//! Stage 3 — the real XMA2→PCM decoder.
//!
//! A faithful port of xenia-canary's `apu/xma_context_new.cc` decode pipeline
//! (`Work`/`Decode`/`Consume`/`StoreContextMerged`), adapted to the *mainline*
//! distro FFmpeg `AV_CODEC_ID_XMA2` decoder rather than canary's vendored
//! `AV_CODEC_ID_XMAFRAMES`.
//!
//! ## Determinism
//! There is no host decoder thread. [`super::xma::XmaDecoder::decode_pending`]
//! is invoked from the CPU scheduler's per-round coordinator
//! (`coord_post_round` in xenia-app). FFmpeg decode is itself deterministic
//! (same input bytes → same PCM), so the lockstep golden stays reproducible.
//!
//! ## FFmpeg framing — why this differs from canary
//! Canary feeds FFmpeg one *frame* at a time (it bit-extracts a single 512-
//! sample frame from the guest packet stream and hands it to the vendored
//! `XMAFRAMES` codec with a custom 1-byte padding header). The mainline
//! `xma2` decoder does NOT have `XMAFRAMES`; instead it consumes whole 2 KB
//! XMA2 *packets* (`block_align == 2048`), needs `extradata` declaring the
//! stream/channel layout, and manages frame splitting + a per-stream sample
//! FIFO internally. So this module keeps canary's *guest-facing* contract
//! (the `XMA_CONTEXT_DATA` packet/frame bookkeeping, the 256-byte-block output
//! ring buffer, the field writeback) but replaces canary's per-frame
//! `Decode()` body with: feed the current 2 KB packet to the xma2 decoder,
//! pull any 512-sample PCM frames it emits, convert them to interleaved S16BE,
//! and stage them as the "raw frame" that `Consume()` drains into the output
//! ring.
//!
//! See `xma2_codec.rs` for the unsafe FFmpeg wrapper.

use std::collections::VecDeque;

use xenia_memory::access::MemoryAccess;
use xenia_memory::GuestMemory;

use crate::xma2_codec::Xma2Codec;

// ---- Constants (canary `XmaContext` / `XmaContextNew`).

pub const BYTES_PER_PACKET: u32 = 2048;
pub const BYTES_PER_PACKET_HEADER: u32 = 4;
pub const BYTES_PER_PACKET_DATA: u32 = BYTES_PER_PACKET - BYTES_PER_PACKET_HEADER;
pub const BITS_PER_PACKET: u32 = BYTES_PER_PACKET * 8;
/// Canary `kBitsPerPacketHeader` (in the *new* context) is 32.
pub const BITS_PER_PACKET_HEADER: u32 = 32;
pub const BITS_PER_FRAME_HEADER: u32 = 15;

pub const SAMPLES_PER_FRAME: u32 = 512;
pub const BYTES_PER_SAMPLE: u32 = 2;
pub const BYTES_PER_FRAME_CHANNEL: u32 = SAMPLES_PER_FRAME * BYTES_PER_SAMPLE; // 1024
pub const OUTPUT_BYTES_PER_BLOCK: u32 = 256;
pub const OUTPUT_MAX_SIZE_BYTES: u32 = 31 * OUTPUT_BYTES_PER_BLOCK;

pub const MAX_FRAME_LENGTH: u32 = 0x7FFF;
pub const MAX_FRAME_SIZE_IN_BITS: u32 = 0x4000 - BITS_PER_PACKET_HEADER;

const ID_TO_SAMPLE_RATE: [u32; 4] = [24000, 32000, 44100, 48000];

/// Project a bare-physical XMA buffer pointer (`0x0xxxxxxx`) to the host-backed
/// guest VA used by the rest of the emulator. Identical formula to
/// `xenia_gpu::physical_to_backing` for the physical window; the input/output
/// buffer pointers in the context are always in the low physical window.
#[inline]
pub fn xma_phys_to_backing(p: u32) -> u32 {
    0x4000_0000 | (p & 0x1FFF_FFFF)
}

// ---- XMA_CONTEXT_DATA (canary `xma_context.h`, 64 bytes, 16 dwords).
//
// Stored big-endian in guest memory. We load all 16 dwords (BE) and unpack the
// bitfields exactly per the canary layout (bitfields pack LSB-first within each
// host-order dword). All fields below are kept as plain integers.

