rscrypto 0.1.1

//! rapidhash V3 (**NOT CRYPTO**).
//!
//! Portable scalar implementation.
//!
//! Two variants are provided:
//!
//! - **`RapidHash64` / `RapidHash128`** — the standard V3 algorithm with avalanche finisher (extra
//!   `rapid_mix`). C++-compatible output. Use when you need stable, cross-language hash values.
//!
//! - **`RapidHashFast64` / `RapidHashFast128`** — V3 core with avalanche disabled. Saves one
//!   128-bit multiply at finalization. Use when you only need a fast in-process hash (e.g. hash
//!   maps) and don't need C++ compatibility.

#![allow(clippy::indexing_slicing)] // Tight block parsing

use crate::traits::FastHash;

#[doc(hidden)]
pub(crate) mod dispatch;
#[cfg(all(feature = "rapidhash", any(test, feature = "diag")))]
#[doc(hidden)]
pub(crate) mod dispatch_tables;
#[cfg(all(feature = "rapidhash", any(test, feature = "diag")))]
pub(crate) mod kernels;

/// Standard V3 rapidhash (64-bit) with avalanche finisher.
#[derive(Clone, Debug, Default)]
pub struct RapidHash64;

/// Standard V3 rapidhash (128-bit) with avalanche finisher.
#[derive(Clone, Debug, Default)]
pub struct RapidHash128;

/// Fast V3 rapidhash (64-bit) — avalanche disabled for maximum throughput.
#[derive(Clone, Debug, Default)]
pub struct RapidHashFast64;

/// Fast V3 rapidhash (128-bit) — avalanche disabled for maximum throughput.
#[derive(Clone, Debug, Default)]
pub struct RapidHashFast128;

// rapidhash v3 default secrets (C++ compatible).
const DEFAULT_SECRETS: [u64; 7] = [
  0x2d35_8dcc_aa6c_78a5,
  0x8bb8_4b93_962e_acc9,
  0x4b33_a62e_d433_d4a3,
  0x4d5a_2da5_1de1_aa47,
  0xa076_1d64_78bd_642f,
  0xe703_7ed1_a0b4_28db,
  0x90ed_1765_281c_388c,
];

#[cfg(any(target_arch = "s390x", target_arch = "powerpc64"))]
struct PrecomputedSecrets {
  seed: u64,
  secrets: [u64; 7],
}

// Default paths use referenced precomputed state to match upstream's
// `RapidSecrets` code shape on targets where materializing many 64-bit
// constants is costly.
#[cfg(any(target_arch = "s390x", target_arch = "powerpc64"))]
static V3_DEFAULT_SECRETS: PrecomputedSecrets = PrecomputedSecrets {
  seed: PRECOMPUTED_SEED_0,
  secrets: DEFAULT_SECRETS,
};

#[cfg(any(target_arch = "s390x", target_arch = "powerpc64"))]
static V3_DEFAULT_HI_SECRETS: PrecomputedSecrets = PrecomputedSecrets {
  seed: PRECOMPUTED_SEED_HI,
  secrets: DEFAULT_SECRETS,
};

#[cfg(any(target_arch = "s390x", target_arch = "powerpc64"))]
static FAST_DEFAULT_SECRETS: [u64; 7] = DEFAULT_SECRETS;

#[inline(always)]
fn default_fast_secret<const IDX: usize>() -> u64 {
  debug_assert!(IDX < DEFAULT_SECRETS.len());

  #[cfg(any(target_arch = "s390x", target_arch = "powerpc64"))]
  {
    FAST_DEFAULT_SECRETS[IDX]
  }

  #[cfg(not(any(target_arch = "s390x", target_arch = "powerpc64")))]
  {
    DEFAULT_SECRETS[IDX]
  }
}

#[inline(always)]
#[cfg(any(target_arch = "s390x", target_arch = "powerpc64"))]
fn v3_secret<const IDX: usize>(secrets: &[u64; 7]) -> u64 {
  debug_assert!(IDX < DEFAULT_SECRETS.len());
  secrets[IDX]
}

#[inline(always)]
fn read_u32_le(input: &[u8], offset: usize) -> u32 {
  debug_assert!(offset + 4 <= input.len());
  // SAFETY: caller ensures `offset + 4 <= input.len()`, and `read_unaligned` supports unaligned
  // loads.
  let v = unsafe { core::ptr::read_unaligned(input.as_ptr().add(offset) as *const u32) };
  u32::from_le(v)
}

#[inline(always)]
fn read_u64_le(input: &[u8], offset: usize) -> u64 {
  debug_assert!(offset + 8 <= input.len());
  // SAFETY: caller ensures `offset + 8 <= input.len()`, and `read_unaligned` supports unaligned
  // loads.
  let v = unsafe { core::ptr::read_unaligned(input.as_ptr().add(offset) as *const u64) };
  u64::from_le(v)
}

/// Native-endian u32 read. Used by the Fast variant only — the Fast variant is
/// for in-process hashing and does not need cross-platform byte-order consistency.
/// Skipping the byte-swap eliminates LRVG overhead on big-endian platforms (s390x,
/// POWER), matching the competitor's `read_u32_np`.
#[inline(always)]
fn read_u32_np(input: &[u8], offset: usize) -> u32 {
  debug_assert!(offset + 4 <= input.len());
  // SAFETY: caller ensures `offset + 4 <= input.len()`, and `read_unaligned` supports unaligned
  // loads.
  unsafe { core::ptr::read_unaligned(input.as_ptr().add(offset) as *const u32) }
}

/// Native-endian u64 read. Used by the Fast variant only — see `read_u32_np`.
#[inline(always)]
fn read_u64_np(input: &[u8], offset: usize) -> u64 {
  debug_assert!(offset + 8 <= input.len());
  // SAFETY: caller ensures `offset + 8 <= input.len()`, and `read_unaligned` supports unaligned
  // loads.
  unsafe { core::ptr::read_unaligned(input.as_ptr().add(offset) as *const u64) }
}

#[inline(always)]
const fn rapid_mum(a: u64, b: u64) -> (u64, u64) {
  let r = (a as u128).wrapping_mul(b as u128);
  (r as u64, (r >> 64) as u64)
}

#[inline(always)]
const fn rapid_mix(a: u64, b: u64) -> u64 {
  let r = (a as u128).wrapping_mul(b as u128);
  (r as u64) ^ ((r >> 64) as u64)
}

/// Hint that `cond` is likely true. Uses `#[cold]` to nudge branch layout.
#[inline(always)]
fn likely(cond: bool) -> bool {
  if !cond {
    cold_path();
  }
  cond
}

/// Hint that `cond` is likely false.
#[inline(always)]
fn unlikely(cond: bool) -> bool {
  if cond {
    cold_path();
  }
  cond
}

#[cold]
#[inline(always)]
fn cold_path() {}

#[inline(always)]
const fn rapidhash_seed_cpp(seed: u64) -> u64 {
  seed ^ rapid_mix(seed ^ DEFAULT_SECRETS[2], DEFAULT_SECRETS[1])
}

const V3_HI_SEED: u64 = 0x9E37_79B9_7F4A_7C15;

/// Precomputed seed for the default case (seed = 0). Eliminates one 128-bit
/// multiply per call when using `FastHash::hash()` (the common path).
const PRECOMPUTED_SEED_0: u64 = rapidhash_seed_cpp(0);
const PRECOMPUTED_SEED_HI: u64 = rapidhash_seed_cpp(V3_HI_SEED);

