ganler
commented
3 years ago Task: convert LLVM-IR into embedding vectors.
Symbolic level
- Instruction Level:
- map each component from an inst into a triplet<h, r, t> standing for
opcode(head),relation(relation), andtype(tail). (learned from knowledge graph community) - perform unsupervised learning such that |h+r-t| = 0; (means
tis the most important) - now we have a basic triplet embedding for each inst;
- final inst vector = LINEAR TRANSFORMATION OF:
- opcode vector, type vector, and {variable0, 1, 2, ...} (all values from the triplets)
- map each component from an inst into a triplet<h, r, t> standing for
Flow-aware (w/o GNN structure)
- Intra-BB Level:
- opcode vector, type vector, and {variable0, 1, 2, ...} (all values from the triplets)
In addition, add instruction embedding according to USE-DEF chain;
- Inter-BB level (CFG): Still UD chain.
Result
Faster (training time) & higher acc.
Key design choices
- Model relations using types and UD chain;
- NLP & Knowledge graph dark magic: |h+r-t| = 0 triplets which makes the embedding distributed and continuous;
https://dl.acm.org/doi/pdf/10.1145/3418463