "LTO" as it usually is run on the final linked wasm.

Advanced optimization techniques in the Binaryen optimizer include SSAification, Flat IR, and Stack/Poppy IR.

See the Optimizer Cookbook wiki page for more on how to use the optimizer effectively.

Binaryen also contains various passes that do other things than optimizations, like legalization for JavaScript, Asyncify, etc.

Building

Binaryen uses git submodules (at time of writing just for gtest), so before you build you will have to initialize the submodules:

git submodule init
git submodule update

After that you can build with CMake:

cmake . && make

A C++17 compiler is required. On macOS, you need to install cmake, for example, via brew install cmake. Note that you can also use ninja as your generator: cmake -G Ninja . && ninja.

To avoid the gtest dependency, you can pass -DBUILD_TESTS=OFF to cmake.

Binaryen.js can be built using Emscripten, which can be installed via the SDK.

  • Building for Node.js: 
    emcmake cmake . && emmake make binaryen_js
  • Building for the browser: 
    emcmake cmake -DBUILD_FOR_BROWSER=ON . && emmake make

Visual C++

  1. Using the Microsoft Visual Studio Installer, install the "Visual C++ tools for CMake" component.

  2. Generate the projects:

    mkdir build
cd build
"%VISUAL_STUDIO_ROOT%\Common7\IDE\CommonExtensions\Microsoft\CMake\CMake\bin\cmake.exe" ..

    Substitute VISUAL_STUDIO_ROOT with the path to your Visual Studio installation. In case you are using the Visual Studio Build Tools, the path will be "C:\Program Files (x86)\Microsoft Visual Studio\2017\BuildTools".

  3. From the Developer Command Prompt, build the desired projects:

    msbuild binaryen.vcxproj

    CMake generates a project named "ALL_BUILD.vcxproj" for conveniently building all the projects.

Releases

Builds are distributed by the various toolchains that use Binaryen, like Emscripten, wasm-pack, etc. There are also official releases on GitHub:

https://github.com/WebAssembly/binaryen/releases

Currently builds of the following platforms are included:

  • Linux-x86_64
  • Linux-arm64
  • MacOS-x86_64
  • MacOS-arm64
  • Windows-x86_64
  • Node.js (experimental): A port of wasm-opt to JavaScript+WebAssembly. Run node wasm-opt.js as a drop-in replacement for a native build of wasm-opt, on any platform that Node.js runs on. Requires Node.js 18+ (for Wasm EH and Wasm Threads). (Note that this build may also run in Deno, Bun, or other JavaScript+WebAssembly environments, but is tested only on Node.js.)

Running

wasm-opt

Run

bin/wasm-opt [.wasm or .wat file] [options] [passes, see --help] [--help]

The wasm optimizer receives WebAssembly as input, and can run transformation passes on it, as well as print it (before and/or after the transformations). For example, try

bin/wasm-opt test/lit/passes/name-types.wast -all -S -o -

That will output one of the test cases in the test suite. To run a transformation pass on it, try

bin/wasm-opt test/lit/passes/name-types.wast --name-types -all -S -o -

The name-types pass ensures each type has a name and renames exceptionally long type names. You can see the change the transformation causes by comparing the output of the two commands.

It's easy to add your own transformation passes to the shell, just add .cpp files into src/passes, and rebuild the shell. For example code, take a look at the name-types pass.

Some more notes:

  • See bin/wasm-opt --help for the full list of options and passes.
  • Passing --debug will emit some debugging info. Individual debug channels (defined in the source code via #define DEBUG_TYPE xxx) can be enabled by passing them as list of comma-separated strings. For example: bin/wasm-opt --debug=binary. These debug channels can also be enabled via the BINARYEN_DEBUG environment variable.

wasm2js

Run

bin/wasm2js [input.wasm file]

This will print out JavaScript to the console.

For example, try

bin/wasm2js test/hello_world.wat

That output contains

 function add(x, y) {
  x = x | 0;
  y = y | 0;
  return x + y | 0 | 0;
 }

as a translation of

 (func $add (; 0 ;) (type $0) (param $x i32) (param $y i32) (result i32)
  (i32.add
   (local.get $x)
   (local.get $y)
  )
 )

wasm2js's output is in ES6 module format - basically, it converts a wasm module into an ES6 module (to run on older browsers and Node.js versions you can use Babel etc. to convert it to ES5). Let's look at a full example of calling that hello world wat; first, create the main JS file:

// main.mjs
import { add } from "./hello_world.mjs";
console.log('the sum of 1 and 2 is:', add(1, 2));

The run this (note that you need a new enough Node.js with ES6 module support):

$ bin/wasm2js test/hello_world.wat -o hello_world.mjs
$ node --experimental-modules main.mjs
the sum of 1 and 2 is: 3

