[llvm] [SPIR-V] Type inference must realize that a <1 x Type> vector type is not a legal vector type in LLT (PR #124560)

[Dmitry Sidorov via llvm-commits]

https://github.com/MrSidims approved this pull request.

It's a good started, thanks!
There are several other issues to fix:
1. Apart of insert/extract instructions we should handle `shufflevector` instruction, please consider it fixing within this patch;
2. Apart of insert/extract instructions is legal (though not recommended) to use GEP on vectors. In practice with the switch to opaque pointers it should never happen, but theoretically there can be a GEP(0, 0) to such vector followed by load. In this case backend should remove zero-GEP and replace load with the scalar register allocation.
3. This is a grey area, but I do believe, that if a function declaration with external linkage has <1 x ...> parameter type, then instead of lowering this type to scalar the backend should emit an error. The reason is the following: lets imagine we have some language builtin which implementation we expect to be linked way after the SPIR-V backend (lets say after SPIR-V was consumed by llvm-spirv emitting LLVM IR). Compiler's library if written on LLVM IR then can expect parameter type to be a vector, while it's not a vector anymore, resulting in an error, which is harder to debug, then an error, that SPIR-V backend would emit.

https://github.com/llvm/llvm-project/pull/124560