#[inline(always)]
const fn rapidhash_fast_default_empty(seed: u64) -> u64 {
  rapid_mix(DEFAULT_SECRETS[0], seed)
}

const FAST_EMPTY_HASH_SEED_0: u64 = rapidhash_fast_default_empty(PRECOMPUTED_SEED_0);
const FAST_EMPTY_HASH_SEED_HI: u64 = rapidhash_fast_default_empty(PRECOMPUTED_SEED_HI);

#[inline(always)]
#[cfg(any(target_arch = "s390x", target_arch = "powerpc64"))]
fn rapidhash_finish(a: u64, b: u64, remainder: u64, secrets: &[u64; 7]) -> u64 {
  rapid_mix(a ^ 0xaaaa_aaaa_aaaa_aaaa, b ^ v3_secret::<1>(secrets) ^ remainder)
}

#[inline(always)]
#[cfg(not(any(target_arch = "s390x", target_arch = "powerpc64")))]
fn rapidhash_finish_default(a: u64, b: u64, remainder: u64) -> u64 {
  rapid_mix(a ^ 0xaaaa_aaaa_aaaa_aaaa, b ^ DEFAULT_SECRETS[1] ^ remainder)
}

#[inline(always)]
fn rapidhash_v3_with_seed(data: &[u8], seed: u64) -> u64 {
  rapidhash_core_default::<true>(data, rapidhash_seed_cpp(seed))
}

#[inline(always)]
fn rapidhash_fast_with_seed(data: &[u8], seed: u64) -> u64 {
  rapidhash_fast_core(data, rapidhash_seed_cpp(seed))
}

#[inline(always)]
fn rapidhash_fast_128_with_seed(data: &[u8], seed: u64) -> u128 {
  let seed_lo = rapidhash_seed_cpp(seed);
  let seed_hi = rapidhash_seed_cpp(seed ^ V3_HI_SEED);
  rapidhash_fast_128_core(data, seed_lo, seed_hi)
}

#[inline(always)]
#[cfg(any(target_arch = "s390x", target_arch = "powerpc64"))]
fn rapidhash_core<const AVALANCHE: bool>(data: &[u8], mut seed: u64, secrets: &[u64; 7]) -> u64 {
  let mut a = 0u64;
  let mut b = 0u64;

  if likely(data.len() <= 16) {
    // Small path: 0-16 bytes. Fully inline for minimum overhead.
    if data.len() >= 4 {
      seed ^= data.len() as u64;
      if data.len() >= 8 {
        a = read_u64_le(data, 0);
        b = read_u64_le(data, data.len() - 8);
      } else {
        a = read_u32_le(data, 0) as u64;
        b = read_u32_le(data, data.len() - 4) as u64;
      }
    } else if !data.is_empty() {
      a = ((data[0] as u64) << 45) | (data[data.len() - 1] as u64);
      b = data[data.len() >> 1] as u64;
    }
  } else {
    // Medium/large path: >16 bytes. Keep the 7-stream bulk loop out of the
    // 17-112B path so fixed medium inputs don't pay large-handler setup.
    // SAFETY: we just verified data.len() > 16.
    unsafe {
      if data.len() <= 112 {
        return rapidhash_core_medium::<AVALANCHE>(data, seed, secrets);
      }
      return rapidhash_core_large::<AVALANCHE>(data, seed, secrets);
    }
  }

  let remainder = data.len() as u64;
  a ^= v3_secret::<1>(secrets);
  b ^= seed;
  (a, b) = rapid_mum(a, b);
  rapidhash_final::<AVALANCHE>(a, b, remainder, secrets)
}

#[inline(always)]
fn rapidhash_core_default<const AVALANCHE: bool>(data: &[u8], seed: u64) -> u64 {
  #[cfg(any(target_arch = "s390x", target_arch = "powerpc64"))]
  {
    rapidhash_core::<AVALANCHE>(data, seed, &V3_DEFAULT_SECRETS.secrets)
  }

  #[cfg(not(any(target_arch = "s390x", target_arch = "powerpc64")))]
  {
    let mut seed = seed;
    let mut a = 0u64;
    let mut b = 0u64;

    if likely(data.len() <= 16) {
      if data.len() >= 4 {
        seed ^= data.len() as u64;
        if data.len() >= 8 {
          a = read_u64_le(data, 0);
          b = read_u64_le(data, data.len() - 8);
        } else {
          a = read_u32_le(data, 0) as u64;
          b = read_u32_le(data, data.len() - 4) as u64;
        }
      } else if !data.is_empty() {
        a = ((data[0] as u64) << 45) | (data[data.len() - 1] as u64);
        b = data[data.len() >> 1] as u64;
      }

      let remainder = data.len() as u64;
      a ^= DEFAULT_SECRETS[1];
      b ^= seed;
      (a, b) = rapid_mum(a, b);
      return rapidhash_final_default::<AVALANCHE>(a, b, remainder);
    }

    // SAFETY: data.len() > 16 verified above.
    unsafe {
      if data.len() <= 112 {
        return rapidhash_core_medium_default::<AVALANCHE>(data, seed);
      }
      rapidhash_core_large_default::<AVALANCHE>(data, seed)
    }
  }
}

#[inline(always)]
#[cfg(any(target_arch = "s390x", target_arch = "powerpc64"))]
fn rapidhash_final<const AVALANCHE: bool>(a: u64, b: u64, remainder: u64, secrets: &[u64; 7]) -> u64 {
  if AVALANCHE {
    rapidhash_finish(a, b, remainder, secrets)
  } else {
    a ^ b
  }
}

#[inline(always)]
#[cfg(not(any(target_arch = "s390x", target_arch = "powerpc64")))]
fn rapidhash_final_default<const AVALANCHE: bool>(a: u64, b: u64, remainder: u64) -> u64 {
  if AVALANCHE {
    rapidhash_finish_default(a, b, remainder)
  } else {
    a ^ b
  }
}

/// Handles inputs 17-112 bytes. This keeps fixed-size medium inputs out of the
/// 7-stream large handler without changing V3 output.
///
/// # Safety
///
/// Caller must guarantee `16 < data.len() <= 112`.
#[inline(always)]
#[cfg(not(any(target_arch = "s390x", target_arch = "powerpc64")))]
unsafe fn rapidhash_core_medium_default<const AVALANCHE: bool>(data: &[u8], mut seed: u64) -> u64 {
  // SAFETY: caller guarantees 16 < data.len() <= 112. This eliminates redundant
  // bounds checks in the tail-read section below.
  unsafe {
    core::hint::assert_unchecked(data.len() > 16);
    core::hint::assert_unchecked(data.len() <= 112);
  }

  let mut a = 0u64;
  let mut b = 0u64;

  seed = rapid_mix(read_u64_le(data, 0) ^ DEFAULT_SECRETS[2], read_u64_le(data, 8) ^ seed);
  if data.len() > 32 {
    seed = rapid_mix(read_u64_le(data, 16) ^ DEFAULT_SECRETS[2], read_u64_le(data, 24) ^ seed);
    if data.len() > 48 {
      seed = rapid_mix(read_u64_le(data, 32) ^ DEFAULT_SECRETS[1], read_u64_le(data, 40) ^ seed);
      if data.len() > 64 {
        seed = rapid_mix(read_u64_le(data, 48) ^ DEFAULT_SECRETS[1], read_u64_le(data, 56) ^ seed);
        if data.len() > 80 {
          seed = rapid_mix(read_u64_le(data, 64) ^ DEFAULT_SECRETS[2], read_u64_le(data, 72) ^ seed);
          if data.len() > 96 {
            seed = rapid_mix(read_u64_le(data, 80) ^ DEFAULT_SECRETS[1], read_u64_le(data, 88) ^ seed);
          }
        }
      }
    }
  }

  let remainder = data.len() as u64;
  a ^= read_u64_le(data, data.len() - 16) ^ remainder;
  b ^= read_u64_le(data, data.len() - 8);

  a ^= DEFAULT_SECRETS[1];
  b ^= seed;