Things keep to in mind with wasm2js's output:

  • You should run wasm2js with optimizations for release builds, using -O or another optimization level. That will optimize along the entire pipeline (wasm and JS). It won't do everything a JS minifer would, though, like minify whitespace, so you should still run a normal JS minifer afterwards.
  • It is not possible to match WebAssembly semantics 100% precisely with fast JavaScript code. For example, every load and store may trap, and to make JavaScript do the same we'd need to add checks everywhere, which would be large and slow. Instead, wasm2js assumes loads and stores do not trap, that int/float conversions do not trap, and so forth. There may also be slight differences in corner cases of conversions, like non-trapping float to int.

wasm-ctor-eval

wasm-ctor-eval executes functions, or parts of them, at compile time. After doing so it serializes the runtime state into the wasm, which is like taking a "snapshot". When the wasm is later loaded and run in a VM, it will continue execution from that point, without re-doing the work that was already executed.

For example, consider this small program:

(module
 ;; A global variable that begins at 0.
 (global $global (mut i32) (i32.const 0))

 (import "import" "import" (func $import))

 (func "main"
  ;; Set the global to 1.
  (global.set $global
   (i32.const 1))

  ;; Call the imported function. This *cannot* be executed at
  ;; compile time.
  (call $import)

  ;; We will never get to this point, since we stop at the
  ;; import.
  (global.set $global
   (i32.const 2))
 )
)

We can evaluate part of it at compile time like this:

wasm-ctor-eval input.wat --ctors=main -S -o -

This tells it that there is a single function that we want to execute ("ctor" is short for "global constructor", a name that comes from code that is executed before a program's entry point) and then to print it as text to stdout. The result is this:

trying to eval main
  ...partial evalling successful, but stopping since could not eval: call import: import.import
  ...stopping
(module
 (type $none_=>_none (func))
 (import "import" "import" (func $import))
 (global $global (mut i32) (i32.const 1))
 (export "main" (func $0_0))
 (func $0_0
  (call $import)
  (global.set $global
   (i32.const 2)
  )
 )
)

The logging shows us managing to eval part of main(), but not all of it, as expected: We can eval the first global.get, but then we stop at the call to the imported function (because we don't know what that function will be when the wasm is actually run in a VM later). Note how in the output wasm the global's value has been updated from 0 to 1, and that the first global.get has been removed: the wasm is now in a state that, when we run it in a VM, will seamlessly continue to run from the point at which wasm-ctor-eval stopped.

In this tiny example we just saved a small amount of work. How much work can be saved depends on your program. (It can help to do pure computation up front, and leave calls to imports to as late as possible.)

Note that wasm-ctor-eval's name is related to global constructor functions, as mentioned earlier, but there is no limitation on what you can execute here. Any export from the wasm can be executed, if its contents are suitable. For example, in Emscripten wasm-ctor-eval is even run on main() when possible.

wasm-merge

wasm-merge combines wasm files together. For example, imagine you have a project that uses wasm files from multiple toolchains. Then it can be helpful to merge them all into a single wasm file before shipping, since in a single wasm file the calls between the modules become just normal calls inside a module, which allows them to be inlined, dead code eliminated, and so forth, potentially improving speed and size.

wasm-merge operates on normal wasm files. It differs from wasm-ld in that respect, as wasm-ld operates on wasm object files. wasm-merge can help in multi-toolchain situations where at least one of the toolchains does not use wasm object files.

For example, imagine we have these two wasm files:

;; a.wasm
(module
  (import "second" "bar" (func $second.bar))

  (export "main" (func $func))

  (func $func
    (call $second.bar)
  )
)
;; b.wasm
(module
  (import "outside" "log" (func $log (param i32)))

  (export "bar" (func $func))

  (func $func
    (call $log
      (i32.const 42)
    )
  )
)

The filenames on your local drive are a.wasm and b.wasm, but for merging / bundling purposes let's say that the first is known as "first" and the second as "second". That is, we want the first module's import of "second.bar" to call the function $func in the second module. Here is a wasm-merge command for that:

wasm-merge a.wasm first b.wasm second -o output.wasm

We give it the first wasm file, then its name, and then the second wasm file and then its name. The merged output is this:

(module
  (import "outside" "log" (func $log (param i32)))

  (export "main" (func $func))
  (export "bar" (func $func_2))

  (func $func
    (call $func_2)
  )

  (func $func_2
    (call $log
      (i32.const 42)
    )
  )
)

wasm-merge combined the two files into one, merging their functions, imports, etc., all while fixing up name conflicts and connecting corresponding imports to exports. In particular, note how $func calls $func_2, which is exactly what we wanted: $func_2 is the function from the second module (renamed to avoid a name collision).

