ladybird/Userland/Libraries/LibGfx/ImageFormats/JBIG2Loader.cpp

/*
 * Copyright (c) 2024, Nico Weber <thakis@chromium.org>
 *
 * SPDX-License-Identifier: BSD-2-Clause
 */

#include <AK/Debug.h>
#include <AK/Utf16View.h>
#include <LibGfx/ImageFormats/CCITTDecoder.h>
#include <LibGfx/ImageFormats/JBIG2Loader.h>
#include <LibTextCodec/Decoder.h>

// Spec: ITU-T_T_88__08_2018.pdf in the zip file here:
// https://www.itu.int/rec/T-REC-T.88-201808-I
// Annex H has a datastream example.

namespace Gfx {

namespace JBIG2 {

// Table E.1 – Qe values and probability estimation process
// See also E.1.2 Coding conventions and approximations
// and E.2.5 Probability estimation.
struct QeEntry {
    u16 qe;          // Sub-interval for the less probable symbol.
    u16 nmps;        // Next index if the more probable symbol is decoded
    u16 nlps;        // Next index if the less probable symbol is decoded
    u16 switch_flag; // See second-to-last paragraph in E.1.2.
};
constexpr auto qe_table = to_array<QeEntry>({
    { 0x5601, 1, 1, 1 },
    { 0x3401, 2, 6, 0 },
    { 0x1801, 3, 9, 0 },
    { 0x0AC1, 4, 12, 0 },
    { 0x0521, 5, 29, 0 },
    { 0x0221, 38, 33, 0 },
    { 0x5601, 7, 6, 1 },
    { 0x5401, 8, 14, 0 },
    { 0x4801, 9, 14, 0 },
    { 0x3801, 10, 14, 0 },
    { 0x3001, 11, 17, 0 },
    { 0x2401, 12, 18, 0 },
    { 0x1C01, 13, 20, 0 },
    { 0x1601, 29, 21, 0 },
    { 0x5601, 15, 14, 1 },
    { 0x5401, 16, 14, 0 },
    { 0x5101, 17, 15, 0 },
    { 0x4801, 18, 16, 0 },
    { 0x3801, 19, 17, 0 },
    { 0x3401, 20, 18, 0 },
    { 0x3001, 21, 19, 0 },
    { 0x2801, 22, 19, 0 },
    { 0x2401, 23, 20, 0 },
    { 0x2201, 24, 21, 0 },
    { 0x1C01, 25, 22, 0 },
    { 0x1801, 26, 23, 0 },
    { 0x1601, 27, 24, 0 },
    { 0x1401, 28, 25, 0 },
    { 0x1201, 29, 26, 0 },
    { 0x1101, 30, 27, 0 },
    { 0x0AC1, 31, 28, 0 },
    { 0x09C1, 32, 29, 0 },
    { 0x08A1, 33, 30, 0 },
    { 0x0521, 34, 31, 0 },
    { 0x0441, 35, 32, 0 },
    { 0x02A1, 36, 33, 0 },
    { 0x0221, 37, 34, 0 },
    { 0x0141, 38, 35, 0 },
    { 0x0111, 39, 36, 0 },
    { 0x0085, 40, 37, 0 },
    { 0x0049, 41, 38, 0 },
    { 0x0025, 42, 39, 0 },
    { 0x0015, 43, 40, 0 },
    { 0x0009, 44, 41, 0 },
    { 0x0005, 45, 42, 0 },
    { 0x0001, 45, 43, 0 },
    { 0x5601, 46, 46, 0 },
});

ErrorOr<ArithmeticDecoder> ArithmeticDecoder::initialize(ReadonlyBytes data)
{
    ArithmeticDecoder decoder { data };
    decoder.INITDEC();
    return decoder;
}

bool ArithmeticDecoder::get_next_bit(Context& context)
{
    CX = &context;
    // Useful for comparing to Table H.1 – Encoder and decoder trace data.
    // dbg("I={} MPS={} A={:#x} C={:#x} CT={} B={:#x}", I(CX), MPS(CX), A, C, CT, B());
    u8 D = DECODE();
    // dbgln(" -> D={}", D);
    return D;
}

u16 ArithmeticDecoder::Qe(u16 index) { return qe_table[index].qe; }
u8 ArithmeticDecoder::NMPS(u16 index) { return qe_table[index].nmps; }
u8 ArithmeticDecoder::NLPS(u16 index) { return qe_table[index].nlps; }
u8 ArithmeticDecoder::SWITCH(u16 index) { return qe_table[index].switch_flag; }

u8 ArithmeticDecoder::B(size_t offset) const
{
    // E.2.10 Minimization of the compressed data
    // "the convention is used in the decoder that when a marker code is encountered,
    //  1-bits (without bit stuffing) are supplied to the decoder until the coding interval is complete."
    if (BP + offset >= m_data.size())
        return 0xFF;
    return m_data[BP + offset];
}

void ArithmeticDecoder::INITDEC()
{
    // E.3.5 Initialization of the decoder (INITDEC)
    // Figure G.1 – Initialization of the software conventions decoder

    // "BP, the pointer to the compressed data, is initialized to BPST (pointing to the first compressed byte)."
    auto const BPST = 0;
    BP = BPST;
    C = (B() ^ 0xFF) << 16;

    BYTEIN();

    C = C << 7;
    CT = CT - 7;
    A = 0x8000;
}

u8 ArithmeticDecoder::DECODE()
{
    // E.3.2 Decoding a decision (DECODE)
    // Figure G.2 – Decoding an MPS or an LPS in the software-conventions decoder
    u8 D;
    A = A - Qe(I(CX));
    if (C < ((u32)A << 16)) { // `(C_high < A)` in spec
        if ((A & 0x8000) == 0) {
            D = MPS_EXCHANGE();
            RENORMD();
        } else {
            D = MPS(CX);
        }
    } else {
        C = C - ((u32)A << 16); // `C_high = C_high - A` in spec
        D = LPS_EXCHANGE();
        RENORMD();
    }
    return D;
}

u8 ArithmeticDecoder::MPS_EXCHANGE()
{
    // Figure E.16 – Decoder MPS path conditional exchange procedure
    u8 D;
    if (A < Qe(I(CX))) {
        D = 1 - MPS(CX);
        if (SWITCH(I(CX)) == 1) {
            MPS(CX) = 1 - MPS(CX);
        }
        I(CX) = NLPS(I(CX));
    } else {
        D = MPS(CX);
        I(CX) = NMPS(I(CX));
    }
    return D;
}

u8 ArithmeticDecoder::LPS_EXCHANGE()
{
    // Figure E.17 – Decoder LPS path conditional exchange procedure
    u8 D;
    if (A < Qe(I(CX))) {
        A = Qe(I(CX));
        D = MPS(CX);
        I(CX) = NMPS(I(CX));
    } else {
        A = Qe(I(CX));
        D = 1 - MPS(CX);
        if (SWITCH(I(CX)) == 1) {
            MPS(CX) = 1 - MPS(CX);
        }
        I(CX) = NLPS(I(CX));
    }
    return D;
}

void ArithmeticDecoder::RENORMD()
{
    // E.3.3 Renormalization in the decoder (RENORMD)
    // Figure E.18 – Decoder renormalization procedure
    do {
        if (CT == 0)
            BYTEIN();
        A = A << 1;
        C = C << 1;
        CT = CT - 1;
    } while ((A & 0x8000) == 0);
}

void ArithmeticDecoder::BYTEIN()
{
    // E.3.4 Compressed data input (BYTEIN)
    // Figure G.3 – Inserting a new byte into the C register in the software-conventions decoder
    if (B() == 0xFF) {
        if (B(1) > 0x8F) {
            CT = 8;
        } else {
            BP = BP + 1;
            C = C + 0xFE00 - (B() << 9);
            CT = 7;
        }
    } else {
        BP = BP + 1;
        C = C + 0xFF00 - (B() << 8);
        CT = 8;
    }
}

// Annex A, Arithmetic integer decoding procedure
class ArithmeticIntegerDecoder {
public:
    ArithmeticIntegerDecoder(ArithmeticDecoder&);

    // A.2 Procedure for decoding values (except IAID)
    // Returns OptionalNone for OOB.
    Optional<i32> decode();

private:
    ArithmeticDecoder& m_decoder;
    u16 PREV { 0 };
    Vector<ArithmeticDecoder::Context> contexts;
};

ArithmeticIntegerDecoder::ArithmeticIntegerDecoder(ArithmeticDecoder& decoder)
    : m_decoder(decoder)
{
    contexts.resize(1 << 9);
}

Optional<int> ArithmeticIntegerDecoder::decode()
{
    // A.2 Procedure for decoding values (except IAID)
    // "1) Set:
    //    PREV = 1"
    u16 PREV = 1;

    // "2) Follow the flowchart in Figure A.1. Decode each bit with CX equal to "IAx + PREV" where "IAx" represents the identifier
    //     of the current arithmetic integer decoding procedure, "+" represents concatenation, and the rightmost 9 bits of PREV are used."
    auto decode_bit = [&]() {
        bool D = m_decoder.get_next_bit(contexts[PREV & 0x1FF]);
        // "3) After each bit is decoded:
        //     If PREV < 256 set:
        //         PREV = (PREV << 1) OR D
        //     Otherwise set:
        //         PREV = (((PREV << 1) OR D) AND 511) OR 256
        //     where D represents the value of the just-decoded bit.
        if (PREV < 256)
            PREV = (PREV << 1) | (u16)D;
        else
            PREV = (((PREV << 1) | (u16)D) & 511) | 256;
        return D;
    };

    auto decode_bits = [&](int n) {
        u32 result = 0;
        for (int i = 0; i < n; ++i)
            result = (result << 1) | decode_bit();
        return result;
    };

    // Figure A.1 – Flowchart for the integer arithmetic decoding procedures (except IAID)
    u8 S = decode_bit();
    u32 V;
    if (!decode_bit())
        V = decode_bits(2);
    else if (!decode_bit())
        V = decode_bits(4) + 4;
    else if (!decode_bit())
        V = decode_bits(6) + 20;
    else if (!decode_bit())
        V = decode_bits(8) + 84;
    else if (!decode_bit())
        V = decode_bits(12) + 340;
    else
        V = decode_bits(32) + 4436;

    // "4) The sequence of bits decoded, interpreted according to Table A.1, gives the value that is the result of this invocation
    //     of the integer arithmetic decoding procedure."
    if (S == 1 && V == 0)
        return {};
    return S ? -V : V;
}

}

// JBIG2 spec, Annex D, D.4.1 ID string
static constexpr u8 id_string[] = { 0x97, 0x4A, 0x42, 0x32, 0x0D, 0x0A, 0x1A, 0x0A };

// 7.3 Segment types
enum SegmentType {
    SymbolDictionary = 0,
    IntermediateTextRegion = 4,
    ImmediateTextRegion = 6,
    ImmediateLosslessTextRegion = 7,
    PatternDictionary = 16,
    IntermediateHalftoneRegion = 20,
    ImmediateHalftoneRegion = 22,
    ImmediateLosslessHalftoneRegion = 23,
    IntermediateGenericRegion = 36,
    ImmediateGenericRegion = 38,
    ImmediateLosslessGenericRegion = 39,
    IntermediateGenericRefinementRegion = 40,
    ImmediateGenericRefinementRegion = 42,
    ImmediateLosslessGenericRefinementRegion = 43,
    PageInformation = 48,
    EndOfPage = 49,
    EndOfStripe = 50,
    EndOfFile = 51,
    Profiles = 52,
    Tables = 53,
    ColorPalette = 54,
    Extension = 62,
};

// Annex D
enum class Organization {
    // D.1 Sequential organization
    Sequential,

    // D.2 Random-access organization
    RandomAccess,

    // D.3 Embedded organization
    Embedded,
};

struct SegmentHeader {
    u32 segment_number { 0 };
    SegmentType type { SegmentType::Extension };
    Vector<u32> referred_to_segment_numbers;