  (a, b) = rapid_mum(a, b);
  rapidhash_final_default::<AVALANCHE>(a, b, remainder)
}

/// Handles inputs >16 bytes. Kept as a separate function so the small-path
/// entry point stays lean for inlining. LLVM decides whether to inline this.
///
/// # Safety
///
/// Caller must guarantee `data.len() > 16`.
#[inline]
#[cfg(not(any(target_arch = "s390x", target_arch = "powerpc64")))]
unsafe fn rapidhash_core_large_default<const AVALANCHE: bool>(data: &[u8], mut seed: u64) -> u64 {
  // SAFETY: caller guarantees data.len() > 16. This eliminates redundant
  // bounds checks in the tail-read section below.
  unsafe { core::hint::assert_unchecked(data.len() > 16) };

  let mut a = 0u64;
  let mut b = 0u64;
  let mut slice = data;

  if unlikely(slice.len() > 112) {
    let mut see1 = seed;
    let mut see2 = seed;
    let mut see3 = seed;
    let mut see4 = seed;
    let mut see5 = seed;
    let mut see6 = seed;

    while slice.len() > 224 {
      seed = rapid_mix(read_u64_le(slice, 0) ^ DEFAULT_SECRETS[0], read_u64_le(slice, 8) ^ seed);
      see1 = rapid_mix(
        read_u64_le(slice, 16) ^ DEFAULT_SECRETS[1],
        read_u64_le(slice, 24) ^ see1,
      );
      see2 = rapid_mix(
        read_u64_le(slice, 32) ^ DEFAULT_SECRETS[2],
        read_u64_le(slice, 40) ^ see2,
      );
      see3 = rapid_mix(
        read_u64_le(slice, 48) ^ DEFAULT_SECRETS[3],
        read_u64_le(slice, 56) ^ see3,
      );
      see4 = rapid_mix(
        read_u64_le(slice, 64) ^ DEFAULT_SECRETS[4],
        read_u64_le(slice, 72) ^ see4,
      );
      see5 = rapid_mix(
        read_u64_le(slice, 80) ^ DEFAULT_SECRETS[5],
        read_u64_le(slice, 88) ^ see5,
      );
      see6 = rapid_mix(
        read_u64_le(slice, 96) ^ DEFAULT_SECRETS[6],
        read_u64_le(slice, 104) ^ see6,
      );

      seed = rapid_mix(
        read_u64_le(slice, 112) ^ DEFAULT_SECRETS[0],
        read_u64_le(slice, 120) ^ seed,
      );
      see1 = rapid_mix(
        read_u64_le(slice, 128) ^ DEFAULT_SECRETS[1],
        read_u64_le(slice, 136) ^ see1,
      );
      see2 = rapid_mix(
        read_u64_le(slice, 144) ^ DEFAULT_SECRETS[2],
        read_u64_le(slice, 152) ^ see2,
      );
      see3 = rapid_mix(
        read_u64_le(slice, 160) ^ DEFAULT_SECRETS[3],
        read_u64_le(slice, 168) ^ see3,
      );
      see4 = rapid_mix(
        read_u64_le(slice, 176) ^ DEFAULT_SECRETS[4],
        read_u64_le(slice, 184) ^ see4,
      );
      see5 = rapid_mix(
        read_u64_le(slice, 192) ^ DEFAULT_SECRETS[5],
        read_u64_le(slice, 200) ^ see5,
      );
      see6 = rapid_mix(
        read_u64_le(slice, 208) ^ DEFAULT_SECRETS[6],
        read_u64_le(slice, 216) ^ see6,
      );

      let (_, rest) = slice.split_at(224);
      slice = rest;
    }

    if slice.len() > 112 {
      seed = rapid_mix(read_u64_le(slice, 0) ^ DEFAULT_SECRETS[0], read_u64_le(slice, 8) ^ seed);
      see1 = rapid_mix(
        read_u64_le(slice, 16) ^ DEFAULT_SECRETS[1],
        read_u64_le(slice, 24) ^ see1,
      );
      see2 = rapid_mix(
        read_u64_le(slice, 32) ^ DEFAULT_SECRETS[2],
        read_u64_le(slice, 40) ^ see2,
      );
      see3 = rapid_mix(
        read_u64_le(slice, 48) ^ DEFAULT_SECRETS[3],
        read_u64_le(slice, 56) ^ see3,
      );
      see4 = rapid_mix(
        read_u64_le(slice, 64) ^ DEFAULT_SECRETS[4],
        read_u64_le(slice, 72) ^ see4,
      );
      see5 = rapid_mix(
        read_u64_le(slice, 80) ^ DEFAULT_SECRETS[5],
        read_u64_le(slice, 88) ^ see5,
      );
      see6 = rapid_mix(
        read_u64_le(slice, 96) ^ DEFAULT_SECRETS[6],
        read_u64_le(slice, 104) ^ see6,
      );
      let (_, rest) = slice.split_at(112);
      slice = rest;
    }

    seed ^= see1;
    see2 ^= see3;
    see4 ^= see5;
    seed ^= see6;
    see2 ^= see4;
    seed ^= see2;
  }

  if slice.len() > 16 {
    seed = rapid_mix(read_u64_le(slice, 0) ^ DEFAULT_SECRETS[2], read_u64_le(slice, 8) ^ seed);
    if slice.len() > 32 {
      seed = rapid_mix(
        read_u64_le(slice, 16) ^ DEFAULT_SECRETS[2],
        read_u64_le(slice, 24) ^ seed,
      );
      if slice.len() > 48 {
        seed = rapid_mix(
          read_u64_le(slice, 32) ^ DEFAULT_SECRETS[1],
          read_u64_le(slice, 40) ^ seed,
        );
        if slice.len() > 64 {
          seed = rapid_mix(
            read_u64_le(slice, 48) ^ DEFAULT_SECRETS[1],
            read_u64_le(slice, 56) ^ seed,
          );
          if slice.len() > 80 {
            seed = rapid_mix(
              read_u64_le(slice, 64) ^ DEFAULT_SECRETS[2],
              read_u64_le(slice, 72) ^ seed,
            );
            if slice.len() > 96 {
              seed = rapid_mix(
                read_u64_le(slice, 80) ^ DEFAULT_SECRETS[1],
                read_u64_le(slice, 88) ^ seed,
              );
            }
          }
        }
      }
    }
  }

  let remainder = slice.len() as u64;
  a ^= read_u64_le(data, data.len() - 16) ^ remainder;
  b ^= read_u64_le(data, data.len() - 8);

  a ^= DEFAULT_SECRETS[1];
  b ^= seed;

