//! A difference library similar to `rsync`.
//!
//! # Performance
//!
//! `den` is performant enough for large files and files, exceeding 10MB (when using rolling hashes).
//! Please compile with the **release** preset for **10X** the performance.
//!
//! Allocating data and keeping it in memory is very fast compared to hashing.
//! In the future, Den will support reading data bit by bit, greatly reducing the memory usage.
//!
//! # How-to & examples
//!
//! Keep in mind this isn't guaranteed to give the exact same data.
//! Please check the data with a secure hashing algorithm (e.g. SHA-3) to ensure consistency.
//!
//! Sending the data is possible due to [`serde`](https://docs.rs/serde/) providing serialization and deserialization.
//! This requires the cargo feature `serde` to be enabled.
//! You serialize all the structs in this library to any format.
//!
//! > These examples should cover what rsync does.
//!
//! ## Get a remote's data
//!
//! To get someone else's data, we construct a [`Signature`] and send it.
//! The remote calculates a [`Difference`] using [`Signature::diff`].
//! The remote sends back the [`Difference`] which we [`Difference::apply`].
//!
//! ## Push my data to remote
//!
//! Request the remote's [`Signature`].
//! They calculate it and respond.
//! We calculate a [`Signature::diff`] and send it to them.
//! They [`Difference::apply`] it. Their data should now be equal to mine.
//!
//! ## Get the difference of a local file
//!
//! Gets a small diff to send to others, almost like how `git` works.
//!
//! `base_data` is considered prior knowledge. `target_data` is the modified data.
//!
//! The data segments can be any size. Performance should still be good.
//!
//! ```
//! # use den::*;
//! let base_data = b"This is a document everyone has. It's about some new difference library.";
//! let target_data = b"This is a document only I have. It's about some new difference library.";
//!
//! let mut signature = Signature::new(128);
//! signature.write(base_data);
//! let signature = signature.finish();
//!
//! let diff = signature.diff(target_data);
//!
//! // This is the small diff you could serialize with Serde and send.
//! let minified = diff.minify(8, base_data)
//!     .expect("This won't panic, as the data hasn't changed from calling the other functions.");
//! ```
//!
//! # Future improvements
//!
//! - [x] Rolling hash
//! - [x] ~~Multi-threaded [`Signature::diff`].~~
//!       There is no feasible way to implement this, as we look ahead and change which window
//!       we're looking at after each iteration. Now with rolling hash, the performance is great.
//! - [ ] Support read/write
//!     - [ ] Support to diff a reader
//!     - [ ] Support to apply to a writer
//!     - [ ] Fetch API for apply to get data on demand.
//!         - This could slow things down dramatically.
//!     - [ ] Implement Write for `HashBuilder`.
//! - [x] ~~Use SHA(1|256?) to verify integrity of data. Bundled with the [`Signature`].~~ The
//!       implementer should provide this.

#![deny(
    clippy::all,
    clippy::pedantic,
    unreachable_pub,
    missing_debug_implementations,
    missing_docs
)]

use rustc_hash::FxHasher;
#[cfg(feature = "serde")]
use serde::{Deserialize, Serialize};
use std::cmp::{self, Ordering};
use std::collections::{HashMap, VecDeque};
use std::fmt::{self, Debug};
use std::hash::{BuildHasher, Hasher};
use std::num::NonZeroUsize;
use std::ops::Range;
use xxhash_rust::xxh3::Xxh3;
use xxhash_rust::xxh64::Xxh64;

pub mod parallel;

#[derive(PartialEq, Eq, PartialOrd, Ord)]
enum LargeHash<K, V> {
    Single(K, V),
    Multiple(Vec<(K, V)>),
}
#[derive(PartialEq, Eq, PartialOrd, Ord)]
struct LargeHashKey([u8; 8]);
struct HashMap128Hasher([u8; 8]);
impl Hasher for HashMap128Hasher {
    fn write(&mut self, bytes: &[u8]) {
        // for 32-bit architectures
        self.0[..bytes.len()].copy_from_slice(bytes);
    }
    fn finish(&self) -> u64 {
        u64::from_le_bytes(self.0)
    }
}
struct HashMap128HashBuilder;
impl BuildHasher for HashMap128HashBuilder {
    type Hasher = HashMap128Hasher;
    fn build_hasher(&self) -> Self::Hasher {
        HashMap128Hasher([0; 8])
    }
}
/// Hash map for 128-bit hashes as keys. This means we don't have to hash the value, improving
/// performance.
struct HashMap128 {
    map: HashMap<[u8; 8], LargeHash<[u8; 16], BlockData>, HashMap128HashBuilder>,
}
impl HashMap128 {
    fn new() -> Self {
        Self {
            map: HashMap::with_hasher(HashMap128HashBuilder),
        }
    }
    fn get(&self, key: [u8; 16], target: usize) -> Option<BlockData> {
        let mut bytes = [0; 8];
        bytes.copy_from_slice(&key[..8]);
        if let Some(option) = self.map.get(&bytes) {
            match option {
                LargeHash::Single(_key, block) => Some(*block),
                LargeHash::Multiple(list) => list
                    .binary_search_by(|probe| probe.0.cmp(&key))
                    .ok()
                    .map(|mut idx| {
                        let mut best = list[idx].1;
                        let mut best_dist = usize::MAX;
                        loop {
                            match best.start.cmp(&target) {
                                Ordering::Less => {
                                    idx += 1;
                                    if idx >= list.len() {
                                        return best;
                                    };
                                    let new = list[idx].1;
                                    let new_dist = new.start.abs_diff(target);
                                    if new_dist < best_dist && list[idx].0 == key {
                                        best = new;
                                        best_dist = new_dist;
                                    } else {
                                        return best;
                                    }
                                }
                                Ordering::Greater => {
                                    if idx == 0 {
                                        return best;
                                    };
                                    idx -= 1;
                                    let new = list[idx].1;
                                    let new_dist = new.start.abs_diff(target);
                                    if new_dist < best_dist && list[idx].0 == key {
                                        best = new;
                                        best_dist = new_dist;
                                    } else {
                                        return best;
                                    }
                                }
                                Ordering::Equal => return best,
                            }
                        }
                    }),
            }
        } else {
            None
        }
    }
    fn insert(&mut self, key: [u8; 16], value: BlockData) {
        let mut bytes = [0; 8];
        bytes.copy_from_slice(&key[..8]);
        if let Some(option) = self.map.get_mut(&bytes) {
            match option {
                LargeHash::Single(old_key, block) => {
                    let mut list = Vec::with_capacity(2);
                    list.push((*old_key, *block));
                    *option = LargeHash::Multiple(list);
                }
                LargeHash::Multiple(_) => {}
            }
            let list = match option {
                LargeHash::Single(_, _) => {
                    unreachable!()
                }
                LargeHash::Multiple(list) => list,
            };

            list.push((key, value));
        } else {
            self.map.insert(bytes, LargeHash::Single(key, value));
        }
    }
}

/// The algorithms which can be used for hashing the data.
#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
#[derive(Debug, PartialEq, Eq, Clone, Copy)]
#[must_use]
#[allow(missing_docs)]
pub enum HashAlgorithm {
    None4,
    None8,
    None16,
    XXH64,
    XXH3_64,
    XXH3_128,
    CyclicPoly32,
    CyclicPoly64,
    Adler32,
    Fx64,
}
impl HashAlgorithm {
    #[inline]
    fn builder(self, block_size: usize) -> HashBuilder {
        match self {
            Self::None4 => HashBuilder::None4(StackSlice::default()),
            Self::None8 => HashBuilder::None8(StackSlice::default()),
            Self::None16 => HashBuilder::None16(StackSlice::default()),
            Self::XXH64 => HashBuilder::XXH64(Xxh64::new(42)),
            Self::XXH3_64 => HashBuilder::XXH3_64(Xxh3::default()),
            Self::XXH3_128 => HashBuilder::XXH3_128(Xxh3::default()),
            Self::CyclicPoly32 => HashBuilder::CyclicPoly32(CyclicPoly32::new(block_size)),
            Self::CyclicPoly64 => HashBuilder::CyclicPoly64(CyclicPoly64::new(block_size)),
            Self::Adler32 => HashBuilder::Adler32(Adler32::new(block_size)),
            Self::Fx64 => HashBuilder::Fx64(FxHasher::default()),
        }
    }
}

/// Used to make [`RollingHash`] generic.
///
/// This is implemented by the [Adler32](https://docs.rs/adler32/latest/adler32/)
/// algorithm and both variants of the
/// [cyclic poly 32](https://docs.rs/cyclic-poly-23/latest/cyclic_poly_23/) algorithm. The
/// implementation of cyclic poly 23 is in a [`Box`] due to the large size of the struct
/// (1KiB and 2KiB for the 32- and 64-bit implementations respectively).
pub trait RollingHasher {
    /// The hash returned from this hasher.
    type Hash;

