Building to WebAssembly

WebAssembly is a binary format for executing code on the web, allowing fast start times (smaller download and much faster parsing in browsers when compared to JS or asm.js). Emscripten compiles to WebAssembly by default, but you can also compile to JS for older browsers.

For some historical background, see

Setup

WebAssembly is emitted by default, without the need for any special flags.

Note

If you don’t want WebAssembly, you can disable it with something like

emcc [..args..] -s WASM=0

Note

Emscripten’s WebAssembly support depends on Binaryen, which is provided by the emsdk (if you don’t use the emsdk, you need to build it and set it up in your .emscripten file).

Note

Deciding to compile to wasm or JS can be done at the linking stage: it doesn’t affect the object files.

Backends

Emscripten emits WebAssembly using the upstream LLVM wasm backend, since version 1.39.0 (October 2019), and the old fastcomp backend is deprecated (you can still use the deprecated fastcomp backend by getting latest-fastcomp instead of the normal latest, or 1.39.0-fastcomp instead of 1.39.0, etc.).

There are some differences you may notice between the two backends, if you upgrade from fastcomp to upstream:

  • The wasm backend is strict about linking files with different features sets - for example, if one file was built with atomics but another was not, it will error at link time. This prevents possible bugs, but may mean you need to make some build system fixes.
  • WASM=0 behaves differently in the two backends. In fastcomp we emit asm.js, while in upstream we emit JS (since not all wasm constructs can be expressed in asm.js). Also, the JS support implements the same external WebAssembly.* API, so in particular startup will be async just like wasm by default, and you can control that with WASM_ASYNC_COMPILATION (even though WASM=0).
  • The wasm backend uses wasm object files by default. That means that it does codegen at the compile step, which makes the link step much faster - like a normal native compiler. For comparison, in fastcomp the compile step emits LLVM IR in object files.
    • You normally wouldn’t notice this, but some compiler flags affect codegen, like DISABLE_EXCEPTION_CATCHING. Such flags must be passed during codegen. The simple and safe thing is to pass all -s flags at both compile and link time.
    • You can enable LTO object files with the usual llvm compiler flags (-flto, -flto=full, -flto=thin, -emit-llvm). These flags will make the wasm backend behave more like fastcomp. Neither fastcomp nor the wasm backend without wasm object files will run the LLVM optimization passes by default, even if using LLVM IR in object files; for that you must pass --llvm-lto 1.
    • Another thing you might notice is that fastcomp’s link stage is able to perform some types of link time optimization by default that the LLVM backend requires flags for (for example, propagate the “noexcept” property), which can have an impact on performance. Building with LTO in the wasm backend can achieve similar things, see the previous bullet point.
  • wasm-ld, the linker used by the wasm backend, requires libraries (.a archives) to contain symbol indexes. This matches the behaviour the native GNU linker. While emar will create such indexes by default, native tools such as GNU ar and GNU strip are not aware of the WebAssembly object format and cannot create archive indexes. In particular, if you run GNU strip on an archive file that contains WebAssembly object files it will remove the index which makes the archive unusable at link time.
  • Fastcomp emits asm.js and so has some limitations on function pointers. For example, the RESERVED_FUNCTION_POINTERS setting exists there to work around the fact that we can’t grow the table. In the upstream backend table growth is easy, and you can just enable ALLOW_TABLE_GROWTH.
  • Also see the blocker bugs on the wasm backend, and the wasm backend tagged issues.

Binaryen codegen options

Trapping

WebAssembly can trap - throw an exception - on things like division by zero, rounding a very large float to an int, and so forth. In asm.js such things were silently ignored, as in JavaScript they do not throw, so this is a difference between JavaScript and WebAssembly that you may notice, with the browser reporting an error like float unrepresentable in integer range, integer result unrepresentable, integer overflow, or Out of bounds Trunc operation.

Fastcomp/asm2wasm

In fastcomp/asm2wasm, emscripten will emit code that is optimized for size and speed, which means it emits code that may trap on the things mentioned before. That mode is called allow. The other modes are clamp, which will avoid traps by clamping values to a reasonable range, and js, which ensures the exact same behavior as JavaScript does (which also does clamping, but makes sure to clamp exactly like JavaScript does, and also do other things JavaScript would).

In general, using clamp is safest, as whether such a trap occurs depends on how the LLVM optimizer optimizes code. In other words, there is no guarantee that this will not be an issue, and updating LLVM can make a problem appear or vanish (the wasm spec process has recognized this problem and intends to standardize new operations that avoid it). Also, there is not a big downside to using clamp: it is only slightly larger and slower than the default allow, in most cases. To do so, build with

-s "BINARYEN_TRAP_MODE='clamp'"

However, if the default (to allow traps) works in your codebase, then it may be worth keeping it that way, for the (small) benefits. Note that js, which preserves the exact same behavior as JavaScript does, adds a large amount of overhead, so unless you really need that, use clamp (js is often useful for debugging, though).

LLVM wasm backend

The LLVM wasm backend avoids traps by adding more code around each possible trap (basically clamping the value if it would trap). This can increase code size and decrease speed, if you don’t need that extra code. The proper solution for this is to use newer wasm instructions that do not trap, by calling emcc or clang with -mnontrapping-fptoint. That code may not run in older VMs, though.