  (a, b) = rapid_mum(a, b);
  rapidhash_final_default::<AVALANCHE>(a, b, remainder)
}
/// Handles inputs 17-112 bytes. This keeps fixed-size medium inputs out of the
/// 7-stream large handler without changing V3 output.
///
/// # Safety
///
/// Caller must guarantee `16 < data.len() <= 112`.
#[inline(always)]
#[cfg(any(target_arch = "s390x", target_arch = "powerpc64"))]
unsafe fn rapidhash_core_medium<const AVALANCHE: bool>(data: &[u8], mut seed: u64, secrets: &[u64; 7]) -> u64 {
  // SAFETY: caller guarantees 16 < data.len() <= 112. This eliminates redundant
  // bounds checks in the tail-read section below.
  unsafe {
    core::hint::assert_unchecked(data.len() > 16);
    core::hint::assert_unchecked(data.len() <= 112);
  }

  let mut a = 0u64;
  let mut b = 0u64;

  seed = rapid_mix(
    read_u64_le(data, 0) ^ v3_secret::<2>(secrets),
    read_u64_le(data, 8) ^ seed,
  );
  if data.len() > 32 {
    seed = rapid_mix(
      read_u64_le(data, 16) ^ v3_secret::<2>(secrets),
      read_u64_le(data, 24) ^ seed,
    );
    if data.len() > 48 {
      seed = rapid_mix(
        read_u64_le(data, 32) ^ v3_secret::<1>(secrets),
        read_u64_le(data, 40) ^ seed,
      );
      if data.len() > 64 {
        seed = rapid_mix(
          read_u64_le(data, 48) ^ v3_secret::<1>(secrets),
          read_u64_le(data, 56) ^ seed,
        );
        if data.len() > 80 {
          seed = rapid_mix(
            read_u64_le(data, 64) ^ v3_secret::<2>(secrets),
            read_u64_le(data, 72) ^ seed,
          );
          if data.len() > 96 {
            seed = rapid_mix(
              read_u64_le(data, 80) ^ v3_secret::<1>(secrets),
              read_u64_le(data, 88) ^ seed,
            );
          }
        }
      }
    }
  }

  let remainder = data.len() as u64;
  a ^= read_u64_le(data, data.len() - 16) ^ remainder;
  b ^= read_u64_le(data, data.len() - 8);

  a ^= v3_secret::<1>(secrets);
  b ^= seed;

  (a, b) = rapid_mum(a, b);
  rapidhash_final::<AVALANCHE>(a, b, remainder, secrets)
}

/// Handles inputs >16 bytes. Kept as a separate function so the small-path
/// entry point stays lean for inlining. LLVM decides whether to inline this.
///
/// # Safety
///
/// Caller must guarantee `data.len() > 16`.
#[inline]
#[cfg(any(target_arch = "s390x", target_arch = "powerpc64"))]
unsafe fn rapidhash_core_large<const AVALANCHE: bool>(data: &[u8], mut seed: u64, secrets: &[u64; 7]) -> u64 {
  // SAFETY: caller guarantees data.len() > 16. This eliminates redundant
  // bounds checks in the tail-read section below.
  unsafe { core::hint::assert_unchecked(data.len() > 16) };

  let mut a = 0u64;
  let mut b = 0u64;
  let mut slice = data;

  if unlikely(slice.len() > 112) {
    let mut see1 = seed;
    let mut see2 = seed;
    let mut see3 = seed;
    let mut see4 = seed;
    let mut see5 = seed;
    let mut see6 = seed;

    while slice.len() > 224 {
      seed = rapid_mix(
        read_u64_le(slice, 0) ^ v3_secret::<0>(secrets),
        read_u64_le(slice, 8) ^ seed,
      );
      see1 = rapid_mix(
        read_u64_le(slice, 16) ^ v3_secret::<1>(secrets),
        read_u64_le(slice, 24) ^ see1,
      );
      see2 = rapid_mix(
        read_u64_le(slice, 32) ^ v3_secret::<2>(secrets),
        read_u64_le(slice, 40) ^ see2,
      );
      see3 = rapid_mix(
        read_u64_le(slice, 48) ^ v3_secret::<3>(secrets),
        read_u64_le(slice, 56) ^ see3,
      );
      see4 = rapid_mix(
        read_u64_le(slice, 64) ^ v3_secret::<4>(secrets),
        read_u64_le(slice, 72) ^ see4,
      );
      see5 = rapid_mix(
        read_u64_le(slice, 80) ^ v3_secret::<5>(secrets),
        read_u64_le(slice, 88) ^ see5,
      );
      see6 = rapid_mix(
        read_u64_le(slice, 96) ^ v3_secret::<6>(secrets),
        read_u64_le(slice, 104) ^ see6,
      );

      seed = rapid_mix(
        read_u64_le(slice, 112) ^ v3_secret::<0>(secrets),
        read_u64_le(slice, 120) ^ seed,
      );
      see1 = rapid_mix(
        read_u64_le(slice, 128) ^ v3_secret::<1>(secrets),
        read_u64_le(slice, 136) ^ see1,
      );
      see2 = rapid_mix(
        read_u64_le(slice, 144) ^ v3_secret::<2>(secrets),
        read_u64_le(slice, 152) ^ see2,
      );
      see3 = rapid_mix(
        read_u64_le(slice, 160) ^ v3_secret::<3>(secrets),
        read_u64_le(slice, 168) ^ see3,
      );
      see4 = rapid_mix(
        read_u64_le(slice, 176) ^ v3_secret::<4>(secrets),
        read_u64_le(slice, 184) ^ see4,
      );
      see5 = rapid_mix(
        read_u64_le(slice, 192) ^ v3_secret::<5>(secrets),
        read_u64_le(slice, 200) ^ see5,
      );
      see6 = rapid_mix(
        read_u64_le(slice, 208) ^ v3_secret::<6>(secrets),
        read_u64_le(slice, 216) ^ see6,
      );

      let (_, rest) = slice.split_at(224);
      slice = rest;
    }

    if slice.len() > 112 {
      seed = rapid_mix(
        read_u64_le(slice, 0) ^ v3_secret::<0>(secrets),
        read_u64_le(slice, 8) ^ seed,
      );
      see1 = rapid_mix(
        read_u64_le(slice, 16) ^ v3_secret::<1>(secrets),
        read_u64_le(slice, 24) ^ see1,
      );
      see2 = rapid_mix(
        read_u64_le(slice, 32) ^ v3_secret::<2>(secrets),
        read_u64_le(slice, 40) ^ see2,
      );
      see3 = rapid_mix(
        read_u64_le(slice, 48) ^ v3_secret::<3>(secrets),
        read_u64_le(slice, 56) ^ see3,
      );
      see4 = rapid_mix(
        read_u64_le(slice, 64) ^ v3_secret::<4>(secrets),
        read_u64_le(slice, 72) ^ see4,
      );
      see5 = rapid_mix(
        read_u64_le(slice, 80) ^ v3_secret::<5>(secrets),
        read_u64_le(slice, 88) ^ see5,
      );
      see6 = rapid_mix(
        read_u64_le(slice, 96) ^ v3_secret::<6>(secrets),
        read_u64_le(slice, 104) ^ see6,
      );
      let (_, rest) = slice.split_at(112);
      slice = rest;
    }

    seed ^= see1;
    see2 ^= see3;
    see4 ^= see5;
    seed ^= see6;
    see2 ^= see4;
    seed ^= see2;
  }

  if slice.len() > 16 {
    seed = rapid_mix(
      read_u64_le(slice, 0) ^ v3_secret::<2>(secrets),
      read_u64_le(slice, 8) ^ seed,
    );
    if slice.len() > 32 {
      seed = rapid_mix(
        read_u64_le(slice, 16) ^ v3_secret::<2>(secrets),
        read_u64_le(slice, 24) ^ seed,
      );
      if slice.len() > 48 {
        seed = rapid_mix(
          read_u64_le(slice, 32) ^ v3_secret::<1>(secrets),
          read_u64_le(slice, 40) ^ seed,
        );
        if slice.len() > 64 {
          seed = rapid_mix(
            read_u64_le(slice, 48) ^ v3_secret::<1>(secrets),
            read_u64_le(slice, 56) ^ seed,
          );
          if slice.len() > 80 {
            seed = rapid_mix(
              read_u64_le(slice, 64) ^ v3_secret::<2>(secrets),
              read_u64_le(slice, 72) ^ seed,
            );
            if slice.len() > 96 {
              seed = rapid_mix(
                read_u64_le(slice, 80) ^ v3_secret::<1>(secrets),
                read_u64_le(slice, 88) ^ seed,
              );
            }
          }
        }
      }
    }
  }

  let remainder = slice.len() as u64;
  a ^= read_u64_le(data, data.len() - 16) ^ remainder;
  b ^= read_u64_le(data, data.len() - 8);

  a ^= v3_secret::<1>(secrets);
  b ^= seed;

  (a, b) = rapid_mum(a, b);
  rapidhash_final::<AVALANCHE>(a, b, remainder, secrets)
}

// ---------------------------------------------------------------------------
// Inner-algorithm core for RapidHashFast64 / RapidHashFast128
// ---------------------------------------------------------------------------
//
// Distinct from V3: size-tuned dispatch (3-stream mid-range, 7-stream large),
// cold-path separation for codegen quality, and inner-module finalization.
// Output is NOT V3-compatible — use RapidHash64/128 for cross-language hashes.