    // 7.2.6 Segment page association
    // "The first page must be numbered "1". This field may contain a value of zero; this value indicates that this segment is not associated with any page."
    u32 page_association { 0 };

    Optional<u32> data_length;
};

class BitBuffer {
public:
    static ErrorOr<NonnullOwnPtr<BitBuffer>> create(size_t width, size_t height);
    bool get_bit(size_t x, size_t y) const;
    void set_bit(size_t x, size_t y, bool b);
    void fill(bool b);

    ErrorOr<NonnullRefPtr<Gfx::Bitmap>> to_gfx_bitmap() const;
    ErrorOr<ByteBuffer> to_byte_buffer() const;

    size_t width() const { return m_width; }
    size_t height() const { return m_height; }

private:
    BitBuffer(ByteBuffer, size_t width, size_t height, size_t pitch);

    ByteBuffer m_bits;
    size_t m_width { 0 };
    size_t m_height { 0 };
    size_t m_pitch { 0 };
};

ErrorOr<NonnullOwnPtr<BitBuffer>> BitBuffer::create(size_t width, size_t height)
{
    size_t pitch = ceil_div(width, 8ull);
    auto bits = TRY(ByteBuffer::create_uninitialized(pitch * height));
    return adopt_nonnull_own_or_enomem(new (nothrow) BitBuffer(move(bits), width, height, pitch));
}

bool BitBuffer::get_bit(size_t x, size_t y) const
{
    VERIFY(x < m_width);
    VERIFY(y < m_height);
    size_t byte_offset = x / 8;
    size_t bit_offset = x % 8;
    u8 byte = m_bits[y * m_pitch + byte_offset];
    byte = (byte >> (8 - 1 - bit_offset)) & 1;
    return byte != 0;
}

void BitBuffer::set_bit(size_t x, size_t y, bool b)
{
    VERIFY(x < m_width);
    VERIFY(y < m_height);
    size_t byte_offset = x / 8;
    size_t bit_offset = x % 8;
    u8 byte = m_bits[y * m_pitch + byte_offset];
    u8 mask = 1u << (8 - 1 - bit_offset);
    if (b)
        byte |= mask;
    else
        byte &= ~mask;
    m_bits[y * m_pitch + byte_offset] = byte;
}

void BitBuffer::fill(bool b)
{
    u8 fill_byte = b ? 0xff : 0;
    for (auto& byte : m_bits.bytes())
        byte = fill_byte;
}

ErrorOr<NonnullRefPtr<Gfx::Bitmap>> BitBuffer::to_gfx_bitmap() const
{
    auto bitmap = TRY(Gfx::Bitmap::create(Gfx::BitmapFormat::BGRx8888, { m_width, m_height }));
    for (size_t y = 0; y < m_height; ++y) {
        for (size_t x = 0; x < m_width; ++x) {
            auto color = get_bit(x, y) ? Color::Black : Color::White;
            bitmap->set_pixel(x, y, color);
        }
    }
    return bitmap;
}

ErrorOr<ByteBuffer> BitBuffer::to_byte_buffer() const
{
    return ByteBuffer::copy(m_bits);
}

BitBuffer::BitBuffer(ByteBuffer bits, size_t width, size_t height, size_t pitch)
    : m_bits(move(bits))
    , m_width(width)
    , m_height(height)
    , m_pitch(pitch)
{
}

class Symbol : public RefCounted<Symbol> {
public:
    static NonnullRefPtr<Symbol> create(NonnullOwnPtr<BitBuffer> bitmap)
    {
        return adopt_ref(*new Symbol(move(bitmap)));
    }

    BitBuffer const& bitmap() const { return *m_bitmap; }

private:
    Symbol(NonnullOwnPtr<BitBuffer> bitmap)
        : m_bitmap(move(bitmap))
    {
    }

    NonnullOwnPtr<BitBuffer> m_bitmap;
};

struct SegmentData {
    SegmentHeader header;
    ReadonlyBytes data;

    // Set on dictionary segments after they've been decoded.
    Optional<Vector<NonnullRefPtr<Symbol>>> symbols;
};

// 7.4.8.5 Page segment flags
enum class CombinationOperator {
    Or = 0,
    And = 1,
    Xor = 2,
    XNor = 3,
    Replace = 4,
};

static void composite_bitbuffer(BitBuffer& out, BitBuffer const& bitmap, Gfx::IntPoint position, CombinationOperator operator_)
{
    size_t start_x = 0, end_x = bitmap.width();
    size_t start_y = 0, end_y = bitmap.height();
    if (position.x() < 0) {
        start_x = -position.x();
        position.set_x(0);
    }
    if (position.y() < 0) {
        start_y = -position.y();
        position.set_y(0);
    }
    if (position.x() + bitmap.width() > out.width())
        end_x = out.width() - position.x();
    if (position.y() + bitmap.height() > out.height())
        end_y = out.height() - position.y();

    for (size_t y = start_y; y < end_y; ++y) {
        for (size_t x = start_x; x < end_x; ++x) {
            bool bit = bitmap.get_bit(x, y);
            switch (operator_) {
            case CombinationOperator::Or:
                bit = bit || out.get_bit(position.x() + x, position.y() + y);
                break;
            case CombinationOperator::And:
                bit = bit && out.get_bit(position.x() + x, position.y() + y);
                break;
            case CombinationOperator::Xor:
                bit = bit ^ out.get_bit(position.x() + x, position.y() + y);
                break;
            case CombinationOperator::XNor:
                bit = !(bit ^ out.get_bit(position.x() + x, position.y() + y));
                break;
            case CombinationOperator::Replace:
                // Nothing to do.
                break;
            }
            out.set_bit(position.x() + x, position.y() + y, bit);
        }
    }
}

struct Page {
    IntSize size;

    // This is never CombinationOperator::Replace for Pages.
    CombinationOperator default_combination_operator { CombinationOperator::Or };

    OwnPtr<BitBuffer> bits;
};

struct JBIG2LoadingContext {
    enum class State {
        NotDecoded = 0,
        Error,
        Decoded,
    };
    State state { State::NotDecoded };

    Organization organization { Organization::Sequential };
    Page page;

    Optional<u32> number_of_pages;

    Vector<SegmentData> segments;
    HashMap<u32, u32> segments_by_number;
};

static ErrorOr<void> decode_jbig2_header(JBIG2LoadingContext& context, ReadonlyBytes data)
{
    if (!JBIG2ImageDecoderPlugin::sniff(data))
        return Error::from_string_literal("JBIG2LoadingContext: Invalid JBIG2 header");

    FixedMemoryStream stream(data.slice(sizeof(id_string)));

    // D.4.2 File header flags
    u8 header_flags = TRY(stream.read_value<u8>());
    if (header_flags & 0b11110000)
        return Error::from_string_literal("JBIG2LoadingContext: Invalid header flags");
    context.organization = (header_flags & 1) ? Organization::Sequential : Organization::RandomAccess;
    dbgln_if(JBIG2_DEBUG, "JBIG2LoadingContext: Organization: {} ({})", (int)context.organization, context.organization == Organization::Sequential ? "Sequential" : "Random-access");
    bool has_known_number_of_pages = (header_flags & 2) ? false : true;
    bool uses_templates_with_12_AT_pixels = (header_flags & 4) ? true : false;
    bool contains_colored_region_segments = (header_flags & 8) ? true : false;

    // FIXME: Do something with these?
    (void)uses_templates_with_12_AT_pixels;
    (void)contains_colored_region_segments;

    // D.4.3 Number of pages
    if (has_known_number_of_pages) {
        context.number_of_pages = TRY(stream.read_value<BigEndian<u32>>());
        dbgln_if(JBIG2_DEBUG, "JBIG2LoadingContext: Number of pages: {}", context.number_of_pages.value());
    }

    return {};
}

static ErrorOr<SegmentHeader> decode_segment_header(SeekableStream& stream)
{
    // 7.2.2 Segment number
    u32 segment_number = TRY(stream.read_value<BigEndian<u32>>());
    dbgln_if(JBIG2_DEBUG, "Segment number: {}", segment_number);

    // 7.2.3 Segment header flags
    u8 flags = TRY(stream.read_value<u8>());
    SegmentType type = static_cast<SegmentType>(flags & 0b11'1111);
    dbgln_if(JBIG2_DEBUG, "Segment type: {}", (int)type);
    bool segment_page_association_size_is_32_bits = (flags & 0b100'0000) != 0;
    bool segment_retained_only_by_itself_and_extension_segments = (flags & 0b1000'00000) != 0;

    // FIXME: Do something with these.
    (void)segment_page_association_size_is_32_bits;
    (void)segment_retained_only_by_itself_and_extension_segments;

    // 7.2.4 Referred-to segment count and retention flags
    u8 referred_to_segment_count_and_retention_flags = TRY(stream.read_value<u8>());
    u32 count_of_referred_to_segments = referred_to_segment_count_and_retention_flags >> 5;
    if (count_of_referred_to_segments == 5 || count_of_referred_to_segments == 6)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid count_of_referred_to_segments");
    u32 extra_count = 0;
    if (count_of_referred_to_segments == 7) {
        TRY(stream.seek(-1, SeekMode::FromCurrentPosition));
        count_of_referred_to_segments = TRY(stream.read_value<BigEndian<u32>>()) & 0x1FFF'FFFF;
        extra_count = ceil_div(count_of_referred_to_segments + 1, 8);
        TRY(stream.seek(extra_count, SeekMode::FromCurrentPosition));
    }
    dbgln_if(JBIG2_DEBUG, "Referred-to segment count: {}", count_of_referred_to_segments);

    // 7.2.5 Referred-to segment numbers
    Vector<u32> referred_to_segment_numbers;
    for (u32 i = 0; i < count_of_referred_to_segments; ++i) {
        u32 referred_to_segment_number;
        if (segment_number <= 256)
            referred_to_segment_number = TRY(stream.read_value<u8>());
        else if (segment_number <= 65536)
            referred_to_segment_number = TRY(stream.read_value<BigEndian<u16>>());
        else
            referred_to_segment_number = TRY(stream.read_value<BigEndian<u32>>());
        referred_to_segment_numbers.append(referred_to_segment_number);
        dbgln_if(JBIG2_DEBUG, "Referred-to segment number: {}", referred_to_segment_number);
    }

    // 7.2.6 Segment page association
    u32 segment_page_association;
    if (segment_page_association_size_is_32_bits) {
        segment_page_association = TRY(stream.read_value<BigEndian<u32>>());
    } else {
        segment_page_association = TRY(stream.read_value<u8>());
    }
    dbgln_if(JBIG2_DEBUG, "Segment page association: {}", segment_page_association);

    // 7.2.7 Segment data length
    u32 data_length = TRY(stream.read_value<BigEndian<u32>>());
    dbgln_if(JBIG2_DEBUG, "Segment data length: {}", data_length);

    // FIXME: Add some validity checks:
    // - check type is valid
    // - check referred_to_segment_numbers are smaller than segment_number
    // - 7.3.1 Rules for segment references
    // - 7.3.2 Rules for page associations

