[llvm] [WIP][AMDGPU] combine uniform AMDGPU lane Intrinsics (PR #116953)

[Sameer Sahasrabuddhe via llvm-commits]

ssahasra wrote:

> ```
>   if (ballot_all(x)) {
>     speculatable_use(x)
>      ...
>   }
> ```
> 
> Which may then later be incorrectly hoisted:
> 
> ```
>   speculatable_use(x) // x is not actually uniform anymore
>   if (ballot_all(x)) { ... }
> ```
> 
> There needs to be something tying the uniformity to the use position

Actually there is more going on here than just a lack of `convergent`. `ballot_all(x)` is a uniform condition. It will allow convergent operations to be hoisted over it. So if `speculatable_use(x)` really does depend on `x` being uniform, then it `convergent` is not enough to stop it. But should this use have `speculatable` in the first place? By explicitly putting the `speculatable` attribute, the user is saying that the compiler is free to move hoist this use wherever other conditions are met. Perhaps they shouldn't be using `speculatable` in the absence of nuance about uniformity?

https://github.com/llvm/llvm-project/pull/116953