Building large projects with Emscripten is very easy. Emscripten provides two simple scripts that configure your makefiles to use emcc as a drop-in replacement for gcc — in most cases the rest of your project’s current build system remains unchanged.
To build using Emscripten you need to replace gcc with emcc in your makefiles. This is done using emconfigure, which sets the appropriate environment variables like CXX (C++ compiler) and CC (the compiler).
Consider the case where you normally build with the following commands:
./configure
make
To build with Emscripten, you would instead use the following commands:
# Run emconfigure with the normal configure command as an argument.
./emconfigure ./configure
# Run emmake with the normal make to generate linked LLVM bitcode.
./emmake make
# Compile the linked code generated by make to JavaScript + WebAssembly.
# 'project.o' should be replaced with the make output for your project, and
# you may need to rename it if it isn't something emcc recognizes
# (for example, it might have a different suffix like 'project.so' or
# 'project.so.1', or no suffix like just 'project' for an executable).
# If the project output is a library, you may need to add your 'main.c' file
# here as well.
# [-Ox] represents build optimisations (discussed in the next section).
./emcc [-Ox] project.o -o project.js
emconfigure is called with the normal configure as an argument (in configure-based build systems), and emmake with make as an argument. If your build system uses CMake, replace ./configure
with cmake .
etc. in the above example. If your build system doesn’t use configure or CMake, then you can omit the first step and just run make
(although then you may need to edit the Makefile
manually).
Tip
We recommend you call both emconfigure and emmake scripts in configure- and CMake-based build systems. Whether you actually need to call both tools depends on the build system (some systems will store the environment variables in the configure step, and others will not).
Make generates linked LLVM bitcode. It does not automatically generate JavaScript during linking because all the files must be compiled using the same optimizations and compiler options — and it makes sense to do this in the final conversion from bitcode to JavaScript.
Note
The file output from make might have a different suffix: .a for a static library archive, .so for a shared library, .o for object files (these file extensions are the same as gcc would use for the different types). Irrespective of the file extension, these files contain something that emcc can compile into the final JavaScript + WebAssembly (typically the contents will be wasm object files, but if you build with LTO then they will contain LLVM bitcode).
Note
Some build systems may not properly emit bitcode using the above procedure,
and you may see is not a valid input file
warnings. You can run file
to
check what a file contains (also you can manually check if the contents
start with \0asm
to see if they are wasm object files, or BC
if they
are LLVM bitcode). It is also worth running emmake make VERBOSE=1
which
will print out the commands it runs - you should see emcc being used, and
not the native system compiler. If emcc is not used, you may need to modify
the configure or cmake scripts.
Unless run with certain specific flags (such as -c
, -S
, -r
, or
-shared
) emcc
will run the link phase which can produce more than just
one file. The set of produced files changes depending on the final flags passed
to emcc
and the name of the specified output file. Here is a cheat sheet of
which files are produced under which conditions:
emcc ... -o output.html
builds a output.html
file as an output, as well as an accompanying output.js
launcher file, and a output.wasm
WebAssembly file.
emcc ... -o output.js
omits generating a HTML launcher file (expecting you to provide it yourself if you plan to run in browser), and produces two files, output.js
and output.wasm
. (that can be run in e.g. node.js shell)
emcc ... -o output.wasm
omits generating either JavaScript or HTML launcher file, and produces a single Wasm file built in standalone mode as if the -s STANDALONE_WASM
settting had been used.
emcc ... -o output.{html,js} -s WASM=0
causes the compiler to target JavaScript, and therefore a .wasm
file is not produced.
emcc ... -o output.{html,js} --emit-symbol-map
produces a file output.{html,js}.symbols
if WebAssembly is being targeted (-s WASM=0
not specified), or if JavaScript is being targeted and -Os
, -Oz
or -O2
or higher is specified, but debug level setting is -g1
or lower (i.e. if symbols minification did occur).
emcc ... -o output.{html,js} -s WASM=0 --memory-init-file 1
causes the generation of output.{html,js}.mem
memory initializer file. Pasing -O2
, -Os
or -Oz
also implies --memory-init-file 1
.
emcc ... -o output.{html,js} -g4
generates a source map file output.wasm.map
. If targeting JavaScript with -s WASM=0
, the filename is output.{html,js}.map
.
emcc ... -o output.{html,js} --preload-file xxx
directive generates a preloaded MEMFS filesystem file output.data
.
emcc ... -o output.{html,js} -s WASM={0,1} -s SINGLE_FILE=1
merges JavaScript and WebAssembly code in the single output file output.{html,js}
(in base64) to produce only one file for deployment. (If paired with --preload-file
, the preloaded .data
file still exists as a separate file)
This list is not exhaustive, but illustrates most commonly used combinations.