    /// Create a new hasher.
    fn new(block_size: usize) -> Self;
    /// Reset the inner state. If the struct provides no such functionality, consider overriding
    /// the current value with a `new` instance.
    fn reset(&mut self, block_size: usize);
    /// Write the `block` to the hasher.
    fn update(&mut self, block: &[u8], block_size: usize);
    /// Remove `old` and add `new`.
    fn rotate(&mut self, old: u8, new: u8, block_size: usize);
    /// Get the current internal hash.
    fn value(&self) -> Self::Hash;
}
impl RollingHasher for adler32::RollingAdler32 {
    type Hash = u32;

    fn new(_: usize) -> Self {
        Self::new()
    }
    fn reset(&mut self, _: usize) {
        *self = Self::new();
    }
    fn update(&mut self, block: &[u8], _: usize) {
        self.update_buffer(block);
    }
    fn rotate(&mut self, old: u8, new: u8, block_size: usize) {
        self.remove(block_size, old);
        self.update(new);
    }
    fn value(&self) -> Self::Hash {
        self.hash()
    }
}
impl RollingHasher for Box<cyclic_poly_23::CyclicPoly32> {
    type Hash = u32;

    fn new(block_size: usize) -> Self {
        Box::new(cyclic_poly_23::CyclicPoly32::new(block_size))
    }
    fn reset(&mut self, _: usize) {
        self.reset_hash();
    }
    fn update(&mut self, block: &[u8], _: usize) {
        (**self).update(block);
    }
    fn rotate(&mut self, old: u8, new: u8, _: usize) {
        (**self).rotate(old, new);
    }
    fn value(&self) -> Self::Hash {
        (**self).value()
    }
}
impl RollingHasher for Box<cyclic_poly_23::CyclicPoly64> {
    type Hash = u64;

    fn new(block_size: usize) -> Self {
        Box::new(cyclic_poly_23::CyclicPoly64::new(block_size))
    }
    fn reset(&mut self, _: usize) {
        self.reset_hash();
    }
    fn update(&mut self, block: &[u8], _: usize) {
        (**self).update(block);
    }
    fn rotate(&mut self, old: u8, new: u8, _: usize) {
        (**self).rotate(old, new);
    }
    fn value(&self) -> Self::Hash {
        (**self).value()
    }
}

/// A generic rolling hash implementation.
#[derive(Debug, Clone)]
pub struct RollingHash<T: RollingHasher> {
    inner: T,
    data: VecDeque<u8>,
    block_size: usize,
    last_position: usize,
    write_data: bool,
    reset_data: bool,
}
impl<T: RollingHasher> RollingHash<T> {
    fn new(block_size: usize) -> Self {
        Self {
            inner: T::new(block_size),
            data: VecDeque::with_capacity(block_size),
            block_size,
            last_position: 0,
            write_data: false,
            reset_data: false,
        }
    }
    fn write_data(&mut self) {
        if self.write_data {
            let data = self.data.make_contiguous();
            self.inner.reset(self.block_size);
            self.inner.update(data, self.block_size);
            self.write_data = false;
        }
    }
    fn write(&mut self, data: &[u8], position: Option<usize>) {
        if self.data.len() == self.block_size && data.len() == self.data.len() {
            if let Some(pos) = position {
                // assuming data is from the same source
                if pos == self.last_position + 1 {
                    // flush
                    self.write_data();

                    let first = self.data.pop_front().unwrap();
                    let new = *data.last().unwrap();
                    self.data.push_back(new);
                    self.inner.rotate(first, new, self.block_size);
                    self.last_position = pos;

                    return;
                }
            }
        }

        if self.reset_data {
            self.data.clear();
        }
        // normal execution
        self.data.extend(data);
        self.last_position = position.unwrap_or(0);
        self.write_data = true;
    }
    fn finish_reset(&mut self) -> T::Hash {
        let wrote_data = self.write_data;
        self.write_data();
        if wrote_data {
            self.reset_data = true;
        }
        self.inner.value()
    }
}

/// [`RollingHash`] using the 32-bit cyclic poly 23 algorithm.
///
/// Implements the trait [`RollingHasher`].
pub type CyclicPoly32 = RollingHash<Box<cyclic_poly_23::CyclicPoly32>>;
/// [`RollingHash`] using the 32-bit cyclic poly 23 algorithm.
///
/// Implements the trait [`RollingHasher`].
pub type CyclicPoly64 = RollingHash<Box<cyclic_poly_23::CyclicPoly64>>;
/// [`RollingHash`] using the Adler32 algorithm.
///
/// Implements the trait [`RollingHasher`].
pub type Adler32 = RollingHash<adler32::RollingAdler32>;

enum HashBuilder {
    None4(StackSlice<4>),
    None8(StackSlice<8>),
    None16(StackSlice<16>),
    XXH64(Xxh64),
    XXH3_64(Xxh3),
    XXH3_128(Xxh3),
    CyclicPoly32(CyclicPoly32),
    CyclicPoly64(CyclicPoly64),
    Adler32(Adler32),
    Fx64(FxHasher),
}

impl Clone for HashBuilder {
    fn clone(&self) -> Self {
        match self {
            Self::None4(h) => Self::None4(h.clone()),
            Self::None8(h) => Self::None8(h.clone()),
            Self::None16(h) => Self::None16(h.clone()),
            Self::XXH64(_) => Self::XXH64(Xxh64::new(42)),
            Self::XXH3_64(h) => Self::XXH3_64(h.clone()),
            Self::XXH3_128(h) => Self::XXH3_128(h.clone()),
            Self::CyclicPoly32(h) => Self::CyclicPoly32(h.clone()),
            Self::CyclicPoly64(h) => Self::CyclicPoly64(h.clone()),
            Self::Adler32(h) => Self::Adler32(h.clone()),
            Self::Fx64(_) => Self::Fx64(FxHasher::default()),
        }
    }
}
impl HashBuilder {
    #[inline]
    fn finish_reset(&mut self) -> HashResult {
        match self {
            Self::None4(h) => {
                let r = HashResult::B4(h.finish().to_le_bytes());
                *h = StackSlice::default();
                r
            }
            Self::None8(h) => {
                let r = HashResult::B8(h.finish().to_le_bytes());
                *h = StackSlice::default();
                r
            }
            Self::None16(h) => {
                let r = HashResult::B16(h.finish().to_le_bytes());
                *h = StackSlice::default();
                r
            }
            Self::XXH64(h) => {
                let r = HashResult::B8(h.digest().to_le_bytes());
                h.reset(42);
                r
            }
            Self::XXH3_64(h) => {
                let r = HashResult::B8(h.digest().to_le_bytes());
                h.reset();
                r
            }
            Self::XXH3_128(h) => {
                let r = HashResult::B16(h.digest128().to_le_bytes());
                h.reset();
                r
            }
            Self::CyclicPoly32(h) => HashResult::B4(h.finish_reset().to_le_bytes()),
            Self::CyclicPoly64(h) => HashResult::B8(h.finish_reset().to_le_bytes()),
            Self::Adler32(h) => HashResult::B4(h.finish_reset().to_le_bytes()),
            Self::Fx64(fx) => HashResult::B8(core::mem::take(fx).finish().to_le_bytes()),
        }
    }
    #[inline]
    fn finish(self) -> HashResult {
        match self {
            Self::None4(h) => HashResult::B4(h.finish().to_le_bytes()),
            Self::None8(h) => HashResult::B8(h.finish().to_le_bytes()),
            Self::None16(h) => HashResult::B16(h.finish().to_le_bytes()),
            Self::XXH64(h) => HashResult::B8(h.digest().to_le_bytes()),
            Self::XXH3_64(h) => HashResult::B8(h.digest().to_le_bytes()),
            Self::XXH3_128(h) => HashResult::B16(h.digest128().to_le_bytes()),
            Self::CyclicPoly32(mut h) => HashResult::B4(h.finish_reset().to_le_bytes()),
            Self::CyclicPoly64(mut h) => HashResult::B8(h.finish_reset().to_le_bytes()),
            Self::Adler32(mut h) => HashResult::B4(h.finish_reset().to_le_bytes()),
            Self::Fx64(fx) => HashResult::B8(fx.finish().to_le_bytes()),
        }
    }
    #[inline]
    fn write(&mut self, data: &[u8], position: Option<usize>) {
        match self {
            Self::None4(h) => h.write(data),
            Self::None8(h) => h.write(data),
            Self::None16(h) => h.write(data),
            Self::XXH64(h) => h.update(data),
            Self::XXH3_64(h) | Self::XXH3_128(h) => h.update(data),
            Self::CyclicPoly32(h) => h.write(data, position),
            Self::CyclicPoly64(h) => h.write(data, position),
            Self::Adler32(h) => h.write(data, position),
            Self::Fx64(h) => h.write(data),
        }
    }
}

impl Debug for HashBuilder {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.write_str("HashBuilder (internal hasher data)")
    }
}

#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
#[derive(PartialEq, Eq, PartialOrd, Ord, Copy, Clone)]
#[must_use]
enum HashResult {
    B4([u8; 4]),
    B8([u8; 8]),
    B16([u8; 16]),
}
impl HashResult {
    fn to_bytes(self) -> [u8; 16] {
        match self {
            Self::B4(bytes) => to_16_le_bytes(&bytes),
            Self::B8(bytes) => to_16_le_bytes(&bytes),
            Self::B16(bytes) => to_16_le_bytes(&bytes),
        }
    }
}
impl Debug for HashResult {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(
            f,
            "HashResult({:0<16X})",
            u128::from_le_bytes(self.to_bytes())
        )
    }
}

#[derive(Debug, Clone)]
struct StackSlice<const SIZE: usize> {
    data: [u8; SIZE],
    len: u8,
}
impl<const SIZE: usize> Default for StackSlice<SIZE> {
    fn default() -> Self {
        Self {
            data: zeroed(),
            len: 0,
        }
    }
}
impl<const SIZE: usize> StackSlice<SIZE> {
    /// # Panics
    ///
    /// Panics if `data.len()` > [`Self::available`].
    #[allow(clippy::cast_possible_truncation)]
    #[inline]
    fn write(&mut self, data: &[u8]) {
        // SAFETY: Needed for the block below.
        assert!(
            data.len() <= self.available(),
            "Length ({}) is greater than what's available ({})",
            data.len(),
            self.available()
        );
        // SAFETY: We've checked above the guarantees hold up.
        unsafe {
            self.data
                .get_unchecked_mut(self.len as usize..self.len as usize + data.len())
        }
        .copy_from_slice(data);

