[llvm] [AMDGPU] Fix LDS allocation granularity in occupancy computation for RDNA 2 (PR #219495)

[Frederik Harwath via llvm-commits]

================

----------------
frederik-h wrote:

> These has*(ST) predicates should never be used, directly check the feature bit.

Ok.

> It would also be better to properly track this instead of basing this on the only happens to be related instruction feature

I completely agree, but this is the preexisting way to handle the LDS granularity. Should we perhaps introduce separate features for this? 


https://github.com/llvm/llvm-project/pull/219495