/// @file dpf/geneval.hpp
/// @brief Fused generation and evaluation (Doerner–Shelat on the eval trie).
/// @details `make_dpf` / `make_dpf_doerner_shelat` build a reusable key, then
///          `eval_*` walks it. `geneval_*` does both at once: one correction
///          word per level, opened from the XOR-reduction of the nodes the
///          public query actually expands. While the secret path's parent is
///          still in that trie the word matches the reusable key byte for
///          byte (same roots, same Beaver tape). After the path leaves, the
///          word is uniform and later outputs still reconstruct — off-path
///          nodes are identical across the two parties, so a dummy word
///          cancels.
///
///          Default calls take XOR shares of the point. Tagged with
///          `arith_input`, the point is the ring sum of the two shares; path
///          bits are opened by a carry chain inside the local CW protocol so
///          the words match `make_dpf(x0 + x1)` at the caller's query.
///
///          `geneval_cmp` is the comparison-channel form. The value-correction
///          word is a function of the secret path at every level, so the walk
///          stays live for the whole depth and the opened words match a
///          Doerner–Shelat comparison key. Prefix shares are
///          `eval_point(cmp, ...)` at each endpoint. Piecewise-cubic evaluation
///          on top of that is `grotto::geneval_offset_horner`.
/// @copyright Copyright (c) 2019-2026 Ryan Henry and [others](@ref authors)
/// @license Released under a GNU General Public v2.0 (GPLv2) license;
///          see [LICENSE.md](@ref license) for details.

#ifndef LIBDPF_INCLUDE_DPF_GENEVAL_HPP__
#define LIBDPF_INCLUDE_DPF_GENEVAL_HPP__

#include <algorithm>
#include <cstddef>
#include <cstdint>
#include <cstring>
#include <iterator>
#include <stdexcept>
#include <tuple>
#include <type_traits>
#include <utility>
#include <vector>

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

#include "dpf/aligned_allocator.hpp"
#include "dpf/doerner_shelat.hpp"
#include "dpf/eval_target.hpp"
#include "dpf/leaf_node.hpp"

namespace dpf
{

/// @brief Shares and the correction words opened along the query trie.
/// @details `correction_words[i]` / `correction_advice[i]` match a reusable key at
/// the same target for every `i < live_levels`. `leaf_live` means the
/// target's leaf was in the trie, so `leaf` is that key's leaf word.
/// @tparam Output output
/// @tparam Leaf leaf
template <typename Output, typename Leaf>
struct geneval_result
{
    std::vector<Output> party0;
    std::vector<Output> party1;
    HEDLEY_PRAGMA(GCC diagnostic push)
    HEDLEY_PRAGMA(GCC diagnostic ignored "-Wignored-attributes")
    std::vector<simde__m128i, aligned_allocator<simde__m128i>> correction_words;
    HEDLEY_PRAGMA(GCC diagnostic pop)
    std::vector<uint8_t> correction_advice;
    std::size_t live_levels = 0;
    bool leaf_live = false;
    Leaf leaf{};
};

namespace detail
{

template <typename T>
HEDLEY_ALWAYS_INLINE
HEDLEY_NO_THROW
T geneval_mod_add(T a, T b) noexcept
{
    using U = std::make_unsigned_t<T>;
    U sum = static_cast<U>(static_cast<U>(a) + static_cast<U>(b));
    T out;
    std::memcpy(&out, &sum, sizeof(out));
    return out;
}

template <typename T>
T geneval_flipped(T x)
{
    utils::flip_msb_if_signed_integral(x);
    return x;
}

/// @brief Leaf-node id of an already MSB-flipped input. The id is the high
/// `depth` bits; the low `lg(outputs_per_leaf)` bits select the lane.
/// @tparam Dpf dpf
/// @param x the `x`
/// @return Leaf-node id of an already MSB-flipped input
template <typename Dpf>
uint64_t geneval_leaf_id(typename Dpf::input_type x)
{
    return static_cast<uint64_t>(utils::get_from_node<Dpf>(x));
}

inline uint64_t geneval_prefix(uint64_t leaf, std::size_t depth, std::size_t bits)
{
    if (bits == 0)
        return 0;
    if (bits >= depth)
        return leaf;
    return leaf >> (depth - bits);
}

inline bool geneval_any_prefix(const std::vector<uint64_t> & leaves,
    std::size_t depth, uint64_t id, std::size_t bits)
{
    if (leaves.empty())
        return false;
    if (bits == 0)
        return true;
    const std::size_t sh = depth - bits;
    const uint64_t lo = (sh >= 64) ? 0 : (id << sh);
    auto it = std::lower_bound(leaves.begin(), leaves.end(), lo);
    if (it == leaves.end())
        return false;
    return geneval_prefix(*it, depth, bits) == id;
}