Note that the wasm output in this example could benefit from additional optimization. First, the call to $func_2 can now be easily inlined, so we can run wasm-opt -O3 to do that for us. Also, we may not need all the imports and exports, for which we can run wasm-metadce. A good workflow could be to run wasm-merge, then wasm-metadce, then finish with wasm-opt.

wasm-merge is kind of like a bundler for wasm files, in the sense of a "JS bundler" but for wasm. That is, with the wasm files above, imagine that we had this JS code to instantiate and connect them at runtime:

// Compile the first module.
var first = await fetch("a.wasm");
first = new WebAssembly.Module(first);

// Compile the first module.
var second = await fetch("b.wasm");
second = new WebAssembly.Module(second);

// Instantiate the second, with a JS import.
second = new WebAssembly.Instance(second, {
  outside: {
    log: (value) => {
      console.log('value:', value);
    }
  }
});

// Instantiate the first, importing from the second.
first = new WebAssembly.Instance(first, {
  second: second.exports
});

// Call the main function.
first.exports.main();

What wasm-merge does is basically what that JS does: it hooks up imports to exports, resolving names using the module names you provided. That is, by running wasm-merge we are moving the work of connecting the modules from runtime to compile time. As a result, after running wasm-merge we need a lot less JS to get the same result:

// Compile the single module.
var merged = await fetch("merged.wasm");
merged = new WebAssembly.Module(merged);

// Instantiate it with a JS import.
merged = new WebAssembly.Instance(merged, {
  outside: {
    log: (value) => {
      console.log('value:', value);
    }
  }
});

// Call the main function.
merged.exports.main();

We still need to fetch and compile the merged wasm, and to provide it the JS import, but the work to connect two wasm modules is not needed any more.

Handling exports

By default wasm-merge errors if there are overlapping export names. That is, wasm-merge will automatically handle overlapping function names and so forth, because those are not externally visible (the code still behaves the same), but if we renamed exports then the outside would need to be modified to expect the new export names, and so we error instead on such name conflicts.

If you do want exports to be renamed, run wasm-merge with --rename-export-conflicts. Later exports will have a suffix appended to them to ensure they do not overlap with previous exports. The suffixes are deterministic, so once you see what they are you can call them from the outside.

Another option is to use --skip-export-conflicts which will simply skip later exports that have conflicting names. For example, this can be useful in the case where the first module is the only one that interacts with the outside and the later modules just interact with the first module.

Features

wasm-merge uses the multi-memory and multi-table features. That is, if multiple input modules each have a memory then the output wasm will have several memories, and will depend on the multi-memory feature, which means that older wasm VMs might not be able to run the wasm. (As a workaround for such older VMs you can run wasm-opt --multi-memory-lowering to lower multiple memories into a single one.)

Testing

./check.py

(or python check.py) will run wasm-shell, wasm-opt, etc. on the testcases in test/, and verify their outputs.

The check.py script supports some options:

./check.py [--interpreter=/path/to/interpreter] [TEST1] [TEST2]..
  • If an interpreter is provided, we run the output through it, checking for parse errors.
  • If tests are provided, we run exactly those. If none are provided, we run them all. To see what tests are available, run ./check.py --list-suites.
  • Some tests require emcc or nodejs in the path. They will not run if the tool cannot be found, and you'll see a warning.
  • We have tests from upstream in tests/spec, in git submodules. Running ./check.py should update those.

Note that we are trying to gradually port the legacy wasm-opt tests to use lit and filecheck as we modify them. For passes tests that output wast, this can be done automatically with scripts/port_passes_tests_to_lit.py and for non-passes tests that output wast, see #4779 for an example of how to do a simple manual port.

For lit tests the test expectations (the CHECK lines) can often be automatically updated as changes are made to binaryen. See scripts/update_lit_checks.py.

Non-lit tests can also be automatically updated in most cases. See scripts/auto_update_tests.py.

Setting up dependencies

./third_party/setup.py [mozjs|v8|wabt|all]

(or python third_party/setup.py) installs required dependencies like the SpiderMonkey JS shell, the V8 JS shell and WABT in third_party/. Other scripts automatically pick these up when installed.

Run pip3 install -r requirements-dev.txt to get the requirements for the lit tests. Note that you need to have the location pip installs to in your $PATH (on linux, ~/.local/bin).

Fuzzing

./scripts/fuzz_opt.py [--binaryen-bin=build/bin]

(or python scripts/fuzz_opt.py) will run various fuzzing modes on random inputs with random passes until it finds a possible bug. See the wiki page for all the details.

