[kibwen]

I would expect a CRC to add a negligible number of lines of code. The reason that production-grade decompressors are tens of thousands of LOC is likely attributable to extreme manual optimization. For example, I wouldn't be surprised if a measurable fraction of those lines are actually inline assembly.

[nayuki]

True. A most basic CRC implementation is about 7 lines of code: (presented in Java to avoid some C/C++ footguns)

    int crc32(byte[] data) {
        int crc = ~0;
        for (byte b : data) {
            crc ^= b & 0xFF;
            for (int i = 0; i < 8; i++)
                crc = (crc >>> 1) ^ ((crc & 1) * 0xEDB88320);
        }
        return ~crc;
    }

Or smooshed down slightly (with caveats):

    int crc32(byte[] data) {
        int crc = ~0;
        for (int i = 0; i < data.length * 8; i++) {
            crc ^= (data[i / 8] >> (i % 8)) & 1;
            crc = (crc >>> 1) ^ ((crc & 1) * 0xEDB88320);
        }
        return ~crc;
    }

But one reason that many CRC implementations are large is because they include a pre-computed table of 256× 32-bit constants so that one byte can processed at a time. For example: https://github.com/madler/zlib/blob/7cdaaa09095e9266dee21314...

[xxs]

That's java code, though... bit weird, esp. i % 8 (which is just i & 7). The compiler should be able to optimize it since 'i' is guaranteed to be non-negative, still awkward.

Java CRC32 nowadays uses intrinsics and avx128 for crc32.

[kevin_thibedeau]

With C++20 you can use consteval to compute the table(s) at compile time from template parameters.

[ack_complete]

Doesn't need to be inline assembly, just pre-encoded lookup tables and intrinsics-based vectorized CRC alone will add quite a lot of code. Most multi-platform CRC algorithms tend to have at least a few paths for byte/word/dword at a time, hardware CRC, and hardware GF(2) multiply. It's not really extreme optimization, just better algorithms to match better hardware capabilities.

The Huffman decoding implementation is also bigger in production implementations for both speed and error checking. Two Huffman trees need to be exactly complete except in the special case of a single code, and in most cases they are flattened to two-level tables for speed (though the latest desktop CPUs have enough L1 cache to use single-level).

Finally, the LZ copy typically has special cases added for using wider than byte copies for non-overlapping, non-wrapping runs. This is a significant decoding speed optimization.

[tyingq]

Yes, there's subdirs with language bindings for many non-C langs, an examples folder with example C code, win32 specific C code, test code, etc.

More reasons it's an odd comparison.