[llvm] [AMDGPU] Added hot-block-rematerialize pass (PR #136631)

[Adam Yang via llvm-commits]

adam-yang wrote:

> Is there any data supporting this improves performance? How much does it hurt and where?

For our hw's, the default, non-aggressive SGPR code path (which is included in this change) is enabled by default. If we turn it off, across our retail shader test set, occupancy goes down across the board. Among the shaders that showed any change in performance, occupancy lowered by ~1 on average.

For real timing measurement on PIX captures, the total frame time of all our performance suite goes down by 1% with the pass turned off.

https://github.com/llvm/llvm-project/pull/136631