    Optional<u32> opt_data_length;
    if (data_length != 0xffff'ffff)
        opt_data_length = data_length;
    else if (type != ImmediateGenericRegion)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Unknown data length only allowed for ImmediateGenericRegion");

    return SegmentHeader { segment_number, type, move(referred_to_segment_numbers), segment_page_association, opt_data_length };
}

static ErrorOr<size_t> scan_for_immediate_generic_region_size(ReadonlyBytes data)
{
    // 7.2.7 Segment data length
    // "If the segment's type is "Immediate generic region", then the length field may contain the value 0xFFFFFFFF.
    //  This value is intended to mean that the length of the segment's data part is unknown at the time that the segment header is written (...).
    //  In this case, the true length of the segment's data part shall be determined through examination of the data:
    //  if the segment uses template-based arithmetic coding, then the segment's data part ends with the two-byte sequence 0xFF 0xAC followed by a four-byte row count.
    //  If the segment uses MMR coding, then the segment's data part ends with the two-byte sequence 0x00 0x00 followed by a four-byte row count.
    //  The form of encoding used by the segment may be determined by examining the eighteenth byte of its segment data part,
    //  and the end sequences can occur anywhere after that eighteenth byte."
    // 7.4.6.4 Decoding a generic region segment
    // "NOTE – The sequence 0x00 0x00 cannot occur within MMR-encoded data; the sequence 0xFF 0xAC can occur only at the end of arithmetically-coded data.
    //  Thus, those sequences cannot occur by chance in the data that is decoded to generate the contents of the generic region."
    dbgln_if(JBIG2_DEBUG, "(Unknown data length, computing it)");

    if (data.size() < 19 + sizeof(u32))
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Data too short to contain segment data header and end sequence");

    // Per 7.4.6.1 Generic region segment data header, this starts with the 17 bytes described in
    // 7.4.1 Region segment information field, followed the byte described in 7.4.6.2 Generic region segment flags.
    // That byte's lowest bit stores if the segment uses MMR.
    u8 flags = data[17];
    bool uses_mmr = (flags & 1) != 0;
    auto end_sequence = uses_mmr ? to_array<u8>({ 0x00, 0x00 }) : to_array<u8>({ 0xFF, 0xAC });
    u8 const* end = static_cast<u8 const*>(memmem(data.data() + 19, data.size() - 19 - sizeof(u32), end_sequence.data(), end_sequence.size()));
    if (!end)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Could not find end sequence in segment data");

    size_t size = end - data.data() + end_sequence.size() + sizeof(u32);
    dbgln_if(JBIG2_DEBUG, "(Computed size is {})", size);
    return size;
}

static ErrorOr<void> decode_segment_headers(JBIG2LoadingContext& context, ReadonlyBytes data)
{
    FixedMemoryStream stream(data);

    Vector<ReadonlyBytes> segment_datas;
    auto store_and_skip_segment_data = [&](SegmentHeader const& segment_header) -> ErrorOr<void> {
        size_t start_offset = TRY(stream.tell());
        u32 data_length = TRY(segment_header.data_length.try_value_or_lazy_evaluated([&]() {
            return scan_for_immediate_generic_region_size(data.slice(start_offset));
        }));

        if (start_offset + data_length > data.size()) {
            dbgln_if(JBIG2_DEBUG, "JBIG2ImageDecoderPlugin: start_offset={}, data_length={}, data.size()={}", start_offset, data_length, data.size());
            return Error::from_string_literal("JBIG2ImageDecoderPlugin: Segment data length exceeds file size");
        }
        ReadonlyBytes segment_data = data.slice(start_offset, data_length);
        segment_datas.append(segment_data);

        TRY(stream.seek(data_length, SeekMode::FromCurrentPosition));
        return {};
    };

    Vector<SegmentHeader> segment_headers;
    while (!stream.is_eof()) {
        auto segment_header = TRY(decode_segment_header(stream));
        segment_headers.append(segment_header);

        if (context.organization != Organization::RandomAccess)
            TRY(store_and_skip_segment_data(segment_header));

        // Required per spec for files with RandomAccess organization.
        if (segment_header.type == SegmentType::EndOfFile)
            break;
    }

    if (context.organization == Organization::RandomAccess) {
        for (auto const& segment_header : segment_headers)
            TRY(store_and_skip_segment_data(segment_header));
    }

    if (segment_headers.size() != segment_datas.size())
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Segment headers and segment datas have different sizes");
    for (size_t i = 0; i < segment_headers.size(); ++i) {
        context.segments.append({ segment_headers[i], segment_datas[i], {} });
        context.segments_by_number.set(segment_headers[i].segment_number, context.segments.size() - 1);
    }

    return {};
}

// 7.4.1 Region segment information field
struct [[gnu::packed]] RegionSegmentInformationField {
    BigEndian<u32> width;
    BigEndian<u32> height;
    BigEndian<u32> x_location;
    BigEndian<u32> y_location;
    u8 flags;

    CombinationOperator external_combination_operator() const
    {
        VERIFY((flags & 0x7) <= 4);
        return static_cast<CombinationOperator>(flags & 0x7);
    }

    bool is_color_bitmap() const
    {
        return (flags & 0x8) != 0;
    }
};
static_assert(AssertSize<RegionSegmentInformationField, 17>());

static ErrorOr<RegionSegmentInformationField> decode_region_segment_information_field(ReadonlyBytes data)
{
    // 7.4.8 Page information segment syntax
    if (data.size() < sizeof(RegionSegmentInformationField))
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid region segment information field size");
    auto result = *(RegionSegmentInformationField const*)data.data();
    if ((result.flags & 0b1111'0000) != 0)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid region segment information field flags");
    if ((result.flags & 0x7) > 4)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid region segment information field operator");

    // NOTE 3 – If the colour extension flag (COLEXTFLAG) is equal to 1, the external combination operator must be REPLACE.
    if (result.is_color_bitmap() && result.external_combination_operator() != CombinationOperator::Replace)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid colored region segment information field operator");

    return result;
}

// 7.4.8 Page information segment syntax
struct [[gnu::packed]] PageInformationSegment {
    BigEndian<u32> bitmap_width;
    BigEndian<u32> bitmap_height;
    BigEndian<u32> page_x_resolution; // In pixels/meter.
    BigEndian<u32> page_y_resolution; // In pixels/meter.
    u8 flags;
    BigEndian<u16> striping_information;
};
static_assert(AssertSize<PageInformationSegment, 19>());

static ErrorOr<PageInformationSegment> decode_page_information_segment(ReadonlyBytes data)
{
    // 7.4.8 Page information segment syntax
    if (data.size() != sizeof(PageInformationSegment))
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid page information segment size");
    return *(PageInformationSegment const*)data.data();
}

static ErrorOr<void> scan_for_page_size(JBIG2LoadingContext& context)
{
    // We only decode the first page at the moment.
    bool found_size = false;
    for (auto const& segment : context.segments) {
        if (segment.header.type != SegmentType::PageInformation || segment.header.page_association != 1)
            continue;
        auto page_information = TRY(decode_page_information_segment(segment.data));

        // FIXME: We're supposed to compute this from the striping information if it's not set.
        if (page_information.bitmap_height == 0xffff'ffff)
            return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot handle unknown page height yet");

        context.page.size = { page_information.bitmap_width, page_information.bitmap_height };
        found_size = true;
    }
    if (!found_size)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: No page information segment found for page 1");
    return {};
}

static ErrorOr<void> warn_about_multiple_pages(JBIG2LoadingContext& context)
{
    HashTable<u32> seen_pages;
    Vector<u32> pages;

    for (auto const& segment : context.segments) {
        if (segment.header.page_association == 0)
            continue;
        if (seen_pages.contains(segment.header.page_association))
            continue;
        seen_pages.set(segment.header.page_association);
        pages.append(segment.header.page_association);
    }

    // scan_for_page_size() already checked that there's a page 1.
    VERIFY(seen_pages.contains(1));
    if (pages.size() == 1)
        return {};

    StringBuilder builder;
    builder.appendff("JBIG2 file contains {} pages ({}", pages.size(), pages[0]);
    size_t i;
    for (i = 1; i < min(pages.size(), 10); ++i)
        builder.appendff(" {}", pages[i]);
    if (i != pages.size())
        builder.append(" ..."sv);
    builder.append("). We will only render page 1."sv);
    dbgln("JBIG2ImageDecoderPlugin: {}", TRY(builder.to_string()));

    return {};
}

struct AdaptiveTemplatePixel {
    i8 x { 0 };
    i8 y { 0 };
};

// 6.2.2 Input parameters
// Table 2 – Parameters for the generic region decoding procedure
struct GenericRegionDecodingInputParameters {
    bool is_modified_modified_read { false }; // "MMR" in spec.
    u32 region_width { 0 };                   // "GBW" in spec.
    u32 region_height { 0 };                  // "GBH" in spec.
    u8 gb_template { 0 };
    bool is_typical_prediction_used { false };          // "TPGDON" in spec.
    bool is_extended_reference_template_used { false }; // "EXTTEMPLATE" in spec.
    Optional<NonnullOwnPtr<BitBuffer>> skip_pattern;    // "USESKIP", "SKIP" in spec.

    Array<AdaptiveTemplatePixel, 12> adaptive_template_pixels; // "GBATX" / "GBATY" in spec.
    // FIXME: GBCOLS, GBCOMBOP, COLEXTFLAG

    // If is_modified_modified_read is false, generic_region_decoding_procedure() reads data off this decoder.
    JBIG2::ArithmeticDecoder* arithmetic_decoder { nullptr };
};

// 6.2 Generic region decoding procedure
static ErrorOr<NonnullOwnPtr<BitBuffer>> generic_region_decoding_procedure(GenericRegionDecodingInputParameters const& inputs, ReadonlyBytes data, Vector<JBIG2::ArithmeticDecoder::Context>& contexts)
{
    if (inputs.is_modified_modified_read) {
        dbgln_if(JBIG2_DEBUG, "JBIG2ImageDecoderPlugin: MMR image data");

        // 6.2.6 Decoding using MMR coding
        auto buffer = TRY(CCITT::decode_ccitt_group4(data, inputs.region_width, inputs.region_height));
        auto result = TRY(BitBuffer::create(inputs.region_width, inputs.region_height));
        size_t bytes_per_row = ceil_div(inputs.region_width, 8);
        if (buffer.size() != bytes_per_row * inputs.region_height)
            return Error::from_string_literal("JBIG2ImageDecoderPlugin: Decoded MMR data has wrong size");

        // FIXME: Could probably just copy the ByteBuffer directly into the BitBuffer's internal ByteBuffer instead.
        for (size_t y = 0; y < inputs.region_height; ++y) {
            for (size_t x = 0; x < inputs.region_width; ++x) {
                bool bit = buffer[y * bytes_per_row + x / 8] & (1 << (7 - x % 8));
                result->set_bit(x, y, bit);
            }
        }
        return result;
    }

    // 6.2.5 Decoding using a template and arithmetic coding
    if (inputs.gb_template != 0)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode GBTEMPLATE != 0 yet");

    if (inputs.adaptive_template_pixels[0].x != 3 || inputs.adaptive_template_pixels[0].y != -1
        || inputs.adaptive_template_pixels[1].x != -3 || inputs.adaptive_template_pixels[1].y != -1
        || inputs.adaptive_template_pixels[2].x != 2 || inputs.adaptive_template_pixels[2].y != -2
        || inputs.adaptive_template_pixels[3].x != -2 || inputs.adaptive_template_pixels[3].y != -2)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot handle custom adaptive pixels yet");

    if (inputs.is_extended_reference_template_used)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode EXTTEMPLATE yet");

    if (inputs.skip_pattern.has_value())
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode USESKIP yet");

    auto result = TRY(BitBuffer::create(inputs.region_width, inputs.region_height));

    auto get_pixel = [&inputs](NonnullOwnPtr<BitBuffer> const& buffer, int x, int y) -> bool {
        if (x < 0 || x >= (int)inputs.region_width || y < 0)
            return false;
        return buffer->get_bit(x, y);
    };

    // Figure 3(a) – Template when GBTEMPLATE = 0 and EXTTEMPLATE = 0,
    auto compute_context = [&get_pixel](NonnullOwnPtr<BitBuffer> const& buffer, int x, int y) -> u16 {
        u16 result = 0;
        for (int i = 0; i < 5; ++i)
            result = (result << 1) | (u16)get_pixel(buffer, x - 2 + i, y - 2);
        for (int i = 0; i < 7; ++i)
            result = (result << 1) | (u16)get_pixel(buffer, x - 3 + i, y - 1);
        for (int i = 0; i < 4; ++i)
            result = (result << 1) | (u16)get_pixel(buffer, x - 4 + i, y);
        return result;
    };

    // "The values of the pixels in this neighbourhood define a context. Each context has its own adaptive probability estimate
    //  used by the arithmetic coder (see Annex E)."
    // "* Decode the current pixel by invoking the arithmetic entropy decoding procedure, with CX set to the value formed by
    //    concatenating the label "GB" and the 10-16 pixel values gathered in CONTEXT."
    // Implementor's note: What this is supposed to mean is that we have a bunch of independent contexts, and we pick the
    // context for the current pixel based on pixel values in the neighborhood. The "GB" part just means this context is
    // independent from other contexts in the spec. They are passed in to this function.

