// Copyright 2022 The Wuffs Authors.
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// https://www.apache.org/licenses/LICENSE-2.0> or the MIT license
// <LICENSE-MIT or https://opensource.org/licenses/MIT>, at your
// option. This file may not be copied, modified, or distributed
// except according to those terms.
//
// SPDX-License-Identifier: Apache-2.0 OR MIT

pub status "#bad header"
pub status "#bad run length encoding"
pub status "#truncated input"
pub status "#unsupported TARGA file"

pub const DECODER_WORKBUF_LEN_MAX_INCL_WORST_CASE : base.u64 = 0

pub struct decoder? implements base.image_decoder(
        width  : base.u32[..= 0xFFFF],
        height : base.u32[..= 0xFFFF],

        // The call sequence state machine is discussed in
        // (/doc/std/image-decoders-call-sequence.md).
        call_sequence : base.u8,

        header_id_length                   : base.u8,
        header_color_map_type              : base.u8,
        header_image_type                  : base.u8,
        header_color_map_first_entry_index : base.u16,
        header_color_map_length            : base.u16,
        header_color_map_entry_size        : base.u8,
        header_pixel_depth                 : base.u8,
        header_image_descriptor            : base.u8,

        opaque : base.bool,

        scratch_bytes_per_pixel : base.u32[..= 4],
        src_bytes_per_pixel     : base.u32[..= 4],
        src_pixfmt              : base.u32,

        frame_config_io_position : base.u64,

        swizzler : base.pixel_swizzler,
        util     : base.utility,
) + (
        dst_palette : array[4 * 256] base.u8,
        src_palette : array[4 * 256] base.u8,
        scratch     : array[4] base.u8,
)

pub func decoder.get_quirk(key: base.u32) base.u64 {
    return 0
}

pub func decoder.set_quirk!(key: base.u32, value: base.u64) base.status {
    return base."#unsupported option"
}

pub func decoder.decode_image_config?(dst: nptr base.image_config, src: base.io_reader) {
    var status : base.status

    while true {
        status =? this.do_decode_image_config?(dst: args.dst, src: args.src)
        if (status == base."$short read") and args.src.is_closed() {
            return "#truncated input"
        }
        yield? status
    }
}

pri func decoder.do_decode_image_config?(dst: nptr base.image_config, src: base.io_reader) {
    var c32 : base.u32
    var c5  : base.u32[..= 0x1F]
    var i   : base.u32

    if this.call_sequence <> 0x00 {
        return base."#bad call sequence"
    }

    this.header_id_length = args.src.read_u8?()
    this.header_color_map_type = args.src.read_u8?()
    if this.header_color_map_type > 1 {
        return "#bad header"
    }

    this.header_image_type = args.src.read_u8?()
    if (this.header_image_type == 0x01) or
            (this.header_image_type == 0x02) or
            (this.header_image_type == 0x03) or
            (this.header_image_type == 0x09) or
            (this.header_image_type == 0x0A) or
            (this.header_image_type == 0x0B) {
        // No-op.
    } else {
        // TODO: 0x20 and 0x21 are invalid, according to the spec, but are
        // apparently unofficial extensions.
        return "#bad header"
    }

    this.header_color_map_first_entry_index = args.src.read_u16le?()
    this.header_color_map_length = args.src.read_u16le?()
    this.header_color_map_entry_size = args.src.read_u8?()
    if this.header_color_map_type <> 0 {
        // We have a color-mapped image (in Wuffs, an indexed pixel format).
        if (this.header_color_map_first_entry_index <> 0) or
                (this.header_color_map_length > 0x100) {
            return "#unsupported TARGA file"
        } else if (this.header_color_map_entry_size <> 0x0F) and
                (this.header_color_map_entry_size <> 0x10) and
                (this.header_color_map_entry_size <> 0x18) and
                (this.header_color_map_entry_size <> 0x20) {
            return "#bad header"
        }
    } else {
        // The color-map fields must be zero.
        if (this.header_color_map_first_entry_index <> 0) or
                (this.header_color_map_length <> 0) or
                (this.header_color_map_entry_size <> 0) {
            return "#bad header"
        }
    }