template <typename Output, typename Leaf>
geneval_result<Output, Leaf> geneval_empty_result()
{
    geneval_result<Output, Leaf> out;
    std::memset(&out.leaf, 0, sizeof(out.leaf));
    return out;
}

template <typename InteriorPRG,
          typename ExteriorPRG,
          typename InputT,
          typename OutputT,
          typename RootSampler,
          typename PadRng>
auto geneval_run(bool arith, bool arith_out, InputT x0, InputT x1,
    const std::vector<InputT> & queries, RootSampler & root_sampler,
    PadRng & pads, OutputT y0, OutputT y1 = OutputT{})
{
    static_assert(std::is_integral_v<InputT>,
        "geneval input shares are an integral domain");
    static_assert(!dpf::is_wildcard_v<OutputT>,
        "geneval output is concrete; assign a wildcard leaf on a key");
    static_assert(utils::bitlength_of_v<InputT> <= 64,
        "geneval leaf ids are 64-bit");

    using dpf_type = utils::dpf_type_t<InteriorPRG, ExteriorPRG, InputT, OutputT>;
    using node = typename dpf_type::interior_node;
    HEDLEY_PRAGMA(GCC diagnostic push)
    HEDLEY_PRAGMA(GCC diagnostic ignored "-Wignored-attributes")
    using leaf_node = leaf_node_t<node, OutputT>;
    HEDLEY_PRAGMA(GCC diagnostic pop)
    using outputs_tuple = std::tuple<OutputT>;
    constexpr std::size_t depth = dpf_type::depth;

    if (queries.empty())
        return geneval_empty_result<OutputT, leaf_node>();

    if (queries.size() > (std::size_t{1} << 22))
        throw std::length_error("geneval query is too large");

    local_cw_protocol<PadRng> proto{pads};
    InputT x0c = x0;
    InputT x1c = x1;
    proto.encode_walk_shares(x0c, x1c, arith);
    const InputT alpha = utils::xor_input_shares(x0c, x1c);

    std::vector<InputT> flipped;
    flipped.reserve(queries.size());
    std::vector<uint64_t> leaves;
    leaves.reserve(queries.size());
    for (const InputT & q : queries)
    {
        InputT fq = geneval_flipped(q);
        flipped.push_back(fq);
        leaves.push_back(geneval_leaf_id<dpf_type>(fq));
    }
    std::vector<uint64_t> unique_leaves = leaves;
    std::sort(unique_leaves.begin(), unique_leaves.end());
    unique_leaves.erase(std::unique(unique_leaves.begin(), unique_leaves.end()),
        unique_leaves.end());
    if (unique_leaves.size() > (std::size_t{1} << 20))
        throw std::length_error("geneval trie is too large");

    const uint64_t secret_leaf = geneval_leaf_id<dpf_type>(alpha);

    constexpr auto to_int = utils::to_integral_type<InputT>{};

    using tree = dpf::tree_traits<InteriorPRG>;
HEDLEY_PRAGMA(GCC diagnostic push)
HEDLEY_PRAGMA(GCC diagnostic ignored "-Wignored-attributes")
    node roots[2];
HEDLEY_PRAGMA(GCC diagnostic pop)
    tree::root_init(roots, [&]() -> node {
        return static_cast<node>(root_sampler());
    });
    const node root0 = roots[0];
    const node root1 = roots[1];

    struct slot
    {
        uint64_t id;
        node s0;
        node s1;
    };
    std::vector<slot> frontier;
    frontier.push_back(slot{0, root0, root1});

    HEDLEY_PRAGMA(GCC diagnostic push)
    HEDLEY_PRAGMA(GCC diagnostic ignored "-Wignored-attributes")
    geneval_result<OutputT, leaf_node> result;
    HEDLEY_PRAGMA(GCC diagnostic pop)
    std::memset(&result.leaf, 0, sizeof(result.leaf));
    result.correction_words.reserve(depth);
    result.correction_advice.reserve(depth);

    auto mask = dpf_type::msb_mask;
    bool still_live = true;
    for (std::size_t level = 0; level < depth; ++level, mask >>= 1)
    {
        const uint8_t bit0 = static_cast<uint8_t>(!!(to_int(mask) & to_int(x0c)));
        const uint8_t bit1 = static_cast<uint8_t>(!!(to_int(mask) & to_int(x1c)));
        const uint64_t parent_id = geneval_prefix(secret_leaf, depth, level);
        const bool is_last = tree::is_last_level(level, depth);

