[llvm] [WebAssembly] Fix lowering of (extending) loads from addrspace(1) globals (PR #155937)

[Derek Schuff via llvm-commits]

dschuff wrote:

Thanks for the ping,  I did actually take a look at it, and in particular the issue with the simd tests (since as you both observed, the load+extend is not getting turned into load_s). It's not quite obvious to me yet what is happening since I would have thought that if this custom lowering declines to actually lower the instruction, that it would fall back on the default lowering path including all of the DAG combines. But obviously that's not what's happening. I didn't get as far as trying to figure out exactly which code is running in each case and why.

https://github.com/llvm/llvm-project/pull/155937