Note
Regardless of the name of the output file emcc
will always perform
linking and produce a final exectuable, unless a specific flags (e.g. -c
)
direct it do something else. This differs to previous behaviour where
emcc
would default to combining object files (essentially assuming
-r
) unless given a specific executable extension (e.g. .js
or
.html
).
Emscripten performs compiler optimization at two levels: each source file is optimized by LLVM as it is compiled into an object file, and then JavaScript/WebAssembly-specific optimizations are applied when converting object files into the final JavaScript/WebAssembly.
In order to properly optimize code, it is usually best to use the same optimization flags and other compiler options when compiling source to object code, and object code to JavaScript (or HTML).
Consider the examples below:
# Sub-optimal - JavaScript/WebAssembly optimizations are omitted
./emcc -O2 a.cpp -c -o a.o
./emcc -O2 b.cpp -c -o b.o
./emcc a.o b.o -o project.js
# Sub-optimal - LLVM optimizations omitted
./emcc a.cpp -c -o a.o
./emcc b.cpp -c -o b.o
./emcc -O2 a.o b.o -o project.js
# Usually the right thing: The same options are provided at compile and link.
./emcc -O2 a.cpp -c -o a.o
./emcc -O2 b.cpp -c -o b.o
./emcc -O2 a.o b.o -o project.js
However, sometimes you may want slightly different optimizations on certain files:
# Optimize the first file for size, and the rest using `-O2`.
./emcc -Oz a.cpp -c -o a.o
./emcc -O2 b.cpp -c -o b.o
./emcc -O2 a.o b.o -o project.js
Note
Unfortunately each build-system defines its own mechanisms for setting compiler and optimization methods. You will need to work out the correct approach to set the LLVM optimization flags for your system.
Some build systems have a flag like ./configure --enable-optimize
.
You can control whether LLVM optimizations are run using --llvm-opts N
where N is an integer in the range 0-3. Sending -O2 --llvm-opts 0
to emcc during all compilation stages will disable LLVM optimizations but utilize JavaScript optimizations. This can be useful when debugging a build failure.
JavaScript/WebAssembly optimizations are specified in the final step (sometimes called “link”, as that step typically also links together a bunch of files that are all compiled together into one JavaScript/WebAssembly output). For example, to compile with -O1:
# Compile the linked bitcode to JavaScript with -O1 optimizations.
./emcc -O1 project.o -o project.js
Building a project containing debug information requires that debug flags are specified for both the LLVM and JavaScript compilation phases.
To make Clang and LLVM emit debug information in the bitcode files you need to compile the sources with -g (exactly the same as with clang or gcc normally). To get emcc to include the debug information when compiling the bitcode to JavaScript, specify -g or one of the -gN
debug level options.
Note
Each build-system defines its own mechanisms for setting debug flags. To get Clang to emit LLVM debug information, you will need to work out the correct approach for your system.
Some build systems have a flag like ./configure --enable-debug
.
The flags for emitting debug information when compiling from bitcode to JavaScript are specified as an emcc option in the final step:
# Compile the linked bitcode to JavaScript.
# -g or -gN can be used to set the debug level (N)
./emcc -g project.o -o project.js
For more general information, see the topic Debugging.
Built-in support is available for a number of standard libraries: libc, libc++ and SDL. These will automatically be linked when you compile code that uses them (you do not even need to add -lSDL
, but see below for more SDL-specific details).
If your project uses other libraries, for example zlib or glib, you will need to build and link them. The normal approach is to build the libraries to bitcode and then compile library and main program bitcode together to JavaScript.
For example, consider the case where a project “project” uses a library “libstuff”:
# Compile libstuff to bitcode
./emconfigure ./configure
./emmake make
# Compile project to bitcode
./emconfigure ./configure
./emmake make
# Compile the library and code together to HTML
emcc project.o libstuff.a -o final.html
Emscripten Ports is a collection of useful libraries, ported to Emscripten. They reside on github, and have integration support in emcc. When you request that a port be used, emcc will fetch it from the remote server, set it up and build it locally, then link it with your project, add necessary include to your build commands, etc. For example, SDL2 is in ports, and you can request that it be used with -s USE_SDL=2
. For example,
./emcc tests/sdl2glshader.c -s USE_SDL=2 -s LEGACY_GL_EMULATION=1 -o sdl2.html
You should see some notifications about SDL2 being used, and built if it wasn’t previously. You can then view sdl2.html
in your browser.
