[llvm] [AMDGPU] Allocate i1 argument to SGPRs (PR #72461)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> @arsenm In the following case where the func returns "zeroext i1",
> 
> ```
> define zeroext i1 @i1_zeroext_func_void() {
>   %val = load i1, ptr addrspace(1) undef
>   ret i1 %val
> }
> ```
> 
> because of the zeroext, the return type is updated to i32 and return val allocated to a VGPR. However, on the caller side, the return val is still obtained from a SGPR. How do you want the behavior to be, using VGPR or SGPR?

I expect this to be the boolean SGPR, and ignore the zeroext. Is inreg allowed on return values? If so, I would expect the promote-to-32-bit VGPR case 

https://github.com/llvm/llvm-project/pull/72461