[llvm] [AMDGPU] Add intrinsic readanylane (PR #115696)

[Jay Foad via llvm-commits]

jayfoad wrote:

Thanks for working on this!

> It doesn't convergent, so can be moved between control flows.

Unfortunately the new intrinsic still has to be `convergent`.

Consider these transformations:
```
A: x = def();                            B: x = def();
   y = read*lane(x);                        if (divergentcondition) {
   if (divergentcondition) {                  y = read*lane(x);
     use(y);                                  use(y);
   }                                        }
```
Transforming A-->B is not safe for `readfirstlane`, but it is safe for `readanylane` (because if x is uniform outside the `if` then it must also be uniform inside the `if`). Transforming B-->A is not safe for `readfirstlane` and not safe for `readanylane`.

Unfortunately the `convergent` attribute currently restricts both of these transforms, so we need `readanylane` to be convergent to prevent B-->A.

Maybe in future we can have some better way to express this which will allow transforming A-->B but prevent B-->A.

Cc: @nhaehnle @ssahasra

https://github.com/llvm/llvm-project/pull/115696