How would setjmp/longjmp be implemented in WebAssembly?

Question

I've just started looking at WebAssembly MVP, and notice that there is no access to the stack and stack pointer, or indeed any structured exception handling support (throw/catch) etc.

Given it is supposed to be a C compilation target, it surely must be possible to implement setjmp and longjmp, but I can't fathom how this can be done nicely. How does this construct look in wast?

Answer 1

Zero-cost exception handling isn't supported in the WebAssembly MVP.

C++ exception handling and setjmp/longjmp are currently implemented through Emscripten by having each try or "invoke" perform a call to JavaScript with a continuation to WebAssembly's C++ code. A throw then throws a JavaScript exception, which unwinds the stack and handles the "landingpad" where unwinding code lives (usually destructor calls and catch blocks). This means that every continuation receives a boolean value: exception path or regular path.

This is super expensive! If LLVM can't prove that a function call can't throw then its IR contains an invoke instruction, and Emscripten relies on this to insert exception handling code. The default in C++ is that anything can throw, so if you look at LLVM IR there are invokes all over the place when you compile with exceptions.

Zero-cost exception handling is being worked on at the moment, so this situation should resolve eventually. This will be used to implement setjmp/longjmp. This can enable all of the defined behavior of setjmp/longjmp, namely unwinding the stack without calling C++ destructors. It does not, however, allow the undefined behavior case of jumping forward to a stack that was already unwound which is sometimes used to implement coroutines.