[llvm] [CodeGen] Renumber slot indexes before register allocation (PR #66334)

[Carl Ritson via llvm-commits]

perlfu wrote:

This seems reasonable to me.

For AMDGPU graphics, this seems like it yields a very minor reduction in mean VGPR usage, and slightly greater reduction in mean SGPR usage.  There are some a handful of edges case where VGPR usage increased significantly (20 VGPRs).  I guess this is unsurprising given it tweaks a heuristic input.


https://github.com/llvm/llvm-project/pull/66334