        // The assert above guarantees we never reach this point if the input is too large.
        self.len += data.len() as u8;
    }
    #[inline]
    fn available(&self) -> usize {
        SIZE - self.len as usize
    }
}
impl StackSlice<4> {
    fn finish(&self) -> u32 {
        u32::from_ne_bytes(self.data)
    }
}
impl StackSlice<8> {
    fn finish(&self) -> u64 {
        u64::from_ne_bytes(self.data)
    }
}
impl StackSlice<16> {
    fn finish(&self) -> u128 {
        u128::from_ne_bytes(self.data)
    }
}

#[must_use]
fn to_16_le_bytes<const SIZE: usize>(bytes: &[u8; SIZE]) -> [u8; 16] {
    let mut bytes_fixed = zeroed();

    bytes_fixed[..SIZE].copy_from_slice(bytes);

    bytes_fixed
}

#[must_use]
const fn zeroed<const SIZE: usize>() -> [u8; SIZE] {
    [0; SIZE]
}

/// A trait to fill extend a [`Vec`] with more data.
///
/// This is by default implemented by all types which can cohere to a byte slice `&[u8]`.
pub trait ExtendVec: Debug {
    /// Extend `vec` with our data.
    /// This must extend with [`ExtendVec::len`] bytes.
    fn extend(&self, vec: &mut Vec<u8>);
    /// Copy the bytes of this struct to `position` in `vec`, overriding any data there.
    /// This must replace [`ExtendVec::len`] bytes.
    fn replace(&self, vec: &mut Vec<u8>, position: usize);
    /// If our data equals `bytes`.
    #[must_use]
    fn equals(&self, bytes: &[u8]) -> bool;
    /// The length of the data of this struct.
    fn len(&self) -> usize;
    /// If no data is available.
    #[inline]
    fn is_empty(&self) -> bool {
        self.len() == 0
    }
}
impl<T: AsRef<[u8]> + Debug> ExtendVec for T {
    #[inline]
    fn extend(&self, vec: &mut Vec<u8>) {
        let slice = self.as_ref();
        // SAFETY: This guarantees `vec.capacity()` >= `vec.len() + slice.len()`
        vec.reserve(slice.len());
        let len = vec.len();
        // SAFETY: We get uninitialized bytes and write to them. This is fine.
        // The length is guaranteed to be allocated from above.
        let destination = unsafe { vec.get_unchecked_mut(len..len + slice.len()) };
        destination.copy_from_slice(slice);
        // SAFETY: We set the length to that we've written to above.
        unsafe { vec.set_len(vec.len() + slice.len()) };
    }
    #[allow(clippy::uninit_vec)] // we know what we're doing
    #[inline]
    fn replace(&self, vec: &mut Vec<u8>, position: usize) {
        let slice = self.as_ref();
        let new_len = (position + slice.len()).max(vec.len());
        // SAFETY: This guarantees `vec.capacity()` >= `vec.len() + slice.len()`
        vec.reserve(new_len - vec.len());
        // SAFETY: We set the length to what we write below.
        unsafe { vec.set_len(new_len) };
        let destination = &mut vec[position..(position + slice.len())];
        destination.copy_from_slice(slice);
    }

    #[inline]
    fn equals(&self, bytes: &[u8]) -> bool {
        self.as_ref() == bytes
    }

    #[inline]
    fn len(&self) -> usize {
        self.as_ref().len()
    }
}

/// Builder of a [`Signature`].
/// Created using constructors on [`Signature`] (e.g. [`Signature::with_algorithm`]);
///
/// You [`Self::write`] data and then [`Self::finish`] to get a [`Signature`].
#[derive(Debug)]
#[must_use]
pub struct SignatureBuilder {
    algo: HashAlgorithm,
    blocks: Vec<HashResult>,
    block_size: usize,

    current: HashBuilder,
    len: usize,
    total: usize,
}
impl SignatureBuilder {
    /// The `hasher` is used as the template hasher from which all other hashers are cloned.
    ///
    /// A good block size is `1024`.
    fn new(algo: HashAlgorithm, block_size: usize) -> Self {
        Self {
            algo,
            blocks: Vec::new(),
            block_size,

            current: algo.builder(block_size),
            len: 0,
            total: 0,
        }
    }
    const fn block_available(&self) -> usize {
        self.block_size - self.len
    }
    fn finish_hash(&mut self) {
        if self.len != 0 {
            let result = self.current.finish_reset();
            self.blocks.push(result);
        }
        self.len = 0;
    }
    /// Appends data to the hasher.
    ///
    /// This can be called multiple times to write the resource bit-by-bit.
    pub fn write(&mut self, data: &[u8]) {
        let mut data = data;
        let mut position = 0;

        self.total += data.len();

        while data.len() >= self.block_available() {
            let bytes = &data[..self.block_available()];
            self.current.write(bytes, Some(position));

            data = &data[self.block_available()..];
            position += bytes.len();
            self.len += bytes.len();
            self.finish_hash();
        }
        self.len += data.len();
        // the data is now less than `self.block_available()`.
        self.current.write(data, Some(position));
    }
    /// Flushes the data from [`Self::write`] and prepares a [`Signature`].
    pub fn finish(mut self) -> Signature {
        self.finish_hash();

        let Self {
            algo,
            blocks,
            block_size,
            ..
        } = self;

        Signature {
            algo,
            blocks,
            block_size,
            original_data_len: self.total,
        }
    }
}
impl PartialEq for SignatureBuilder {
    #[allow(clippy::nonminimal_bool)] // you are incorrect
    fn eq(&self, other: &Self) -> bool {
        self.algo == other.algo
            && self.blocks == other.blocks
            && self.blocks == other.blocks
            && self.len == other.len
            && self.total == other.total
    }
}

/// A identifier of a file, much smaller than the file itself.
///
/// See [crate-level documentation](crate) for more details.
#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
#[derive(PartialEq, Eq, Clone)]
#[must_use]
pub struct Signature {
    algo: HashAlgorithm,
    blocks: Vec<HashResult>,
    block_size: usize,
    original_data_len: usize,
}
impl Signature {
    /// Creates a new [`SignatureBuilder`] in which data can be added to create a list of hashes
    /// for blocks the size of `block_size`.
    ///
    /// I recommend block size `4096` for remote transfers.
    /// If small documents are what's mostly being transmitted, consider `512`.
    /// Consider running [`Difference::minify`] if getting the smallest diff is your concern.
    ///
    /// This creates a signature of a resource. The signature takes up much less space than the
    /// resource.
    ///
    /// Larger `block_size`s will take more less to compute but signatures become larger, as if a
    /// block's contents isn't found, it has to send `block_size` bytes of data.
    /// Smaller `block_size`s takes more time to compute, but send data in the [`Signature`].
    /// Diffs with smaller `block_size`s also have the advantage of interfering less with changes
    /// if applied to data other than what's [written to the signature](SignatureBuilder::write).
    ///
    /// The [`HashAlgorithm`] is chosen using experience with hasher's performance and heuristics.
    /// You can query the algorithm using [`Self::algorithm`].
    ///
    /// # Panics
    ///
    /// Panics if `block_size` is `0`.
    #[allow(clippy::new_ret_no_self)] // This returns a builder.
    pub fn new(block_size: usize) -> SignatureBuilder {
        match block_size {
            0..=4 => Signature::with_algorithm(HashAlgorithm::None4, block_size),
            5..=8 => Signature::with_algorithm(HashAlgorithm::None8, block_size),
            9..=16 => Signature::with_algorithm(HashAlgorithm::None16, block_size),
            17..=128 => Signature::with_algorithm(HashAlgorithm::Fx64, block_size),
            // according to imperical results, XXHash is just that much faster than CyclicPoly -
            // XXH3 performs better than CyclicPoly even for rolling hash applications.
            //
            // This is probably due to the requirement of buffers when working with rolling hashes.
            // `TODO`: If we digest the whole data block in one go, we could remove the need for buffers.
            129..=4096 => Signature::with_algorithm(HashAlgorithm::XXH3_64, block_size),
            // 4096..
            _ => Signature::with_algorithm(HashAlgorithm::XXH3_128, block_size),
        }
    }
    /// This will create a new [`SignatureBuilder`] with the `algorithm`. Consider using [`Self::new`]
    /// if you don't know *exactly* what you are doing, as it sets the algorithm for you.
    /// You can query the algorithm using [`Self::algorithm`].
    ///
    /// See [`Self::new`] for more details and insights into how to choose `block_size`.
    ///
    /// # Panics
    ///
    /// Will panic if [`HashAlgorithm`] is of type `None*` and `block_size` is larger than the
    /// `HashAlgorithm`.
    /// Also panics if `block_size` is `0`.
    pub fn with_algorithm(algorithm: HashAlgorithm, block_size: usize) -> SignatureBuilder {
        assert_ne!(block_size, 0, "Block size cannot be 0.");
        match algorithm {
            HashAlgorithm::None4 => assert!(block_size <= 4),
            HashAlgorithm::None8 => assert!(block_size <= 8),
            HashAlgorithm::None16 => assert!(block_size <= 16),
            _ => {}
        }
        SignatureBuilder::new(algorithm, block_size)
    }

    /// Get the algorithm used by this signature.
    pub const fn algorithm(&self) -> HashAlgorithm {
        self.algo
    }
    /// Returns the block size of this signature.
    #[must_use]
    pub const fn block_size(&self) -> usize {
        self.block_size
    }
    /// Gets the hashes for all the blocks of the input resource.
    pub(crate) fn blocks(&self) -> &[HashResult] {
        &self.blocks
    }

    /// Get the [`Difference`] between the data the [`Signature`] represents and the local `data`.
    ///
    /// This will return a struct which when serialized (using e.g. `bincode`) is much smaller than
    /// `data`.
    pub fn diff(&self, data: &[u8]) -> Difference {
        let mut map = HashMap128::new();

        let block_size = self.block_size();

