this seems like a strange assertion to me. why not? you can do amazing things with static analysis and runtime optimization, and that operates on the AST / some intermediate state, not the source code. if anything, higher-level languages (ie, JS as opposed to bytecode) lend themselves to more automatic-optimizations than lower-level ones because intent is encoded better.
Asm.js is lower level than LLVM IR. In LLVM IR the data structures and memory allocation are still visible, in asm.js it's just opaque code working on a giant heap array. For example think how you'd do pointer analysis on asm.js...
The optimizations are already done ahead of time, by Emscripten. There is no need to do them on the client, and doing so just increases compilation time for no reason.
In fact, shipping the types to the client is harmful, because it increases the download size. This is one of the ways asm.js is, IMHO, a better bytecode than LLVM IR…
