[llvm] [SLPVectorizer][AMDGPU] Look-through vectorization of AMDGPU intrinsics (PR #175821)

[via llvm-commits]

xgxanq wrote:

> > > 3. `<user subtree>` is profitable, `<operand subtree>` is not profitable for the vectorization.I don't understand if mix the user subtree and operand subtree, the answer can also be not profitable or profitable. I can't understand broke the tree to consider is better.
> 
> 3rd case can be handled by the tree trimming (throttling), but it affects compile time. That's another reason why the 2-independent-trees approach is better.

I still don't fully see why this would be better, but I'll keep learning and working to understand this part better.

https://github.com/llvm/llvm-project/pull/175821