[llvm] [AARCH64][CostModel] - Amortize sdiv costs to allow for loops with predicated sdivs to vectorize (PR #201296)

[David Sherwood via llvm-commits]

================
@@ -4594,17 +4594,7 @@ InstructionCost AArch64TTIImpl::getArithmeticInstrCost(
             return (4 + DivCost) * VTy->getNumElements();
           }
         }
-        // On AArch64, without SVE, vector divisions are expanded
----------------
david-arm wrote:

I think the AArch64 cost model changes should probably live in a different PR to the vplan cost model changes. At the moment it's hard to tell which test changes are due to which bits. It will also reduce the risk of either PR introducing regressions. Conceptually, these two changes are independent I think?

https://github.com/llvm/llvm-project/pull/201296