        node L0 = simde_mm_setzero_si128();
        node R0 = simde_mm_setzero_si128();
        node L1 = simde_mm_setzero_si128();
        node R1 = simde_mm_setzero_si128();
        bool level_live = false;

        struct exp
        {
            uint64_t id;
            node s0, s1, L0, R0, L1, R1;
        };
        std::vector<exp> exps;
        exps.reserve(frontier.size());
        for (const slot & n : frontier)
        {
            if (n.id == parent_id)
                level_live = true;
            const auto c0 = tree::expand(n.s0, is_last);
            const auto c1 = tree::expand(n.s1, is_last);
            L0 = ds_xor(L0, c0[0]);
            R0 = ds_xor(R0, c0[1]);
            L1 = ds_xor(L1, c1[0]);
            R1 = ds_xor(R1, c1[1]);
            exps.push_back(exp{n.id, n.s0, n.s1, c0[0], c0[1], c1[0], c1[1]});
        }

        node cw;
        uint8_t advice;
        if (still_live && level_live)
        {
            auto blinds = proto.prepare_level(L0, R0, bit0, L1, R1, bit1);
            auto opened = proto.open_cw(blinds);
            cw = opened.first;
            advice = opened.second;
            if constexpr (tree::is_half_tree)
            {
                if (!is_last)
                    advice = 0;
            }
            ++result.live_levels;
        }
        else
        {
            still_live = false;
            cw = pads.block();
            if constexpr (tree::is_half_tree)
            {
                if (!is_last)
                {
                    advice = 0;
                }
                else
                {
                    const uint8_t t0 = static_cast<uint8_t>(pads.bit() & 1u);
                    const uint8_t t1 = static_cast<uint8_t>(pads.bit() & 1u);
                    advice = static_cast<uint8_t>((t1 << 1) | t0);
                }
            }
            else
            {
                const uint8_t t0 = static_cast<uint8_t>(pads.bit() & 1u);
                const uint8_t t1 = static_cast<uint8_t>(pads.bit() & 1u);
                advice = static_cast<uint8_t>((t1 << 1) | t0);
            }
        }
        result.correction_words.push_back(cw);
        result.correction_advice.push_back(advice);

        const node cw0 = tree::pack_cw(cw, advice, false, is_last);
        const node cw1 = tree::pack_cw(cw, advice, true, is_last);
        const std::size_t child_bits = level + 1;
        std::vector<slot> next;
        next.reserve(exps.size() * 2);
        for (const exp & e : exps)
        {
            const uint64_t left = e.id << 1;
            const uint64_t right = left | 1ull;
            if (geneval_any_prefix(unique_leaves, depth, left, child_bits))
            {
                next.push_back(slot{left,
                    dpf::xor_if_lo_bit(e.L0, cw0, e.s0),
                    dpf::xor_if_lo_bit(e.L1, cw0, e.s1)});
            }
            if (geneval_any_prefix(unique_leaves, depth, right, child_bits))
            {
                next.push_back(slot{right,
                    dpf::xor_if_lo_bit(e.R0, cw1, e.s0),
                    dpf::xor_if_lo_bit(e.R1, cw1, e.s1)});
            }
        }
        frontier = std::move(next);
    }

    result.leaf_live = geneval_any_prefix(unique_leaves, depth, secret_leaf, depth);
    if (result.leaf_live)
    {
        const slot * on = nullptr;
        for (const slot & n : frontier)
        {
            if (n.id == secret_leaf)
            {
                on = &n;
                break;
            }
        }
        if (on == nullptr)
            throw std::logic_error("geneval: secret leaf missing from trie");
        if (arith_out)
        {
            const uint8_t t0 = static_cast<uint8_t>(dpf::get_lo_bit(on->s0));
            const uint8_t t1 = static_cast<uint8_t>(dpf::get_lo_bit(on->s1));
            const std::size_t lane = static_cast<std::size_t>(to_int(alpha));
            result.leaf = proto.template open_arith_leaf<ExteriorPRG, 0, outputs_tuple>(
                dpf::unset_lo_2bits(on->s0), dpf::unset_lo_2bits(on->s1), t0, t1,
                y0, y1, std::size_t{0}, lane);
        }
        else
        {
            const bool sign0 = dpf::get_lo_bit(on->s0);
            auto built = dpf::make_leaves<ExteriorPRG>(alpha,
                dpf::unset_lo_2bits(on->s0), dpf::unset_lo_2bits(on->s1), sign0,
                std::size_t{0}, y0);
            result.leaf = std::get<0>(built.first.first);
        }
    }