/// Inputs above this use the 7-stream large kernel. Below it, the 3-stream
/// mid-range loop has lower register pressure and avoids stack spills.
const FAST_COLD_CUTOFF: usize = 400;

#[inline(always)]
fn rapidhash_fast_finish(a: u64, b: u64, seed: u64) -> u64 {
  rapid_mix(a ^ default_fast_secret::<0>(), b ^ seed)
}

#[inline(always)]
fn rapidhash_fast_small_parts(data: &[u8], mut seed: u64) -> (u64, u64, u64) {
  let mut a = 0u64;
  let mut b = 0u64;

  if data.len() == 1 {
    let byte = data[0] as u64;
    return ((byte << 45) | byte, byte, seed.wrapping_add(1));
  }

  if likely(data.len() >= 8) {
    a = read_u64_np(data, 0);
    b = read_u64_np(data, data.len() - 8);
  } else if likely(data.len() >= 4) {
    a = read_u32_np(data, 0) as u64;
    b = read_u32_np(data, data.len() - 4) as u64;
  } else if !data.is_empty() {
    a = ((data[0] as u64) << 45) | data[data.len() - 1] as u64;
    b = data[data.len() >> 1] as u64;
  }

  seed = seed.wrapping_add(data.len() as u64);
  (a, b, seed)
}

/// Inner-algorithm entry point for the Fast variants.
///
/// 0–16B path is always inlined. Larger inputs dispatch to `#[cold]` paths
/// that are never inlined, keeping this entry point lean for the common case.
///
/// Uses native-endian reads (`read_u64_np` / `read_u32_np`) — the Fast
/// variant is for in-process hashing only and does not need cross-platform
/// byte-order consistency. This eliminates byte-swap overhead on big-endian
/// platforms (s390x, POWER).
#[inline(always)]
fn rapidhash_fast_core(data: &[u8], mut seed: u64) -> u64 {
  if likely(data.len() <= 16) {
    let mut a = 0u64;
    let mut b = 0u64;

    if data.len() == 1 {
      let byte = data[0] as u64;
      return rapidhash_fast_finish((byte << 45) | byte, byte, seed.wrapping_add(1));
    }

    if likely(data.len() >= 8) {
      a = read_u64_np(data, 0);
      b = read_u64_np(data, data.len() - 8);
    } else if likely(data.len() >= 4) {
      a = read_u32_np(data, 0) as u64;
      b = read_u32_np(data, data.len() - 4) as u64;
    } else if !data.is_empty() {
      a = ((data[0] as u64) << 45) | data[data.len() - 1] as u64;
      b = data[data.len() >> 1] as u64;
    }

    seed = seed.wrapping_add(data.len() as u64);
    rapidhash_fast_finish(a, b, seed)
  } else {
    // SAFETY: data.len() > 16 verified above.
    unsafe { rapidhash_fast_core_medium(data, seed) }
  }
}

#[inline(always)]
fn rapidhash_fast_default_core(data: &[u8], seed: u64) -> u64 {
  if likely(data.len() <= 16) {
    let (a, b, seed) = rapidhash_fast_small_parts(data, seed);
    rapidhash_fast_finish(a, b, seed)
  } else if data.len() <= 96 {
    // SAFETY: data.len() > 16 and <= 96 verified above.
    unsafe { rapidhash_fast_core_inline_mid(data, seed) }
  } else {
    rapidhash_fast_core(data, seed)
  }
}

#[inline(always)]
fn rapidhash_fast_128_core(data: &[u8], seed_lo: u64, seed_hi: u64) -> u128 {
  if likely(data.len() > 16) {
    let lo = rapidhash_fast_default_core(data, seed_lo) as u128;
    let hi = rapidhash_fast_default_core(data, seed_hi) as u128;
    lo | (hi << 64)
  } else {
    let (a, b, seed_lo) = rapidhash_fast_small_parts(data, seed_lo);
    let seed_hi = seed_hi.wrapping_add(data.len() as u64);
    let lo = rapidhash_fast_finish(a, b, seed_lo) as u128;
    let hi = rapidhash_fast_finish(a, b, seed_hi) as u128;
    lo | (hi << 64)
  }
}

#[inline(always)]
unsafe fn rapidhash_fast_core_inline_mid(data: &[u8], mut seed: u64) -> u64 {
  // SAFETY: caller guarantees 16 < data.len() <= 96.
  unsafe {
    core::hint::assert_unchecked(data.len() > 16);
    core::hint::assert_unchecked(data.len() <= 96);
  }

  if data.len() <= 48 {
    // SAFETY: data.len() > 16 guaranteed by caller.
    return unsafe { rapidhash_fast_tail(seed, data, data) };
  }

  let mut slice = data;
  let mut see1 = seed;
  let mut see2 = seed;

  while slice.len() >= 48 {
    seed = rapid_mix(
      read_u64_np(slice, 0) ^ default_fast_secret::<0>(),
      read_u64_np(slice, 8) ^ seed,
    );
    see1 = rapid_mix(
      read_u64_np(slice, 16) ^ default_fast_secret::<1>(),
      read_u64_np(slice, 24) ^ see1,
    );
    see2 = rapid_mix(
      read_u64_np(slice, 32) ^ default_fast_secret::<2>(),
      read_u64_np(slice, 40) ^ see2,
    );
    let (_, rest) = slice.split_at(48);
    slice = rest;
  }

  seed ^= see1 ^ see2;

  // SAFETY: data.len() > 16 guaranteed by caller.
  unsafe { rapidhash_fast_tail(seed, slice, data) }
}

/// Final ≤48B tail processing shared by medium and large paths.
///
/// Reads up to two 16B pairs from `slice` (the unprocessed remainder), then
/// reads the final 16B from `data` (the original input) for overlap handling.
///
/// # Safety
///
/// Caller must guarantee `data.len() > 16`.
#[inline(always)]
unsafe fn rapidhash_fast_tail(mut seed: u64, slice: &[u8], data: &[u8]) -> u64 {
  // SAFETY: caller guarantees data.len() > 16.
  unsafe { core::hint::assert_unchecked(data.len() > 16) };

  if likely(slice.len() > 16) {
    seed = rapid_mix(
      read_u64_np(slice, 0) ^ default_fast_secret::<0>(),
      read_u64_np(slice, 8) ^ seed,
    );
    if likely(slice.len() > 32) {
      seed = rapid_mix(
        read_u64_np(slice, 16) ^ default_fast_secret::<0>(),
        read_u64_np(slice, 24) ^ seed,
      );
    }
  }

  let a = read_u64_np(data, data.len() - 16);
  let b = read_u64_np(data, data.len() - 8);
  seed = seed.wrapping_add(data.len() as u64);
  rapidhash_fast_finish(a, b, seed)
}

