Home
» Optimizing Code
Generally you should first compile and run your code without optimizations (the default). Once you are sure that the code runs correctly, you can use the techniques in this article to make it load and run faster.
Code is optimized by specifying optimization flags when running emcc. The levels include: -O0 (no optimization), -O1, -O2, -Os, -Oz, and -O3.
For example, to compile with optimization level -O2
:
emcc -O2 file.cpp
The higher optimization levels introduce progressively more aggressive optimization, resulting in improved performance and code size at the cost of increased compilation time. The levels can also highlight different issues related to undefined behavior in code.
The optimization level you should use depends mostly on the current stage of development:
When first porting code, run emcc on your code using the default settings (without optimization). Check that your code works and debug and fix any issues before continuing.
Build with lower optimization levels during development for a shorter compile/test iteration cycle (-O0
or -O1
).
Build with -O2
to get a well-optimized build.
Building with -O3
or -Os
can produce an ever better build than -O2
, and are worth considering for release builds. -O3
builds are even more optimized than -O2
, but at the cost of significantly longer compilation time and potentially larger code size. -Os
is similar in increasing compile times, but focuses on reducing code size while doing additional optimization. It’s worth trying these different optimization options to see what works best for your application.
Other optimizations are discussed in the following sections.
Note
The meanings of the emcc optimization flags (-O1, -O2
etc.) are similar to gcc, clang, and other compilers, but also different because optimizing WebAssembly includes some additional types of optimizations. The mapping of the emcc levels to the LLVM bitcode optimization levels is documented in the reference.
Compiling source files to object files works as you’d expect in a native build system that uses clang and LLVM. When linking object files to the final executable, Emscripten does additional optimizations as well depending on the optimization level:
For wasm, the Binaryen optimizer is run. Binaryen does both general-purpose optimizations to the wasm that LLVM does not, and also does some whole-program optimization. (Note that Binaryen’s whole-program optimizations may do things like inlining, which can be surprising in some cases as LLVM IR attributes like noinline
have been lost at this point.)
JavaScript is generated at this phase, and is optimized by Emscripten’s JS optimizer. Optionally you can also run the closure compiler, which is highly recommended for code size.
Emscripten also optimizes the combined wasm+JS, by minifying imports and exports between them, and by running meta-dce which removes unused code in cycles that span the two worlds.
To skip extra optimization work at link time, link with -O0
or -O1
. In
those modes Emscripten focuses on faster iteration times. (Note that it is ok
to link with those flags even if the source files were compiled with a different
optimization level.)
To also skip non-optimization work at link time, link with -s WASM_BIGINT
.
Enabling BigInt support removes the need for Emscripten to “legalize” the wasm
to handle i64
values on the JS/Wasm boundary (as with BigInts i64
values
are legal, and require no extra processing).
Some link flags add additional work at the link stage that can slow things down.
For example -g
enables DWARF support, flags like -s SAFE_HEAP
will require
JS post-processing, and flags like -s ASYNCIFY
will require wasm
post-processing. To ensure your flags allow the fastest possible link, in which
the wasm is not modified after wasm-ld
, build with
-s ERROR_ON_WASM_CHANGES_AFTER_LINK
. With that option you will get an error
during link if Emscripten must perform changes to the Wasm. For example, if you
didn’t pass -s WASM_BIGINT
then it will tell you that legalization forces
it to change the Wasm. You will also get an error if you build with -O2
or
above, as the Binaryen optimizer would normally be run.
There are several flags you can pass to the compiler to affect code generation, which will also affect performance — for example DISABLE_EXCEPTION_CATCHING. These are documented in src/settings.js.
Emscripten will emit WebAssembly by default. You can switch that off with -s WASM=0
(in which case emscripten emit JavaScript), which is necessary if you want the output to run in places where wasm support is not present yet, but the downside is larger and slower code.
This section describes optimisations and issues that are relevant to code size. They are useful both for small projects or libraries where you want the smallest footprint you can achieve, and in large projects where the sheer size may cause issues (like slow startup speed) that you want to avoid.
You may wish to build the less performance-sensitive source files in your project using -Os or -Oz and the remainder using -O2 (-Os and -Oz are similar to -O2, but reduce code size at the expense of performance. -Oz reduces code size more than -Os.)
