[llvm] [LoopUnroll] Consider simplified operands while retrieving TTI instruction cost (PR #70929)

[Sergey Kachkov via llvm-commits]

skachkov-sc wrote:

Sorry, missed one test that is affected by this change (didn't have AMDGPU backend in local build). @arsenm , could you please review this change? It looks like the cost of such GEPs:
```
%arrayidx.out = getelementptr inbounds float, ptr addrspace(3) %out, i32 %indvars.iv
```
is considered as zero, so it unrolls loop in test_func_addrspacecast_cost_nonfree.

https://github.com/llvm/llvm-project/pull/70929