I doubt you get the same performance as direct compilation. LLVM ir carries lots of annotations useful to optimizing passes, e.g. aliasing information.
The twist is, you can produce wasm from LLVM, including running the full mid-level optimizer first, which is the part of LLVM where those annotations and aliasing information are most valuable.
If this turns out to be useful then what's stopping us adding similar annotations in a custom section for Webassembly?