Code-mix is widespread in domains like sentiment classification, polarity identification, dialect identification, question answering, part-of-speech tagging, named entity recognition, speech technologies, and other emerging domains. Studying the impact of pre-processing steps has been limited to the monolinguals. The emerging contributions in various applications for code-mix data lack such a dedicated study. Our study fills the gap. In this paper, we conduct an empirical study involving 512 combinations of pre-processing steps with 7 different tokenizers applied to 3 multi-domain datasets. We have investigated the impact of these pre-processing steps and tokenizers on the performance of three conventional neural networks and five machine-learning techniques. Our study includes a detailed analysis of 3,44,064 scores. Our study shows mean differences of 0.3276, 0.2178, 0.1999, and 0.1467 over the Matthews Correlation Coefficient, Macro-F1, Weighted-F1, and Accuracy, respectively , due to the choice of pre-processing steps. We propose to choose the best-performing set of pre-processing steps for good performance.