#[derive(Clone, Copy, Debug, Default)]
pub struct XmaContextData {
    // DWORD 0
    pub input_buffer_0_packet_count: u32, // :12
    pub loop_count: u32,                  // :8
    pub input_buffer_0_valid: u32,        // :1
    pub input_buffer_1_valid: u32,        // :1
    pub output_buffer_block_count: u32,   // :5
    pub output_buffer_write_offset: u32,  // :5
    // DWORD 1
    pub input_buffer_1_packet_count: u32, // :12
    pub loop_subframe_start: u32,         // :2
    pub loop_subframe_end: u32,           // :3
    pub loop_subframe_skip: u32,          // :3
    pub subframe_decode_count: u32,       // :4
    pub output_buffer_padding: u32,       // :3
    pub sample_rate: u32,                 // :2
    pub is_stereo: u32,                   // :1
    pub unk_dword_1_c: u32,               // :1
    pub output_buffer_valid: u32,         // :1
    // DWORD 2
    pub input_buffer_read_offset: u32, // :26
    pub error_status: u32,             // :5
    pub error_set: u32,                // :1
    // DWORD 3
    pub loop_start: u32,          // :26
    pub parser_error_status: u32, // :5
    pub parser_error_set: u32,    // :1
    // DWORD 4
    pub loop_end: u32,        // :26
    pub packet_metadata: u32, // :5
    pub current_buffer: u32,  // :1
    // DWORD 5..8
    pub input_buffer_0_ptr: u32,
    pub input_buffer_1_ptr: u32,
    pub output_buffer_ptr: u32,
    pub work_buffer_ptr: u32,
    // DWORD 9
    pub output_buffer_read_offset: u32, // :5
    pub stop_when_done: u32,            // :1 (bit 30)
    pub interrupt_when_done: u32,       // :1 (bit 31)
}

#[inline]
fn bits(v: u32, shift: u32, width: u32) -> u32 {
    (v >> shift) & ((1u32 << width) - 1)
}

impl XmaContextData {
    /// Read the 64-byte context struct from guest VA `ctx_va` (already a VA,
    /// not a physical ptr). Each dword is read big-endian via `read_u32`.
    pub fn read(mem: &GuestMemory, ctx_va: u32) -> Self {
        let mut d = [0u32; 16];
        for (i, w) in d.iter_mut().enumerate() {
            *w = mem.read_u32(ctx_va + (i as u32) * 4);
        }
        let mut c = Self::default();
        // DWORD 0
        c.input_buffer_0_packet_count = bits(d[0], 0, 12);
        c.loop_count = bits(d[0], 12, 8);
        c.input_buffer_0_valid = bits(d[0], 20, 1);
        c.input_buffer_1_valid = bits(d[0], 21, 1);
        c.output_buffer_block_count = bits(d[0], 22, 5);
        c.output_buffer_write_offset = bits(d[0], 27, 5);
        // DWORD 1
        c.input_buffer_1_packet_count = bits(d[1], 0, 12);
        c.loop_subframe_start = bits(d[1], 12, 2);
        c.loop_subframe_end = bits(d[1], 14, 3);
        c.loop_subframe_skip = bits(d[1], 17, 3);
        c.subframe_decode_count = bits(d[1], 20, 4);
        c.output_buffer_padding = bits(d[1], 24, 3);
        c.sample_rate = bits(d[1], 27, 2);
        c.is_stereo = bits(d[1], 29, 1);
        c.unk_dword_1_c = bits(d[1], 30, 1);
        c.output_buffer_valid = bits(d[1], 31, 1);
        // DWORD 2
        c.input_buffer_read_offset = bits(d[2], 0, 26);
        c.error_status = bits(d[2], 26, 5);
        c.error_set = bits(d[2], 31, 1);
        // DWORD 3
        c.loop_start = bits(d[3], 0, 26);
        c.parser_error_status = bits(d[3], 26, 5);
        c.parser_error_set = bits(d[3], 31, 1);
        // DWORD 4
        c.loop_end = bits(d[4], 0, 26);
        c.packet_metadata = bits(d[4], 26, 5);
        c.current_buffer = bits(d[4], 31, 1);
        // DWORD 5..8
        c.input_buffer_0_ptr = d[5];
        c.input_buffer_1_ptr = d[6];
        c.output_buffer_ptr = d[7];
        c.work_buffer_ptr = d[8];
        // DWORD 9
        c.output_buffer_read_offset = bits(d[9], 0, 5);
        c.stop_when_done = bits(d[9], 30, 1);
        c.interrupt_when_done = bits(d[9], 31, 1);
        c
    }

