/// @file dpf/dcf.hpp
/// @brief Comparison-channel specs and GGM path-sum helpers for libdpf.
/// @details `lt`/`leq`/`gt`/`geq` (+ `_at`) take `(if_true, if_false=0)`.
///          Eval walks the same GGM tree as the DPF (per-level value CWs).
///          `eq` / `eq_at` are synonyms for ordinary point placements.
/// @copyright Copyright (c) 2019-2026 Ryan Henry and [others](@ref authors)
/// @license Released under a GNU General Public v2.0 (GPLv2) license.

#ifndef LIBDPF_INCLUDE_DPF_DCF_HPP__
#define LIBDPF_INCLUDE_DPF_DCF_HPP__

#include <cstddef>
#include <cstdint>
#include <type_traits>
#include <utility>
#include <limits>

#include "hedley/hedley.h"
#include "simde/simde/x86/avx2.h"

#include "dpf/utils.hpp"
#include "dpf/bit.hpp"
#include "dpf/xor_wrapper.hpp"
#include "dpf/twiddle.hpp"

namespace dpf
{

/// Comparison kind for the optional DCF channel on a key.
enum class cmp_kind : uint8_t
{
    lt = 0,
    leq = 1,
    gt = 2,
    geq = 3
};

enum class cmp_trivial : uint8_t
{
    none = 0,
    always_true = 1,
    always_false = 2
};

namespace detail
{
namespace dcf_impl
{

template <typename Beta>
Beta default_false() noexcept
{
    if constexpr (std::is_same_v<Beta, dpf::bit>)
        return dpf::bit::zero;
    else
        return Beta{};
}

template <typename Beta>
uint64_t beta_delta_u64(const Beta & if_true, const Beta & if_false,
    uint64_t mask) noexcept
{
    if constexpr (std::is_same_v<Beta, dpf::bit>)
    {
        const uint64_t t = static_cast<bool>(if_true) ? 1ULL : 0ULL;
        const uint64_t f = static_cast<bool>(if_false) ? 1ULL : 0ULL;
        return (t ^ f) & mask;
    }
    else if constexpr (dpf::utils::is_xor_wrapper_v<Beta>)
    {
        return (static_cast<uint64_t>(if_true) ^ static_cast<uint64_t>(if_false))
            & mask;
    }
    else
    {
        return (static_cast<uint64_t>(if_true)
            - static_cast<uint64_t>(if_false)) & mask;
    }
}

template <typename Beta>
uint64_t beta_to_u64_simple(const Beta & beta, uint64_t mask) noexcept
{
    if constexpr (std::is_same_v<Beta, dpf::bit>)
        return (static_cast<bool>(beta) ? 1ULL : 0ULL) & mask;
    else
        return static_cast<uint64_t>(beta) & mask;
}

template <typename Beta>
Beta sub_beta(const Beta & a, const Beta & b) noexcept
{
    if constexpr (std::is_same_v<Beta, dpf::bit>)
        return dpf::bit{static_cast<bool>(a) ^ static_cast<bool>(b)};
    else if constexpr (dpf::utils::is_xor_wrapper_v<Beta>)
        return Beta{static_cast<uint64_t>(a) ^ static_cast<uint64_t>(b)};
    else
        return static_cast<Beta>(a - b);
}

template <typename Beta>
Beta u64_to_beta(uint64_t v) noexcept
{
    if constexpr (std::is_same_v<Beta, dpf::bit>)
        return dpf::bit{static_cast<bool>(v & 1u)};
    else
        return static_cast<Beta>(v);
}

inline uint64_t default_mask_for_bits(std::size_t out_bits) noexcept
{
    if (out_bits >= 64)
        return ~0ULL;
    if (out_bits == 0)
        return 0ULL;
    return (1ULL << out_bits) - 1ULL;
}

HEDLEY_ALWAYS_INLINE
uint64_t neg_m(uint64_t x, uint64_t mask) noexcept
{
    return (0ULL - x) & mask;
}

HEDLEY_ALWAYS_INLINE
uint64_t sgn_m(uint8_t t1, uint64_t x, uint64_t mask) noexcept
{
    return t1 ? neg_m(x, mask) : x;
}

/// Convert a GGM node to a group element (low 64 bits, control bits cleared).
HEDLEY_ALWAYS_INLINE
uint64_t convert_node(simde__m128i n, uint64_t mask) noexcept
{
    return static_cast<uint64_t>(
        simde_mm_cvtsi128_si64(dpf::unset_lo_2bits(n))) & mask;
}

