asm.js is a subset of JavaScript designed to allow computer software written in languages such as C to be run as web applications while maintaining performance characteristics considerably better than standard JavaScript, which is the typical language used for such applications.
asm.js consists of a strict subset of JavaScript, to which code written in statically typed languages with manual memory management (such as C) is translated by a source-to-source compiler such as Emscripten (based on LLVM). Performance is improved by limiting language features to those amenable to ahead-of-time optimization and other performance improvements.
Mozilla Firefox was the first web browser to implement asm.js-specific optimizations, starting with version 22.
asm.js is superseded by WebAssembly. See below.
Design
asm.js enables significant performance improvements for web applications, but does not aim to improve the performance of hand-written JavaScript code, nor does it enable anything other than enhanced performance.
It is intended to have performance characteristics closer to that of native code than standard JavaScript by limiting language features to those amenable to ahead-of-time optimization and other performance improvements. By using a subset of JavaScript, asm.js is largely supported by all major web browsers, unlike alternative approaches such as Google Native Client.
Code generation
asm.js is not typically written directly: instead, as an intermediate language, it is generated through the use of a compiler that takes source code in a language such as C++ and outputs asm.js.
For example, given the following C code:
int f(int i) {
return i + 1; }
Emscripten would output the following JS code: function f(i) { i = i|0; return (i + 1)|0; }
Note the addition of |0 and the lack of type specifiers. In JavaScript, bitwise operators convert their operands to 32-bit signed integers and give integer results. This means that a bitwise OR with zero converts a value to an integer (a very simple "conceptual" presentation of bitwise operators may not deal with type conversion at all, but every programming language defines operators for its own convenience, as Javascript does here). By doing this for each parameter, this ensures that if the function is called from outside code, the value will be converted to the correct type. This is also used on the return value, in this case to ensure that the result of adding 1 to i will be an integer (as otherwise it could become too large), and to mark the return type of the function. These conversions are required by asm.js, so that an optimising compiler can produce highly efficient native code ahead-of-time. In such an optimising compiler, no conversions are performed when asm.js code calls other asm.js code, as the required type specifiers mean it is guaranteed that values will already have the correct type. Furthermore, rather than performing a floating-point addition and converting to an integer, it can simply do a native integer operation. Together, this leads to significant performance benefits.
Here is another example to calculate the length of a string:
size_t strlen(char *ptr) {
char *curr = ptr; while (*curr != 0) { curr++; } return (curr - ptr); }
This would result in the following asm.js code:
function strlen(ptr) {
ptr = ptr|0; var curr = 0; curr = ptr; while ((MEM8[curr>>0]|0) != 0) { curr = (curr + 1)|0; } return (curr - ptr)|0; }
In the generated code, the variable MEM8 is actually a byte-by-byte "view" of a typed buffer, which serves as the "heap" of the asm.js code.
Performance
Since asm.js runs in a browser, the performance heavily depends on both the browser and hardware. Preliminary benchmarks of C programs compiled to asm.js are usually within a factor of 2 slower than native compilation with Clang.
Much of this performance gain over normal JavaScript is due to 100% type consistency and virtually no garbage collection (memory is manually managed in a large typed array). This simpler model with no dynamic behavior, no memory allocation or deallocation, just a narrow set of well-defined integer and floating point operations enables much greater performance and potential for optimization.