I want to know the specific principle of parsing ramp tags into LLVM IR during vectorization. Is there any paper I can read.