/// Draw the group-width blind `r` used to split the `cmp_addend` share.
/// `sample` yields one interior block; only `popcount(mask)` live bits are
/// kept, so the blind (and thus the addend share) never needs a full padded
/// `uint64_t` on the wire. Dealer and Doerner–Shelat gen call this with the
/// same block source so their keys stay byte-identical (matched tapes).
template <typename BlockSampler>
HEDLEY_ALWAYS_INLINE
uint64_t sample_addend_blind(uint64_t mask, BlockSampler && sample) noexcept
{
    return convert_node(dpf::unset_lo_2bits(sample()), mask);
}

/// One level of value CW on GGM children. Updates running `Va`.
/// `ai` is the keep-path bit of the (effective) threshold.
inline uint64_t make_value_cw(simde__m128i c0L, simde__m128i c0R,
    simde__m128i c1L, simde__m128i c1R, uint8_t t0, uint8_t t1, int ai,
    uint64_t & Va, uint64_t beta, uint64_t mask) noexcept
{
    (void)t0;
    uint64_t v0K, v1K, v0Lo, v1Lo;
    if (ai == 0)
    {
        v0K = convert_node(c0L, mask);
        v1K = convert_node(c1L, mask);
        v0Lo = convert_node(c0R, mask);
        v1Lo = convert_node(c1R, mask);
    }
    else
    {
        v0K = convert_node(c0R, mask);
        v1K = convert_node(c1R, mask);
        v0Lo = convert_node(c0L, mask);
        v1Lo = convert_node(c1L, mask);
    }
    uint64_t vcw = sgn_m(t1,
        (v1Lo + neg_m(v0Lo, mask) + neg_m(Va, mask)) & mask, mask);
    // Lose-left (ai==1) is the x<α diverge: plant β there.
    if (ai == 1)
        vcw = (vcw + sgn_m(t1, beta, mask)) & mask;
    Va = (Va + neg_m(v1K, mask) + v0K + sgn_m(t1, vcw, mask)) & mask;
    return vcw;
}

/// Final leaf value CW. `on_path` is the payload reconstructed when the query
/// stays on α's path through all levels (0 for strict lt/geq; β for leq/gt).
inline uint64_t make_final_cw(simde__m128i s0, simde__m128i s1, uint8_t t1,
    uint64_t Va, uint64_t mask, uint64_t on_path = 0) noexcept
{
    uint64_t c0 = convert_node(s0, mask);
    uint64_t c1 = convert_node(s1, mask);
    return sgn_m(t1,
        (c1 + neg_m(c0, mask) + neg_m(Va, mask) + on_path) & mask, mask);
}

}  // namespace dcf_impl

/// Comparison metadata on an incremental key (value CWs live on the key).
/// Payload δ = if_true − if_false is dealer-known and baked into `value_cw` /
/// `cw_last` only — never stored clear on the key (traditional DPF hiding).
/// The second output value (`if_false`) is held as a per-party additive share
/// on the key (`cmp_addend`), not as a public constant.
struct cmp_meta
{
    int nbits = 0;                 // comparison prefix length
    uint64_t mask = 0;
    cmp_kind kind = cmp_kind::lt;
    cmp_trivial trivial = cmp_trivial::none;
    bool eval_as_ge = false;       // invert path-sum (geq / gt)
    bool include_eq = false;       // plant δ on the α-path leaf (leq / gt)
    bool active = false;

    bool empty() const noexcept { return !active; }
};

/// Backward-compatible alias while call sites migrate.
using cmp_channel = cmp_meta;

}  // namespace detail

// ---------------------------------------------------------------------------
// Comparison specs: lt/leq/gt/geq (+ _at)
// ---------------------------------------------------------------------------

template <cmp_kind Kind, typename Beta>
struct cmp_pack
{
    static constexpr bool is_cmp = true;
    static constexpr cmp_kind kind = Kind;
    static constexpr std::size_t prefix = 0;
    using beta_type = Beta;
    Beta if_true;
    Beta if_false;

    explicit cmp_pack(Beta t, Beta f = detail::dcf_impl::default_false<Beta>())
      : if_true{std::move(t)}, if_false{std::move(f)} { }
};

template <std::size_t N, cmp_kind Kind, typename Beta>
struct cmp_at_pack
{
    static constexpr bool is_cmp = true;
    static constexpr cmp_kind kind = Kind;
    static constexpr std::size_t prefix = N;
    using beta_type = Beta;
    Beta if_true;
    Beta if_false;

    explicit cmp_at_pack(Beta t, Beta f = detail::dcf_impl::default_false<Beta>())
      : if_true{std::move(t)}, if_false{std::move(f)} { }
};

