The current training data generator applies random pass sequences to valid bitcode. While the sequences are syntactically correct , they may induce invalid IR states or violate implicit pass dependencies that the LLVM Pass Manager would normally reject or assert on.
Generating data from invalid IR states introduces noise and "poisoned" samples into the ML dataset, as the model may learn to optimize broken code or associate metrics with effective crashes.
Goal: Ensure that only semantically valid, verified optimization sequences contribute to the training dataset.
The current training data generator applies random pass sequences to valid bitcode. While the sequences are syntactically correct , they may induce invalid IR states or violate implicit pass dependencies that the LLVM Pass Manager would normally reject or assert on.
Generating data from invalid IR states introduces noise and "poisoned" samples into the ML dataset, as the model may learn to optimize broken code or associate metrics with effective crashes.
Goal: Ensure that only semantically valid, verified optimization sequences contribute to the training dataset.