[llvm] [WebAssembly] Split separate component LiveIntervals for TEEs (PR #131561)

[Quentin Colombet via llvm-commits]

qcolombet wrote:

> @qcolombet I think `if (LIS->shrinkToUses(LI, Dead))` condition in
> 
> https://github.com/llvm/llvm-project/blob/9d92d4b01ce0ddfdcbb2bd5ec0ec3208ad480ccc/llvm/lib/CodeGen/RegisterCoalescer.cpp#L341-L353
> 
> 
> was first written in [0ddd315](https://github.com/llvm/llvm-project/commit/0ddd315db0990d11e9f608a86b489b8da8d8c4f6) authored by you. It's a long time ago, but in case you remember the context, could you explain what this condition means, and why it does not return true in some cases when there are unconnected components (as described in the test case of this PR)?

`shrinkToUses` is supposed to return `true` when there are unconnected components as described in the comments of this method (https://github.com/llvm/llvm-project/blob/77648476f28c4ee50eb2ffca952453a210386d7d/llvm/include/llvm/CodeGen/LiveIntervals.h#L186)

If this is not the case, then there may be a bug in `computeDeadValues`.
In any case, I don't believe your patch is the correct fix. Although this will fix the problem, this is a big hammer because this has a compile time cost that we shouldn't have to pay in most cases (hence why the code was only running when the method returns true.)

> And it turns out our TEE generation in RegStackify can create virtual registers with LiveInterval with multiple disconnected segments.

I'm wondering if you're trying to apply the same recipe as the register coalescer on something that doesn't follow the same rules.

Could you check (e.g., by inserting calls to the verifier at different stages of the reg stackifier) whether your live intervals are "correct" when you enter into `shrinkToUsers`.

https://github.com/llvm/llvm-project/pull/131561