    result.party0.reserve(flipped.size());
    result.party1.reserve(flipped.size());
    for (std::size_t i = 0; i < flipped.size(); ++i)
    {
        const uint64_t id = leaves[i];
        const slot * n = nullptr;
        for (const slot & s : frontier)
        {
            if (s.id == id)
            {
                n = &s;
                break;
            }
        }
        if (n == nullptr)
            throw std::logic_error("geneval: query leaf missing from trie");
        auto share0 = dpf_type::template traverse_exterior<0>(n->s0, result.leaf);
        auto share1 = dpf_type::template traverse_exterior<0>(n->s1, result.leaf);
        const auto lane = static_cast<std::size_t>(to_int(flipped[i]));
        result.party0.push_back(extract_leaf<node, OutputT>(share0, lane));
        result.party1.push_back(extract_leaf<node, OutputT>(share1, lane));
    }
    return result;
}

template <typename InteriorPRG,
          typename ExteriorPRG,
          typename InputT,
          typename OutputT,
          typename RootSampler,
          typename PadRng>
auto geneval_run(bool arith, InputT x0, InputT x1,
    const std::vector<InputT> & queries, RootSampler & root_sampler,
    PadRng & pads, OutputT y)
{
    return geneval_run<InteriorPRG, ExteriorPRG>(arith, false, x0, x1, queries,
        root_sampler, pads, y, OutputT{});
}

template <typename InteriorPRG,
          typename ExteriorPRG,
          typename InputT,
          typename OutputT,
          typename RootSampler,
          typename PadRng>
auto geneval_run(InputT x0, InputT x1, const std::vector<InputT> & queries,
    RootSampler & root_sampler, PadRng & pads, OutputT y)
{
    return geneval_run<InteriorPRG, ExteriorPRG>(false, false, x0, x1, queries,
        root_sampler, pads, y, OutputT{});
}

template <typename InputT>
InputT geneval_from_bits(uint64_t bits)
{
    using U = std::make_unsigned_t<InputT>;
    U u = static_cast<U>(bits);
    InputT out;
    std::memcpy(&out, &u, sizeof(out));
    return out;
}

template <typename InputT>
bool geneval_out_of_order(InputT from, InputT to)
{
    // Numeric order. An unsigned compare of a signed value treats a negative
    // `from` as larger than a positive `to`, and would reject `[-1, 1]`.
    if constexpr (std::is_signed_v<InputT>)
        return from > to;
    else
        return utils::to_integral_type<InputT>{}(from)
            > utils::to_integral_type<InputT>{}(to);
}

template <typename InputT>
std::vector<InputT> geneval_full_domain()
{
    constexpr std::size_t bitlen = utils::bitlength_of_v<InputT>;
    if (bitlen > 20)
        throw std::length_error("geneval_full domain is too large");
    const uint64_t n = uint64_t{1} << bitlen;
    std::vector<InputT> qs(static_cast<std::size_t>(n));
    // Index `i` is the input's bit pattern, including the sign bit. A
    // narrowing cast of `i` to a signed type is implementation-defined.
    for (uint64_t i = 0; i < n; ++i)
        qs[static_cast<std::size_t>(i)] = geneval_from_bits<InputT>(i);
    return qs;
}

template <typename InputT>
std::vector<InputT> geneval_inclusive(InputT from, InputT to)
{
    if (geneval_out_of_order(from, to))
    {
        throw std::invalid_argument("geneval_interval: from > to");
    }
    std::vector<InputT> qs;
    InputT q = from;
    const InputT one = utils::make_from_integral_value<InputT>{}(1);
    for (;;)
    {
        qs.push_back(q);
        if (q == to)
            break;
        q = geneval_mod_add(q, one);
        if (qs.size() > (std::size_t{1} << 22))
            throw std::length_error("geneval_interval is too large");
    }
    return qs;
}

}  // namespace detail

/// @name Point geneval
/// @tparam InteriorPRG PRG that expands interior nodes. Defaults to `dpf::prg::aes128`
/// @tparam ExteriorPRG PRG that expands the root. Defaults to `InteriorPRG`
/// @tparam InputT input domain type
/// @tparam OutputT output type
/// @tparam RootSampler sampler for the Doerner–Shelat root seed
/// @tparam PadRng pad stream for the Doerner–Shelat protocol
/// @param x0 party 0's share of the secret point
/// @param x1 party 1's share of the secret point
/// @param query the query point
/// @param rng the Doerner–Shelat randomness tapes
/// @{

