Potentially expensive trimming of Chunk bytes

[bbannier]

A Chunk currently directly wraps containers holding its data. When calling Chunk::trim we directly trim these wrapped values,

spicy/hilti/runtime/src/types/stream.cc

Line 56 in 4970368

memmove(a->second.data(), begin, a->first.Ref());

spicy/hilti/runtime/src/types/stream.cc

Line 60 in 4970368

v.erase(v.begin(), v.begin() + static_cast<Vector::difference_type>((o - _offset).Ref()));

For the Zeek case we typically receive larger chunks and extract small bits from them, e.g., imagine a unit parsing a uint8[]. The current trim implementation makes sure that we free as much memory as possible (by removing each individual byte as soon as possible), but potentially incurs a performance overhead for that (e..g, std::vector::erase requires touching all bytes in the Chunk). The current implementation shows pathological performance for e.g., inputs generated by oss-fuzz.

We should consider using a cheaper implementation where Chunk::trim just moves the begin of a Chunk, but does not free the memory anymore.