Design Principles

  • Interned strings for names: It's very convenient to have names on nodes, instead of just numeric indices etc. To avoid most of the performance difference between strings and numeric indices, all strings are interned, which means there is a single copy of each string in memory, string comparisons are just a pointer comparison, etc.
  • Allocate in arenas: Based on experience with other optimizing/transformating toolchains, it's not worth the overhead to carefully track memory of individual nodes. Instead, we allocate all elements of a module in an arena, and the entire arena can be freed when the module is no longer needed.

Debug Info Support

Source Maps

Binaryen can read and write source maps (see the -ism and -osm flags to wasm-opt). It can also read and read source map annotations in the text format, that is,

;;@ src.cpp:100:33
(i32.const 42)

That 42 constant is annotated as appearing in a file called src.cpp at line 100 and column 33. Source maps and text format annotations are interchangeable, that is, they both lead to the same IR representation, so you can start with an annotated wat and have Binaryen write that to a binary + a source map file, or read a binary + source map file and print text which will contain those annotations.

The IR representation of source map info is simple: in each function we have a map of expressions to their locations. Optimization passes should update the map as relevant. Often this "just works" because the optimizer tries to reuse nodes when possible, so they keep the same debug info.

Shorthand notation

The text format annotations support a shorthand in which repeated annotations are not necessary. For example, children are tagged with the debug info of the parent, if they have no annotation of their own:

;;@ src.cpp:100:33
(i32.add
  (i32.const 41)      ;; This receives an annotation of src.cpp:100:33
  ;;@ src.cpp:111:44
  (i32.const 1)
)

The first const will have debug info identical to the parent, because it has none specified, and generally such nesting indicates a "bundle" of instructions that all implement the same source code.

Note that text printing will not emit such repeated annotations, which can be confusing. To print out all the annotations, set BINARYEN_PRINT_FULL=1 in the environment. That will print this for the above add:

[i32] ;;@ src.cpp:100:33
(i32.add
 [i32] ;;@ src.cpp:100:33
 (i32.const 41)
 [i32] ;;@ src.cpp:111:44
 (i32.const 1)
)

(full print mode also adds a [type] for each expression, right before the debug location).

The debug information is also propagated from an expression to its next sibling:

;;@ src.cpp:100:33
(local.set $x
 (i32.const 0)
)
(local.set $y ;; This receives an annotation of src.cpp:100:33
 (i32.const 0)
)

You can prevent the propagation of debug info by explicitly mentioning that an expression has not debug info using the annotation ;;@ with nothing else:

;;@ src.cpp:100:33
(local.set $x
 ;;@
 (i32.const 0) ;; This does not receive any annotation
)
;;@
(local.set $y ;; This does not receive any annotation
 (i32.const 7)
)

This stops the propagatation to children and siblings as well. So, expression (i32.const 7) does not have any debug info either.

There is no shorthand in the binary format. That is, roundtripping (writing and reading) through a binary + source map should not change which expressions have debug info on them or the contents of that info.

Implementation Details

The source maps format defines a mapping using segments, that is, if a segment starts at binary offset 10 then it applies to all instructions at that offset and until another segment begins (or the end of the input is reached). Binaryen's IR represents a mapping from expressions to locations, as mentioned, so we need to map to and from the segment-based format when writing and reading source maps.

That is mostly straightforward, but one thing we need to do is to handle the lack of debug info in between things that have it. If we have A B C where B lacks debug info, then just emitting a segment for A and C would lead A's segment to also cover B, since in source maps segments do not have a size - rather they end when a new segment begins. To avoid B getting smeared in this manner, we emit a source maps entry to B of size 1, which just marks the binary offset it has, and without the later 3 fields of the source file, line number, and column. (This appears to be the intent of the source maps spec, and works in browsers and tools.)

DWARF

Binaryen also has optional support for DWARF. This primarily just tracks the locations of expressions and rewrites the DWARF's locations accordingly; it does not handle things like re-indexing of locals, and so passes that might break DWARF are disabled by default. As a result, this mode is not suitable for a fully optimized release build, but it can be useful for local debugging.

FAQ

  • Why the weird name for the project?

Binaryen's name was inspired by Emscripten's: Emscripten's name suggests it converts something into a script - specifically JavaScript - and Binaryen's suggests it converts something into a binary - specifically WebAssembly. Binaryen began as Emscripten's WebAssembly generation and optimization tool, so the name fit as it moved Emscripten from something that emitted the text-based format JavaScript (as it did from its early days) to the binary format WebAssembly (which it has done since WebAssembly launched).

"Binaryen" is pronounced in the same manner as "Targaryen".

  • Does it compile under Windows and/or Visual Studio?

Yes, it does. Here's a step-by-step tutorial on how to compile it under Windows 10 x64 with with CMake and Visual Studio 2015. However, Visual Studio 2017 may now be required. Help would be appreciated on Windows and OS X as most of the core devs are on Linux.