/// @brief The secret point is `x0 XOR x1`.
/// @param x0 party 0's share of the secret point
/// @param x1 party 1's share of the secret point
/// @param query the query point
/// @param rng the Doerner–Shelat randomness tapes
/// @param y the payload
/// @return the opened shares and correction words
template <typename InteriorPRG = dpf::prg::aes128,
          typename ExteriorPRG = InteriorPRG,
          typename InputT,
          typename OutputT,
          typename RootSampler,
          typename PadRng>
HEDLEY_WARN_UNUSED_RESULT
auto geneval_point(InputT x0, InputT x1, InputT query,
    ds_randomness<RootSampler, PadRng> rng, OutputT y)
{
    return detail::geneval_run<InteriorPRG, ExteriorPRG>(false, false, x0, x1,
        std::vector<InputT>{query}, rng.root, rng.pad, y, OutputT{});
}

/// @brief The secret point is `x0 + x1`.
/// @param x0 party 0's share of the secret point
/// @param x1 party 1's share of the secret point
/// @param query the query point
/// @param rng the Doerner–Shelat randomness tapes
/// @param y the payload
/// @return the opened shares and correction words
template <typename InteriorPRG = dpf::prg::aes128,
          typename ExteriorPRG = InteriorPRG,
          typename InputT,
          typename OutputT,
          typename RootSampler,
          typename PadRng>
HEDLEY_WARN_UNUSED_RESULT
auto geneval_point(arith_input_t, InputT x0, InputT x1, InputT query,
    ds_randomness<RootSampler, PadRng> rng, OutputT y)
{
    return detail::geneval_run<InteriorPRG, ExteriorPRG>(true, false, x0, x1,
        std::vector<InputT>{query}, rng.root, rng.pad, y, OutputT{});
}

/// @brief XOR-index shares, additively shared payload `y0 + y1 = β`.
/// @param x0 party 0's share of the secret point
/// @param x1 party 1's share of the secret point
/// @param query the query point
/// @param rng the Doerner–Shelat randomness tapes
/// @param y0 party 0's share of the payload
/// @param y1 party 1's share of the payload
/// @return the opened shares and correction words
template <typename InteriorPRG = dpf::prg::aes128,
          typename ExteriorPRG = InteriorPRG,
          typename InputT,
          typename OutputT,
          typename RootSampler,
          typename PadRng>
HEDLEY_WARN_UNUSED_RESULT
auto geneval_point(arith_output_t, InputT x0, InputT x1, InputT query,
    ds_randomness<RootSampler, PadRng> rng, OutputT y0, OutputT y1)
{
    return detail::geneval_run<InteriorPRG, ExteriorPRG>(false, true, x0, x1,
        std::vector<InputT>{query}, rng.root, rng.pad, y0, y1);
}

/// @brief Additive index and additive payload shares.
/// @param x0 party 0's share of the secret point
/// @param x1 party 1's share of the secret point
/// @param query the query point
/// @param rng the Doerner–Shelat randomness tapes
/// @param y0 party 0's share of the payload
/// @param y1 party 1's share of the payload
/// @return the opened shares and correction words
template <typename InteriorPRG = dpf::prg::aes128,
          typename ExteriorPRG = InteriorPRG,
          typename InputT,
          typename OutputT,
          typename RootSampler,
          typename PadRng>
HEDLEY_WARN_UNUSED_RESULT
auto geneval_point(arith_input_t, arith_output_t, InputT x0, InputT x1,
    InputT query, ds_randomness<RootSampler, PadRng> rng, OutputT y0, OutputT y1)
{
    return detail::geneval_run<InteriorPRG, ExteriorPRG>(true, true, x0, x1,
        std::vector<InputT>{query}, rng.root, rng.pad, y0, y1);
}

/// @}

/// @brief Geneval on the inclusive interval `[from, to]`.
/// @tparam InteriorPRG PRG that expands interior nodes. Defaults to `dpf::prg::aes128`
/// @tparam ExteriorPRG PRG that expands the root. Defaults to `InteriorPRG`
/// @tparam InputT input domain type
/// @tparam OutputT output type
/// @tparam RootSampler sampler for the Doerner–Shelat root seed
/// @tparam PadRng pad stream for the Doerner–Shelat protocol
/// @param x0 the `x0`
/// @param x1 the `x1`
/// @param from the inclusive start of the range
/// @param to the `to`
/// @param rng the Doerner–Shelat randomness tapes
/// @param y the `y`
/// @return Geneval on the inclusive interval `[from, to]`
template <typename InteriorPRG = dpf::prg::aes128,
          typename ExteriorPRG = InteriorPRG,
          typename InputT,
          typename OutputT,
          typename RootSampler,
          typename PadRng>
HEDLEY_WARN_UNUSED_RESULT
auto geneval_interval(InputT x0, InputT x1, InputT from, InputT to,
    ds_randomness<RootSampler, PadRng> rng, OutputT y)
{
    return detail::geneval_run<InteriorPRG, ExteriorPRG>(false, x0, x1,
        detail::geneval_inclusive(from, to), rng.root, rng.pad, y);
}