Compiler output

When using emcc to build to WebAssembly, you will see a .wasm file containing that code, as well as the usual .js file that is the main target of compilation. Those two are built to work together: run the .js (or .html, if that’s what you asked for) file, and it will load and set up the WebAssembly code for you, properly setting up imports and exports for it, etc. Basically, you don’t need to care about whether the compiled code is asm.js or WebAssembly, it’s just a compiler flag, and otherwise everything should just work (except the WebAssembly should be faster).

  • Note that the .wasm file is not standalone - it’s not easy to manually run it without that .js code, as it depends on getting the proper imports that integrate with JS. For example, it receives imports for syscalls so that it can do things like print to the console. There is work in progress towards ways to create standalone .wasm files, see the WebAssembly Standalone page.

You may also see additional files generated, like a .data file if you are preloading files into the virtual filesystem. All that is exactly the same as when building to asm.js. One difference you may notice is the lack of a .mem file, which for asm.js contains the static memory initialization data, which in WebAssembly we can pack more efficiently into the WebAssembly binary itself.

Testing native WebAssembly in browsers

WebAssembly support is enabled by default as of Firefox 52, Chrome 57 and Opera 44. On Edge 15 you can enable it via “Experimental JavaScript Features” flag.

Debugging

asm.js support is considered very stable now, and you can change between it and wasm with -s WASM=0, so if you see something odd in a wasm build, comparing to a parallel asm.js build can help. In general, any difference between the two could be a compiler bug or browser bug, but there are a few legitimate causes of different behavior between the two, that you may want to rule out:

  • wasm allows unaligned accesses, i.e. it will load 4 bytes from an unaligned address the same way x86 does (it doesn’t care it’s unaligned). asm.js works more like ARM CPUs which mostly don’t accept such things (but they often trap, while asm.js just returns a wrong result). To rule this out, you can build with -s SAFE_HEAP=1, that will catch all such invalid accesses.
  • Timing issues - wasm might run faster or slower. To some extent you can mitigate that by building with -s DETERMINISTIC=1.
  • Trap mode. As mentioned above, we can generate wasm that traps or that avoids traps. Make sure the trap mode is "js" when comparing builds. The "js" trap mode is also useful in a single build, as otherwise operations like division or float-to-int may trap, and the optimizer may happen to change whether a trap occurs or not, which can be confusing (for example, enabling SAFE_HEAP may prevent some optimizations, and a trap may start to occur). Instead, in the "js" trap mode there are no traps and all operations are deterministically defined as identical to JavaScript.
  • Minor libc and runtime differences exist between wasm and asm.js. We used to have a way to emit more compatable builds (-s "BINARYEN_METHOD='asmjs,native-wasm'" etc.) but due to its complexity and low value it was removed.
  • Floating-point behavior: WebAssembly uses 32-bit floats in a standard way, while asm.js by default implements floats using doubles. That can lead to differences in the precision of results. You can force 32-bit float behavior in asm.js with -s PRECISE_F32=1, in which case it should be identical to wasm.
  • Browser instability: It’s worth testing multiple browsers, as one might have a wasm bug that another doesn’t. You can also test the Binaryen interpreter (e.g. using the interpret-binary method, as discussed above).

If you find that an asm.js build has the same behavior as a wasm one, then it is currently easier to debug the asm.js build: you can edit the source easily (add debug printouts, etc.), there is debug info and source maps support, etc.

Debugging WebAssembly

When you do need to debug a WebAssembly build, the following tips might help you.

WebAssembly doesn’t have source maps support yet, but building with -g will emit both a text and a binary wasm, and it will include function names in both, and also include source file and line number information in the text, for example, building hello world might produce this .wat:

;; tests/hello_world.c:4
(drop
  (call $_printf
    (i32.const 1144)
    (get_local $$vararg_buffer)
  )
)
;; tests/hello_world.c:5
(return
  (i32.const 0)
)

This indicates that the printf call comes from line 4, and the return from line 5, of hello_world.c.

.wasm files and compilation

WebAssembly code is prepared somewhat differently than asm.js. asm.js can be bundled inside the main JS file, while as mentioned earlier WebAssembly is a binary file on the side, so you will have more than one file to distribute.

Another noticeable effect is that WebAssembly is compiled asynchronously by default, which means you must wait for compilation to complete before calling compiled code (by waiting for main(), or the onRuntimeInitialized callback, etc., which you also need to do when you have anything else that makes startup async, like a .mem file for asm.js, or preloaded file data, etc.). You can turn off async compilation by setting WASM_ASYNC_COMPILATION=0, but that may not work in Chrome due to current limitations there.

  • Note that even with async compilation turned off, fetching the WebAssembly binary may need to be an asynchronous operation (since the Web does not allow synchronous binary downloads on the main thread). If you can fetch the binary yourself, you can set Module['wasmBinary'] and it will be used from there, and then (with async compilation off) compilation should be synchronous.

Web server setup

To serve wasm in the most efficient way over the network, make sure your web server has the proper MIME time for .wasm files, which is application/wasm. That will allow streaming compilation, where the browser can start to compile code as it downloads.

In Apache, you can do this with

AddType application/wasm .wasm

Also make sure that gzip is enabled:

AddOutputFilterByType DEFLATE application/wasm

If you serve large .wasm files, the webserver will consume CPU compressing them on the fly at each request. Instead you can pre-compress them to .wasm.gz and use content negotiation:

Options Multiviews
RemoveType .gz
AddEncoding x-gzip .gz
AddType application/wasm .wasm