Okay, so after replacing all Hash{Map,Set}s in cli-support with BTree, the problem persists. So it is not caused by any of those data structures (at least, directly).

And that was all Hash{Map,Set}s in the entire repo replaced. The issue persists. So it is not caused by any of the hashmaps in this repo.

The issue might step from external deps or cfgs. Not sure.

Alright, I analyzed the code and narrow it down to all changes in raw.rs being caused by the JS code for Intrisic::ObjectDropRef and Intrisic::Throw being generated in the "wrong" order. (Well, order doesn't matter, it just needs to be the same order everywhere.)

To be specific:

• Intrisic::ObjectDropRef causes Instruction::ExternrefLoadOwned, which in turn generates all globals from const heap to function takeObject(idx).
• Intrisic::Throw causes Instruction::MemoryToString, which in turn generates all globals from const lTextDecoder to function getStringFromWasm0(ptr, len).

This explains why the order of export function __wbindgen_object_drop_ref and export function __wbindgen_throw correlates with the order of the globals I mentioned above.

Thank you for looking into this, it has been terribly annoying so far.
My observation so far is that it is actually deterministic (in the CI), as long as nothing changes.
This is still annoying because every PR has to adjust the raw.js file.

Closed in favor of #4190