[llvm] [AMDGPU][LSV] Restrict large vectors in graphics (PR #92540)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> > > We definitely should not be using the OS as a hint on usage or to change any properties like this
> > 
> > 
> > Fine - would adding a flag be better?
> 
> Not really. It would be better to just unconditionally change this to be 128 for any of the global-like address spaces. Ultimately this hook is supposed to be a "give me this property", and the constant address spaces are a weak heuristic for using scalar instructions. The problem would be in the use context for this

Plus I think we should have better support for split and partial rematerialized of the large SMRD loads 

https://github.com/llvm/llvm-project/pull/92540