Separately, you can do the final link/build command with -Os
or -Oz
to make the compiler focus more on code size when generating WebAssembly module.
In addition to the above, the following tips can help to reduce code size:
Use the closure compiler on the non-compiled code: --closure 1
. This can hugely reduce the size of the support JavaScript code, and is highly recommended. However, if you add your own additional JavaScript code (in a --pre-js
, for example) then you need to make sure it uses closure annotations properly.
Floh’s blogpost on this topic is very helpful.
Make sure to use gzip compression on your webserver, which all browsers now support.
The following compiler settings can help (see src/settings.js
for more details):
Disable inlining when possible, using -s INLINING_LIMIT=1
. Compiling with -Os or -Oz generally avoids inlining too. (Inlining can make code faster, though, so use this carefully.)
You can use the -s FILESYSTEM=0
option to disable bundling of filesystem support code (the compiler should optimize it out if not used, but may not always succeed). This can be useful if you are building a pure computational library, for example.
The ENVIRONMENT
flag lets you specify that the output will only run on the web, or only run in node.js, etc. This prevents the compiler from emitting code to support all possible runtime environments, saving ~2KB.
Link Time Optimization (LTO) lets the compiler do more optimizations, as it can
inline across separate compilation units, and even with system libraries.
LTO is enabled by compiling objects files with -flto
. The effect of this
flag is to emit LTO object files (technically this means emitting bitcode). The
linker can handle a mix wasm object files and LTO object files. Passing
-flto
at link time will also trigger LTO system libraries to be used.
Thus, to allow maximal LTO opportunities with the LLVM wasm backend, build all
source files with -flto
and also link with flto
.
The previous section on reducing code size can be helpful on very large codebases. In addition, here are some other topics that might be useful.
If you hit memory limits in browsers, it can help to run your project by itself, as opposed to inside a web page containing other content. If you open a new web page (as a new tab, or a new window) that contains just your project, then you have the best chance at avoiding memory fragmentation issues.
Catching C++ exceptions (specifically, emitting catch blocks) is turned off by default in -O1
(and above). Due to how WebAssembly currently implement exceptions, this makes the code much smaller and faster (eventually, wasm should gain native support for exceptions, and not have this issue).
To re-enable exceptions in optimized code, run emcc with -s DISABLE_EXCEPTION_CATCHING=0
(see src/settings.js).
Note
When exception catching is disabled, a thrown exception terminates the application. In other words, an exception is still thrown, but it isn’t caught.
Note
Even with catch blocks not being emitted, there is some code size overhead unless you build your source files with -fno-exceptions
, which will omit all exceptions support code (for example, it will avoid creating proper C++ exception objects in errors in std::vector, and just abort the application if they occur)
C++ run-time type info support (dynamic casts, etc.) adds overhead that is sometimes not needed. For example, in Box2D neither rtti nor exceptions are needed, and if you build the source files with -fno-rtti -fno-exceptions
then it shrinks the output by 15% (!).
Building with -s ALLOW_MEMORY_GROWTH=1
allows the total amount of memory used to change depending on the demands of the application. This is useful for apps that don’t know ahead of time how much they will need.
Enable Debug mode (EMCC_DEBUG) to output files for each compilation phase, including the main optimization operations.
A few UNSAFE optimizations you might want to try are:
--closure 1
: This can help with reducing the size of the non-generated (support/glue) JS code, and with startup. However it can break if you do not do proper Closure Compiler annotations and exports. But it’s worth it!
Modern browsers have JavaScript profilers that can help find the slower parts in your code. As each browser’s profiler has limitations, profiling in multiple browsers is highly recommended.
To ensure that compiled code contains enough information for profiling, build your project with profiling as well as optimization and other flags:
emcc -O2 --profiling file.cpp
Emscripten-compiled code can currently achieve approximately half the speed of a native build. If the performance is significantly poorer than expected, you can also run through the additional troubleshooting steps below:
Building Projects is a two-stage process: compiling source code files to LLVM and generating JavaScript from LLVM. Did you build using the same optimization values in both steps (-O2
or -O3
)?
Test on multiple browsers. If performance is acceptable on one browser and significantly poorer on another, then file a bug report, noting the problem browser and other relevant information.