    // Figure 8 – Reused context for coding the SLTP value when GBTEMPLATE is 0
    constexpr u16 sltp_context_for_template_0 = 0b10011'0110010'0101;

    // 6.2.5.7 Decoding the bitmap
    JBIG2::ArithmeticDecoder& decoder = *inputs.arithmetic_decoder;
    bool ltp = false; // "LTP" in spec. "Line (uses) Typical Prediction" maybe?
    for (size_t y = 0; y < inputs.region_height; ++y) {
        if (inputs.is_typical_prediction_used) {
            // "SLTP" in spec. "Swap LTP" or "Switch LTP" maybe?
            bool sltp = decoder.get_next_bit(contexts[sltp_context_for_template_0]);
            ltp = ltp ^ sltp;
            if (ltp) {
                for (size_t x = 0; x < inputs.region_width; ++x)
                    result->set_bit(x, y, get_pixel(result, (int)x, (int)y - 1));
                continue;
            }
        }

        for (size_t x = 0; x < inputs.region_width; ++x) {
            u16 context = compute_context(result, x, y);
            bool bit = decoder.get_next_bit(contexts[context]);
            result->set_bit(x, y, bit);
        }
    }

    return result;
}

// 6.4.2 Input parameters
// Table 9 – Parameters for the text region decoding procedure
struct TextRegionDecodingInputParameters {
    bool uses_huffman_encoding { false };     // "SBHUFF" in spec.
    bool uses_refinement_coding { false };    // "SBREFINE" in spec.
    u32 region_width { 0 };                   // "SBW" in spec.
    u32 region_height { 0 };                  // "SBH" in spec.
    u32 number_of_instances { 0 };            // "SBNUMINSTANCES" in spec.
    u32 size_of_symbol_instance_strips { 0 }; // "SBSTRIPS" in spec.
    // "SBNUMSYMS" is `symbols.size()` below.

    // FIXME: SBSYMCODES
    u32 id_symbol_code_length { 0 };       // "SBSYMCODELEN" in spec.
    Vector<NonnullRefPtr<Symbol>> symbols; // "SBNUMSYMS" / "SBSYMS" in spec.
    u8 default_pixel { 0 };                // "SBDEFPIXEL" in spec.

    CombinationOperator operator_ { CombinationOperator::Or }; // "SBCOMBOP" in spec.

    bool is_transposed { false }; // "TRANSPOSED" in spec.

    enum class Corner {
        BottomLeft = 0,
        TopLeft = 1,
        BottomRight = 2,
        TopRight = 3,
    };
    Corner reference_corner { Corner::TopLeft }; // "REFCORNER" in spec.

    i8 delta_s_offset { 0 }; // "SBDSOFFSET" in spec.
    // FIXME: SBHUFFFS, SBHUFFFDS, SBHUFFDT, SBHUFFRDW, SBHUFFRDH, SBHUFFRDX, SBHUFFRDY, SBHUFFRSIZE

    u8 refinement_template { 0 };                                        // "SBRTEMPLATE" in spec.
    Array<AdaptiveTemplatePixel, 2> refinement_adaptive_template_pixels; // "SBRATX" / "SBRATY" in spec.
    // FIXME: COLEXTFLAG, SBCOLS
};

// 6.4 Text Region Decoding Procedure
static ErrorOr<NonnullOwnPtr<BitBuffer>> text_region_decoding_procedure(TextRegionDecodingInputParameters const& inputs, ReadonlyBytes data)
{
    if (inputs.uses_huffman_encoding)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode huffman text regions yet");

    if (inputs.is_transposed)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode transposed text regions yet");

    auto decoder = TRY(JBIG2::ArithmeticDecoder::initialize(data));

    // 6.4.6 Strip delta T
    // "If SBHUFF is 1, decode a value using the Huffman table specified by SBHUFFDT and multiply the resulting value by SBSTRIPS.
    //  If SBHUFF is 0, decode a value using the IADT integer arithmetic decoding procedure (see Annex A) and multiply the resulting value by SBSTRIPS."
    // FIXME: Implement support for SBHUFF = 1.
    JBIG2::ArithmeticIntegerDecoder delta_t_integer_decoder(decoder);
    auto read_delta_t = [&]() -> i32 {
        return delta_t_integer_decoder.decode().value() * inputs.size_of_symbol_instance_strips;
    };

    // 6.4.7 First symbol instance S coordinate
    // "If SBHUFF is 1, decode a value using the Huffman table specified by SBHUFFFS.
    //  If SBHUFF is 0, decode a value using the IAFS integer arithmetic decoding procedure (see Annex A)."
    // FIXME: Implement support for SBHUFF = 1.
    JBIG2::ArithmeticIntegerDecoder first_s_integer_decoder(decoder);
    auto read_first_s = [&]() -> i32 {
        return first_s_integer_decoder.decode().value();
    };

    // 6.4.8 Subsequent symbol instance S coordinate
    // "If SBHUFF is 1, decode a value using the Huffman table specified by SBHUFFDS.
    //  If SBHUFF is 0, decode a value using the IADS integer arithmetic decoding procedure (see Annex A).
    //  In either case it is possible that the result of this decoding is the out-of-band value OOB.""
    // FIXME: Implement support for SBHUFF = 1.
    JBIG2::ArithmeticIntegerDecoder subsequent_s_integer_decoder(decoder);
    auto read_subsequent_s = [&]() -> Optional<i32> {
        return subsequent_s_integer_decoder.decode();
    };

    // 6.4.9 Symbol instance T coordinate
    // "If SBSTRIPS == 1, then the value decoded is always zero. Otherwise:
    //  • If SBHUFF is 1, decode a value by reading ceil(log2(SBSTRIPS)) bits directly from the bitstream.
    //  • If SBHUFF is 0, decode a value using the IAIT integer arithmetic decoding procedure (see Annex A)."
    // FIXME: Implement support for SBHUFF = 1.
    JBIG2::ArithmeticIntegerDecoder instance_t_integer_decoder(decoder);
    auto read_instance_t = [&]() -> i32 {
        if (inputs.size_of_symbol_instance_strips == 1)
            return 0;
        return instance_t_integer_decoder.decode().value();
    };

    // 6.4.10 Symbol instance symbol ID
    // "If SBHUFF is 1, decode a value by reading one bit at a time until the resulting bit string is equal to one of the entries in
    //  SBSYMCODES. The resulting value, which is IDI, is the index of the entry in SBSYMCODES that is read.
    //  If SBHUFF is 0, decode a value using the IAID integer arithmetic decoding procedure (see Annex A). Set IDI to the
    //  resulting value.""
    // FIXME: Implement support for SBHUFF = 1.
    Vector<JBIG2::ArithmeticDecoder::Context> id_contexts;
    id_contexts.resize(1 << (inputs.id_symbol_code_length + 1));
    auto read_id = [&]() -> u32 {
        // A.3 The IAID decoding procedure
        u32 prev = 1;
        for (u8 i = 0; i < inputs.id_symbol_code_length; ++i) {
            bool bit = decoder.get_next_bit(id_contexts[prev]);
            prev = (prev << 1) | bit;
        }
        prev = prev - (1 << inputs.id_symbol_code_length);
        return prev;
    };

    // 6.4.11 Symbol instance bitmap
    JBIG2::ArithmeticIntegerDecoder has_refinement_image_decoder(decoder);
    auto read_bitmap = [&](u32 id) -> ErrorOr<BitBuffer const*> {
        bool has_refinement_image = false; // "R_I" in spec.
        if (inputs.uses_refinement_coding) {
            // "• If SBHUFF is 1, then read one bit and set RI to the value of that bit.
            //  • If SBHUFF is 0, then decode one bit using the IARI integer arithmetic decoding procedure and set RI to the value of that bit."
            // FIXME: Implement support for SBHUFF = 1.
            has_refinement_image = has_refinement_image_decoder.decode().value();
        }

        if (!has_refinement_image) {
            if (id >= inputs.symbols.size())
                return Error::from_string_literal("JBIG2ImageDecoderPlugin: Symbol ID out of range");
            auto const& symbol = inputs.symbols[id]->bitmap();
            return &symbol;
        }

        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode refined text regions yet");
    };

    // 6.4.5 Decoding the text region

    // "1) Fill a bitmap SBREG, of the size given by SBW and SBH, with the SBDEFPIXEL value."
    auto result = TRY(BitBuffer::create(inputs.region_width, inputs.region_height));
    if (inputs.default_pixel != 0)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot handle SBDEFPIXEL not equal to 0 yet");
    result->fill(inputs.default_pixel != 0);

    // "2) Decode the initial STRIPT value as described in 6.4.6. Negate the decoded value and assign this negated value to the variable STRIPT.
    //     Assign the value 0 to FIRSTS. Assign the value 0 to NINSTANCES."
    i32 strip_t = -read_delta_t();
    i32 first_s = 0;
    u32 n_instances = 0;

    // "3) If COLEXTFLAG is 1, decode the colour section as described in 6.4.12."
    // FIXME: Implement support for colors one day.

    // "4) Decode each strip as follows:
    //      a) If NINSTANCES is equal to SBNUMINSTANCES then there are no more strips to decode,
    //         and the process of decoding the text region is complete; proceed to step 4)."
    // Implementor's note. The spec means "proceed to step 5)" at the end of 4a).
    while (n_instances < inputs.number_of_instances) {
        // "b) Decode the strip's delta T value as described in 6.4.6. Let DT be the decoded value. Set:
        //         STRIPT = STRIPT + DT"
        i32 delta_t = read_delta_t();
        strip_t += delta_t;

        i32 cur_s;
        bool is_first_symbol = true;
        while (true) {
            // "c) Decode each symbol instance in the strip as follows:
            //      i) If the current symbol instance is the first symbol instance in the strip, then decode the first
            //         symbol instance's S coordinate as described in 6.4.7. Let DFS be the decoded value. Set:
            //              FIRSTS = FIRSTS + DFS
            //              CURS = FIRSTS
            //      ii) Otherwise, if the current symbol instance is not the first symbol instance in the strip, decode
            //          the symbol instance's S coordinate as described in 6.4.8. If the result of this decoding is OOB
            //          then the last symbol instance of the strip has been decoded; proceed to step 3 d). Otherwise, let
            //          IDS be the decoded value. Set:
            //              CURS = CURS + IDS + SBDSOFFSET"
            // Implementor's note: The spec means "proceed to step 4 d)" in 4c ii).
            if (is_first_symbol) {
                i32 delta_first_s = read_first_s();
                first_s += delta_first_s;
                cur_s = first_s;
                is_first_symbol = false;
            } else {
                auto subsequent_s = read_subsequent_s();
                if (!subsequent_s.has_value())
                    break;
                i32 instance_delta_s = subsequent_s.value();
                cur_s += instance_delta_s + inputs.delta_s_offset;
            }

            //     "iii) Decode the symbol instance's T coordinate as described in 6.4.9. Let CURT be the decoded value. Set:
            //              TI = STRIPT + CURT"
            i32 cur_t = read_instance_t();
            i32 t_instance = strip_t + cur_t;

