MLIR's design prioritizes generality and extensibility, offering highly generic data structures that accommodate a vast set of compiler abstractions. While this flexibility is powerful to share and mix abstractions, it comes at the cost of performance. In contrast, systems like LuaJIT achieve exceptional speed, partly due to their use of highly specialized, compact data structures tailored for their specific abstractions. In this talk, we will explore a few experiments where we reimplement MLIR data structures with some of its functionality stripped, to improve its performance on specific compiler pipelines. Examples of features include preventing operations from having more than one result, stripping operands from their types, or removing parts of the def-use chain. Our experiments suggest that we could potentially provide a better trade-off between speed and generality by tailoring MLIR compilers data structures to the dialects they support.