There are significant shortcomings in estimating the codegen for an IR sequence and the resulting cost of running that sequence on a target architecture/machine. This is particularly attributable to the abstraction between the IR and the codegen layer which means that not all the information from backend can be retrieved easily at the IR level or this may be computationally intensive. We aim to highlight some of the cost-modelling issues on which we have worked on and our efforts to improve the situation in the context of the AArch64 platform.