    /// Repack the bitfields back into the 16 dwords (host order). Only the
    /// decoder-owned fields differ from what was read; callers use
    /// [`store_merged`] to write back without clobbering game-owned fields.
    fn pack(&self) -> [u32; 16] {
        let mut d = [0u32; 16];
        d[0] = (self.input_buffer_0_packet_count & 0xFFF)
            | ((self.loop_count & 0xFF) << 12)
            | ((self.input_buffer_0_valid & 1) << 20)
            | ((self.input_buffer_1_valid & 1) << 21)
            | ((self.output_buffer_block_count & 0x1F) << 22)
            | ((self.output_buffer_write_offset & 0x1F) << 27);
        d[1] = (self.input_buffer_1_packet_count & 0xFFF)
            | ((self.loop_subframe_start & 0x3) << 12)
            | ((self.loop_subframe_end & 0x7) << 14)
            | ((self.loop_subframe_skip & 0x7) << 17)
            | ((self.subframe_decode_count & 0xF) << 20)
            | ((self.output_buffer_padding & 0x7) << 24)
            | ((self.sample_rate & 0x3) << 27)
            | ((self.is_stereo & 1) << 29)
            | ((self.unk_dword_1_c & 1) << 30)
            | ((self.output_buffer_valid & 1) << 31);
        d[2] = (self.input_buffer_read_offset & 0x3FF_FFFF)
            | ((self.error_status & 0x1F) << 26)
            | ((self.error_set & 1) << 31);
        d[3] = (self.loop_start & 0x3FF_FFFF)
            | ((self.parser_error_status & 0x1F) << 26)
            | ((self.parser_error_set & 1) << 31);
        d[4] = (self.loop_end & 0x3FF_FFFF)
            | ((self.packet_metadata & 0x1F) << 26)
            | ((self.current_buffer & 1) << 31);
        d[5] = self.input_buffer_0_ptr;
        d[6] = self.input_buffer_1_ptr;
        d[7] = self.output_buffer_ptr;
        d[8] = self.work_buffer_ptr;
        d[9] = (self.output_buffer_read_offset & 0x1F)
            | ((self.stop_when_done & 1) << 30)
            | ((self.interrupt_when_done & 1) << 31);
        d
    }

    pub fn is_input_buffer_valid(&self, idx: u32) -> bool {
        if idx == 0 {
            self.input_buffer_0_valid != 0
        } else {
            self.input_buffer_1_valid != 0
        }
    }
    pub fn is_current_input_buffer_valid(&self) -> bool {
        self.is_input_buffer_valid(self.current_buffer)
    }
    pub fn is_any_input_buffer_valid(&self) -> bool {
        self.input_buffer_0_valid != 0 || self.input_buffer_1_valid != 0
    }
    pub fn input_buffer_address(&self, idx: u32) -> u32 {
        if idx == 0 {
            self.input_buffer_0_ptr
        } else {
            self.input_buffer_1_ptr
        }
    }
    pub fn current_input_buffer_address(&self) -> u32 {
        self.input_buffer_address(self.current_buffer)
    }
    pub fn input_buffer_packet_count(&self, idx: u32) -> u32 {
        if idx == 0 {
            self.input_buffer_0_packet_count
        } else {
            self.input_buffer_1_packet_count
        }
    }
    pub fn current_input_buffer_packet_count(&self) -> u32 {
        self.input_buffer_packet_count(self.current_buffer)
    }
}

