It appears that we're creating a new PE/ELFFILE object per function, when we should be creating these once per analysis. Commenting out extract_insn_bytes_features (the caller) resulted in a significant performance improvement. @mr-tz this is a good place to start investigating for a performance bump, starting by seeing if we can store the PE/ELFFile object once in the AnalysisContext.

please read thread started by @mike-hunhoff in #1950 (comment)_