            //     "iv) Decode the symbol instance's symbol ID as described in 6.4.10. Let IDI be the decoded value."
            u32 id = read_id();

            //     "v) Determine the symbol instance's bitmap IBI as described in 6.4.11. The width and height of this
            //         bitmap shall be denoted as WI and HI respectively."
            auto const& symbol = *TRY(read_bitmap(id));

            //     "vi) Update CURS as follows:
            //      • If TRANSPOSED is 0, and REFCORNER is TOPRIGHT or BOTTOMRIGHT, set:
            //              CURS = CURS + WI – 1
            //      • If TRANSPOSED is 1, and REFCORNER is BOTTOMLEFT or BOTTOMRIGHT, set:
            //              CURS = CURS + HI –1
            //      • Otherwise, do not change CURS in this step."
            using enum TextRegionDecodingInputParameters::Corner;
            if (!inputs.is_transposed && (inputs.reference_corner == TopRight || inputs.reference_corner == BottomRight))
                cur_s += symbol.width() - 1;
            if (inputs.is_transposed && (inputs.reference_corner == BottomLeft || inputs.reference_corner == BottomRight))
                cur_s += symbol.height() - 1;

            //     "vii) Set:
            //              SI = CURS"
            auto s_instance = cur_s;

            //     "viii) Determine the location of the symbol instance bitmap with respect to SBREG as follows:
            //          • If TRANSPOSED is 0, then:
            //              – If REFCORNER is TOPLEFT then the top left pixel of the symbol instance bitmap
            //                IBI shall be placed at SBREG[SI, TI].
            //              – If REFCORNER is TOPRIGHT then the top right pixel of the symbol instance
            //                bitmap IBI shall be placed at SBREG[SI, TI].
            //              – If REFCORNER is BOTTOMLEFT then the bottom left pixel of the symbol
            //                instance bitmap IBI shall be placed at SBREG[SI, TI].
            //              – If REFCORNER is BOTTOMRIGHT then the bottom right pixel of the symbol
            //                instance bitmap IBI shall be placed at SBREG[SI, TI].
            //          • If TRANSPOSED is 1, then:
            //              – If REFCORNER is TOPLEFT then the top left pixel of the symbol instance bitmap
            //                IBI shall be placed at SBREG[TI, SI].
            //              – If REFCORNER is TOPRIGHT then the top right pixel of the symbol instance
            //                bitmap IBI shall be placed at SBREG[TI, SI].
            //              – If REFCORNER is BOTTOMLEFT then the bottom left pixel of the symbol
            //                instance bitmap IBI shall be placed at SBREG[TI, SI].
            //              – If REFCORNER is BOTTOMRIGHT then the bottom right pixel of the symbol
            //                instance bitmap IBI shall be placed at SBREG[TI, SI].
            //          If any part of IBI, when placed at this location, lies outside the bounds of SBREG, then ignore
            //          this part of IBI in step 3 c) ix)."
            // Implementor's note: The spec means "ignore this part of IBI in step 3 c) x)" in 3c viii)'s last sentence.
            // FIXME: Support all reference corners and transpose values.
            if (!inputs.is_transposed) {
                switch (inputs.reference_corner) {
                case TopLeft:
                    break;
                case TopRight:
                    s_instance -= symbol.width() - 1;
                    break;
                case BottomLeft:
                    t_instance -= symbol.height() - 1;
                    break;
                case BottomRight:
                    s_instance -= symbol.width() - 1;
                    t_instance -= symbol.height() - 1;
                    break;
                }
            } else {
                TODO();
            }

            //     "ix) If COLEXTFLAG is 1, set the colour specified by SBCOLS[SBFGCOLID[NINSTANCES]]
            //          to the foreground colour of the symbol instance bitmap IBI."
            // FIXME: Implement support for colors one day.

            //     "x) Draw IBI into SBREG. Combine each pixel of IBI with the current value of the corresponding
            //         pixel in SBREG, using the combination operator specified by SBCOMBOP. Write the results
            //         of each combination into that pixel in SBREG."
            composite_bitbuffer(*result, symbol, { s_instance, t_instance }, inputs.operator_);

            //     "xi) Update CURS as follows:
            //          • If TRANSPOSED is 0, and REFCORNER is TOPLEFT or BOTTOMLEFT, set:
            //              CURS = CURS + WI –1
            //          • If TRANSPOSED is 1, and REFCORNER is TOPLEFT or TOPRIGHT, set:
            //              CURS = CURS + HI –1
            //          • Otherwise, do not change CURS in this step."
            if (!inputs.is_transposed && (inputs.reference_corner == TopLeft || inputs.reference_corner == BottomLeft))
                cur_s += symbol.width() - 1;
            if (inputs.is_transposed && (inputs.reference_corner == TopLeft || inputs.reference_corner == TopRight))
                cur_s += symbol.height() - 1;

            //      "xii) Set:
            //              NINSTANCES = NINSTANCES + 1"
            ++n_instances;
        }
        //  "d) When the strip has been completely decoded, decode the next strip."
        // (Done in the next loop iteration.)
    }

    //  "5) After all the strips have been decoded, the current contents of SBREG are the results that shall be
    //      obtained by every decoder, whether it performs this exact sequence of steps or not."
    return result;
}

// 6.5.2 Input parameters
// Table 13 – Parameters for the symbol dictionary decoding procedure
struct SymbolDictionaryDecodingInputParameters {

    bool uses_huffman_encoding { false };               // "SDHUFF" in spec.
    bool uses_refinement_or_aggregate_coding { false }; // "SDREFAGG" in spec.

    Vector<NonnullRefPtr<Symbol>> input_symbols; // "SDNUMINSYMS", "SDINSYMS" in spec.

    u32 number_of_new_symbols { 0 };      // "SDNUMNEWSYMS" in spec.
    u32 number_of_exported_symbols { 0 }; // "SDNUMEXSYMS" in spec.

    // FIXME: SDHUFFDH, SDHUFFDW, SDHUFFBMSIZE, SDHUFFAGGINST

    u8 symbol_template { 0 };                                 // "SDTEMPLATE" in spec.
    Array<AdaptiveTemplatePixel, 4> adaptive_template_pixels; // "SDATX" / "SDATY" in spec.

    u8 refinement_template { 0 };                                        // "SDRTEMPLATE" in spec;
    Array<AdaptiveTemplatePixel, 2> refinement_adaptive_template_pixels; // "SDRATX" / "SDRATY" in spec.
};

// 6.5 Symbol Dictionary Decoding Procedure
static ErrorOr<Vector<NonnullRefPtr<Symbol>>> symbol_dictionary_decoding_procedure(SymbolDictionaryDecodingInputParameters const& inputs, ReadonlyBytes data)
{
    if (inputs.uses_huffman_encoding)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode huffman symbol dictionaries yet");

    if (inputs.uses_refinement_or_aggregate_coding)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode SDREFAGG symbol dictionaries yet");

    auto decoder = TRY(JBIG2::ArithmeticDecoder::initialize(data));
    Vector<JBIG2::ArithmeticDecoder::Context> contexts;
    u8 template_size = inputs.symbol_template == 0 ? 16 : (inputs.symbol_template == 1 ? 13 : 10);
    contexts.resize(1 << template_size);

    // 6.5.6 Height class delta height
    // "If SDHUFF is 1, decode a value using the Huffman table specified by SDHUFFDH.
    //  If SDHUFF is 0, decode a value using the IADH integer arithmetic decoding procedure (see Annex A)."
    // FIXME: Implement support for SDHUFF = 1.
    JBIG2::ArithmeticIntegerDecoder delta_height_integer_decoder(decoder);
    auto read_delta_height = [&]() -> i32 {
        // No OOB values for delta height.
        return delta_height_integer_decoder.decode().value();
    };

    // 6.5.7 Delta width
    // "If SDHUFF is 1, decode a value using the Huffman table specified by SDHUFFDW.
    //  If SDHUFF is 0, decode a value using the IADW integer arithmetic decoding procedure (see Annex A).
    //  In either case it is possible that the result of this decoding is the out-of-band value OOB."
    // FIXME: Implement support for SDHUFF = 1.
    JBIG2::ArithmeticIntegerDecoder delta_width_integer_decoder(decoder);
    auto read_delta_width = [&]() -> Optional<i32> {
        return delta_width_integer_decoder.decode();
    };

    // 6.5.8 Symbol bitmap
    // "This field is only present if SDHUFF = 0 or SDREFAGG = 1. This field takes one of two forms; SDREFAGG
    //  determines which form is used."
    // FIXME: Add support for SDEFRAG = 1.
    // 6.5.8.1 Direct-coded symbol bitmap
    // "If SDREFAGG is 0, then decode the symbol's bitmap using a generic region decoding procedure as described in 6.2.
    //  Set the parameters to this decoding procedure as shown in Table 16."
    auto read_bitmap = [&](u32 width, u32 height) {
        GenericRegionDecodingInputParameters generic_inputs;
        generic_inputs.is_modified_modified_read = false;
        generic_inputs.region_width = width;
        generic_inputs.region_height = height;
        generic_inputs.gb_template = inputs.symbol_template;
        for (int i = 0; i < 4; ++i)
            generic_inputs.adaptive_template_pixels[i] = inputs.adaptive_template_pixels[i];
        generic_inputs.arithmetic_decoder = &decoder;
        return generic_region_decoding_procedure(generic_inputs, {}, contexts);
    };

    // 6.5.5 Decoding the symbol dictionary
    // "1) Create an array SDNEWSYMS of bitmaps, having SDNUMNEWSYMS entries."
    Vector<NonnullRefPtr<Symbol>> new_symbols;

    // "2) If SDHUFF is 1 and SDREFAGG is 0, create an array SDNEWSYMWIDTHS of integers, having SDNUMNEWSYMS entries."
    // FIXME: Implement support for SDHUFF = 1.

    // "3) Set:
    //      HCHEIGHT = 0
    //      NSYMSDECODED = 0"
    u32 height_class_height = 0;
    u32 number_of_symbols_decoded = 0;

    // "4) Decode each height class as follows:
    //      a) If NSYMSDECODED == SDNUMNEWSYMS then all the symbols in the dictionary have been decoded; proceed to step 5)."
    while (number_of_symbols_decoded < inputs.number_of_new_symbols) {
        // "b) Decode the height class delta height as described in 6.5.6. Let HCDH be the decoded value. Set:
        //      HCHEIGHT = HCEIGHT + HCDH
        //      SYMWIDTH = 0
        //      TOTWIDTH = 0
        //      HCFIRSTSYM = NSYMSDECODED"
        i32 delta_height = read_delta_height();
        height_class_height += delta_height;
        u32 symbol_width = 0;
        u32 total_width = 0;
        u32 height_class_first_symbol = number_of_symbols_decoded;
        // "c) Decode each symbol within the height class as follows:"
        while (true) {
            // "i) Decode the delta width for the symbol as described in 6.5.7."
            auto opt_delta_width = read_delta_width();
            // "   If the result of this decoding is OOB then all the symbols in this height class have been decoded; proceed to step 4 d)."
            if (!opt_delta_width.has_value())
                break;

            VERIFY(number_of_symbols_decoded < inputs.number_of_new_symbols);
            // "   Otherwise let DW be the decoded value and set:"
            //         SYMWIDTH = SYMWIDTH + DW
            //         TOTWIDTH = TOTWIDTH + SYMWIDTH"
            i32 delta_width = opt_delta_width.value();
            symbol_width += delta_width;
            total_width += symbol_width;

            // "ii) If SDHUFF is 0 or SDREFAGG is 1, then decode the symbol's bitmap as described in 6.5.8.
            //      Let BS be the decoded bitmap (this bitmap has width SYMWIDTH and height HCHEIGHT). Set:
            //          SDNEWSYMS[NSYMSDECODED] = BS"
            // FIXME: Implement support for SDHUFF = 1.
            // FIXME: Doing this eagerly is pretty wasteful. Decode on demand instead?
            auto bitmap = TRY(read_bitmap(symbol_width, height_class_height));
            new_symbols.append(Symbol::create(move(bitmap)));