/// Cold medium path: 17B–~400B. 3-stream accumulation for 49B+.
///
/// `#[cold] #[inline(never)]` keeps the small-path entry point lean for
/// inlining. The 17–48B range short-circuits to the tail handler to avoid
/// 3-stream register setup/stack spill on all platforms.
///
/// # Safety
///
/// Caller must guarantee `data.len() > 16`.
#[cold]
#[inline(never)]
unsafe fn rapidhash_fast_core_medium(data: &[u8], mut seed: u64) -> u64 {
  // SAFETY: caller guarantees data.len() > 16.
  unsafe { core::hint::assert_unchecked(data.len() > 16) };

  let mut slice = data;

  // Short-circuit 17–48B to the tail handler. This avoids the 3-stream
  // register setup that causes stack spills.
  if data.len() <= 48 {
    // SAFETY: data.len() > 16 guaranteed by caller.
    return unsafe { rapidhash_fast_tail(seed, data, data) };
  }

  if unlikely(data.len() > FAST_COLD_CUTOFF) {
    // SAFETY: data.len() > FAST_COLD_CUTOFF > 16.
    return unsafe { rapidhash_fast_core_large(data, seed) };
  }

  // 3-stream mid-range: 49–400B. Three independent accumulators allow
  // out-of-order execution to overlap the multiply latency.
  let mut see1 = seed;
  let mut see2 = seed;

  while slice.len() >= 48 {
    seed = rapid_mix(
      read_u64_np(slice, 0) ^ default_fast_secret::<0>(),
      read_u64_np(slice, 8) ^ seed,
    );
    see1 = rapid_mix(
      read_u64_np(slice, 16) ^ default_fast_secret::<1>(),
      read_u64_np(slice, 24) ^ see1,
    );
    see2 = rapid_mix(
      read_u64_np(slice, 32) ^ default_fast_secret::<2>(),
      read_u64_np(slice, 40) ^ see2,
    );
    let (_, rest) = slice.split_at(48);
    slice = rest;
  }

  seed ^= see1 ^ see2;

  // SAFETY: data.len() > 16 guaranteed by caller.
  unsafe { rapidhash_fast_tail(seed, slice, data) }
}

/// Cold large path: >400B. 7-stream bulk with 3-stream residual.
///
/// # Safety
///
/// Caller must guarantee `data.len() > FAST_COLD_CUTOFF`.
#[cold]
#[inline(never)]
unsafe fn rapidhash_fast_core_large(data: &[u8], mut seed: u64) -> u64 {
  // SAFETY: caller guarantees data.len() > FAST_COLD_CUTOFF.
  unsafe { core::hint::assert_unchecked(data.len() > FAST_COLD_CUTOFF) };

  let mut slice = data;
  let mut see1 = seed;
  let mut see2 = seed;
  let mut see3 = seed;
  let mut see4 = seed;
  let mut see5 = seed;
  let mut see6 = seed;

  // 7-stream bulk: 224B per iteration (2 × 112B half-blocks).
  while slice.len() >= 224 {
    seed = rapid_mix(
      read_u64_np(slice, 0) ^ default_fast_secret::<0>(),
      read_u64_np(slice, 8) ^ seed,
    );
    see1 = rapid_mix(
      read_u64_np(slice, 16) ^ default_fast_secret::<1>(),
      read_u64_np(slice, 24) ^ see1,
    );
    see2 = rapid_mix(
      read_u64_np(slice, 32) ^ default_fast_secret::<2>(),
      read_u64_np(slice, 40) ^ see2,
    );
    see3 = rapid_mix(
      read_u64_np(slice, 48) ^ default_fast_secret::<3>(),
      read_u64_np(slice, 56) ^ see3,
    );
    see4 = rapid_mix(
      read_u64_np(slice, 64) ^ default_fast_secret::<4>(),
      read_u64_np(slice, 72) ^ see4,
    );
    see5 = rapid_mix(
      read_u64_np(slice, 80) ^ default_fast_secret::<5>(),
      read_u64_np(slice, 88) ^ see5,
    );
    see6 = rapid_mix(
      read_u64_np(slice, 96) ^ default_fast_secret::<6>(),
      read_u64_np(slice, 104) ^ see6,
    );

    seed = rapid_mix(
      read_u64_np(slice, 112) ^ default_fast_secret::<0>(),
      read_u64_np(slice, 120) ^ seed,
    );
    see1 = rapid_mix(
      read_u64_np(slice, 128) ^ default_fast_secret::<1>(),
      read_u64_np(slice, 136) ^ see1,
    );
    see2 = rapid_mix(
      read_u64_np(slice, 144) ^ default_fast_secret::<2>(),
      read_u64_np(slice, 152) ^ see2,
    );
    see3 = rapid_mix(
      read_u64_np(slice, 160) ^ default_fast_secret::<3>(),
      read_u64_np(slice, 168) ^ see3,
    );
    see4 = rapid_mix(
      read_u64_np(slice, 176) ^ default_fast_secret::<4>(),
      read_u64_np(slice, 184) ^ see4,
    );
    see5 = rapid_mix(
      read_u64_np(slice, 192) ^ default_fast_secret::<5>(),
      read_u64_np(slice, 200) ^ see5,
    );
    see6 = rapid_mix(
      read_u64_np(slice, 208) ^ default_fast_secret::<6>(),
      read_u64_np(slice, 216) ^ see6,
    );

    let (_, rest) = slice.split_at(224);
    slice = rest;
  }

  // Single 112B half-block if enough remains.
  if likely(slice.len() >= 112) {
    seed = rapid_mix(
      read_u64_np(slice, 0) ^ default_fast_secret::<0>(),
      read_u64_np(slice, 8) ^ seed,
    );
    see1 = rapid_mix(
      read_u64_np(slice, 16) ^ default_fast_secret::<1>(),
      read_u64_np(slice, 24) ^ see1,
    );
    see2 = rapid_mix(
      read_u64_np(slice, 32) ^ default_fast_secret::<2>(),
      read_u64_np(slice, 40) ^ see2,
    );
    see3 = rapid_mix(
      read_u64_np(slice, 48) ^ default_fast_secret::<3>(),
      read_u64_np(slice, 56) ^ see3,
    );
    see4 = rapid_mix(
      read_u64_np(slice, 64) ^ default_fast_secret::<4>(),
      read_u64_np(slice, 72) ^ see4,
    );
    see5 = rapid_mix(
      read_u64_np(slice, 80) ^ default_fast_secret::<5>(),
      read_u64_np(slice, 88) ^ see5,
    );
    see6 = rapid_mix(
      read_u64_np(slice, 96) ^ default_fast_secret::<6>(),
      read_u64_np(slice, 104) ^ see6,
    );
    let (_, rest) = slice.split_at(112);
    slice = rest;
  }

  // 3-stream residual: up to 2 × 48B blocks drain the remainder before the
  // tail, avoiding the serial dependency chain of the V3 nested-if tail.
  if slice.len() >= 48 {
    seed = rapid_mix(
      read_u64_np(slice, 0) ^ default_fast_secret::<0>(),
      read_u64_np(slice, 8) ^ seed,
    );
    see1 = rapid_mix(
      read_u64_np(slice, 16) ^ default_fast_secret::<1>(),
      read_u64_np(slice, 24) ^ see1,
    );
    see2 = rapid_mix(
      read_u64_np(slice, 32) ^ default_fast_secret::<2>(),
      read_u64_np(slice, 40) ^ see2,
    );
    let (_, rest) = slice.split_at(48);
    slice = rest;

    if slice.len() >= 48 {
      seed = rapid_mix(
        read_u64_np(slice, 0) ^ default_fast_secret::<0>(),
        read_u64_np(slice, 8) ^ seed,
      );
      see1 = rapid_mix(
        read_u64_np(slice, 16) ^ default_fast_secret::<1>(),
        read_u64_np(slice, 24) ^ see1,
      );
      see2 = rapid_mix(
        read_u64_np(slice, 32) ^ default_fast_secret::<2>(),
        read_u64_np(slice, 40) ^ see2,
      );
      let (_, rest) = slice.split_at(48);
      slice = rest;
    }
  }