Note
SDL_image has also been added to ports, use it with -s USE_SDL_IMAGE=2
. To see a list of all available ports, run emcc --show-ports
. For SDL2_image to be useful, you generally need to specify the image formats you are planning on using with e.g. -s SDL2_IMAGE_FORMATS='["bmp","png","xpm"]'
(note: jpg support is not available yet as of Jun 22 2018 - libjpg needs to be added to emscripten-ports). This will also ensure that IMG_Init
works properly when you specify those formats. Alternatively, you can use emcc --use-preload-plugins
and --preload-file
your images, so the browser codecs decode them (see Preloading files). A code path in the SDL2_image port will load through emscripten_get_preloaded_image_data()
, but then your calls to IMG_Init
with those image formats will fail (as while the images will work through preloading, IMG_Init reports no support for those formats, as it doesn’t have support compiled in - in other words, IMG_Init does not report support for formats that only work through preloading).```
Note
SDL_net has also been added to ports, use it with -s USE_SDL_NET=2
. To see a list of all available ports, run emcc --show-ports
.
Note
Emscripten also has support for older SDL1, which is built-in. If you do not specify SDL2 as in the command above, then SDL1 is linked in and the SDL1 include paths are used. SDL1 has support for sdl-config, which is present in system/bin. Using the native sdl-config may result in compilation or missing-symbol errors. You will need to modify the build system to look for files in emscripten/system or emscripten/system/bin in order to use the Emscripten sdl-config.
Note
You can also build a library from ports in a manual way if you prefer
that, but then you will need to also apply the python logic that ports does.
That code (under tools/ports/
) may do things like ensure necessary JS
functions are included in the build, add exports, and so forth. In general,
it’s better to use the ports version as it is what is tested and known to
work.
Adding more ports is fairly easy. Basically, the steps are
Make sure the port is open source and has a suitable license.
Add it to emscripten-ports on github. The ports maintainers can create the repo and add the relevant developers to a team for that repo, so they have write access.
Add a script to handle it under
tools/ports/
(see existing code for examples) and use it intools/ports/__init__.py
.Add testing in the test suite.
Some large projects generate executables and run them in order to generate input for later parts of the build process (for example, a parser may be built and then run on a grammar, which then generates C/C++ code that implements that grammar). This sort of build process causes problems when using Emscripten because you cannot directly run the code you are generating.
The simplest solution is usually to build the project twice: once natively, and once to JavaScript. When the JavaScript build procedure fails because a generated executable is not present, you can then copy that executable from the native build, and continue to build normally. This approach was successfully used for compiling Python (see tests/python/readme.md for more details).
In some cases it makes sense to modify the build scripts so that they build the generated executable natively. For example, this can be done by specifying two compilers in the build scripts, emcc and gcc, and using gcc just for generated executables. However, this can be more complicated than the previous solution because you need to modify the project build scripts, and you may have to work around cases where code is compiled and used both for the final result and for a generated executable.
Emscripten’s goal is to generate the fastest and smallest possible code, and for that reason it focuses on generating a single JavaScript file for an entire project. For that reason, dynamic linking should be avoided when possible.
By default, Emscripten .so
files are the same as regular .o
object files. Dynamic libraries that you specify in the final build stage (when generating JavaScript or HTML) are linked in as static libraries. Emcc ignores commands to dynamically link libraries when linking together bitcode (i.e., not in the final build stage). This is to ensure that the same dynamic library is not linked multiple times in intermediate build stages, which would result in duplicate symbol errors.
There is experimental support for true dynamic libraries, loaded as runtime, either via dlopen or as a shared library. See that link for the details and limitations.
Projects that use configure, cmake, or some other portable configuration method may run checks during the configure phase to verify that the toolchain and paths are set up properly. Emcc tries to get checks to pass where possible, but you may need to disable tests that fail due to a “false negative” (for example, tests that would pass in the final execution environment, but not in the shell during configure).
Tip
Ensure that if a check is disabled, the tested functionality does work. This might involve manually adding commands to the make files using a build system-specific method.
Note
In general configure is not a good match for a cross-compiler like Emscripten. configure is designed to build natively for the local setup, and works hard to find the native build system and the local system headers. With a cross-compiler, you are targeting a different system, and ignoring these headers etc.
Emscripten supports .a archive files, which are bundles of object files. This is a simple format for libraries, that has special semantics - for example, the order of linking matters with .a files, but not with plain object files. For the most part those special semantics should work the same in Emscripten as elsewhere.
