[llvm-branch-commits] [llvm] [AMDGPU] Retire dying uses before adding defs in GCNDownwardRPTracker (PR #228294)

[Dhruva Chakrabarti via llvm-branch-commits]

================
@@ -5115,10 +5115,10 @@ define amdgpu_kernel void @constant_zextload_v32i32_to_v32i64(ptr addrspace(1) %
 ; GFX7-HSA-NEXT:    v_mov_b32_e32 v8, s0
----------------
dhruvachak wrote:

Agreed, this function regressed slightly. Register use and instruction counts are unchanged. Number of s_nop increased from 7 to 13. Appears to be a side-effect of this patch, not a direct consequence. 

Before the patch, the estimated RP was higher by 2 which pushed down the computed occupancy from 10 to 9. This led to an extra scheduling step which happened to lead to fewer s_nops. After the patch, occupancy was met and the unclustered schedule stage was not run but this led the hazard recognizer to insert 7 more s_nops.

https://github.com/llvm/llvm-project/pull/228294