        // Special case: Signature contains no hashes.
        // Just send the whole input.
        if self.blocks().is_empty() {
            let segments = if data.is_empty() {
                vec![]
            } else {
                vec![Segment::unknown(data)]
            };
            return Difference {
                segments,
                block_size,
                original_data_len: data.len(),
                parallel_data: None,
            };
        }

        for (nr, block) in self.blocks().iter().enumerate() {
            let bytes = block.to_bytes();

            let start = nr * block_size;
            let block_data = BlockData { start };

            map.insert(bytes, block_data);
        }

        let segments = Self::inner_diff(data, 0, usize::MAX, block_size, &map, self.algorithm());

        Difference {
            segments,
            block_size,
            original_data_len: self.original_data_len,
            parallel_data: None,
        }
    }

    #[allow(clippy::too_many_lines)] // well, this is the main implementation
    fn inner_diff(
        data: &[u8],
        start: usize,
        end: usize,
        block_size: usize,
        map: &HashMap128,
        algorithm: HashAlgorithm,
    ) -> Vec<Segment> {
        #[allow(clippy::inline_always)]
        #[inline(always)]
        fn push_unknown_data(
            data: &[u8],
            last_ref: usize,
            blocks_pos: usize,
            segments: &mut Vec<Segment>,
        ) {
            let unknown_data = &data[last_ref..blocks_pos - 1];
            if !unknown_data.is_empty() {
                segments.push(Segment::unknown(unknown_data));
            }
        }

        let mut blocks = Blocks::new(data, block_size);
        blocks.advance(start);

        let mut segments = Vec::new();
        let mut last_ref_block = start;
        let end_or_left = end.min(data.len());

        #[allow(
            clippy::cast_possible_truncation,
            clippy::cast_precision_loss,
            clippy::cast_sign_loss
        )]
        let lookahead_limit = block_size.min((((block_size / 8) as f64).sqrt() as usize).max(8));
        // look into if it's better to not use this.
        // That'd enable trying to find the best any distans in front, but might produce very long
        // unknown segments.
        let mut lookahead_ignore = 0;

        let mut hasher = algorithm.builder(block_size);

        // Iterate over data, in windows. Find hash.
        let mut unknown = 0;
        while let Some(block) = blocks.next() {
            if blocks.pos() >= end {
                break;
            }
            hasher.write(block, Some(blocks.pos() - 1));
            let hash = hasher.finish_reset().to_bytes();

            // If hash matches, push previous data to unknown, and push a ref. Advance by `block_size` (actually `block_size-1` as the `next` call implicitly increments).
            if let Some(block_data) = map.get(hash, blocks.pos() - unknown) {
                // If we're looking for a ref (after a unknown, which hasn't been "committed" yet),
                // check with a offset of 0..<some small integer> for better matches, which give us
                // longer BlockRefs. This reduces jitter in output.
                if lookahead_ignore == 0 {
                    let mut best: Option<(usize, usize)> = None;
                    for offset in 0..lookahead_limit {
                        let mut blocks = blocks.clone();
                        blocks.go_back(1);
                        blocks.advance(offset);
                        let mut last_end = None;

                        let mut successive_block_count = 0;
                        for i in 0..lookahead_limit {
                            if let Some(block) = blocks.next() {
                                // check hash(clone rolling hash), if matches with end of prior,
                                // add 1 to current score & max = max.max(current)
                                hasher.write(block, Some(blocks.pos() - 1));
                                let hash = hasher.finish_reset().to_bytes();

                                if let Some(block_data) = map.get(hash, blocks.pos() - unknown) {
                                    match &mut last_end {
                                        Some(end) => {
                                            if block_data.start == *end {
                                                // successive block

                                                debug_assert_eq!(i, successive_block_count);
                                                successive_block_count += 1;
                                                *end = block_data.start + block_size;

                                                blocks.advance(block.len() - 1);
                                            } else {
                                                // block is not directly after the other
                                                break;
                                            }
                                        }
                                        None => {
                                            successive_block_count += 1;
                                            last_end = Some(block_data.start + block_size);
                                        }
                                    }
                                } else {
                                    // no match, break
                                    break;
                                }
                            } else {
                                break;
                            }
                        }

                        match &mut best {
                            Some(best) => {
                                let (_iter, successive) = *best;
                                if successive < successive_block_count {
                                    *best = (offset, successive_block_count);
                                } else {
                                    // we're worse than the prior loops
                                }
                            }
                            None => {
                                best = Some((offset, successive_block_count));
                            }
                        }
                    }
                    match best {
                        // the current one is the best, stick with this.
                        Some((0, ignore)) => lookahead_ignore = ignore,
                        // didn't find any better
                        None => {}
                        Some((n, ignore)) => {
                            // -1 as the next should not advance the position, it's naturally the
                            // next.
                            blocks.advance(n - 1);
                            lookahead_ignore = ignore;
                            continue;
                        }
                    }
                }
                // decrement the ignored
                lookahead_ignore = lookahead_ignore.saturating_sub(1);

                push_unknown_data(data, last_ref_block, blocks.pos(), &mut segments);

                if let Some(last) = segments.last_mut() {
                    match last {
                        Segment::Ref(block_ref_segment) => {
                            if block_data.start == block_ref_segment.end(block_size) {
                                block_ref_segment.extend(1);
                            } else {
                                segments.push(Segment::reference(block_data.start));
                            }
                        }
                        Segment::Unknown(_) => {
                            segments.push(Segment::reference(block_data.start));
                        }
                    }
                } else {
                    segments.push(Segment::reference(block_data.start));
                }

                blocks.advance(block.len() - 1);
                last_ref_block = blocks.pos();
            } else {
                unknown += 1;
            }
        }

        // If we want them to get our diff, we send our diff, with the `Unknown`s filled with data.
        // Then, we only send references to their data and our new data.
        if data
            .get(last_ref_block..(blocks.pos() + 1).min(end_or_left))
            .map_or(false, |d| !d.is_empty())
        {
            push_unknown_data(data, last_ref_block, blocks.pos() + 1, &mut segments);
            println!("last push unknown");
        }

        segments
    }
}
impl Debug for Signature {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        let blocks: Vec<String> = self
            .blocks()
            .iter()
            .map(|block| format!("{:0<16X}", u128::from_le_bytes(block.to_bytes())))
            .collect();
        f.debug_struct("Signature")
            .field("algo", &self.algo)
            .field("blocks", &blocks)
            .field("block_size", &self.block_size)
            .field("original_data_len", &self.original_data_len)
            .finish()
    }
}