The Emscripten Tutorial showed how emcc can be used to compile single files into JavaScript. Emcc can also be used in all the other ways you would expect of gcc:
# Generate a.out.js from C++. Can also take .ll (LLVM assembly) or .bc (LLVM bitcode) as input
./emcc src.cpp
# Generate an object file called src.o.
./emcc src.cpp -c
# Generate result.js containing JavaScript.
./emcc src.cpp -o result.js
# Generate an object file called result.o
./emcc src.cpp -c -o result.o
# Generate a.out.js from two C++ sources.
./emcc src1.cpp src2.cpp
# Generate object files src1.o and src2.o
./emcc src1.cpp src2.cpp -c
# Combine two object files into a.out.js
./emcc src1.o src2.o
# Combine two object files into another object file (not normally needed)
./emcc src1.o src2.o -r -o combined.o
# Combine two object files into library file
./emar rcs libfoo.a src1.o src2.o
In addition to the capabilities it shares with gcc, emcc supports options to optimize code, control what debug information is emitted, generate HTML and other output formats, etc. These options are documented in the emcc tool reference (./emcc --help
on the command line).
Emscripten provides the following preprocessor macros that can be used to identify the compiler version and platform:
The preprocessor define
__EMSCRIPTEN__
is always defined when compiling programs with Emscripten.The preprocessor variables
__EMSCRIPTEN_major__
,__EMSCRIPTEN_minor__
and__EMSCRIPTEN_tiny__
specify, as integers, the currently used Emscripten compiler version.Emscripten behaves like a variant of Unix, so the preprocessor defines
unix
,__unix
and__unix__
are always present when compiling code with Emscripten.Emscripten uses Clang/LLVM as its underlying codegen compiler, so the preprocessor defines
__llvm__
and__clang__
are defined, and the preprocessor defines__clang_major__
,__clang_minor__
and__clang_patchlevel__
indicate the version of Clang that is used.Clang/LLVM is GCC-compatible, so the preprocessor defines
__GNUC__
,__GNUC_MINOR__
and__GNUC_PATCHLEVEL__
are also defined to represent the level of GCC compatibility that Clang/LLVM provides.The preprocessor string
__VERSION__
indicates the GCC compatible version, which is expanded to also show Emscripten version information.Likewise,
__clang_version__
is present and indicates both Emscripten and LLVM version information.Emscripten is a 32-bit platform, so
size_t
is a 32-bit unsigned integer,__POINTER_WIDTH__=32
,__SIZEOF_LONG__=4
and__LONG_MAX__
equals2147483647L
.When targeting SSEx SIMD APIs using one of the command line compiler flags
-msse
,-msse2
,-msse3
,-mssse3
, or-msse4.1
, one or more of the preprocessor flags__SSE__
,__SSE2__
,__SSE3__
,__SSSE3__
,__SSE4_1__
will be present to indicate available support for these instruction sets.If targeting the pthreads multithreading support with the compiler & linker flag
-s USE_PTHREADS=1
, the preprocessor define__EMSCRIPTEN_PTHREADS__
will be present.
Sometimes it can be useful to use a compiler wrapper in order to do things like
ccache
, distcc
or gomacc
. For ccache
the normal method of
simply wrapping the entire compiler should work, e.g. ccache emcc
. For
distributed builds it can be beneficial to run the emscripten driver locally and
distribute only the underlying clang commands. If this is desirable, the
COMPILER_WRAPPER
setting in the config file can be used to add a wrapper
around the internal calls to clang. Like other config settings this can also be
set via an environment variable. e.g:
EM_COMPILER_WRAPPER=gomacc emcc -c hello.c
The Emscripten test suite (tests/runner.py) contains a number of good examples — large C/C++ projects that are built using their normal build systems as described above: freetype, openjpeg, zlib, bullet and poppler.
It is also worth looking at the build scripts in the ammo.js project.
Make sure to use emar
(which calls llvm-ar
), as the system ar
may
not support our object files. emmake
and emconfigure
set the AR
environment variable correctly, but a build system might incorrectly hardcode
ar
.
Similarly, using the system ranlib
instead of emranlib
(which calls
llvm-ranlib
) may lead to problems, like not supporting our object files
and removing the index, leading to
archive has no index; run ranlib to add one
from wasm-ld
. Again, using
emmake
/emconfigure
should avoid this by setting the env var RANLIB,
but a build system might have it hardcoded, or require you to
pass an option.
The compilation error multiply defined symbol
indicates that the project has linked a particular static library multiple times. The project will need to be changed so that the problem library is linked only once.
Note
You can use llvm-nm
to see which symbols are defined in each bitcode file.
One solution is to use the dynamic-linking approach described above. This ensures that libraries are linked only once, in the final build stage.