  // Merge all accumulators into seed.
  see3 ^= see4;
  see5 ^= see6;
  seed ^= see1;
  see3 ^= see2;
  seed ^= see5;
  seed ^= see3;

  // SAFETY: data.len() > FAST_COLD_CUTOFF > 16.
  unsafe { rapidhash_fast_tail(seed, slice, data) }
}

impl FastHash for RapidHash64 {
  const OUTPUT_SIZE: usize = 8;
  type Output = u64;
  type Seed = u64;

  /// Hash with precomputed default seed — bypasses `rapidhash_seed_cpp(0)`,
  /// saving one 128-bit multiply per call.
  #[inline(always)]
  fn hash(data: &[u8]) -> Self::Output {
    #[cfg(any(target_arch = "s390x", target_arch = "powerpc64"))]
    {
      rapidhash_core::<true>(data, V3_DEFAULT_SECRETS.seed, &V3_DEFAULT_SECRETS.secrets)
    }

    #[cfg(not(any(target_arch = "s390x", target_arch = "powerpc64")))]
    {
      rapidhash_core_default::<true>(data, PRECOMPUTED_SEED_0)
    }
  }

  #[inline]
  fn hash_with_seed(seed: Self::Seed, data: &[u8]) -> Self::Output {
    dispatch::hash64_with_seed(seed, data)
  }
}

impl FastHash for RapidHash128 {
  const OUTPUT_SIZE: usize = 16;
  type Output = u128;
  type Seed = u64;

  /// Hash with precomputed default seeds — bypasses two seed-mixing multiplies
  /// per call.
  #[inline(always)]
  fn hash(data: &[u8]) -> Self::Output {
    #[cfg(any(target_arch = "s390x", target_arch = "powerpc64"))]
    {
      let lo = rapidhash_core::<true>(data, V3_DEFAULT_SECRETS.seed, &V3_DEFAULT_SECRETS.secrets) as u128;
      let hi = rapidhash_core::<true>(data, V3_DEFAULT_HI_SECRETS.seed, &V3_DEFAULT_HI_SECRETS.secrets) as u128;
      lo | (hi << 64)
    }

    #[cfg(not(any(target_arch = "s390x", target_arch = "powerpc64")))]
    {
      let lo = rapidhash_core_default::<true>(data, PRECOMPUTED_SEED_0) as u128;
      let hi = rapidhash_core_default::<true>(data, PRECOMPUTED_SEED_HI) as u128;
      lo | (hi << 64)
    }
  }

  #[inline]
  fn hash_with_seed(seed: Self::Seed, data: &[u8]) -> Self::Output {
    dispatch::hash128_with_seed(seed, data)
  }
}

impl FastHash for RapidHashFast64 {
  const OUTPUT_SIZE: usize = 8;
  type Output = u64;
  type Seed = u64;

  /// Hash with precomputed default seed — bypasses `rapidhash_seed_cpp(0)`,
  /// saving one 128-bit multiply per call.
  #[inline(always)]
  fn hash(data: &[u8]) -> Self::Output {
    if data.is_empty() {
      return FAST_EMPTY_HASH_SEED_0;
    }

    rapidhash_fast_default_core(data, PRECOMPUTED_SEED_0)
  }

  #[inline(always)]
  fn hash_with_seed(seed: Self::Seed, data: &[u8]) -> Self::Output {
    dispatch::hash64_fast_with_seed(seed, data)
  }
}

impl FastHash for RapidHashFast128 {
  const OUTPUT_SIZE: usize = 16;
  type Output = u128;
  type Seed = u64;

  /// Hash with precomputed default seed — bypasses `rapidhash_seed_cpp(0)`,
  /// saving one 128-bit multiply per call.
  #[inline(always)]
  fn hash(data: &[u8]) -> Self::Output {
    if data.is_empty() {
      return (FAST_EMPTY_HASH_SEED_0 as u128) | ((FAST_EMPTY_HASH_SEED_HI as u128) << 64);
    }

    rapidhash_fast_128_core(data, PRECOMPUTED_SEED_0, PRECOMPUTED_SEED_HI)
  }

  #[inline(always)]
  fn hash_with_seed(seed: Self::Seed, data: &[u8]) -> Self::Output {
    dispatch::hash128_fast_with_seed(seed, data)
  }
}

// ─── BuildHasher support ──────────────────────────────────────────────────

/// Streaming [`core::hash::Hasher`] backed by rapidhash-64.
///
/// Created by [`RapidBuildHasher`]. Buffers input and computes the hash
/// on [`finish`](core::hash::Hasher::finish).
#[cfg(feature = "alloc")]
#[cfg_attr(docsrs, doc(cfg(feature = "alloc")))]
pub struct RapidHasher {
  buf: alloc::vec::Vec<u8>,
  seed: u64,
}

#[cfg(feature = "alloc")]
impl core::fmt::Debug for RapidHasher {
  fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
    f.debug_struct("RapidHasher")
      .field("seed", &self.seed)
      .field("buffered", &self.buf.len())
      .finish()
  }
}

#[cfg(feature = "alloc")]
impl core::hash::Hasher for RapidHasher {
  #[inline]
  fn write(&mut self, bytes: &[u8]) {
    self.buf.extend_from_slice(bytes);
  }

  #[inline]
  fn finish(&self) -> u64 {
    RapidHash64::hash_with_seed(self.seed, &self.buf)
  }
}

/// [`BuildHasher`](core::hash::BuildHasher) producing [`RapidHasher`] instances.
///
/// # Examples
///
/// ```
/// use std::collections::HashMap;
///
/// use rscrypto::hashes::fast::rapidhash::RapidBuildHasher;
///
/// let mut map: HashMap<&str, i32, RapidBuildHasher> = HashMap::with_hasher(RapidBuildHasher::new());
/// map.insert("hello", 42);
/// assert_eq!(map["hello"], 42);
/// ```
#[cfg(feature = "alloc")]
#[cfg_attr(docsrs, doc(cfg(feature = "alloc")))]
#[derive(Clone, Debug)]
pub struct RapidBuildHasher {
  seed: u64,
}

#[cfg(feature = "alloc")]
impl RapidBuildHasher {
  /// Create a builder with the default seed (0).
  #[inline]
  #[must_use]
  pub const fn new() -> Self {
    Self { seed: 0 }
  }

  /// Create a builder with a custom seed.
  #[inline]
  #[must_use]
  pub const fn with_seed(seed: u64) -> Self {
    Self { seed }
  }
}

#[cfg(feature = "alloc")]
impl Default for RapidBuildHasher {
  #[inline]
  fn default() -> Self {
    Self::new()
  }
}

#[cfg(feature = "alloc")]
impl core::hash::BuildHasher for RapidBuildHasher {
  type Hasher = RapidHasher;

  #[inline]
  fn build_hasher(&self) -> Self::Hasher {
    RapidHasher {
      buf: alloc::vec::Vec::new(),
      seed: self.seed,
    }
  }
}

#[cfg(test)]
mod tests {
  use alloc::vec::Vec;

  #[cfg(not(miri))]
  use proptest::prelude::*;

  use super::{RapidHash64, RapidHash128, RapidHashFast64, RapidHashFast128, V3_HI_SEED};
  use crate::traits::FastHash;