#[derive(Debug, Clone, Copy)]
struct BlockData {
    start: usize,
}
/// One or more successive blocks found in the common data.
#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
#[derive(PartialEq, Eq, Clone, Copy)]
#[must_use]
pub struct SegmentRef {
    /// Start of segment with a length of [`Self::block_count`]*[`Signature::block_size`].
    pub start: usize,
    /// The number of blocks this segment references.
    pub block_count: usize,
}
impl SegmentRef {
    /// The start in the base resource this reference is pointing to.
    #[must_use]
    pub fn start(self) -> usize {
        self.start
    }
    /// The number of blocks this reference covers.
    #[must_use]
    pub fn block_count(self) -> usize {
        self.block_count
    }
    /// The length of the referred data.
    #[must_use]
    #[allow(clippy::len_without_is_empty)] // that's impossible
    pub fn len(self, block_size: usize) -> usize {
        self.block_count() * block_size
    }
    /// The end of this segment.
    ///
    /// The same as [`Self::start`] + [`Self::block_count`] * `block_size`.
    #[inline]
    #[must_use]
    pub fn end(self, block_size: usize) -> usize {
        self.start + self.len(block_size)
    }
    /// Get the length to the end of `diff`. If this [`Segment::Ref`] doesn't reach `diff`'s end,
    /// [`None`] is returned. Else, the count of bytes until the end of copying is returned.
    #[inline]
    #[must_use]
    pub fn len_to_end(self, diff: &Difference<impl ExtendVec + 'static>) -> Option<usize> {
        let len = diff.original_data_len;
        let end = self.end(diff.block_size());
        if len > end {
            None
        } else {
            Some(end - len)
        }
    }
    /// Get a new [`SegmentRef`] with [`Self::start`] set to `start`.
    #[inline]
    pub fn with_start(mut self, start: usize) -> Self {
        self.start = start;
        self
    }
    /// Get a new [`SegmentRef`] with [`Self::block_count`] set to `n`.
    #[inline]
    pub fn with_blocks(mut self, n: usize) -> Self {
        self.block_count = n;
        self
    }
    /// Add `n` to [`Self::block_count`].
    #[inline]
    pub fn extend(&mut self, n: usize) {
        *self = self.with_blocks(self.block_count() + n);
    }
    /// Multiplies the count of blocks.
    /// Can be useful if the block size changes.
    #[inline]
    pub fn multiply(&mut self, n: usize) {
        *self = self.with_blocks(self.block_count() * n);
    }
}
impl Debug for SegmentRef {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        if self.start == 0 {
            write!(f, "..{}..", self.block_count)
        } else {
            write!(f, "{}..{}..", self.start, self.block_count)
        }
    }
}
/// A segment with unknown contents. This will transmit the data.
#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
#[derive(PartialEq, Eq, Clone)]
#[must_use]
pub struct SegmentUnknown<S: ExtendVec = Vec<u8>> {
    /// The data source to fill the unknown with.
    pub source: S,
}
impl<S: ExtendVec> SegmentUnknown<S> {
    /// Create a new [`SegmentUnknown`] from `source`.
    /// The methods on [`ExtendVec`] is then used to fill the target.
    pub fn new(source: S) -> Self {
        Self { source }
    }
    /// Get a reference to the data source.
    pub fn source(&self) -> &S {
        &self.source
    }
}
impl SegmentUnknown {
    /// Get a reference to the data transmitted.
    #[must_use]
    pub fn data(&self) -> &[u8] {
        &self.source
    }
    /// Get a mutable reference to the internal data [`Vec`].
    pub fn data_mut(&mut self) -> &mut Vec<u8> {
        &mut self.source
    }
    /// Takes the internal data of this segment.
    #[must_use]
    pub fn into_data(self) -> Vec<u8> {
        self.source
    }
}
impl<S: ExtendVec + AsRef<[u8]>> Debug for SegmentUnknown<S> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        write!(f, "{:?}", String::from_utf8_lossy(self.source.as_ref()))
    }
}
/// A segment of data corresponding to a multiple of [`Difference::block_size`].
#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
#[derive(PartialEq, Eq, Clone)]
#[must_use]
pub enum Segment<S: ExtendVec + 'static = Vec<u8>> {
    /// Reference to successive block(s) of data.
    Ref(SegmentRef),
    /// Data unknown to the one who sent the [`Signature`].
    Unknown(SegmentUnknown<S>),
}
impl<S: ExtendVec> Segment<S> {
    /// Create a new [`Segment::Ref`] from `start`.
    ///
    /// [`SegmentRef::block_count`] is set to 1.
    #[inline]
    pub fn reference(start: usize) -> Self {
        Self::Ref(SegmentRef {
            start,
            block_count: 1,
        })
    }
}
impl<S: ExtendVec + AsRef<[u8]>> Debug for Segment<S> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        match self {
            Segment::Ref(seg) => seg.fmt(f),
            Segment::Unknown(seg) => seg.fmt(f),
        }
    }
}
impl Segment {
    /// Creates a [`Segment::Unknown`] with `data` as the unknown part.
    ///
    /// Should mainly be used in testing.
    #[inline]
    pub fn unknown(data: impl Into<Vec<u8>>) -> Self {
        Self::Unknown(SegmentUnknown {
            source: data.into(),
        })
    }
}
#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
#[derive(PartialEq, Eq, Clone, Debug, Copy)]
struct ParallelData {
    /// Clamp last ref segment in each block so the block length becomes this.
    parallel_block_size: NonZeroUsize,
    last_block_segment_length: usize,
}
/// A delta between the local data and the data the [`Signature`] represents.
#[allow(clippy::unsafe_derive_deserialize)] // See SAFETY notes.
#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
#[derive(PartialEq, Eq, Clone)]
#[must_use]
pub struct Difference<S: ExtendVec + 'static = Vec<u8>> {
    segments: Vec<Segment<S>>,
    block_size: usize,
    original_data_len: usize,
    parallel_data: Option<ParallelData>,
}
impl Difference {
    /// Changes the block size to `block_size`, shrinking the [`Segment::Unknown`]s in the process.
    /// This results in a smaller diff.
    ///
    /// Consider [`Self::minify_with_builder`] for using custom signature settings.
    ///
    /// # Errors
    ///
    /// tl;dr, you can [`Option::unwrap`] this if it comes straight from [`Signature::diff`] or
    /// this very function. If it's untrusted data, handle the errors.
    ///
    /// Returns [`MinifyError::NewLarger`] if `block_size` >= [`Self::block_size`] and
    /// [`MinifyError::NotMultiple`] if [`Self::block_size`] is not a multiple of `block_size`.
    /// [`MinifyError::SuccessiveUnknowns`] is returned if two [`Segment::Unknown`] are after
    /// each other.
    /// Returns [`MinifyError::Zero`] if `block_size == 0`.
    pub fn minify(&self, block_size: usize, base: &[u8]) -> Result<Self, MinifyError> {
        self.minify_with_builder(block_size, base, |block_size: usize| {
            Signature::new(block_size)
        })
    }
    /// Changes the block size to `block_size`, shrinking the [`Segment::Unknown`]s in the process.
    /// This results in a smaller diff.
    ///
    /// `signature_builder` takes the block size and gives the signature builder to use.
    ///
    /// # Panics
    ///
    /// Panics if `signature_builder` returns a [`SignatureBuilder`] with a different `block_size`
    /// from what's supplied.
    ///
    /// # Errors
    ///
    /// tl;dr, you can [`Option::unwrap`] this if it comes straight from [`Signature::diff`] or
    /// this very function. If it's untrusted data, handle the errors.
    ///
    /// Returns [`MinifyError::NewLarger`] if `block_size` >= [`Self::block_size`] and
    /// [`MinifyError::NotMultiple`] if [`Self::block_size`] is not a multiple of `block_size`.
    /// [`MinifyError::SuccessiveUnknowns`] is returned if two [`Segment::Unknown`] are after
    /// each other.
    /// Returns [`MinifyError::Zero`] if `block_size == 0`.
    #[allow(clippy::too_many_lines)]
    pub fn minify_with_builder(
        &self,
        block_size: usize,
        base: &[u8],
        mut signature_builder: impl FnMut(usize) -> SignatureBuilder,
    ) -> Result<Self, MinifyError> {
        fn push_segment(segments: &mut Vec<Segment>, item: Segment, block_size: usize) {
            match item {
                Segment::Ref(item) => {
                    if let Some(last) = segments.last_mut() {
                        match last {
                            Segment::Ref(seg) => {
                                if seg.end(block_size) == item.start {
                                    // The previous end is this start.
                                    seg.extend(item.block_count());
                                } else {
                                    segments.push(Segment::Ref(item));
                                }
                            }
                            Segment::Unknown(_) => segments.push(Segment::Ref(item)),
                        }
                    } else {
                        segments.push(Segment::Ref(item));
                    }
                }
                Segment::Unknown(_) => segments.push(item),
            }
        }

        if block_size == 0 {
            return Err(MinifyError::Zero);
        }
        if self.block_size() <= block_size {
            return Err(MinifyError::NewLarger);
        }
        let block_size_shrinkage = self.block_size() / block_size;
        let block_size_shrinkage_remainder = self.block_size() % block_size;
        if block_size_shrinkage_remainder != 0 {
            return Err(MinifyError::NotMultiple);
        }

        // `self` can be applied to `base` to get `target`.
        // We want to minify the `unknown` segments of `self`.

        // Use heuristics to allocate what we have * 1.2, if any more segments are added.
        let mut segments = Vec::with_capacity(self.segments().len() * 6 / 5);
        for (last, current, next) in PrePostWindow::new(self.segments()) {
            match current {
                Segment::Ref(seg) => {
                    let mut seg = *seg;
                    seg.multiply(block_size_shrinkage);
                    push_segment(&mut segments, Segment::Ref(seg), block_size);
                }
                Segment::Unknown(seg) => {
                    let target_data = seg.data();
                    let base_data = {
                        // Start = previous end
                        // end = next's start.
                        // If it's a reference, that's start.
                        //
                        // Get a peek to the next segment.
                        let start = if let Some(last) = last {
                            match last {
                                Segment::Ref(seg) => seg.end(block_size),
                                // Improbable, two Unknowns can't be after each other.
                                // Only through external modification of data.
                                Segment::Unknown(_) => return Err(MinifyError::SuccessiveUnknowns),
                            }
                        } else {
                            0
                        };
                        let mut end = match next {
                            Some(Segment::Ref(seg)) => seg.start,
                            Some(Segment::Unknown(_)) => {
                                return Err(MinifyError::SuccessiveUnknowns)
                            }
                            None => base.len(),
                        };
                        if end < start {
                            end = start + seg.data().len();
                        }
                        end = cmp::min(end, base.len());

                        base.get(start..end).map(|slice| (slice, start))
                    };

                    if let Some((base_data, start)) = base_data {
                        let mut builder = signature_builder(block_size);
                        assert_eq!(
                            builder.block_size, block_size,
                            "signature built from `signature_builder` \
                            callback has to have the given block size"
                        );
                        builder.write(base_data);
                        let signature = builder.finish();

                        let diff = signature.diff(target_data);

                        for mut segment in diff.segments {
                            match &mut segment {
                                Segment::Ref(seg) => seg.start += start,
                                Segment::Unknown(_) => {}
                            }
                            push_segment(&mut segments, segment, block_size);
                        }
                    } else {
                        segments.push(Segment::Unknown(seg.clone()));
                    }
                }
            }
        }

        for segment in &mut segments {
            match segment {
                Segment::Ref(seg) => {
                    let new = seg.end(block_size);
                    if base.len() <= new {
                        let diff = new - base.len();
                        let remove_blocks = (diff) / block_size;
                        seg.block_count -= remove_blocks;
                    }
                }
                Segment::Unknown(_) => {}
            }
        }

