At the meeting yesterday between @jensdietrich, @mmabdpr and myself, the topic came up that compiled bytecode can be normalised in various ways. E.g., the constant pool could be sorted lexicographically (and references to it renumbered accordingly). This will increase the proportion of semantically equivalent compiled classes that are bitwise equivalent -- which is useful for the purposes of someone actually wanting to determine whether two bytecode sequences are semantically equivalent, but it also throws away some natural variability.
Should we do this in our oracle datasets? If we do, it seems that we would then be obliged to filter out identical results, which would reduce dataset size.
Alternatively, should we keep the original non-normalised bytecode sequences, but record their normalised forms (or summaries thereof, e.g., hashes) in additional metadata fields? This would allow future users of the data to come up with their own normalisation schemes, informed by but not restricted to our own.
At the meeting yesterday between @jensdietrich, @mmabdpr and myself, the topic came up that compiled bytecode can be normalised in various ways. E.g., the constant pool could be sorted lexicographically (and references to it renumbered accordingly). This will increase the proportion of semantically equivalent compiled classes that are bitwise equivalent -- which is useful for the purposes of someone actually wanting to determine whether two bytecode sequences are semantically equivalent, but it also throws away some natural variability.
Should we do this in our oracle datasets? If we do, it seems that we would then be obliged to filter out identical results, which would reduce dataset size.
Alternatively, should we keep the original non-normalised bytecode sequences, but record their normalised forms (or summaries thereof, e.g., hashes) in additional metadata fields? This would allow future users of the data to come up with their own normalisation schemes, informed by but not restricted to our own.