    // Ignore the X-Origin and Y-Origin fields.
    args.src.skip_u32?(n: 4)

    this.width = args.src.read_u16le_as_u32?()
    this.height = args.src.read_u16le_as_u32?()

    this.header_pixel_depth = args.src.read_u8?()
    if (this.header_pixel_depth <> 0x01) and
            (this.header_pixel_depth <> 0x08) and
            (this.header_pixel_depth <> 0x0F) and
            (this.header_pixel_depth <> 0x10) and
            (this.header_pixel_depth <> 0x18) and
            (this.header_pixel_depth <> 0x20) {
        return "#bad header"
    }

    if (this.header_image_type | 8) == 0x09 {
        this.scratch_bytes_per_pixel = 1
        this.src_bytes_per_pixel = 1
        this.src_pixfmt = base.PIXEL_FORMAT__INDEXED__BGRA_NONPREMUL
        this.opaque =
                (this.header_color_map_entry_size == 0x0F) or
                (this.header_color_map_entry_size == 0x18)

    } else if (this.header_image_type | 8) == 0x0A {
        if (this.header_pixel_depth == 0x0F) or
                (this.header_pixel_depth == 0x10) {
            // Wuffs' base.pixel_swizzler doesn't support BGRX5551, so
            // scratch_bytes_per_pixel and src_bytes_per_pixel are different.
            this.scratch_bytes_per_pixel = 4
            this.src_bytes_per_pixel = 0
            this.src_pixfmt = base.PIXEL_FORMAT__BGRA_NONPREMUL
        } else if this.header_pixel_depth == 0x18 {
            this.scratch_bytes_per_pixel = 3
            this.src_bytes_per_pixel = 3
            this.src_pixfmt = base.PIXEL_FORMAT__BGR
            this.opaque = true
        } else if this.header_pixel_depth == 0x20 {
            this.scratch_bytes_per_pixel = 4
            this.src_bytes_per_pixel = 4
            this.src_pixfmt = base.PIXEL_FORMAT__BGRA_NONPREMUL
        } else {
            return "#unsupported TARGA file"
        }

    } else {
        if this.header_pixel_depth == 0x08 {
            this.scratch_bytes_per_pixel = 1
            this.src_bytes_per_pixel = 1
            this.src_pixfmt = base.PIXEL_FORMAT__Y
            this.opaque = true
        } else {
            return "#unsupported TARGA file"
        }
    }

    this.header_image_descriptor = args.src.read_u8?()
    if (this.header_image_descriptor & 0x10) <> 0 {
        // We don't support right-to-left, only left-to-right pixel order.
        return "#unsupported TARGA file"
    }

    // Skip the Image ID.
    args.src.skip_u32?(n: this.header_id_length as base.u32)

    // Read the color map.
    if this.header_color_map_type <> 0 {
        while i < (this.header_color_map_length as base.u32) {
            assert i <= 0xFFFF via "a <= b: a <= c; c <= b"(c: this.header_color_map_length as base.u32)
            if this.header_color_map_entry_size == 0x18 {
                c32 = args.src.read_u24le_as_u32?()
                this.src_palette[((i & 0xFF) * 4) + 0] = ((c32 >> 0x00) & 0xFF) as base.u8
                this.src_palette[((i & 0xFF) * 4) + 1] = ((c32 >> 0x08) & 0xFF) as base.u8
                this.src_palette[((i & 0xFF) * 4) + 2] = ((c32 >> 0x10) & 0xFF) as base.u8
                this.src_palette[((i & 0xFF) * 4) + 3] = 0xFF
            } else if this.header_color_map_entry_size == 0x20 {
                c32 = args.src.read_u32le?()
                this.src_palette[((i & 0xFF) * 4) + 0] = ((c32 >> 0x00) & 0xFF) as base.u8
                this.src_palette[((i & 0xFF) * 4) + 1] = ((c32 >> 0x08) & 0xFF) as base.u8
                this.src_palette[((i & 0xFF) * 4) + 2] = ((c32 >> 0x10) & 0xFF) as base.u8
                this.src_palette[((i & 0xFF) * 4) + 3] = ((c32 >> 0x18) & 0xFF) as base.u8
            } else {
                // Expand 15-bit or 16-bit color map entries.
                c32 = args.src.read_u16le_as_u32?()
                c5 = 0x1F & (c32 >> 0)
                this.src_palette[((i & 0xFF) * 4) + 0] = ((c5 << 3) | (c5 >> 2)) as base.u8
                c5 = 0x1F & (c32 >> 5)
                this.src_palette[((i & 0xFF) * 4) + 1] = ((c5 << 3) | (c5 >> 2)) as base.u8
                c5 = 0x1F & (c32 >> 10)
                this.src_palette[((i & 0xFF) * 4) + 2] = ((c5 << 3) | (c5 >> 2)) as base.u8
                // TODO: can the alpha value be zero (BGRA5551 not BGRX5551)?
                this.src_palette[((i & 0xFF) * 4) + 3] = 0xFF
            }
            i += 1
        }
        while i < 0x100 {
            this.src_palette[(i * 4) + 0] = 0x00
            this.src_palette[(i * 4) + 1] = 0x00
            this.src_palette[(i * 4) + 2] = 0x00
            this.src_palette[(i * 4) + 3] = 0xFF
            i += 1
        }
    }