        Ok(Difference {
            segments,
            block_size,
            original_data_len: self.original_data_len,
            parallel_data: None,
        })
    }
}
impl<S: ExtendVec> Difference<S> {
    /// Create an empty diff that does nothing when applied.
    ///
    /// `len` is the length of the data, both before and after.
    /// If this is smaller than the target, it'll truncate it
    /// (in whole `block_size` segments). If `len` is longer,
    /// applying will return an error.
    ///
    /// `block_size` is simply the size of the blocks.
    /// Not that relevant here, but used by other methods on [`Difference`].
    ///
    /// # Panics
    ///
    /// Panics if `block_size` is `0`.
    pub fn empty(len: usize, block_size: usize) -> Self {
        assert!(block_size > 0);
        // (len/block_size).ceil
        let block_count = (len + block_size - 1) / block_size;
        Self {
            block_size,
            original_data_len: len,
            segments: vec![Segment::Ref(SegmentRef {
                start: 0,
                block_count,
            })],
            parallel_data: None,
        }
    }
    /// Returns a reference to all the internal [`Segment`]s.
    ///
    /// This can be used for implementing algorithms other than [`Self::apply`] to apply the data.
    ///
    /// > `agde` uses this to convert from this format to their `Section` style.
    pub fn segments(&self) -> &[Segment<S>] {
        &self.segments
    }
    /// Returns a mutable reference to all the internal [`Segment`]s.
    ///
    /// Don't use this unless you know what you're doing.
    ///
    /// Using this function, you can change the building blocks of the diff.
    /// This can be useful for transforming it to use another [`ExtendVec`] for compact metadata
    /// storage (which can later be used to revert).
    #[must_use]
    pub fn segments_mut(&mut self) -> &mut Vec<Segment<S>> {
        &mut self.segments
    }
    /// Turns this difference into it's list of segments.
    ///
    /// Prefer to use [`Self::segments`] if you don't plan on consuming the internal data.
    #[must_use]
    pub fn into_segments(self) -> Vec<Segment<S>> {
        self.segments
    }
    /// Map the [`ExtendVec`]s of the [`Segment::Unknown`]s with `f`.
    ///
    /// The second argument of `f` is the start of `S` the applied data.
    ///
    /// Creates a new difference with the same length and values.
    /// Consider using [`Difference::map`].
    ///
    /// # Panics
    ///
    /// Panics if `NS` doesn't have the same [`ExtendVec::len`] as `S`.
    pub fn map_ref<NS: ExtendVec + 'static>(
        &self,
        mut f: impl FnMut(&S, usize) -> NS,
    ) -> Difference<NS> {
        let Self {
            segments,
            block_size,
            original_data_len,
            parallel_data: clamp_ref_lengths_to_parallel_block_size,
        } = self;

        let mut counter = 0;
        Difference {
            segments: segments
                .iter()
                .map(|seg| match seg {
                    Segment::Unknown(seg) => {
                        let len = seg.source().len();
                        let new = f(seg.source(), counter);
                        assert_eq!(len, new.len());
                        counter += len;
                        Segment::Unknown(SegmentUnknown { source: new })
                    }
                    Segment::Ref(r) => {
                        counter += r.len(*block_size);
                        Segment::Ref(*r)
                    }
                })
                .collect(),
            block_size: *block_size,
            original_data_len: *original_data_len,
            parallel_data: *clamp_ref_lengths_to_parallel_block_size,
        }
    }
    /// Map the [`ExtendVec`]s of the [`Segment::Unknown`]s with `f`.
    ///
    /// The second argument of `f` is the start of `S` the applied data.
    ///
    /// Creates a new difference with the same length and values.
    /// Still allocates [`Difference::segments`] vector.
    /// Consider using [`Difference::map_ref`].
    ///
    /// # Panics
    ///
    /// Panics if `NS` doesn't have the same [`ExtendVec::len`] as `S`.
    pub fn map<NS: ExtendVec + 'static>(self, mut f: impl FnMut(S, usize) -> NS) -> Difference<NS> {
        let Self {
            segments,
            block_size,
            original_data_len,
            parallel_data: clamp_ref_lengths_to_parallel_block_size,
        } = self;

        let mut counter = 0;

        Difference {
            segments: segments
                .into_iter()
                .map(|seg| match seg {
                    Segment::Unknown(seg) => {
                        let len = seg.source().len();
                        let new = f(seg.source, counter);
                        assert_eq!(len, new.len());
                        counter += len;
                        Segment::Unknown(SegmentUnknown { source: new })
                    }
                    Segment::Ref(r) => {
                        counter += r.len(block_size);
                        Segment::Ref(r)
                    }
                })
                .collect(),
            block_size,
            original_data_len,
            parallel_data: clamp_ref_lengths_to_parallel_block_size,
        }
    }
    /// The block size used by this diff.
    #[must_use]
    #[inline]
    pub fn block_size(&self) -> usize {
        self.block_size
    }
    /// Approximates the size this takes up in memory or when serializing it with a binary
    /// serializer.
    /// The returned value is in bytes.
    #[must_use]
    pub fn approximate_binary_size(&self) -> usize {
        let mut total = 0;
        // The vec is 3 usizes, and we store two more. On 64-bit platforms (which I assume), these
        // take up 8 bytes each.
        total += (3 + 2) * 8;
        for seg in &self.segments {
            // 2 because `&[u8]` is 2 bytes wide and
            // SegmentRef contains two usizes.
            total += 8 * 2;
            match seg {
                Segment::Ref(_) => {}
                Segment::Unknown(seg) => total += seg.source().len(),
            }
        }
        total
    }
    /// Get the length of the original data - the data fed to [`SignatureBuilder::write`].
    #[must_use]
    #[inline]
    pub fn original_data_len(&self) -> usize {
        self.original_data_len
    }
    /// Set the length of the original data.
    ///
    /// See [`Self::original_data_len`] for more details.
    #[inline]
    pub fn set_original_data_len(&mut self, original_data_len: usize) {
        self.original_data_len = original_data_len;
    }

    /// Set the block size.
    ///
    /// Use [`Difference::minify`] to also trim down the size of the contained data.
    ///
    /// `block_size` must be small than [`Self::block_size`].
    /// [`Self::block_size`] must also be dividable by it.
    ///
    /// # Errors
    ///
    /// Returns [`MinifyError::NewLarger`] if `block_size` >= [`Self::block_size`] and
    /// [`MinifyError::NotMultiple`] if [`Self::block_size`] is not a multiple of `block_size`.
    /// Returns [`MinifyError::Zero`] if `block_size == 0`.
    pub fn with_block_size(&mut self, block_size: usize) -> Result<(), MinifyError> {
        if block_size == 0 {
            return Err(MinifyError::Zero);
        }
        if self.block_size() == block_size {
            return Ok(());
        }
        if self.block_size() <= block_size {
            return Err(MinifyError::NewLarger);
        }
        let block_size_shrinkage = self.block_size() / block_size;
        let block_size_shrinkage_remainder = self.block_size() % block_size;
        if block_size_shrinkage_remainder != 0 {
            return Err(MinifyError::NotMultiple);
        }

        for seg in &mut self.segments {
            match seg {
                Segment::Ref(ref_seg) => {
                    ref_seg.multiply(block_size_shrinkage);
                }
                Segment::Unknown(_) => {}
            }
        }

        self.block_size = block_size;

        Ok(())
    }

    /// Returns whether or not applying this diff is impossible to do on a single [`Vec`].
    ///
    /// `base_len` is the length of `base` passed to [`Self::apply`] or [`Self::apply_in_place`].
    ///
    /// If the returned value is `true`, the apply function will try to read data from parts of the
    /// `Vec` already overridden.
    /// You should be able to use a single `Vec` if the returned value is `false`.
    ///
    /// # Examples
    ///
    /// ```
    /// # use den::*;
    /// let base_data = b"This is a document everyone has. It's about some new difference library.";
    /// let target_data = b"This is a document only I have. It's about some new difference library.";
    /// let mut base_data = base_data.to_vec();
    ///
    /// let mut signature = Signature::new(128);
    /// signature.write(&base_data);
    /// let signature = signature.finish();
    ///
    /// let diff = signature.diff(target_data);
    ///
    /// // This is the small diff you could serialize with Serde and send.
    /// let minified = diff.minify(8, &base_data)
    ///     .expect("This won't panic, as the data hasn't changed from calling the other functions.");
    ///
    /// let data = if minified.apply_overlaps(base_data.len()) {
    ///     let mut data = Vec::new();
    ///     minified.apply(&base_data, &mut data);
    ///     data
    /// } else {
    ///     minified.apply_in_place(&mut base_data);
    ///     base_data
    /// };
    ///
    /// assert_eq!(data, target_data);
    /// ```
    #[must_use]
    pub fn apply_overlaps(&self, base_len: usize) -> bool {
        self._apply_overlaps(base_len, false)
    }
    /// Returns whether or not applying this diff is impossible to do on a single [`Vec`].
    ///
    /// `base_len` is the length of `base` passed to [`Self::apply`] or [`Self::apply_in_place`].
    ///
    /// If the returned value is `true`, the apply function will try to read data from parts of the
    /// `Vec` already overridden.
    /// You should be able to use a single `Vec` if the returned value is `false`.
    ///
    /// The adaptive end part means that if `base` isn't the same as fed to
    /// [`SignatureBuilder::write`], we'll try to write any additional written data to the end.
    /// This makes it feasible to apply diffs after modifying `base`, without all
    /// modifications being overridden.
    ///
    /// # Examples
    ///
    /// See [`Self::apply_overlaps`].
    #[must_use]
    pub fn apply_overlaps_adaptive_end(&self, base_len: usize) -> bool {
        self._apply_overlaps(base_len, true)
    }
    #[allow(clippy::inline_always)]
    #[inline(always)]
    fn _apply_overlaps(&self, base_len: usize, adaptive_end: bool) -> bool {
        let previous_data_end = self.original_data_len();
        let mut position = 0;
        let mut block_len = 0;
        let mut found_end = false;

        for (seg_idx, segment) in self.segments().iter().enumerate() {
            match segment {
                Segment::Ref(seg) => {
                    if seg.start() < position {
                        return true;
                    }
                    let mut seg_end = seg.end(self.block_size());

                    block_len += seg_end.saturating_sub(seg.start());
                    if let Some(clamp) = self.parallel_data {
                        if block_len >= clamp.parallel_block_size.get()
                            && seg_idx + clamp.last_block_segment_length < self.segments.len()
                        {
                            seg_end -= block_len - clamp.parallel_block_size.get();
                            block_len = 0;
                        }
                    }

                    if adaptive_end {
                        // `ref_segment` is to the very end.
                        if seg_end >= previous_data_end {
                            // Then, max out end, even if new end is after the previous data's end.
                            let end = seg_end.min(base_len);
                            let additional = base_len - end;
                            found_end = true;
                            position += additional;
                        }
                    }

                    position += seg_end - seg.start();
                }
                Segment::Unknown(seg) => {
                    position += seg.source().len();
                    block_len += seg.source().len();
                }
            }
        }
        if adaptive_end && !found_end {
            // only if position is after previous_data_end, we can't copy within slice.
            return position > previous_data_end;
        }
        false
    }