/// Merge decoder-owned fields back into guest memory (canary `StoreContextMerged`).
/// Re-reads the current context (game may have raced an update), overwrites only
/// the fields the decoder owns, and writes all 16 dwords back BE.
fn store_merged(
    mem: &GuestMemory,
    ctx_va: u32,
    data: &XmaContextData,
    initial: &XmaContextData,
) {
    let mut fresh = XmaContextData::read(mem, ctx_va);
    // DWORD 0
    fresh.loop_count = data.loop_count;
    fresh.output_buffer_write_offset = data.output_buffer_write_offset;
    if initial.input_buffer_0_valid != 0 && data.input_buffer_0_valid == 0 {
        fresh.input_buffer_0_valid = 0;
    }
    if initial.input_buffer_1_valid != 0 && data.input_buffer_1_valid == 0 {
        fresh.input_buffer_1_valid = 0;
    }
    // DWORD 1
    if initial.output_buffer_valid != 0 && data.output_buffer_valid == 0 {
        fresh.output_buffer_valid = 0;
    }
    // DWORD 2
    fresh.input_buffer_read_offset = data.input_buffer_read_offset;
    fresh.error_status = data.error_status;
    // DWORD 4
    fresh.current_buffer = data.current_buffer;
    // DWORD 9
    fresh.output_buffer_read_offset = data.output_buffer_read_offset;

    let d = fresh.pack();
    for (i, w) in d.iter().enumerate() {
        mem.write_u32(ctx_va + (i as u32) * 4, *w);
    }
}

/// Public wrapper for [`store_merged`] (called from the orchestrator in xma.rs).
pub fn store_merged_pub(
    mem: &GuestMemory,
    ctx_va: u32,
    data: &XmaContextData,
    initial: &XmaContextData,
) {
    store_merged(mem, ctx_va, data, initial);
}

/// Free byte count in a ring buffer from `write_off` to `read_off`
/// (canary `RingBuffer::write_count`).
pub fn ring_write_count(read_off: u32, write_off: u32, capacity: u32) -> u32 {
    if read_off == write_off {
        capacity
    } else if write_off < read_off {
        read_off - write_off
    } else {
        (capacity - write_off) + read_off
    }
}

/// Write `bytes` into the guest ring buffer at `backing + write_off`, wrapping
/// at `capacity`. Returns the new write offset (canary `RingBuffer::Write`).
pub fn ring_write(
    mem: &GuestMemory,
    backing: u32,
    capacity: u32,
    write_off: u32,
    bytes: &[u8],
) -> u32 {
    let count = (bytes.len() as u32).min(capacity);
    if count == 0 {
        return write_off;
    }
    if write_off + count < capacity {
        mem.write_bytes(backing + write_off, &bytes[..count as usize]);
        write_off + count
    } else {
        let left = capacity - write_off;
        mem.write_bytes(backing + write_off, &bytes[..left as usize]);
        let right = count - left;
        mem.write_bytes(backing, &bytes[left as usize..(left + right) as usize]);
        right
    }
}

// ---- BitStream (port of canary `base/bit_stream.cc`). Big-endian source.

pub struct BitStream<'a> {
    buf: &'a [u8],
    offset_bits: usize,
    size_bits: usize,
}

impl<'a> BitStream<'a> {
    pub fn new(buf: &'a [u8], size_bits: usize) -> Self {
        Self { buf, offset_bits: 0, size_bits }
    }
    pub fn offset_bits(&self) -> usize {
        self.offset_bits
    }
    pub fn set_offset(&mut self, off: usize) {
        self.offset_bits = off.min(self.size_bits);
    }
    pub fn advance(&mut self, n: usize) {
        self.set_offset(self.offset_bits + n);
    }
    pub fn bits_remaining(&self) -> usize {
        self.size_bits - self.offset_bits
    }
    /// Peek up to 57 bits (canary contract). Reads 8 bytes BE then shifts.
    pub fn peek(&self, num_bits: usize) -> u64 {
        debug_assert!(num_bits <= 57);
        // offset_bytes = min(offset>>3, (size-64)>>3), matching canary so an
        // 8-byte load near the buffer end stays in range.
        let max_byte = if self.size_bits >= 64 {
            (self.size_bits - 64) >> 3
        } else {
            0
        };
        let offset_bytes = (self.offset_bits >> 3).min(max_byte);
        let rel = self.offset_bits - (offset_bytes << 3);
        let mut tmp = [0u8; 8];
        let avail = self.buf.len().saturating_sub(offset_bytes).min(8);
        tmp[..avail].copy_from_slice(&self.buf[offset_bytes..offset_bytes + avail]);
        let mut value = u64::from_be_bytes(tmp);
        value >>= 64 - (rel + num_bits);
        value &= (1u64 << num_bits) - 1;
        value
    }
    pub fn read(&mut self, num_bits: usize) -> u64 {
        let v = self.peek(num_bits);
        self.advance(num_bits);
        v
    }
    /// Copy `num_bits` from the stream into `dest` (bit-packed, MSB-first within
    /// each byte). Returns the starting bit offset within the first byte
    /// (canary returns `rel_offset_bits` — the frame's intra-byte alignment).
    pub fn copy(&mut self, dest: &mut [u8], num_bits: usize) -> usize {
        let offset_bytes = self.offset_bits >> 3;
        let rel = self.offset_bits - (offset_bytes << 3);
        let mut bits_left = num_bits;
        let mut out = 0usize;