template <typename InteriorPRG = dpf::prg::aes128,
          typename ExteriorPRG = InteriorPRG,
          typename InputT,
          typename OutputT,
          typename RootSampler,
          typename PadRng>
HEDLEY_WARN_UNUSED_RESULT
auto geneval_interval(arith_input_t, InputT x0, InputT x1, InputT from,
    InputT to, ds_randomness<RootSampler, PadRng> rng, OutputT y)
{
    return detail::geneval_run<InteriorPRG, ExteriorPRG>(true, x0, x1,
        detail::geneval_inclusive(from, to), rng.root, rng.pad, y);
}

/// @brief Geneval on the whole domain. Refuses a domain above 2^20 inputs.
/// @tparam InteriorPRG PRG that expands interior nodes. Defaults to `dpf::prg::aes128`
/// @tparam ExteriorPRG PRG that expands the root. Defaults to `InteriorPRG`
/// @tparam InputT input domain type
/// @tparam OutputT output type
/// @tparam RootSampler sampler for the Doerner–Shelat root seed
/// @tparam PadRng pad stream for the Doerner–Shelat protocol
/// @param x0 the `x0`
/// @param x1 the `x1`
/// @param rng the Doerner–Shelat randomness tapes
/// @param y the `y`
/// @return Geneval on the whole domain
template <typename InteriorPRG = dpf::prg::aes128,
          typename ExteriorPRG = InteriorPRG,
          typename InputT,
          typename OutputT,
          typename RootSampler,
          typename PadRng>
HEDLEY_WARN_UNUSED_RESULT
auto geneval_full(InputT x0, InputT x1,
    ds_randomness<RootSampler, PadRng> rng, OutputT y)
{
    return detail::geneval_run<InteriorPRG, ExteriorPRG>(false, x0, x1,
        detail::geneval_full_domain<InputT>(), rng.root, rng.pad, y);
}

template <typename InteriorPRG = dpf::prg::aes128,
          typename ExteriorPRG = InteriorPRG,
          typename InputT,
          typename OutputT,
          typename RootSampler,
          typename PadRng>
HEDLEY_WARN_UNUSED_RESULT
auto geneval_full(arith_input_t, InputT x0, InputT x1,
    ds_randomness<RootSampler, PadRng> rng, OutputT y)
{
    return detail::geneval_run<InteriorPRG, ExteriorPRG>(true, x0, x1,
        detail::geneval_full_domain<InputT>(), rng.root, rng.pad, y);
}

/// @brief Geneval on a public sequence, in the order given.
/// @tparam InteriorPRG PRG that expands interior nodes. Defaults to `dpf::prg::aes128`
/// @tparam ExteriorPRG PRG that expands the root. Defaults to `InteriorPRG`
/// @tparam InputT input domain type
/// @tparam OutputT output type
/// @tparam RootSampler sampler for the Doerner–Shelat root seed
/// @tparam PadRng pad stream for the Doerner–Shelat protocol
/// @tparam ForwardIterator forward iterator type
/// @param x0 the `x0`
/// @param x1 the `x1`
/// @param begin the iterator to the first query
/// @param end the iterator past the last query
/// @param rng the Doerner–Shelat randomness tapes
/// @param y the `y`
/// @return Geneval on a public sequence, in the order given
template <typename InteriorPRG = dpf::prg::aes128,
          typename ExteriorPRG = InteriorPRG,
          typename InputT,
          typename OutputT,
          typename RootSampler,
          typename PadRng,
          typename ForwardIterator>
HEDLEY_WARN_UNUSED_RESULT
auto geneval_sequence(InputT x0, InputT x1, ForwardIterator begin,
    ForwardIterator end, ds_randomness<RootSampler, PadRng> rng, OutputT y)
{
    std::vector<InputT> qs(begin, end);
    return detail::geneval_run<InteriorPRG, ExteriorPRG>(false, x0, x1,
        std::move(qs), rng.root, rng.pad, y);
}

template <typename InteriorPRG = dpf::prg::aes128,
          typename ExteriorPRG = InteriorPRG,
          typename InputT,
          typename OutputT,
          typename RootSampler,
          typename PadRng,
          typename ForwardIterator>
HEDLEY_WARN_UNUSED_RESULT
auto geneval_sequence(arith_input_t, InputT x0, InputT x1,
    ForwardIterator begin, ForwardIterator end,
    ds_randomness<RootSampler, PadRng> rng, OutputT y)
{
    std::vector<InputT> qs(begin, end);
    return detail::geneval_run<InteriorPRG, ExteriorPRG>(true, x0, x1,
        std::move(qs), rng.root, rng.pad, y);
}