    /// Apply `diff` to the `base` data base, appending the result to `out`.
    ///
    /// Consider checking [`Self::apply_overlaps`] and calling [`Self::apply_in_place`]
    /// to remove the need for the [`Vec`].
    ///
    /// # Security
    ///
    /// The `diff` should be sanitized if input is suspected to be malicious.
    ///
    /// # Errors
    ///
    /// Returns [`ApplyError::RefOutOfBounds`] if a reference is out of bounds of the `base`.
    /// If `base` is the same data written to [`SignatureBuilder::write`], this error will not
    /// occur, granted this diff is derived from that [`Signature`].
    ///
    /// Ok to unwrap if [`Self::in_bounds`].
    pub fn apply(&self, base: &[u8], out: &mut Vec<u8>) -> Result<(), ApplyError> {
        self._apply(base, out, false)
    }
    /// Apply `diff` to the `base` data base, appending the result to `out`.
    ///
    /// Consider checking [`Self::apply_overlaps`] and calling [`Self::apply_in_place`]
    /// to remove the need for the [`Vec`].
    ///
    /// The adaptive end part means that if `base` isn't the same as fed to
    /// [`SignatureBuilder::write`], we'll try to write any additional written data to the end.
    /// This makes it feasible to apply diffs after modifying `base`, without all
    /// modifications being overridden.
    ///
    /// # Security
    ///
    /// The `diff` should be sanitized if input is suspected to be malicious.
    ///
    /// # Errors
    ///
    /// Returns [`ApplyError::RefOutOfBounds`] if a reference is out of bounds of the `base`.
    /// If `base` is the same data written to [`SignatureBuilder::write`], this error will not
    /// occur, granted this diff is derived from that [`Signature`].
    ///
    /// Ok to unwrap if [`Self::in_bounds`].
    pub fn apply_adaptive_end(&self, base: &[u8], out: &mut Vec<u8>) -> Result<(), ApplyError> {
        self._apply(base, out, true)
    }
    #[inline(always)]
    fn _apply(&self, base: &[u8], out: &mut Vec<u8>, adaptive_end: bool) -> Result<(), ApplyError> {
        use ApplyError::RefOutOfBounds as Roob;

        let block_size = self.block_size();
        let previous_data_end = self.original_data_len();
        let mut found_end = false;

        let mut block_len = 0;

        for (seg_idx, segment) in self.segments().iter().enumerate() {
            match segment {
                Segment::Ref(ref_segment) => {
                    let start = ref_segment.start;
                    let mut seg_end = ref_segment.end(block_size);

                    block_len += seg_end.saturating_sub(start);
                    if let Some(clamp) = self.parallel_data {
                        if block_len >= clamp.parallel_block_size.get()
                            && seg_idx + clamp.last_block_segment_length < self.segments.len()
                        {
                            seg_end -= block_len - clamp.parallel_block_size.get();
                            block_len = 0;
                        }
                    }

                    let mut end = seg_end.min(base.len());

                    if adaptive_end {
                        // `ref_segment` is to the very end.
                        if seg_end >= previous_data_end {
                            // Then, max out end, even if new end is after the previous data's end.
                            end = base.len();
                            found_end = true;
                        }
                    }

                    let data = base.get(start..end).ok_or(Roob)?;
                    data.extend(out);
                }
                Segment::Unknown(unknown_segment) => {
                    let data = &unknown_segment.source;
                    block_len += data.len();
                    data.extend(out);
                }
            }
        }

        if adaptive_end {
            // previous_data_end is the length of the data written to Signature,
            // base is the new, potentially modified data.
            //
            // If we've added to the end, push the additional bytes to `out`, as a last-effort
            // to not lose any data.
            //
            // This will only occur if no other ref covered the end
            if !found_end && base.len() > previous_data_end {
                out.extend_from_slice(&base[previous_data_end..]);
            }
        }

        Ok(())
    }
    /// Apply `diff` to the `base` data base, mutating the `base` data.
    /// You **MUST** check that this is possible using [`Self::apply_overlaps`].
    /// Neglecting that step will result in data loss and erroneous data.
    /// This WILL NOT give an error in case of breaking that contract.
    ///
    /// # Security
    ///
    /// The `diff` should be sanitized if input is suspected to be malicious.
    ///
    /// # Errors
    ///
    /// If this returns an error, consider `base` to be bogus data.
    ///
    /// Returns [`ApplyError::RefOutOfBounds`] if a reference is out of bounds of the `base`.
    ///
    /// Ok to unwrap if [`Self::in_bounds`].
    ///
    /// # Examples
    ///
    /// See [`Self::apply_overlaps`].
    pub fn apply_in_place(&self, base: &mut Vec<u8>) -> Result<(), ApplyError> {
        self._apply_in_place(base, false)
    }
    /// Apply `diff` to the `base` data base, mutating the `base` data.
    /// You **MUST** check that this is possible using [`Self::apply_overlaps_adaptive_end`].
    /// Neglecting that step will result in data loss and erroneous data.
    /// This WILL NOT give an error in case of breaking that contract.
    ///
    /// The adaptive end part means that if `base` isn't the same as fed to
    /// [`SignatureBuilder::write`], we'll try to write any additional written data to the end.
    /// This makes it feasible to apply diffs after modifying `base`, without all
    /// modifications being overridden.
    ///
    /// # Security
    ///
    /// The `diff` should be sanitized if input is suspected to be malicious.
    ///
    /// # Errors
    ///
    /// If this returns an error, consider `base` to be bogus data.
    ///
    /// Returns [`ApplyError::RefOutOfBounds`] if a reference is out of bounds of the `base`.
    ///
    /// Ok to unwrap if [`Self::in_bounds`].
    ///
    /// # Examples
    ///
    /// See [`Self::apply_overlaps`].
    pub fn apply_in_place_adaptive_end(&self, base: &mut Vec<u8>) -> Result<(), ApplyError> {
        self._apply_in_place(base, true)
    }
    #[inline(always)]
    fn _apply_in_place(&self, base: &mut Vec<u8>, adaptive_end: bool) -> Result<(), ApplyError> {
        #[allow(clippy::uninit_vec)] // we know what we're doing
        fn copy_within_vec<T: Copy>(vec: &mut Vec<T>, range: Range<usize>, position: usize) {
            let range_len = range.len();
            let new_len = (position + range_len).max(vec.len());
            // SAFETY: This guarantees `vec.capacity()` >= `vec.len() + slice.len()`
            vec.reserve(new_len - vec.len());
            // SAFETY: We set the length to what we write below.
            unsafe { vec.set_len(new_len) };
            vec.copy_within(range, position);
        }
        use ApplyError::RefOutOfBounds as Roob;

        let needed = self.applied_len(base);
        if needed > base.len() {
            base.reserve(needed - base.len());
        }

        let block_size = self.block_size();
        let previous_data_end = self.original_data_len();

        let mut position = 0;
        let mut block_len = 0;
        let mut found_end = false;

        for (seg_idx, segment) in self.segments().iter().enumerate() {
            match segment {
                Segment::Ref(ref_segment) => {
                    let start = ref_segment.start;
                    let mut seg_end = ref_segment.end(block_size);

                    block_len += seg_end.saturating_sub(start);
                    if let Some(clamp) = self.parallel_data {
                        if block_len >= clamp.parallel_block_size.get()
                            && seg_idx + clamp.last_block_segment_length < self.segments.len()
                        {
                            seg_end -= block_len - clamp.parallel_block_size.get();
                            block_len = 0;
                        }
                    }

                    let mut end = seg_end.min(base.len());

                    if adaptive_end {
                        // `ref_segment` is to the very end.
                        if seg_end >= previous_data_end {
                            // Then, max out end, even if new end is after the previous data's end.
                            end = base.len();
                            found_end = true;
                        }
                    }

                    let range = start..end;
                    base.get(range.clone()).ok_or(Roob)?;
                    copy_within_vec(base, range, position);
                    position += end - start;
                }
                Segment::Unknown(unknown_segment) => {
                    let data = &unknown_segment.source;
                    block_len += data.len();
                    data.replace(base, position);
                    position += data.len();
                }
            }
        }

        if !found_end && position < previous_data_end && previous_data_end < Vec::len(base) {
            let len = Vec::len(base);
            copy_within_vec(base, previous_data_end..len, position);
            position += len - previous_data_end;
        }

        // shorten the vec if the new length is less than old
        base.truncate(position);

        Ok(())
    }

    /// Reverts `current` to what it would have been before applying this diff.
    /// This completely overrides `target`.
    ///
    /// Say I have the data: `hello there - from vim`
    /// and the diff is `[Segment::Ref(0..8)]`, the new data (`current`) will be
    /// `hello th`. The length of the original data is stored, but the lost bytes are
    /// unrecoverable. They are filled with `fill_byte`. That should probably be `b' '` for text
    /// applications.
    ///
    /// # Security
    ///
    /// The `diff` should be sanitized if input is suspected to be malicious.
    ///
    /// # Errors
    ///
    /// Returns [`ApplyError::RefOutOfBounds`] if a reference is out of bounds of the `base`.
    #[allow(clippy::uninit_vec)] // we know what we're doing
    pub fn revert(
        &self,
        current: &[u8],
        target: &mut Vec<u8>,
        fill_byte: u8,
    ) -> Result<(), ApplyError> {
        use ApplyError::RefOutOfBounds as Roob;