        if rel != 0 {
            let bits = self.peek(8 - rel) as u8;
            let clear_mask = !(((1u8 << rel) - 1)) as u8;
            dest[out] &= clear_mask;
            dest[out] |= bits;
            bits_left -= 8 - rel;
            self.advance(8 - rel);
            out += 1;
        }
        if bits_left >= 8 {
            let nbytes = bits_left / 8;
            let src_off = (self.offset_bits >> 3).min(self.buf.len());
            let copy = nbytes.min(self.buf.len().saturating_sub(src_off));
            dest[out..out + copy]
                .copy_from_slice(&self.buf[src_off..src_off + copy]);
            out += nbytes;
            self.advance(nbytes * 8);
            bits_left -= nbytes * 8;
        }
        if bits_left != 0 {
            let mut b = self.peek(bits_left) as u8;
            b <<= 8 - bits_left;
            let clear_mask = ((1u16 << bits_left) - 1) as u8;
            dest[out] &= clear_mask;
            dest[out] |= b;
            self.advance(bits_left);
        }
        rel
    }
}

// ---- XMA packet header helpers (canary `xma_helpers.h`).

#[inline]
pub fn packet_frame_count(packet: &[u8]) -> u8 {
    packet[0] >> 2
}
#[inline]
pub fn packet_metadata(packet: &[u8]) -> u8 {
    packet[2] & 0x7
}
#[inline]
pub fn is_packet_xma2(packet: &[u8]) -> bool {
    packet_metadata(packet) == 1
}
#[inline]
pub fn packet_skip_count(packet: &[u8]) -> u8 {
    packet[3]
}
/// First frame offset in bits (canary `GetPacketFrameOffset`): a 15-bit value
/// across bytes 0..2, plus the 32-bit header.
#[inline]
pub fn packet_frame_offset(packet: &[u8]) -> u32 {
    let val = (((packet[0] as u32 & 0x3) << 13)
        | ((packet[1] as u32) << 5)
        | ((packet[2] as u32) >> 3))
        & 0xFFFF;
    val + 32
}

/// Sample-rate id → Hz.
pub fn sample_rate_hz(id: u32) -> u32 {
    ID_TO_SAMPLE_RATE[id.min(3) as usize]
}

// ---- Packet-walk for faithful input-offset advance (canary `GetPacketInfo`,
// `GetNextPacketReadOffset`, and the offset arithmetic at the tail of
// `XmaContextNew::Decode`). These let us advance `input_buffer_read_offset` one
// *frame* at a time at canary's exact cadence — independent of the mainline
// xma2 decoder's whole-packet/burst framing — so the offset crosses packet and
// buffer boundaries (and triggers SwapInputBuffer) at the true input-drain
// rate the guest's WMV demuxer polls.

/// Info about the frame at a given bit offset within a packet (canary
/// `kPacketInfo` / `GetPacketInfo`). `frame_count_` is the number of frames
/// that begin in the packet; `current_frame_size_` is the compressed bit size
/// of the frame at `frame_offset` (0 if it can't be resolved within this
/// packet — a split header).
#[derive(Default, Clone, Copy)]
pub struct PacketInfo {
    pub frame_count: u32,
    pub current_frame: u32,
    pub current_frame_size: u32,
}

impl PacketInfo {
    pub fn is_last_frame_in_packet(&self) -> bool {
        self.current_frame + 1 == self.frame_count
    }
}

/// Faithful port of canary `XmaContextNew::GetPacketInfo`.
pub fn get_packet_info(packet: &[u8], frame_offset: u32) -> PacketInfo {
    let mut info = PacketInfo::default();
    let first_frame_offset = packet_frame_offset(packet);
    let mut stream = BitStream::new(packet, BITS_PER_PACKET as usize);
    stream.set_offset(first_frame_offset as usize);

