It identifies similarities between functions, and if profitable, merges them into a single function, replacing the original ones. Functions do not need to be identical to be merged. In fact, there is very little restriction to merge two function, however, the produced merged function can be larger than the two original functions together. For that reason, it uses the TargetTransformInfo analysis to estimate the code-size costs of instructions in order to estimate the profitability of merging two functions.
This function merging transformation has three major parts:
- The input functions are linearized, representing their CFGs as sequences of labels and instructions.
- We apply a sequence alignment algorithm, namely, the Needleman-Wunsch algorithm, to identify similar code between the two linearized functions.
- We use the aligned sequences to perform code generate, producing the new merged function, using an extra parameter to represent the function identifier.
This pass integrates the function merging transformation with an exploration framework. For every function, the other functions are ranked based their degree of similarity, which is computed from the functions' fingerprints. Only the top candidates are analyzed in a greedy manner and if one of them produces a profitable result, the merged function is taken.
Rodrigo C. O. Rocha, Pavlos Petoumenos, Zheng Wang, Murray Cole, Hugh Leather; "Function Merging by Sequence Alignment". International Symposium on Code Generation and Optimization, 2019 (CGO'19).