Why does Emsctipten use wasm-opt given the great optimization from LLVM?

[Shawdox]

Since Emscripten uses clang to generate object/wasm files at compile time (e.g. C) and uses wasm-opt to optimize the code at linking time, and since both clang and wasm-opt are given the same optimization level (e.g. O2), why do we still need to use wasm-opt to optimize the Wasm binary after optimizing on LLVM-IR?

1. Is there some optimization that can only be done on Wasm binary?
2. In addition to Wasm specificed passes, wasm-opt also provides some general passes, are these general passes not possible with LLVM's optimization?
3. For the same optimization level, which passes are used by clang and wasm-opt, and do any of these passes have the same functionality?

[kripken]

One major reason is LTO-style optimizations like inlining between compilation units. clang only does that when LTO is enabled, but wasm has enough information for it, so wasm-opt can always do it.

Otherwise, there is a lot of overlap between the two optimizers in basic passes like DCE etc. clang can do many things wasm-opt can't, like type-based aliasing, but to get the most out of it you need LLVM LTO, which can be slow. OTOH wasm-opt does have some wasm-specific optimizations that clang doesn't do yet (like sorting globals, packing segments, the if structure, etc., things that are easier to do directly on wasm).

(wasm-opt also has optimizations for wasm things that LLVM doesn't emit, like wasm GC. that can matter if you link wasm from clang with a GC language, for example)