    // Split frame from previous packet.
    if frame_offset < first_frame_offset {
        info.current_frame = 0;
        info.current_frame_size = first_frame_offset - frame_offset;
    }

    loop {
        if stream.bits_remaining() < BITS_PER_FRAME_HEADER as usize {
            break;
        }
        let frame_size = stream.peek(BITS_PER_FRAME_HEADER as usize) as u32;
        if frame_size == 0 || frame_size == MAX_FRAME_LENGTH {
            break;
        }
        if stream.offset_bits() == frame_offset as usize {
            info.current_frame = info.frame_count;
            info.current_frame_size = frame_size;
        }
        info.frame_count += 1;
        if frame_size as usize > stream.bits_remaining() {
            // Last frame.
            break;
        }
        stream.advance((frame_size - 1) as usize);
        // Trailing continuation bit.
        if stream.read(1) == 0 {
            break;
        }
    }

    if is_packet_xma2(packet) {
        let xma2_frame_count = packet_frame_count(packet) as u32;
        if xma2_frame_count > info.frame_count {
            if info.current_frame_size == 0 {
                info.current_frame = info.frame_count;
            }
            info.frame_count = xma2_frame_count;
        }
    }
    info
}

/// Packet number for a bit offset (canary `GetPacketNumber`). Returns None when
/// the offset is in the header or past the buffer.
pub fn packet_number(size_bytes: u32, bit_offset: u32) -> Option<u32> {
    if bit_offset < BITS_PER_PACKET_HEADER {
        return None;
    }
    if bit_offset >= size_bytes * 8 {
        return None;
    }
    Some((bit_offset >> 3) / BYTES_PER_PACKET)
}

/// min(remaining_stream_bits, frame_size) (canary `GetAmountOfBitsToRead`).
pub fn amount_of_bits_to_read(remaining_stream_bits: u32, frame_size: u32) -> u32 {
    remaining_stream_bits.min(frame_size)
}

// ---- Per-context decode state (lives in the XmaDecoder, one per ctx).

#[derive(Default)]
pub struct ContextDecodeState {
    /// FFmpeg xma2 codec for this context (lazily created / reconfigured).
    pub codec: Option<Xma2Codec>,
    pub codec_rate: u32,
    pub codec_channels: u32,
    /// Staged interleaved S16BE PCM for the current decoded frame
    /// (`raw_frame_`), drained by Consume in 256-byte blocks.
    pub raw_frame: Vec<u8>,
    /// Decoded interleaved S16BE PCM not yet split into per-frame `raw_frame`s.
    /// The mainline xma2 decoder emits bursts of many 512-sample frames at once
    /// (internal FIFO + 4096-sample lookahead); we queue the bytes here and
    /// hand the guest exactly one 512-sample frame per `produce_frame`.
    pub pcm_queue: VecDeque<u8>,
    pub current_frame_remaining_subframes: u8,
    pub remaining_subframe_blocks_in_output: i32,
    /// Total 512-sample frames decoded for this context (diagnostic).
    pub frames_decoded: u64,
    /// Whether a "first frame" diagnostic has been emitted.
    pub first_frame_logged: bool,
    /// FFmpeg feed cursor: the next packet index (within the *current* input
    /// buffer at feed time) we will hand to FFmpeg. This is the decoder's
    /// internal intake position and is intentionally decoupled from the
    /// guest-visible `input_buffer_read_offset` (which advances per *emitted*
    /// frame via the faithful packet-walk). We feed ahead so FFmpeg always has
    /// enough buffered input to satisfy the guest's drain, while the guest sees
    /// the read offset move at canary's true per-frame cadence.
    pub feed_packet_index: u32,
    /// `current_buffer` the feed cursor is reading from; reset on swap so the
    /// feed follows the same ping-pong as the guest-visible buffer.
    pub feed_buffer: u32,
}

#[cfg(test)]
mod tests {
    use super::*;

    /// The bitfield unpack/pack must round-trip every decoder-relevant field at
    /// the exact canary offsets (regression against a shifted bit).
    #[test]
    fn context_bitfields_round_trip() {
        let mut c = XmaContextData::default();
        c.input_buffer_0_packet_count = 632;
        c.loop_count = 0;
        c.input_buffer_0_valid = 1;
        c.input_buffer_1_valid = 0;
        c.output_buffer_block_count = 30;
        c.output_buffer_write_offset = 5;
        c.subframe_decode_count = 8;
        c.output_buffer_padding = 1;
        c.sample_rate = 3;
        c.is_stereo = 1;
        c.output_buffer_valid = 1;
        c.input_buffer_read_offset = 16416;
        c.error_status = 4;
        c.current_buffer = 1;
        c.input_buffer_0_ptr = 0x0b9f_d000;
        c.output_buffer_ptr = 0x01f6_6e00;
        c.output_buffer_read_offset = 7;
        c.interrupt_when_done = 1;

