reformat_arrow.py

"""Re-format the idioms dataset to use its non-compact form in a json format friendly for uptake by pyarrow.

Also handy: 

for BIN in "--" "--binary"; for OPT in O0 O{1,2,3}_noinline; echo $OPT $BIN; python reformat_arrow.py $BIN ~/Projects/idioms-data/published/idioms_dataset_{$OPT}_opt_parity; end; end

for opt in O0 O{1,2,3}_noinline; for t in binary function; for split in train test valid; echo $opt $t $split; set dir ~/Projects/idioms-realtype/by-{$t}-hex-rays-parity-{$opt}; mkdir -p "$dir"; cat arrow-idioms_dataset_{$opt}_opt_parity-{$t}.json | pv | jq -c .$split"[]" | zstd > "$dir/$split.jsonl.zst"; end; end; end
"""

from pathlib import Path
import argparse
import json
from typing import Optional, Any

from idioms.data.types import TypeInfo
from idioms.dataiter import MatchedFunctionDataset, MatchedBinaryDataset
from idioms.data.dataset import MatchedBinary, MatchedFunction

def to_json(fn, type2id: Optional[dict[TypeInfo, int]] = None) -> dict[str, Any]:
    # If given a MatchedBinary, return a binary-level JSON containing its matched functions
    if isinstance(fn, MatchedBinary):
        return {
            "binary_hash": fn.binary_hash,
            "repo": fn.repo,
            # datasets/arrow really struggles with nested dicts
            "call_graph": [([k], vs) for k, vs in fn.call_graph.items()],
            "unmatched": [(k, v) for k, v in fn.unmatched.items()],
            "matched_functions": [to_json(f, type2id) for f in fn.functions],
        }

    variable_types = {
        name: (typ.declaration("") if type2id is None else type2id[typ])
        for name, typ in fn.variable_types.items()
    }
    user_defined_types = [
        (typ.declaration("") if type2id is None else type2id[typ]) for typ in fn.user_defined_types
    ]
    return {
        "name": fn.name,
        "canonical_name": fn.canonical_name,
        "repo": fn.repo,
        "decompiled_code": fn.decompiled_code,
        "canonical_decompiled_code": fn.canonical_decompiled_code,
        "original_code": fn.original_code,
        "canonical_original_code": fn.canonical_original_code,
        # Ignore code tokens for now; we'll use just unigram tokenization
        # "memory_layout": {loc.json_key(): var.to_json() for loc, var in self.memory_layout.items()},
        # datasets/arrow really struggles with nested dicts
        "variable_types": [(k,v) for k, v in variable_types.items()],
        "return_type": (fn.return_type.declaration("") if hasattr(fn.return_type, "declaration") else str(fn.return_type)),
        "user_defined_types": user_defined_types,
        "function_decls": fn.function_decls,
        "global_decls": fn.global_decls,
        "binary_hash": fn.binary_hash,
    }

def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("dataset")
    parser.add_argument("--binary", action="store_true")
    args = parser.parse_args()
    ds_path = Path(args.dataset)
    
    ds_class = MatchedBinaryDataset if args.binary else MatchedFunctionDataset

    train_set = ds_class(ds_path.glob("train*.tar"), shuffle=False)
    validation_set = ds_class(ds_path.glob("valid*.tar"), shuffle=False)
    test_set = ds_class(ds_path.glob("test*.tar"), shuffle=False)

    arrow = {}
    for name, partition in zip(["train", "valid", "test"], [train_set, validation_set, test_set]):
        arrow[name] = [to_json(fn) for fn in partition]
    
    filename = f"arrow-{ds_path.name}" + ("-binary" if args.binary else "-function") + ".json"
    with open(filename, "w") as fp:
        json.dump(arrow, fp, indent=2)

if __name__ == "__main__":
    main()