    this.frame_config_io_position = args.src.position()

    if args.dst <> nullptr {
        args.dst.set!(
                pixfmt: this.src_pixfmt,
                pixsub: 0,
                width: this.width,
                height: this.height,
                first_frame_io_position: this.frame_config_io_position,
                first_frame_is_opaque: this.opaque)
    }

    this.call_sequence = 0x20
}

pub func decoder.decode_frame_config?(dst: nptr base.frame_config, src: base.io_reader) {
    var status : base.status

    while true {
        status =? this.do_decode_frame_config?(dst: args.dst, src: args.src)
        if (status == base."$short read") and args.src.is_closed() {
            return "#truncated input"
        }
        yield? status
    }
}

pri func decoder.do_decode_frame_config?(dst: nptr base.frame_config, src: base.io_reader) {
    var pixfmt : base.pixel_format

    if this.call_sequence == 0x20 {
        // No-op.
    } else if this.call_sequence < 0x20 {
        this.do_decode_image_config?(dst: nullptr, src: args.src)
    } else if this.call_sequence == 0x28 {
        if this.frame_config_io_position <> args.src.position() {
            return base."#bad restart"
        }
    } else if this.call_sequence == 0x40 {
        this.call_sequence = 0x60
        return base."@end of data"
    } else {
        return base."@end of data"
    }

    if args.dst <> nullptr {
        pixfmt = this.util.make_pixel_format(repr: this.src_pixfmt)
        args.dst.set!(bounds: this.util.make_rect_ie_u32(
                min_incl_x: 0,
                min_incl_y: 0,
                max_excl_x: this.width,
                max_excl_y: this.height),
                duration: 0,
                index: 0,
                io_position: this.frame_config_io_position,
                disposal: 0,
                opaque_within_bounds: this.opaque,
                overwrite_instead_of_blend: false,
                background_color: pixfmt.default_background_color())
    }

    this.call_sequence = 0x40
}

pub func decoder.decode_frame?(dst: ptr base.pixel_buffer, src: base.io_reader, blend: base.pixel_blend, workbuf: slice base.u8, opts: nptr base.decode_frame_options) {
    var status : base.status

    while true {
        status =? this.do_decode_frame?(dst: args.dst, src: args.src, blend: args.blend, workbuf: args.workbuf, opts: args.opts)
        if (status == base."$short read") and args.src.is_closed() {
            return "#truncated input"
        }
        yield? status
    }
}

pri func decoder.do_decode_frame?(dst: ptr base.pixel_buffer, src: base.io_reader, blend: base.pixel_blend, workbuf: slice base.u8, opts: nptr base.decode_frame_options) {
    var status              : base.status
    var dst_pixfmt          : base.pixel_format
    var dst_bits_per_pixel  : base.u32[..= 256]
    var dst_bytes_per_pixel : base.u64[..= 32]
    var dst_x               : base.u32
    var dst_y               : base.u32
    var tab                 : table base.u8
    var dst_palette         : slice base.u8
    var dst                 : slice base.u8
    var dst_start           : base.u64
    var src_palette         : roslice base.u8
    var mark                : base.u64
    var num_pixels64        : base.u64
    var num_pixels32        : base.u32[..= 0xFFFF]
    var lit_length          : base.u32[..= 0xFFFF]
    var run_length          : base.u32[..= 0xFFFF]
    var num_dst_bytes       : base.u64[..= 0x1F_FFE0]
    var num_src_bytes       : base.u32[..= 0x3_FFFC]
    var c32                 : base.u32
    var c5                  : base.u32[..= 0x1F]

    if this.call_sequence == 0x40 {
        // No-op.
    } else if this.call_sequence < 0x40 {
        this.do_decode_frame_config?(dst: nullptr, src: args.src)
    } else {
        return base."@end of data"
    }

    if this.header_color_map_type <> 0 {
        src_palette = this.src_palette[..]
    }
    status = this.swizzler.prepare!(
            dst_pixfmt: args.dst.pixel_format(),
            dst_palette: args.dst.palette_or_else(fallback: this.dst_palette[..]),
            src_pixfmt: this.util.make_pixel_format(repr: this.src_pixfmt),
            src_palette: src_palette,
            blend: args.blend)
    if not status.is_ok() {
        return status
    }

