With today's asmjs and wasm targets, we optimize the LLVM IR ourselves, and pass that IR to emcc. emcc has its own optimization options though, including but not limited to LLVM's optimization passes, so we should probably leave all optimizations to emcc. Likewise LTO should be delegated to emscripten.

So we need to change the backend to treat these targets as unoptimized in LLVM, and instead pass the optimization arguments to emcc during the "link" step. This may involve adding new features to the target spec, or possibly just special casing emscripten in the backend.