        // pack → words → re-read via the same word layout.
        let d = c.pack();
        // Simulate read() decode from the packed words.
        let mut c2 = XmaContextData::default();
        c2.input_buffer_0_packet_count = bits(d[0], 0, 12);
        c2.input_buffer_0_valid = bits(d[0], 20, 1);
        c2.output_buffer_block_count = bits(d[0], 22, 5);
        c2.output_buffer_write_offset = bits(d[0], 27, 5);
        c2.subframe_decode_count = bits(d[1], 20, 4);
        c2.output_buffer_padding = bits(d[1], 24, 3);
        c2.sample_rate = bits(d[1], 27, 2);
        c2.is_stereo = bits(d[1], 29, 1);
        c2.output_buffer_valid = bits(d[1], 31, 1);
        c2.input_buffer_read_offset = bits(d[2], 0, 26);
        c2.error_status = bits(d[2], 26, 5);
        c2.current_buffer = bits(d[4], 31, 1);
        c2.output_buffer_read_offset = bits(d[9], 0, 5);
        c2.interrupt_when_done = bits(d[9], 31, 1);

        assert_eq!(c2.input_buffer_0_packet_count, 632);
        assert_eq!(c2.input_buffer_0_valid, 1);
        assert_eq!(c2.output_buffer_block_count, 30);
        assert_eq!(c2.output_buffer_write_offset, 5);
        assert_eq!(c2.subframe_decode_count, 8);
        assert_eq!(c2.output_buffer_padding, 1);
        assert_eq!(c2.sample_rate, 3);
        assert_eq!(c2.is_stereo, 1);
        assert_eq!(c2.output_buffer_valid, 1);
        assert_eq!(c2.input_buffer_read_offset, 16416);
        assert_eq!(c2.error_status, 4);
        assert_eq!(c2.current_buffer, 1);
        assert_eq!(c2.output_buffer_read_offset, 7);
        assert_eq!(c2.interrupt_when_done, 1);
    }

    #[test]
    fn phys_to_backing_projects_physical_window() {
        assert_eq!(xma_phys_to_backing(0x0b9f_d000), 0x4b9f_d000);
        assert_eq!(xma_phys_to_backing(0x01f6_6e00), 0x41f6_6e00);
    }

    #[test]
    fn ring_write_count_matches_canary() {
        // empty (read==write) → full capacity.
        assert_eq!(ring_write_count(0, 0, 7680), 7680);
        // write ahead of read.
        assert_eq!(ring_write_count(0, 256, 7680), 7680 - 256);
        // write wrapped behind read.
        assert_eq!(ring_write_count(512, 256, 7680), 256);
    }

    #[test]
    fn packet_header_helpers() {
        // Matches the observed first packet word 0x08000000: byte0=0x08.
        let pkt = [0x08u8, 0x00, 0x00, 0x00];
        assert_eq!(packet_frame_count(&pkt), 2); // 0x08>>2 = 2
        // frame offset: ((0x08&3)<<13 | 0<<5 | 0x00>>3) + 32 = 32.
        assert_eq!(packet_frame_offset(&pkt), 32);
        // A non-zero byte2 shifts the offset: 0x08>>3 = 1 → +1.
        let pkt2 = [0x08u8, 0x00, 0x08, 0x00];
        assert_eq!(packet_frame_offset(&pkt2), 33);
    }
}

impl ContextDecodeState {
    pub fn new() -> Self {
        Self {
            codec: None,
            codec_rate: 0,
            codec_channels: 0,
            raw_frame: vec![0u8; (BYTES_PER_FRAME_CHANNEL * 2) as usize],
            pcm_queue: VecDeque::new(),
            current_frame_remaining_subframes: 0,
            remaining_subframe_blocks_in_output: 0,
            frames_decoded: 0,
            first_frame_logged: false,
            feed_packet_index: 0,
            feed_buffer: 0,
        }
    }
}