            // "iii) If SDHUFF is 1 and SDREFAGG is 0, then set:
            //      SDNEWSYMWIDTHS[NSYMSDECODED] = SYMWIDTH"
            // FIXME: Implement support for SDHUFF = 1.
            (void)total_width;
            (void)height_class_first_symbol;

            // "iv) Set:
            //      NSYMSDECODED = NSYMSDECODED + 1"
            number_of_symbols_decoded++;
        }
        //  d) If SDHUFF is 1 and SDREFAGG is 0, [...long text elided...]
        // FIXME: Implement support for SDHUFF = 1.
    }

    //  5) Determine which symbol bitmaps are exported from this symbol dictionary, as described in 6.5.10. These
    //     bitmaps can be drawn from the symbols that are used as input to the symbol dictionary decoding
    //     procedure as well as the new symbols produced by the decoding procedure."
    JBIG2::ArithmeticIntegerDecoder export_integer_decoder(decoder);

    // 6.5.10 Exported symbols
    Vector<bool> export_flags;
    export_flags.resize(inputs.input_symbols.size() + inputs.number_of_new_symbols);

    // "1) Set:
    //      EXINDEX = 0
    //      CUREXFLAG = 0"
    u32 exported_index = 0;
    bool current_export_flag = false;

    do {
        // "2) Decode a value using Table B.1 if SDHUFF is 1, or the IAEX integer arithmetic decoding procedure if
        //  SDHUFF is 0. Let EXRUNLENGTH be the decoded value."
        // FIXME: Implement support for SDHUFF = 1.
        i32 export_run_length = export_integer_decoder.decode().value(); // No OOB value.

        // "3) Set EXFLAGS[EXINDEX] through EXFLAGS[EXINDEX + EXRUNLENGTH – 1] to CUREXFLAG.
        //  If EXRUNLENGTH = 0, then this step does not change any values."
        for (int i = 0; i < export_run_length; ++i)
            export_flags[exported_index + i] = current_export_flag;

        // "4) Set:
        //      EXINDEX = EXINDEX + EXRUNLENGTH
        //      CUREXFLAG = NOT(CUREXFLAG)"
        exported_index += export_run_length;
        current_export_flag = !current_export_flag;

        //  5) Repeat steps 2) through 4) until EXINDEX == SDNUMINSYMS + SDNUMNEWSYMS.
    } while (exported_index < inputs.input_symbols.size() + inputs.number_of_new_symbols);

    // "6) The array EXFLAGS now contains 1 for each symbol that is exported from the dictionary, and 0 for each
    //  symbol that is not exported."
    Vector<NonnullRefPtr<Symbol>> exported_symbols;

    // "7) Set:
    //      I = 0
    //      J = 0
    //  8) For each value of I from 0 to SDNUMINSYMS + SDNUMNEWSYMS – 1,"
    for (size_t i = 0; i < inputs.input_symbols.size() + inputs.number_of_new_symbols; ++i) {
        // "if EXFLAGS[I] == 1 then perform the following steps:"
        if (!export_flags[i])
            continue;
        //  "a) If I < SDNUMINSYMS then set:
        //       SDEXSYMS[J] = SDINSYMS[I]
        //       J = J + 1"
        if (i < inputs.input_symbols.size())
            exported_symbols.append(inputs.input_symbols[i]);

        //  "b) If I >= SDNUMINSYMS then set:
        //       SDEXSYMS[J] = SDNEWSYMS[I – SDNUMINSYMS]
        //       J = J + 1"
        if (i >= inputs.input_symbols.size())
            exported_symbols.append(move(new_symbols[i - inputs.input_symbols.size()]));
    }

    if (exported_symbols.size() != inputs.number_of_exported_symbols)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Unexpected number of exported symbols");

    return exported_symbols;
}

static ErrorOr<void> decode_symbol_dictionary(JBIG2LoadingContext&, SegmentData& segment)
{
    // 7.4.2 Symbol dictionary segment syntax

    // 7.4.2.1 Symbol dictionary segment data header
    FixedMemoryStream stream(segment.data);

    // 7.4.2.1.1 Symbol dictionary flags
    u16 flags = TRY(stream.read_value<BigEndian<u16>>());
    bool uses_huffman_encoding = (flags & 1) != 0;               // "SDHUFF" in spec.
    bool uses_refinement_or_aggregate_coding = (flags & 2) != 0; // "SDREFAGG" in spec.

    u8 huffman_table_selection_for_height_differences = (flags >> 2) & 0b11; // "SDHUFFDH" in spec.
    if (huffman_table_selection_for_height_differences == 2)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid huffman_table_selection_for_height_differences");
    if (!uses_huffman_encoding && huffman_table_selection_for_height_differences != 0)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid huffman_table_selection_for_height_differences");

    u8 huffman_table_selection_for_width_differences = (flags >> 4) & 0b11; // "SDHUFFDW" in spec.
    if (huffman_table_selection_for_width_differences == 2)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid huffman_table_selection_for_width_differences");
    if (!uses_huffman_encoding && huffman_table_selection_for_width_differences != 0)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid huffman_table_selection_for_width_differences");

    bool uses_user_supplied_size_table = (flags >> 6) & 1; // "SDHUFFBMSIZE" in spec.
    if (!uses_huffman_encoding && uses_user_supplied_size_table)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid uses_user_supplied_size_table");

    bool uses_user_supplied_aggregate_table = (flags >> 7) & 1; // "SDHUFFAGGINST" in spec.
    if (!uses_huffman_encoding && uses_user_supplied_aggregate_table)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid uses_user_supplied_aggregate_table");

    bool bitmap_coding_context_used = (flags >> 8) & 1;
    if (uses_huffman_encoding && !uses_refinement_or_aggregate_coding && bitmap_coding_context_used)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid bitmap_coding_context_used");

    bool bitmap_coding_context_retained = (flags >> 9) & 1;
    if (uses_huffman_encoding && !uses_refinement_or_aggregate_coding && bitmap_coding_context_retained)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid bitmap_coding_context_retained");

    u8 template_used = (flags >> 10) & 0b11; // "SDTEMPLATE" in spec.
    if (uses_huffman_encoding && template_used != 0)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid template_used");

    u8 refinement_template_used = (flags >> 12) & 0b11; // "SDREFTEMPLATE" in spec.
    if (!uses_refinement_or_aggregate_coding && refinement_template_used != 0)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid refinement_template_used");

    if (flags & 0b1110'0000'0000'0000)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid symbol dictionary flags");

    // 7.4.2.1.2 Symbol dictionary AT flags
    Array<AdaptiveTemplatePixel, 4> adaptive_template {};
    if (!uses_huffman_encoding) {
        int number_of_adaptive_template_pixels = template_used == 0 ? 4 : 1;
        for (int i = 0; i < number_of_adaptive_template_pixels; ++i) {
            adaptive_template[i].x = TRY(stream.read_value<i8>());
            adaptive_template[i].y = TRY(stream.read_value<i8>());
        }
    }

    // 7.4.2.1.3 Symbol dictionary refinement AT flags
    Array<AdaptiveTemplatePixel, 2> adaptive_refinement_template {};
    if (uses_refinement_or_aggregate_coding && refinement_template_used == 0) {
        for (size_t i = 0; i < adaptive_refinement_template.size(); ++i) {
            adaptive_refinement_template[i].x = TRY(stream.read_value<i8>());
            adaptive_refinement_template[i].y = TRY(stream.read_value<i8>());
        }
    }

    // 7.4.2.1.4 Number of exported symbols (SDNUMEXSYMS)
    u32 number_of_exported_symbols = TRY(stream.read_value<BigEndian<u32>>());

    // 7.4.2.1.5 Number of new symbols (SDNUMNEWSYMS)
    u32 number_of_new_symbols = TRY(stream.read_value<BigEndian<u32>>());

    dbgln_if(JBIG2_DEBUG, "Symbol dictionary: uses_huffman_encoding={}", uses_huffman_encoding);
    dbgln_if(JBIG2_DEBUG, "Symbol dictionary: uses_refinement_or_aggregate_coding={}", uses_refinement_or_aggregate_coding);
    dbgln_if(JBIG2_DEBUG, "Symbol dictionary: huffman_table_selection_for_height_differences={}", huffman_table_selection_for_height_differences);
    dbgln_if(JBIG2_DEBUG, "Symbol dictionary: huffman_table_selection_for_width_differences={}", huffman_table_selection_for_width_differences);
    dbgln_if(JBIG2_DEBUG, "Symbol dictionary: uses_user_supplied_size_table={}", uses_user_supplied_size_table);
    dbgln_if(JBIG2_DEBUG, "Symbol dictionary: uses_user_supplied_aggregate_table={}", uses_user_supplied_aggregate_table);
    dbgln_if(JBIG2_DEBUG, "Symbol dictionary: bitmap_coding_context_used={}", bitmap_coding_context_used);
    dbgln_if(JBIG2_DEBUG, "Symbol dictionary: bitmap_coding_context_retained={}", bitmap_coding_context_retained);
    dbgln_if(JBIG2_DEBUG, "Symbol dictionary: template_used={}", template_used);
    dbgln_if(JBIG2_DEBUG, "Symbol dictionary: refinement_template_used={}", refinement_template_used);
    dbgln_if(JBIG2_DEBUG, "Symbol dictionary: number_of_exported_symbols={}", number_of_exported_symbols);
    dbgln_if(JBIG2_DEBUG, "Symbol dictionary: number_of_new_symbols={}", number_of_new_symbols);

    // 7.4.2.1.6 Symbol dictionary segment Huffman table selection
    // FIXME

    // 7.4.2.2 Decoding a symbol dictionary segment
    // "1) Interpret its header, as described in 7.4.2.1."
    // Done!

    // "2) Decode (or retrieve the results of decoding) any referred-to symbol dictionary and tables segments."
    if (segment.header.referred_to_segment_numbers.size() != 0)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode referred-to symbol dictionary segments yet");

    // "3) If the "bitmap coding context used" bit in the header was 1, ..."
    if (bitmap_coding_context_used)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode bitmap coding context segment yet");

    // "4) If the "bitmap coding context used" bit in the header was 0, then, as described in E.3.7,
    //     reset all the arithmetic coding statistics for the generic region and generic refinement region decoding procedures to zero."
    // Nothing to do.

