[llvm] [AMDGPU] Fix LSR cost comparator that regressed GFX9+ occupancy (PR #204344)

[Yuyang Zhang via llvm-commits]

================
@@ -1858,21 +1858,20 @@ InstructionCost GCNTTIImpl::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
 
 bool GCNTTIImpl::isLSRCostLess(const TTI::LSRCost &A,
                                const TTI::LSRCost &B) const {
-  // Favor lower per-iteration work over preheader/setup costs.
-  // AMDGPU lacks rich addressing modes, so ScaleCost is folded into the
-  // effective instruction count (base+scale*index requires a separate ADD).
+  // GFX9+ occupancy is VGPR-bound, so register pressure must dominate the LSR
----------------
yuyzhang512 wrote:

Done.

https://github.com/llvm/llvm-project/pull/204344