        // Make buffer contain `current` and then optionally `fill_byte`s.
        target.clear();
        let new_len = current.len().max(self.original_data_len);
        target.reserve(new_len);
        // SAFETY: we copy the data up to `current.len()`
        // if new_len > current.len, the rest is filled below.
        unsafe { target.set_len(new_len) };
        target[..current.len()].copy_from_slice(current);

        target[current.len()..].fill(fill_byte);

        let block_size = self.block_size();

        let mut cursor = 0;
        let mut block_len = 0;

        for (seg_idx, segment) in self.segments().iter().enumerate() {
            match segment {
                Segment::Unknown(unknown) => {
                    cursor += unknown.source().len();
                    block_len += unknown.source().len();
                }
                Segment::Ref(seg) => {
                    let end = seg.end(block_size);
                    let missing = end.checked_sub(target.len());
                    let mut offset = 0;
                    if let Some(missing) = missing {
                        offset = missing.min(block_size);
                    }
                    let current_end = cursor + seg.len(block_size) - offset;
                    let mut current_end = current_end.min(current.len());
                    block_len += current_end - cursor;
                    if let Some(clamp) = self.parallel_data {
                        if block_len >= clamp.parallel_block_size.get()
                            && seg_idx + clamp.last_block_segment_length < self.segments.len()
                        {
                            current_end -= block_len - clamp.parallel_block_size.get();
                            block_len = 0;
                        }
                    }

                    if current_end > current.len() {
                        return Err(Roob);
                    }
                    target[seg.start()..seg.start() + current_end - cursor]
                        .copy_from_slice(&current[cursor..current_end]);
                    cursor += seg.len(block_size);
                }
            }
        }
        target.truncate(self.original_data_len);
        Ok(())
    }

    /// Returns true if this diff is a no-op.
    #[must_use]
    pub fn is_empty(&self) -> bool {
        if self.segments().len() != 1 {
            return false;
        };
        if let Segment::Ref(seg) = self.segments()[0] {
            seg.start() == 0 && seg.end(self.block_size) >= self.original_data_len()
        } else {
            false
        }
    }

    /// Checks `self` if it gets `base` to `target`, without any allocations.
    ///
    /// Returns `true` if that's the case.
    #[must_use]
    pub fn verify(&self, base: &[u8], target: &[u8]) -> bool {
        let block_size = self.block_size();

        let mut cursor = 0;
        let mut block_len = 0;

        for (seg_idx, segment) in self.segments().iter().enumerate() {
            match segment {
                Segment::Ref(ref_segment) => {
                    let start = ref_segment.start;
                    let seg_end = ref_segment.end(block_size);
                    let mut end = seg_end.min(base.len());

                    block_len += end.saturating_sub(start);
                    if let Some(clamp) = self.parallel_data {
                        if block_len >= clamp.parallel_block_size.get()
                            && seg_idx + clamp.last_block_segment_length < self.segments.len()
                        {
                            end -= block_len - clamp.parallel_block_size.get();
                            block_len = 0;
                        }
                    }

                    let data = if let Some(data) = base.get(start..end) {
                        data
                    } else {
                        return false;
                    };
                    if target.get(cursor..cursor + data.len()) != Some(data) {
                        return false;
                    }
                    cursor += data.len();
                }
                Segment::Unknown(unknown_segment) => {
                    let data = &unknown_segment.source;
                    if target
                        .get(cursor..cursor + data.len())
                        .map_or(true, |target_data| !data.equals(target_data))
                    {
                        return false;
                    }
                    cursor += data.len();
                    block_len += data.len();
                }
            }
        }

        true
    }
    /// Calculates the length of the output of [`Self::apply`] without any allocations.
    #[must_use]
    pub fn applied_len(&self, base: &[u8]) -> usize {
        let block_size = self.block_size();
        let mut len = 0;
        let mut block_len = 0;
        // println!("new");

        for (seg_idx, segment) in self.segments().iter().enumerate() {
            match segment {
                Segment::Ref(ref_segment) => {
                    let start = ref_segment.start;
                    let seg_end = ref_segment.end(block_size);
                    let mut end = seg_end.min(base.len());

                    // let previous_bl = block_len;
                    block_len += end.saturating_sub(start);
                    // println!(
                    // "Adding len {} with previous bl {previous_bl}",
                    // end.saturating_sub(start)
                    // );
                    if let Some(clamp) = self.parallel_data {
                        if block_len >= clamp.parallel_block_size.get()
                            && seg_idx + clamp.last_block_segment_length < self.segments.len()
                        {
                            end -= block_len - clamp.parallel_block_size.get();
                            // println!(
                            // "Reset. Rm: {}, This block: {}",
                            // block_len - clamp.parallel_block_size.get(),
                            // previous_bl + end.saturating_sub(start)
                            // );
                            block_len = 0;
                        }
                    }

                    len += end.saturating_sub(start);
                }
                Segment::Unknown(unknown_segment) => {
                    let data = &unknown_segment.source;
                    block_len += data.len();
                    // println!("Adding len {} unknown", data.len());
                    len += data.len();
                }
            }
        }
        len
    }
    /// Checks `self` only contains valid references to `base`.
    /// There will not occur any error from apply functions if this returns true.
    ///
    /// Returns `true` if that's the case.
    #[must_use]
    pub fn in_bounds(&self, base: &[u8]) -> bool {
        let block_size = self.block_size();

        for segment in self.segments() {
            match segment {
                Segment::Ref(ref_segment) => {
                    let start = ref_segment.start;
                    let seg_end = ref_segment.end(block_size);
                    let end = seg_end.min(base.len());

                    if base.get(start..end).is_none() {
                        return false;
                    };
                }
                Segment::Unknown(_unknown_segment) => {}
            }
        }

        true
    }
}
impl<S: ExtendVec + AsRef<[u8]>> Debug for Difference<S> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        if let Some(parallel_data) = self.parallel_data {
            write!(
                f,
                "parallel diff layout (parallel_block_size: {}, last_block_segment_length: {})",
                parallel_data.parallel_block_size, parallel_data.last_block_segment_length
            )?;
        } else {
            write!(f, "serial diff layout")?;
        }
        if f.alternate() {
            f.write_str(":\n")?;
        } else {
            f.write_str(": ")?;
        }
        f.write_str("[")?;
        if f.alternate() {
            f.write_str("\n")?;
        }
        for (idx, seg) in self.segments.iter().enumerate() {
            if f.alternate() {
                f.write_str("    ")?;
            }
            seg.fmt(f)?;
            if idx + 1 != self.segments.len() {
                if f.alternate() {
                    f.write_str(",\n")?;
                } else {
                    f.write_str(", ")?;
                }
            }
        }
        if f.alternate() {
            f.write_str("\n")?;
        }
        f.write_str("]")?;
        if f.alternate() {
            f.write_str(",\n")?;
        } else {
            f.write_str(", ")?;
        }
        write!(f, "block_size: {}", self.block_size)?;
        if f.alternate() {
            f.write_str(",\n")?;
        } else {
            f.write_str(", ")?;
        }
        write!(f, "original_data_len: {}", self.original_data_len)?;
        Ok(())
    }
}

/// An error during [`Difference::minify`].
#[derive(Debug, PartialEq, Eq)]
pub enum MinifyError {
    /// New block size is larger than previous.
    NewLarger,
    /// Old block size is not a multiple of the new.
    NotMultiple,
    /// Successive unknown segments are not allowed and **SHOULD** never occur.
    SuccessiveUnknowns,
    /// New block size is 0. That isn't possible!
    Zero,
}

/// An error during [`Difference::apply`].
#[derive(Debug, PartialEq, Eq)]
pub enum ApplyError {
    /// The reference is out of bounds.
    ///
    /// The data might be malicious or corrupted or the `base` data has changed from constructing
    /// [`Signature`] and [`Difference::apply`].
    ///
    /// Previously, this was also thrown when [`SegmentRef::block_count`] caused the end to extend
    /// past the data more than `block_size`. This has since been relaxed, but it's a good habit to
    /// only let it extend up to one `block_size`, as it can save space when serializing.
    /// The [`SegmentRef`] can go past the end to accommodate the case where no data in the end has
    /// changed. We then simply continue the reference and use all the available data when calling
    /// [`Difference::apply`].
    RefOutOfBounds,
}

#[derive(Debug, Clone)]
struct Blocks<'a, T> {
    slice: &'a [T],
    block_size: usize,
    pos: usize,
}
impl<'a, T> Blocks<'a, T> {
    fn new(slice: &'a [T], block_size: usize) -> Self {
        Self {
            slice,
            block_size,
            pos: 0,
        }
    }
    #[inline]
    fn advance(&mut self, n: usize) {
        self.pos += n;
    }
    /// Decrease the inner position. If `n == 1`, you revert the `.next()` call, resulting in
    /// repeated slices being returned.
    #[inline]
    fn go_back(&mut self, n: usize) {
        self.pos -= n;
    }
    /// Where we're at in the original slice.
    ///
    /// Clamped to `slice.len()`.
    #[inline]
    fn pos(&self) -> usize {
        cmp::min(self.pos, self.slice.len())
    }
}
impl<'a, T> Iterator for Blocks<'a, T> {
    type Item = &'a [T];
    #[inline]
    fn next(&mut self) -> Option<Self::Item> {
        if self.pos + 1 > self.slice.len() {
            return None;
        }

        let start = self.pos;
        let end = cmp::min(start + self.block_size, self.slice.len());

        self.advance(1);

        Some(&self.slice[start..end])
    }
}
struct PrePostWindow<'a, T> {
    slice: &'a [T],
    pos: usize,
}
impl<'a, T> PrePostWindow<'a, T> {
    fn new(slice: &'a [T]) -> Self {
        Self { slice, pos: 0 }
    }
}
impl<'a, T> Iterator for PrePostWindow<'a, T> {
    type Item = (Option<&'a T>, &'a T, Option<&'a T>);
    fn next(&mut self) -> Option<Self::Item> {
        let current = self.slice.get(self.pos)?;
        let before = self.pos.checked_sub(1).and_then(|pos| self.slice.get(pos));
        let after = self.slice.get(self.pos + 1);
        self.pos += 1;
        Some((before, current, after))
    }
}