template <typename Beta>
inline auto lt(Beta t, Beta f = detail::dcf_impl::default_false<std::decay_t<Beta>>())
{
    return cmp_pack<cmp_kind::lt, std::decay_t<Beta>>(std::move(t), std::move(f));
}
template <typename Beta>
inline auto leq(Beta t, Beta f = detail::dcf_impl::default_false<std::decay_t<Beta>>())
{
    return cmp_pack<cmp_kind::leq, std::decay_t<Beta>>(std::move(t), std::move(f));
}
template <typename Beta>
inline auto gt(Beta t, Beta f = detail::dcf_impl::default_false<std::decay_t<Beta>>())
{
    return cmp_pack<cmp_kind::gt, std::decay_t<Beta>>(std::move(t), std::move(f));
}
template <typename Beta>
inline auto geq(Beta t, Beta f = detail::dcf_impl::default_false<std::decay_t<Beta>>())
{
    return cmp_pack<cmp_kind::geq, std::decay_t<Beta>>(std::move(t), std::move(f));
}

template <std::size_t N, typename Beta>
inline auto lt_at(Beta t, Beta f = detail::dcf_impl::default_false<std::decay_t<Beta>>())
{
    return cmp_at_pack<N, cmp_kind::lt, std::decay_t<Beta>>(std::move(t), std::move(f));
}
template <std::size_t N, typename Beta>
inline auto leq_at(Beta t, Beta f = detail::dcf_impl::default_false<std::decay_t<Beta>>())
{
    return cmp_at_pack<N, cmp_kind::leq, std::decay_t<Beta>>(std::move(t), std::move(f));
}
template <std::size_t N, typename Beta>
inline auto gt_at(Beta t, Beta f = detail::dcf_impl::default_false<std::decay_t<Beta>>())
{
    return cmp_at_pack<N, cmp_kind::gt, std::decay_t<Beta>>(std::move(t), std::move(f));
}
template <std::size_t N, typename Beta>
inline auto geq_at(Beta t, Beta f = detail::dcf_impl::default_false<std::decay_t<Beta>>())
{
    return cmp_at_pack<N, cmp_kind::geq, std::decay_t<Beta>>(std::move(t), std::move(f));
}

// ---------------------------------------------------------------------------
// Equality specs: eq / eq_at
// ---------------------------------------------------------------------------

template <typename Beta>
struct eq_pack
{
    static constexpr bool is_eq = true;
    static constexpr std::size_t prefix = 0;
    using beta_type = Beta;
    Beta if_true;
    Beta if_false;

    explicit eq_pack(Beta t, Beta f = detail::dcf_impl::default_false<Beta>())
      : if_true{std::move(t)}, if_false{std::move(f)} { }
};

template <std::size_t N, typename Beta>
struct eq_at_pack
{
    static constexpr bool is_eq = true;
    static constexpr std::size_t prefix = N;
    using beta_type = Beta;
    Beta if_true;
    Beta if_false;

    explicit eq_at_pack(Beta t, Beta f = detail::dcf_impl::default_false<Beta>())
      : if_true{std::move(t)}, if_false{std::move(f)} { }
};

template <typename Beta>
inline auto eq(Beta t, Beta f = detail::dcf_impl::default_false<std::decay_t<Beta>>())
{
    return eq_pack<std::decay_t<Beta>>(std::move(t), std::move(f));
}
template <std::size_t N, typename Beta>
inline auto eq_at(Beta t, Beta f = detail::dcf_impl::default_false<std::decay_t<Beta>>())
{
    return eq_at_pack<N, std::decay_t<Beta>>(std::move(t), std::move(f));
}

template <typename T> struct is_cmp_spec : std::false_type {};
template <cmp_kind K, typename B> struct is_cmp_spec<cmp_pack<K, B>> : std::true_type {};
template <std::size_t N, cmp_kind K, typename B>
struct is_cmp_spec<cmp_at_pack<N, K, B>> : std::true_type {};
template <typename T>
inline constexpr bool is_cmp_spec_v = is_cmp_spec<T>::value;

template <typename T> struct is_eq_spec : std::false_type {};
template <typename B> struct is_eq_spec<eq_pack<B>> : std::true_type {};
template <std::size_t N, typename B>
struct is_eq_spec<eq_at_pack<N, B>> : std::true_type {};
template <typename T>
inline constexpr bool is_eq_spec_v = is_eq_spec<T>::value;

template <typename T>
inline constexpr bool is_dcf_spec_v = is_cmp_spec_v<T>;

}  // namespace dpf

#endif  // LIBDPF_INCLUDE_DPF_DCF_HPP__