/// @brief Opened comparison key material and one prefix share per endpoint.
/// @details `live_levels` is the full depth: a comparison value word depends on the
/// secret path at every level, so there is no early dummy-word tail.
struct geneval_cmp_result
{
    std::vector<uint64_t> party0;
    std::vector<uint64_t> party1;
    HEDLEY_PRAGMA(GCC diagnostic push)
    HEDLEY_PRAGMA(GCC diagnostic ignored "-Wignored-attributes")
    std::vector<simde__m128i, aligned_allocator<simde__m128i>> correction_words;
    HEDLEY_PRAGMA(GCC diagnostic pop)
    std::vector<uint8_t> correction_advice;
    std::vector<uint64_t> value_cw;
    std::vector<uint64_t> tail_cw;
    uint64_t cw_last = 0;
    uint64_t addend0 = 0;
    uint64_t addend1 = 0;
    uint64_t mask = 0;
    std::size_t live_levels = 0;
};

/// @name Comparison geneval
/// @tparam InputT input domain type
/// @tparam ForwardIterator forward iterator type
/// @tparam RootSampler sampler for the Doerner–Shelat root seed
/// @tparam PadRng pad stream for the Doerner–Shelat protocol
/// @param x0 party 0's share of the secret point
/// @param x1 party 1's share of the secret point
/// @param begin the iterator to the first query
/// @param end the iterator past the last query
/// @param rng the Doerner–Shelat randomness tapes
/// @return the opened comparison shares
/// @{

/// @brief `x0 XOR x1` is the secret point, in the same share convention as
/// `geneval_point`. `spec` is an `lt` / `leq` / `gt` / `geq` pack. Each
/// endpoint is returned in order as the two parties' `eval_point(cmp, ...)`
/// shares. An empty range opens nothing.
/// @tparam Spec comparison or interval specification
/// @param x0 party 0's share of the secret point
/// @param x1 party 1's share of the secret point
/// @param begin the iterator to the first query
/// @param end the iterator past the last query
/// @param rng the Doerner–Shelat randomness tapes
/// @param spec the comparison specification
template <typename InputT,
          typename ForwardIterator,
          typename RootSampler,
          typename PadRng,
          typename Spec>
HEDLEY_WARN_UNUSED_RESULT
geneval_cmp_result geneval_cmp(InputT x0, InputT x1,
    ForwardIterator begin, ForwardIterator end,
    ds_randomness<RootSampler, PadRng> rng, Spec spec)
{
    geneval_cmp_result out;
    if (begin == end)
        return out;

    auto keys = make_dpf_doerner_shelat(std::move(x0), std::move(x1),
        std::move(rng), std::move(spec));
    const auto & k0 = keys.first;
    const auto & k1 = keys.second;
    using key_type = std::decay_t<decltype(k0)>;
    constexpr std::size_t depth = key_type::depth;
    out.live_levels = depth;
    out.mask = k0.cmp().mask;
    out.cw_last = k0.cw_last();
    out.addend0 = k0.cmp_addend().raw();
    out.addend1 = k1.cmp_addend().raw();
    out.correction_words.resize(depth);
    out.correction_advice.resize(depth);
    if constexpr (key_type::cmp_block > 0)
    {
        out.value_cw.resize(key_type::cmp_checkpoints);
        for (std::size_t i = 0; i < key_type::cmp_checkpoints; ++i)
            out.value_cw[i] = k0.value_cw(i);
        out.tail_cw.resize(key_type::cmp_tail);
        for (std::size_t z = 0; z < key_type::cmp_tail; ++z)
            out.tail_cw[z] = k0.tail_cw(z);
    }
    else
        out.value_cw.resize(depth);
    for (std::size_t level = 0; level < depth; ++level)
    {
        out.correction_words[level] = k0.correction_word(level);
        out.correction_advice[level] = static_cast<uint8_t>(k0.correction_advice(level));
        if constexpr (key_type::cmp_block == 0)
            out.value_cw[level] = k0.value_cw(level);
    }
    for (auto it = begin; it != end; ++it)
    {
        out.party0.push_back(eval_point(dpf::cmp, k0, *it).raw());
        out.party1.push_back(eval_point(dpf::cmp, k1, *it).raw());
    }
    return out;
}