    // "5) Reset the arithmetic coding statistics for all the contexts of all the arithmetic integer coders to zero."
    // FIXME

    // "6) Invoke the symbol dictionary decoding procedure described in 6.5, with the parameters to the symbol dictionary decoding procedure set as shown in Table 31."
    SymbolDictionaryDecodingInputParameters inputs;
    inputs.uses_huffman_encoding = uses_huffman_encoding;
    inputs.uses_refinement_or_aggregate_coding = uses_refinement_or_aggregate_coding;
    inputs.input_symbols = {};
    inputs.number_of_new_symbols = number_of_new_symbols;
    inputs.number_of_exported_symbols = number_of_exported_symbols;
    // FIXME: SDHUFFDH, SDHUFFDW, SDHUFFBMSIZE, SDHUFFAGGINST
    inputs.symbol_template = template_used;
    inputs.adaptive_template_pixels = adaptive_template;
    inputs.refinement_template = refinement_template_used;
    inputs.refinement_adaptive_template_pixels = adaptive_refinement_template;
    auto result = TRY(symbol_dictionary_decoding_procedure(inputs, segment.data.slice(TRY(stream.tell()))));

    // "7) If the "bitmap coding context retained" bit in the header was 1, then, as described in E.3.8, preserve the current contents
    //     of the arithmetic coding statistics for the generic region and generic refinement region decoding procedures."
    if (bitmap_coding_context_retained)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot retain bitmap coding context yet");

    segment.symbols = move(result);

    return {};
}

static ErrorOr<void> decode_intermediate_text_region(JBIG2LoadingContext&, SegmentData const&)
{
    return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode intermediate text region yet");
}

static ErrorOr<void> decode_immediate_text_region(JBIG2LoadingContext& context, SegmentData const& segment)
{
    // 7.4.3 Text region segment syntax
    auto data = segment.data;
    auto information_field = TRY(decode_region_segment_information_field(data));
    data = data.slice(sizeof(information_field));

    dbgln_if(JBIG2_DEBUG, "Text region: width={}, height={}, x={}, y={}, flags={:#x}", information_field.width, information_field.height, information_field.x_location, information_field.y_location, information_field.flags);

    FixedMemoryStream stream(data);

    // 7.4.3.1.1 Text region segment flags
    u16 text_region_segment_flags = TRY(stream.read_value<BigEndian<u16>>());
    bool uses_huffman_encoding = (text_region_segment_flags & 1) != 0;  // "SBHUFF" in spec.
    bool uses_refinement_coding = (text_region_segment_flags >> 1) & 1; // "SBREFINE" in spec.
    u8 log_strip_size = (text_region_segment_flags >> 2) & 3;           // "LOGSBSTRIPS" in spec.
    u8 strip_size = 1u << log_strip_size;
    u8 reference_corner = (text_region_segment_flags >> 4) & 3;     // "REFCORNER"
    bool is_transposed = (text_region_segment_flags >> 6) & 1;      // "TRANSPOSED" in spec.
    u8 combination_operator = (text_region_segment_flags >> 7) & 3; // "SBCOMBOP" in spec.
    if (combination_operator > 4)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid text region combination operator");

    u8 default_pixel_value = (text_region_segment_flags >> 9) & 1; // "SBDEFPIXEL" in spec.

    u8 delta_s_offset_value = (text_region_segment_flags >> 10) & 0x1f; // "SBDSOFFSET" in spec.
    i8 delta_s_offset = delta_s_offset_value;
    if (delta_s_offset_value & 0x10) {
        // This is converting a 5-bit two's complement number ot i8.
        // FIXME: There's probably a simpler way to do this? Probably just sign-extend by or-ing in the top 3 bits?
        delta_s_offset_value = (~delta_s_offset_value + 1) & 0x1f;
        delta_s_offset = -delta_s_offset_value;
    }

    u8 refinement_template = (text_region_segment_flags >> 15) != 0; // "SBRTEMPLATE" in spec.
    if (!uses_refinement_coding && refinement_template != 0)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid refinement_template");

    // 7.4.3.1.2 Text region segment Huffman flags
    // "This field is only present if SBHUFF is 1."
    // FIXME: Support this eventually.
    if (uses_huffman_encoding)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode huffman text regions yet");

    // 7.4.3.1.3 Text region refinement AT flags
    // "This field is only present if SBREFINE is 1 and SBRTEMPLATE is 0."
    Array<AdaptiveTemplatePixel, 2> adaptive_refinement_template {};
    if (uses_refinement_coding && refinement_template == 0) {
        for (size_t i = 0; i < adaptive_refinement_template.size(); ++i) {
            adaptive_refinement_template[i].x = TRY(stream.read_value<i8>());
            adaptive_refinement_template[i].y = TRY(stream.read_value<i8>());
        }
    }

    // 7.4.3.1.4 Number of symbol instances (SBNUMINSTANCES)
    u32 number_of_symbol_instances = TRY(stream.read_value<BigEndian<u32>>());

    // 7.4.3.1.5 Text region segment symbol ID Huffman decoding table
    // "It is only present if SBHUFF is 1."
    // FIXME: Support this eventually.

    // 7.4.3.2 Decoding a text region segment
    // "1) Interpret its header, as described in 7.4.3.1."
    // Done!

    // "2) Decode (or retrieve the results of decoding) any referred-to symbol dictionary and tables segments."
    Vector<NonnullRefPtr<Symbol>> symbols;
    for (auto referred_to_segment_number : segment.header.referred_to_segment_numbers) {
        auto opt_referred_to_segment = context.segments_by_number.get(referred_to_segment_number);
        if (!opt_referred_to_segment.has_value())
            return Error::from_string_literal("JBIG2ImageDecoderPlugin: Text segment refers to non-existent segment");
        dbgln_if(JBIG2_DEBUG, "Text segment refers to segment id {} index {}", referred_to_segment_number, opt_referred_to_segment.value());
        auto const& referred_to_segment = context.segments[opt_referred_to_segment.value()];
        if (!referred_to_segment.symbols.has_value())
            return Error::from_string_literal("JBIG2ImageDecoderPlugin: Text segment referred-to segment without symbols");
        symbols.extend(referred_to_segment.symbols.value());
    }

    // "3) As described in E.3.7, reset all the arithmetic coding statistics to zero."
    // FIXME

    // "4) Invoke the text region decoding procedure described in 6.4, with the parameters to the text region decoding procedure set as shown in Table 34."
    TextRegionDecodingInputParameters inputs;
    inputs.uses_huffman_encoding = uses_huffman_encoding;
    inputs.uses_refinement_coding = uses_refinement_coding;
    inputs.default_pixel = default_pixel_value;
    inputs.operator_ = static_cast<CombinationOperator>(combination_operator);
    inputs.is_transposed = is_transposed;
    inputs.reference_corner = static_cast<TextRegionDecodingInputParameters::Corner>(reference_corner);
    inputs.delta_s_offset = delta_s_offset;
    inputs.region_width = information_field.width;
    inputs.region_height = information_field.height;
    inputs.number_of_instances = number_of_symbol_instances;
    inputs.size_of_symbol_instance_strips = strip_size;
    inputs.id_symbol_code_length = ceil(log2(symbols.size()));
    inputs.symbols = move(symbols);
    // FIXME: Huffman tables.
    inputs.refinement_template = refinement_template;
    inputs.refinement_adaptive_template_pixels = adaptive_refinement_template;

    auto result = TRY(text_region_decoding_procedure(inputs, data.slice(TRY(stream.tell()))));

    composite_bitbuffer(*context.page.bits, *result, { information_field.x_location, information_field.y_location }, information_field.external_combination_operator());

    return {};
}

static ErrorOr<void> decode_pattern_dictionary(JBIG2LoadingContext&, SegmentData const&)
{
    return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode pattern dictionary yet");
}

static ErrorOr<void> decode_intermediate_halftone_region(JBIG2LoadingContext&, SegmentData const&)
{
    return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode intermediate halftone region yet");
}

static ErrorOr<void> decode_immediate_halftone_region(JBIG2LoadingContext&, SegmentData const&)
{
    return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode immediate halftone region yet");
}

static ErrorOr<void> decode_immediate_lossless_halftone_region(JBIG2LoadingContext&, SegmentData const&)
{
    return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode immediate lossless halftone region yet");
}

static ErrorOr<void> decode_intermediate_generic_region(JBIG2LoadingContext&, SegmentData const&)
{
    return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode intermediate generic region yet");
}

static ErrorOr<void> decode_immediate_generic_region(JBIG2LoadingContext& context, SegmentData const& segment)
{
    // 7.4.6 Generic region segment syntax
    auto data = segment.data;
    auto information_field = TRY(decode_region_segment_information_field(data));
    data = data.slice(sizeof(information_field));

    dbgln_if(JBIG2_DEBUG, "Generic region: width={}, height={}, x={}, y={}, flags={:#x}", information_field.width, information_field.height, information_field.x_location, information_field.y_location, information_field.flags);

    // 7.4.6.2 Generic region segment flags
    if (data.is_empty())
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: No segment data");
    u8 flags = data[0];
    bool uses_mmr = (flags & 1) != 0;
    u8 arithmetic_coding_template = (flags >> 1) & 3;               // "GBTEMPLATE"
    bool typical_prediction_generic_decoding_on = (flags >> 3) & 1; // "TPGDON"; "TPGD" is short for "Typical Prediction for Generic Direct coding"
    bool uses_extended_reference_template = (flags >> 4) & 1;       // "EXTTEMPLATE"
    if (flags & 0b1110'0000)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid flags");
    data = data.slice(sizeof(flags));

    // 7.4.6.3 Generic region segment AT flags
    Array<AdaptiveTemplatePixel, 12> adaptive_template_pixels {};
    if (!uses_mmr) {
        dbgln_if(JBIG2_DEBUG, "Non-MMR generic region, GBTEMPLATE={} TPGDON={} EXTTEMPLATE={}", arithmetic_coding_template, typical_prediction_generic_decoding_on, uses_extended_reference_template);

        if (arithmetic_coding_template == 0 && uses_extended_reference_template) {
            // This was added in T.88 Amendment 2 (https://www.itu.int/rec/T-REC-T.88-200306-S!Amd2/en) mid-2003.
            // I haven't seen it being used in the wild, and the spec says "32-byte field as shown below" and then shows 24 bytes,
            // so it's not clear how much data to read.
            return Error::from_string_literal("JBIG2ImageDecoderPlugin: GBTEMPLATE=0 EXTTEMPLATE=1 not yet implemented");
        }

        size_t number_of_adaptive_template_pixels = arithmetic_coding_template == 0 ? 4 : 1;
        if (data.size() < 2 * number_of_adaptive_template_pixels)
            return Error::from_string_literal("JBIG2ImageDecoderPlugin: No adaptive template data");
        for (size_t i = 0; i < number_of_adaptive_template_pixels; ++i) {
            adaptive_template_pixels[i].x = static_cast<i8>(data[2 * i]);
            adaptive_template_pixels[i].y = static_cast<i8>(data[2 * i + 1]);
        }
        data = data.slice(2 * number_of_adaptive_template_pixels);
    }

    // 7.4.6.4 Decoding a generic region segment
    // "1) Interpret its header, as described in 7.4.6.1"
    // Done above.
    // "2) As described in E.3.7, reset all the arithmetic coding statistics to zero."
    Vector<JBIG2::ArithmeticDecoder::Context> contexts;
    contexts.resize(1 << 16);

    // "3) Invoke the generic region decoding procedure described in 6.2, with the parameters to the generic region decoding procedure set as shown in Table 37."
    GenericRegionDecodingInputParameters inputs;
    inputs.is_modified_modified_read = uses_mmr;
    inputs.region_width = information_field.width;
    inputs.region_height = information_field.height;
    inputs.gb_template = arithmetic_coding_template;
    inputs.is_typical_prediction_used = typical_prediction_generic_decoding_on;
    inputs.is_extended_reference_template_used = uses_extended_reference_template;
    inputs.skip_pattern = OptionalNone {};
    inputs.adaptive_template_pixels = adaptive_template_pixels;

    Optional<JBIG2::ArithmeticDecoder> decoder;
    if (!uses_mmr) {
        decoder = TRY(JBIG2::ArithmeticDecoder::initialize(data));
        inputs.arithmetic_decoder = &decoder.value();
    }

    auto result = TRY(generic_region_decoding_procedure(inputs, data, contexts));

    // 8.2 Page image composition step 5)
    if (information_field.x_location + information_field.width > (u32)context.page.size.width()
        || information_field.y_location + information_field.height > (u32)context.page.size.height()) {
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Region bounds outsize of page bounds");
    }

    composite_bitbuffer(*context.page.bits, *result, { information_field.x_location, information_field.y_location }, information_field.external_combination_operator());

    return {};
}

static ErrorOr<void> decode_intermediate_generic_refinement_region(JBIG2LoadingContext&, SegmentData const&)
{
    return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode intermediate generic refinement region yet");
}

static ErrorOr<void> decode_immediate_generic_refinement_region(JBIG2LoadingContext&, SegmentData const&)
{
    return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode immediate generic refinement region yet");
}

static ErrorOr<void> decode_immediate_lossless_generic_refinement_region(JBIG2LoadingContext&, SegmentData const&)
{
    return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode immediate lossless generic refinement region yet");
}

static ErrorOr<void> decode_page_information(JBIG2LoadingContext& context, SegmentData const& segment)
{
    // 7.4.8 Page information segment syntax and 8.1 Decoder model steps 1) - 3).