  // -- V3 standard variant oracle (C++-compatible output) --

  #[test]
  fn smoke_v3_empty_matches_oracle() {
    let seed = 0u64;
    let ours = <RapidHash64 as crate::traits::FastHash>::hash_with_seed(seed, b"");
    let secrets = rapidhash::v3::RapidSecrets::seed_cpp(seed);
    let theirs = rapidhash::v3::rapidhash_v3_seeded(b"", &secrets);
    assert_eq!(ours, theirs);
  }

  // -- Fast variant oracle (competitor's inner module) --

  #[test]
  #[cfg(target_endian = "little")]
  fn smoke_fast_empty_matches_oracle() {
    use core::hash::Hasher;
    let seed = 0u64;
    let ours = <RapidHashFast64 as crate::traits::FastHash>::hash_with_seed(seed, b"");
    let mut h = rapidhash::fast::RapidHasher::new(seed);
    h.write(b"");
    let theirs = h.finish();
    assert_eq!(ours, theirs);
  }

  #[cfg(not(miri))]
  proptest! {
    #[test]
    fn rapidhash_v3_64_matches_oracle(seed in any::<u64>(), data in proptest::collection::vec(any::<u8>(), 0..2048)) {
      let ours = <RapidHash64 as crate::traits::FastHash>::hash_with_seed(seed, &data);
      let secrets = rapidhash::v3::RapidSecrets::seed_cpp(seed);
      let theirs = rapidhash::v3::rapidhash_v3_seeded(&data, &secrets);
      prop_assert_eq!(ours, theirs);
    }

    #[test]
    fn rapidhash128_is_two_independent_64bit_hashes(seed in any::<u64>(), data in proptest::collection::vec(any::<u8>(), 0..2048)) {
      let out = <RapidHash128 as crate::traits::FastHash>::hash_with_seed(seed, &data);
      let lo = out as u64;
      let hi = (out >> 64) as u64;

      prop_assert_eq!(lo, <RapidHash64 as crate::traits::FastHash>::hash_with_seed(seed, &data));
      prop_assert_eq!(hi, <RapidHash64 as FastHash>::hash_with_seed(seed ^ V3_HI_SEED, &data));
    }

    /// The fast variant uses the inner-algorithm core (3-stream mid-range,
    /// inner finalization). Oracle: `rapidhash::fast::RapidHasher` (the
    /// competitor's optimized inner module with AVALANCHE=false, SPONGE=true).
    ///
    /// LE-only: the competitor's inner module uses non-portable native-endian
    /// reads (`read_u64_np`), while we use LE reads for consistent hashing
    /// across platforms. On BE targets the outputs diverge by design.
    #[test]
    #[cfg(target_endian = "little")]
    fn rapidhash_fast_64_matches_inner_oracle(seed in any::<u64>(), data in proptest::collection::vec(any::<u8>(), 0..8192)) {
      use core::hash::Hasher;
      let ours = <RapidHashFast64 as crate::traits::FastHash>::hash_with_seed(seed, &data);
      let mut h = rapidhash::fast::RapidHasher::new(seed);
      h.write(&data);
      let theirs = h.finish();
      prop_assert_eq!(ours, theirs);
    }

    #[test]
    fn rapidhash_fast_128_is_two_independent_64bit_hashes(seed in any::<u64>(), data in proptest::collection::vec(any::<u8>(), 0..2048)) {
      let out = <RapidHashFast128 as crate::traits::FastHash>::hash_with_seed(seed, &data);
      let lo = out as u64;
      let hi = (out >> 64) as u64;

      prop_assert_eq!(lo, <RapidHashFast64 as crate::traits::FastHash>::hash_with_seed(seed, &data));
      prop_assert_eq!(hi, <RapidHashFast64 as FastHash>::hash_with_seed(seed ^ V3_HI_SEED, &data));
    }
  }

  fn deterministic_bytes(len: usize) -> Vec<u8> {
    let mut out = alloc::vec![0u8; len];
    let mut x = 0x243f_6a88_85a3_08d3u64;
    for (i, byte) in out.iter_mut().enumerate() {
      x = x.wrapping_mul(6364136223846793005).wrapping_add(1442695040888963407);
      *byte = (x >> 56) as u8 ^ ((i as u8).rotate_left((i & 7) as u32));
    }
    out
  }

  #[test]
  fn default_hash_matches_seed_zero() {
    let sizes = [
      0usize, 1, 2, 3, 4, 7, 8, 15, 16, 17, 31, 32, 33, 47, 48, 49, 63, 64, 65, 127, 128, 129, 241, 399, 400, 401,
      1024, 4096,
    ];

    for &len in &sizes {
      let data = deterministic_bytes(len);

      assert_eq!(
        <RapidHash64 as FastHash>::hash(&data),
        <RapidHash64 as FastHash>::hash_with_seed(0, &data)
      );
      assert_eq!(
        <RapidHash128 as FastHash>::hash(&data),
        <RapidHash128 as FastHash>::hash_with_seed(0, &data)
      );
      assert_eq!(
        <RapidHashFast64 as FastHash>::hash(&data),
        <RapidHashFast64 as FastHash>::hash_with_seed(0, &data)
      );
      assert_eq!(
        <RapidHashFast128 as FastHash>::hash(&data),
        <RapidHashFast128 as FastHash>::hash_with_seed(0, &data)
      );
    }
  }

  #[test]
  fn rapidhash_boundary_paths_match_oracles() {
    let sizes = [
      0usize, 1, 2, 3, 4, 7, 8, 15, 16, 17, 31, 32, 33, 63, 64, 65, 127, 128, 129, 241, 1024, 4096,
    ];
    let seeds = [0u64, 1u64, 0x0123_4567_89ab_cdef];

    for &seed in &seeds {
      for &len in &sizes {
        let data = deterministic_bytes(len);

        let ours_v3_64 = <RapidHash64 as crate::traits::FastHash>::hash_with_seed(seed, &data);
        let secrets = rapidhash::v3::RapidSecrets::seed_cpp(seed);
        let theirs_v3_64 = rapidhash::v3::rapidhash_v3_seeded(&data, &secrets);
        assert_eq!(
          ours_v3_64, theirs_v3_64,
          "RapidHash64 mismatch (seed={seed}, len={len})"
        );

        let ours_v3_128 = <RapidHash128 as crate::traits::FastHash>::hash_with_seed(seed, &data);
        assert_eq!(
          ours_v3_128 as u64, ours_v3_64,
          "RapidHash128 low lane mismatch (seed={seed}, len={len})"
        );
        assert_eq!(
          (ours_v3_128 >> 64) as u64,
          <RapidHash64 as crate::traits::FastHash>::hash_with_seed(seed ^ V3_HI_SEED, &data),
          "RapidHash128 high lane mismatch (seed={seed}, len={len})"
        );

        #[cfg(target_endian = "little")]
        {
          use core::hash::Hasher;

          let ours_fast_64 = <RapidHashFast64 as crate::traits::FastHash>::hash_with_seed(seed, &data);
          let mut h = rapidhash::fast::RapidHasher::new(seed);
          h.write(&data);
          let theirs_fast_64 = h.finish();
          assert_eq!(
            ours_fast_64, theirs_fast_64,
            "RapidHashFast64 mismatch (seed={seed}, len={len})"
          );

          let ours_fast_128 = <RapidHashFast128 as crate::traits::FastHash>::hash_with_seed(seed, &data);
          assert_eq!(
            ours_fast_128 as u64, ours_fast_64,
            "RapidHashFast128 low lane mismatch (seed={seed}, len={len})"
          );
          assert_eq!(
            (ours_fast_128 >> 64) as u64,
            <RapidHashFast64 as crate::traits::FastHash>::hash_with_seed(seed ^ V3_HI_SEED, &data),
            "RapidHashFast128 high lane mismatch (seed={seed}, len={len})"
          );
        }
      }
    }
  }
}