/// @brief Additive shares of the point (`x0 + x1`).
/// @tparam Spec comparison or interval specification
/// @param x0 party 0's share of the secret point
/// @param x1 party 1's share of the secret point
/// @param begin the iterator to the first query
/// @param end the iterator past the last query
/// @param rng the Doerner–Shelat randomness tapes
/// @param spec the comparison specification
template <typename InputT,
          typename ForwardIterator,
          typename RootSampler,
          typename PadRng,
          typename Spec>
HEDLEY_WARN_UNUSED_RESULT
geneval_cmp_result geneval_cmp(arith_input_t, InputT x0, InputT x1,
    ForwardIterator begin, ForwardIterator end,
    ds_randomness<RootSampler, PadRng> rng, Spec spec)
{
    geneval_cmp_result out;
    if (begin == end)
        return out;

    auto keys = make_dpf_doerner_shelat(arith_input, std::move(x0), std::move(x1),
        std::move(rng), std::move(spec));
    const auto & k0 = keys.first;
    const auto & k1 = keys.second;
    using key_type = std::decay_t<decltype(k0)>;
    constexpr std::size_t depth = key_type::depth;
    out.live_levels = depth;
    out.mask = k0.cmp().mask;
    out.cw_last = k0.cw_last();
    out.addend0 = k0.cmp_addend().raw();
    out.addend1 = k1.cmp_addend().raw();
    out.correction_words.resize(depth);
    out.correction_advice.resize(depth);
    if constexpr (key_type::cmp_block > 0)
    {
        out.value_cw.resize(key_type::cmp_checkpoints);
        for (std::size_t i = 0; i < key_type::cmp_checkpoints; ++i)
            out.value_cw[i] = k0.value_cw(i);
        out.tail_cw.resize(key_type::cmp_tail);
        for (std::size_t z = 0; z < key_type::cmp_tail; ++z)
            out.tail_cw[z] = k0.tail_cw(z);
    }
    else
        out.value_cw.resize(depth);
    for (std::size_t level = 0; level < depth; ++level)
    {
        out.correction_words[level] = k0.correction_word(level);
        out.correction_advice[level] = static_cast<uint8_t>(k0.correction_advice(level));
        if constexpr (key_type::cmp_block == 0)
            out.value_cw[level] = k0.value_cw(level);
    }
    for (auto it = begin; it != end; ++it)
    {
        out.party0.push_back(eval_point(dpf::cmp, k0, *it).raw());
        out.party1.push_back(eval_point(dpf::cmp, k1, *it).raw());
    }
    return out;
}

/// @brief `gt(beta)` on XOR shares of the point. `if_false` is 0.
/// @param x0 party 0's share of the secret point
/// @param x1 party 1's share of the secret point
/// @param begin the iterator to the first query
/// @param end the iterator past the last query
/// @param rng the Doerner–Shelat randomness tapes
/// @param beta the true payload
template <typename InputT,
          typename ForwardIterator,
          typename RootSampler,
          typename PadRng>
HEDLEY_WARN_UNUSED_RESULT
geneval_cmp_result geneval_cmp(InputT x0, InputT x1,
    ForwardIterator begin, ForwardIterator end,
    ds_randomness<RootSampler, PadRng> rng, uint64_t beta)
{
    return geneval_cmp(std::move(x0), std::move(x1), begin, end,
        std::move(rng), dpf::gt(beta));
}

/// @brief `gt(beta)` on additive shares of the point. `if_false` is 0.
/// @param x0 party 0's share of the secret point
/// @param x1 party 1's share of the secret point
/// @param begin the iterator to the first query
/// @param end the iterator past the last query
/// @param rng the Doerner–Shelat randomness tapes
/// @param beta the true payload
template <typename InputT,
          typename ForwardIterator,
          typename RootSampler,
          typename PadRng>
HEDLEY_WARN_UNUSED_RESULT
geneval_cmp_result geneval_cmp(arith_input_t, InputT x0, InputT x1,
    ForwardIterator begin, ForwardIterator end,
    ds_randomness<RootSampler, PadRng> rng, uint64_t beta)
{
    return geneval_cmp(arith_input, std::move(x0), std::move(x1), begin, end,
        std::move(rng), dpf::gt(beta));
}

/// @}

}  // namespace dpf

#endif  // LIBDPF_INCLUDE_DPF_GENEVAL_HPP__