    // "1) Decode the page information segment.""
    auto page_information = TRY(decode_page_information_segment(segment.data));

    bool page_is_striped = (page_information.striping_information & 0x80) != 0;
    if (page_information.bitmap_height == 0xffff'ffff && !page_is_striped)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Non-striped bitmaps of indeterminate height not allowed");

    u8 default_color = (page_information.flags >> 2) & 1;
    u8 default_combination_operator = (page_information.flags >> 3) & 3;
    context.page.default_combination_operator = static_cast<CombinationOperator>(default_combination_operator);

    // FIXME: Do something with the other fields in page_information.

    // "2) Create the page buffer, of the size given in the page information segment.
    //
    //     If the page height is unknown, then this is not possible. However, in this case the page must be striped,
    //     and the maximum stripe height specified, and the initial page buffer can be created with height initially
    //     equal to this maximum stripe height."
    size_t height = page_information.bitmap_height;
    if (height == 0xffff'ffff)
        height = page_information.striping_information & 0x7F;
    context.page.bits = TRY(BitBuffer::create(page_information.bitmap_width, height));

    // "3) Fill the page buffer with the page's default pixel value."
    context.page.bits->fill(default_color != 0);

    return {};
}

static ErrorOr<void> decode_end_of_page(JBIG2LoadingContext&, SegmentData const& segment)
{
    // 7.4.9 End of page segment syntax
    if (segment.data.size() != 0)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: End of page segment has non-zero size");
    // FIXME: If the page had unknown height, check that previous segment was end-of-stripe.
    // FIXME: Maybe mark page as completed and error if we see more segments for it?
    return {};
}

static ErrorOr<void> decode_end_of_stripe(JBIG2LoadingContext&, SegmentData const&)
{
    return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode end of stripe yet");
}

static ErrorOr<void> decode_end_of_file(JBIG2LoadingContext&, SegmentData const& segment)
{
    // 7.4.11 End of file segment syntax
    if (segment.data.size() != 0)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: End of file segment has non-zero size");
    return {};
}

static ErrorOr<void> decode_profiles(JBIG2LoadingContext&, SegmentData const&)
{
    return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode profiles yet");
}

static ErrorOr<void> decode_tables(JBIG2LoadingContext&, SegmentData const&)
{
    return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode tables yet");
}

static ErrorOr<void> decode_color_palette(JBIG2LoadingContext&, SegmentData const&)
{
    return Error::from_string_literal("JBIG2ImageDecoderPlugin: Cannot decode color palette yet");
}

static ErrorOr<void> decode_extension(JBIG2LoadingContext&, SegmentData const& segment)
{
    // 7.4.14 Extension segment syntax
    FixedMemoryStream stream { segment.data };

    enum ExtensionType {
        SingleByteCodedComment = 0x20000000,
        MultiByteCodedComment = 0x20000002,
    };
    u32 type = TRY(stream.read_value<BigEndian<u32>>());

    auto read_string = [&]<class T>() -> ErrorOr<Vector<T>> {
        Vector<T> result;
        do {
            result.append(TRY(stream.read_value<BigEndian<T>>()));
        } while (result.last());
        result.take_last();
        return result;
    };

    switch (type) {
    case SingleByteCodedComment: {
        // 7.4.15.1 Single-byte coded comment
        // Pairs of zero-terminated ISO/IEC 8859-1 (latin1) pairs, terminated by another \0.
        while (true) {
            auto first_bytes = TRY(read_string.template operator()<u8>());
            if (first_bytes.is_empty())
                break;

            auto second_bytes = TRY(read_string.template operator()<u8>());

            auto first = TRY(TextCodec::decoder_for("ISO-8859-1"sv)->to_utf8(StringView { first_bytes }));
            auto second = TRY(TextCodec::decoder_for("ISO-8859-1"sv)->to_utf8(StringView { second_bytes }));
            dbgln("JBIG2ImageDecoderPlugin: key '{}', value '{}'", first, second);
        }
        if (!stream.is_eof())
            return Error::from_string_literal("JBIG2ImageDecoderPlugin: Trailing data after SingleByteCodedComment");
        return {};
    }
    case MultiByteCodedComment: {
        // 7.4.15.2 Multi-byte coded comment
        // Pairs of (two-byte-)zero-terminated UCS-2 pairs, terminated by another \0\0.
        while (true) {
            auto first_ucs2 = TRY(read_string.template operator()<u16>());
            if (first_ucs2.is_empty())
                break;

            auto second_ucs2 = TRY(read_string.template operator()<u16>());

            auto first = TRY(Utf16View(first_ucs2).to_utf8());
            auto second = TRY(Utf16View(second_ucs2).to_utf8());
            dbgln("JBIG2ImageDecoderPlugin: key '{}', value '{}'", first, second);
        }
        if (!stream.is_eof())
            return Error::from_string_literal("JBIG2ImageDecoderPlugin: Trailing data after MultiByteCodedComment");
        return {};
    }
    }

    // FIXME: If bit 31 in `type` is not set, the extension isn't necessary, and we could ignore it.
    dbgln("JBIG2ImageDecoderPlugin: Unknown extension type {:#x}", type);
    return Error::from_string_literal("JBIG2ImageDecoderPlugin: Unknown extension type");
}

static ErrorOr<void> decode_data(JBIG2LoadingContext& context)
{
    TRY(warn_about_multiple_pages(context));

    for (size_t i = 0; i < context.segments.size(); ++i) {
        auto& segment = context.segments[i];

        if (segment.header.page_association != 0 && segment.header.page_association != 1)
            continue;

        switch (segment.header.type) {
        case SegmentType::SymbolDictionary:
            TRY(decode_symbol_dictionary(context, segment));
            break;
        case SegmentType::IntermediateTextRegion:
            TRY(decode_intermediate_text_region(context, segment));
            break;
        case SegmentType::ImmediateTextRegion:
        case SegmentType::ImmediateLosslessTextRegion:
            // 7.4.3 Text region segment syntax
            // "The data parts of all three of the text region segment types ("intermediate text region", "immediate text region" and
            //  "immediate lossless text region") are coded identically, but are acted upon differently, see 8.2."
            // But 8.2 only describes a difference between intermediate and immediate regions as far as I can tell,
            // and calling the immediate text region handler for immediate lossless text regions seems to do the right thing (?).
            TRY(decode_immediate_text_region(context, segment));
            break;
        case SegmentType::PatternDictionary:
            TRY(decode_pattern_dictionary(context, segment));
            break;
        case SegmentType::IntermediateHalftoneRegion:
            TRY(decode_intermediate_halftone_region(context, segment));
            break;
        case SegmentType::ImmediateHalftoneRegion:
            TRY(decode_immediate_halftone_region(context, segment));
            break;
        case SegmentType::ImmediateLosslessHalftoneRegion:
            TRY(decode_immediate_lossless_halftone_region(context, segment));
            break;
        case SegmentType::IntermediateGenericRegion:
            TRY(decode_intermediate_generic_region(context, segment));
            break;
        case SegmentType::ImmediateGenericRegion:
        case SegmentType::ImmediateLosslessGenericRegion:
            // 7.4.6 Generic region segment syntax
            // "The data parts of all three of the generic region segment types ("intermediate generic region", "immediate generic region" and
            //  "immediate lossless generic region") are coded identically, but are acted upon differently, see 8.2."
            // But 8.2 only describes a difference between intermediate and immediate regions as far as I can tell,
            // and calling the immediate generic region handler for immediate generic lossless regions seems to do the right thing (?).
            TRY(decode_immediate_generic_region(context, segment));
            break;
        case SegmentType::IntermediateGenericRefinementRegion:
            TRY(decode_intermediate_generic_refinement_region(context, segment));
            break;
        case SegmentType::ImmediateGenericRefinementRegion:
            TRY(decode_immediate_generic_refinement_region(context, segment));
            break;
        case SegmentType::ImmediateLosslessGenericRefinementRegion:
            TRY(decode_immediate_lossless_generic_refinement_region(context, segment));
            break;
        case SegmentType::PageInformation:
            TRY(decode_page_information(context, segment));
            break;
        case SegmentType::EndOfPage:
            TRY(decode_end_of_page(context, segment));
            break;
        case SegmentType::EndOfStripe:
            TRY(decode_end_of_stripe(context, segment));
            break;
        case SegmentType::EndOfFile:
            TRY(decode_end_of_file(context, segment));
            // "If a file contains an end of file segment, it must be the last segment."
            if (i != context.segments.size() - 1)
                return Error::from_string_literal("JBIG2ImageDecoderPlugin: End of file segment not last segment");
            break;
        case SegmentType::Profiles:
            TRY(decode_profiles(context, segment));
            break;
        case SegmentType::Tables:
            TRY(decode_tables(context, segment));
            break;
        case SegmentType::ColorPalette:
            TRY(decode_color_palette(context, segment));
            break;
        case SegmentType::Extension:
            TRY(decode_extension(context, segment));
            break;
        }
    }

    return {};
}

JBIG2ImageDecoderPlugin::JBIG2ImageDecoderPlugin()
{
    m_context = make<JBIG2LoadingContext>();
}

IntSize JBIG2ImageDecoderPlugin::size()
{
    return m_context->page.size;
}

bool JBIG2ImageDecoderPlugin::sniff(ReadonlyBytes data)
{
    return data.starts_with(id_string);
}

ErrorOr<NonnullOwnPtr<ImageDecoderPlugin>> JBIG2ImageDecoderPlugin::create(ReadonlyBytes data)
{
    auto plugin = TRY(adopt_nonnull_own_or_enomem(new (nothrow) JBIG2ImageDecoderPlugin()));
    TRY(decode_jbig2_header(*plugin->m_context, data));

    data = data.slice(sizeof(id_string) + sizeof(u8) + (plugin->m_context->number_of_pages.has_value() ? sizeof(u32) : 0));
    TRY(decode_segment_headers(*plugin->m_context, data));

    TRY(scan_for_page_size(*plugin->m_context));

    return plugin;
}

ErrorOr<ImageFrameDescriptor> JBIG2ImageDecoderPlugin::frame(size_t index, Optional<IntSize>)
{
    // FIXME: Use this for multi-page JBIG2 files?
    if (index != 0)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Invalid frame index");

    if (m_context->state == JBIG2LoadingContext::State::Error)
        return Error::from_string_literal("JBIG2ImageDecoderPlugin: Decoding failed");

    if (m_context->state < JBIG2LoadingContext::State::Decoded) {
        auto result = decode_data(*m_context);
        if (result.is_error()) {
            m_context->state = JBIG2LoadingContext::State::Error;
            return result.release_error();
        }
        m_context->state = JBIG2LoadingContext::State::Decoded;
    }

    auto bitmap = TRY(m_context->page.bits->to_gfx_bitmap());
    return ImageFrameDescriptor { move(bitmap), 0 };
}

ErrorOr<ByteBuffer> JBIG2ImageDecoderPlugin::decode_embedded(Vector<ReadonlyBytes> data)
{
    auto plugin = TRY(adopt_nonnull_own_or_enomem(new (nothrow) JBIG2ImageDecoderPlugin()));
    plugin->m_context->organization = Organization::Embedded;

    for (auto const& segment_data : data)
        TRY(decode_segment_headers(*plugin->m_context, segment_data));

    TRY(scan_for_page_size(*plugin->m_context));
    TRY(decode_data(*plugin->m_context));

    return plugin->m_context->page.bits->to_byte_buffer();
}

}