[llvm] [AMDGPU] Use new CSR cost calculation (PR #219220)

[Lukas Sommer via llvm-commits]

================
@@ -6,66 +6,117 @@
 ; by introducing a copy to AGPR register. The VGPR store to AGPR (v_accvgpr_write_b32) and later the
 ; restore from AGPR (v_accvgpr_read_b32) should be whole-wave operations and hence exec mask should be
 ; manipulated to ensure all lanes are active when these instructions are executed.
+; The profiled loop makes stack spills sufficiently expensive to select this split
+; with the default CSR cost model.
 define void @vector_reg_liverange_split() #0 {
----------------
sommerlukas wrote:

Yes, for this test to still serve its purpose I had to change the test input. With the new normalized cost model, an unprofiled single call is too cold to trigger this split. That is intended: the allocator should choose the split only when the repeated spill cost outweighs the first CSR use. I've clarified the test comment.

https://github.com/llvm/llvm-project/pull/219220