    // TODO: the dst_pixfmt variable shouldn't be necessary. We should be able
    // to chain the two calls: "args.dst.pixel_format().bits_per_pixel()".
    dst_pixfmt = args.dst.pixel_format()
    dst_bits_per_pixel = dst_pixfmt.bits_per_pixel()
    if (dst_bits_per_pixel & 7) <> 0 {
        return base."#unsupported option"
    }
    dst_bytes_per_pixel = (dst_bits_per_pixel / 8) as base.u64

    if (this.header_image_descriptor & 0x20) == 0 {  // Bottom-to-top.
        dst_y = this.height ~mod- 1
    }
    if (this.header_image_type & 8) == 0 {
        // No RLE (run length encoding) means that the entire row is
        // effectively literals.
        lit_length = this.width
    }

    while.resume true {
        tab = args.dst.plane(p: 0)
        dst_palette = args.dst.palette_or_else(fallback: this.dst_palette[..])

        while dst_y < this.height {
            dst = tab.row_u32(y: dst_y)
            dst_start = (dst_x as base.u64) * dst_bytes_per_pixel
            if dst_start <= dst.length() {
                dst = dst[dst_start ..]
            } else {
                dst = this.util.empty_slice_u8()
            }

            while dst_x < this.width {
                assert dst_x <= 0xFFFF via "a <= b: a <= c; c <= b"(c: this.width)

                if this.src_bytes_per_pixel > 0 {
                    if lit_length > 0 {
                        mark = args.src.mark()
                        num_pixels64 = args.src.length() / (this.src_bytes_per_pixel as base.u64)
                        num_pixels32 = num_pixels64.min(no_more_than: lit_length as base.u64) as base.u32
                        num_dst_bytes = (num_pixels32 as base.u64) * dst_bytes_per_pixel
                        num_src_bytes = num_pixels32 * this.src_bytes_per_pixel
                        args.src.skip_u32?(n: num_src_bytes)
                        this.swizzler.swizzle_interleaved_from_slice!(
                                dst: dst,
                                dst_palette: dst_palette,
                                src: args.src.since(mark: mark))
                        if num_dst_bytes <= dst.length() {
                            dst = dst[num_dst_bytes ..]
                        } else {
                            dst = this.util.empty_slice_u8()
                        }
                        dst_x += num_pixels32
                        lit_length = (lit_length ~mod- num_pixels32) & 0xFFFF
                        if lit_length > 0 {
                            yield? base."$short read"
                            continue.resume
                        }

                    } else if run_length > 0 {
                        run_length -= 1
                        this.swizzler.swizzle_interleaved_from_slice!(
                                dst: dst,
                                dst_palette: dst_palette,
                                src: this.scratch[.. this.scratch_bytes_per_pixel])
                        if dst_bytes_per_pixel <= dst.length() {
                            dst = dst[dst_bytes_per_pixel ..]
                        }
                        dst_x += 1

                    } else {
                        // Handle Raw vs RLE packets.
                        if args.src.length() <= 0 {
                            yield? base."$short read"
                            continue.resume
                        }
                        if args.src.peek_u8_as_u32() < 0x80 {
                            lit_length = args.src.peek_u8_as_u32() + 1
                            args.src.skip_u32_fast!(actual: 1, worst_case: 1)
                            if (lit_length + dst_x) > this.width {
                                return "#bad run length encoding"
                            }

                        } else {
                            if this.src_bytes_per_pixel == 1 {
                                if args.src.length() < 2 {
                                    yield? base."$short read"
                                    continue.resume
                                }
                                run_length = (args.src.peek_u8_as_u32() & 0x7F) + 1
                                args.src.skip_u32_fast!(actual: 1, worst_case: 1)
                                this.scratch[0] = args.src.peek_u8()
                                args.src.skip_u32_fast!(actual: 1, worst_case: 1)
                            } else if this.src_bytes_per_pixel == 3 {
                                if args.src.length() < 4 {
                                    yield? base."$short read"
                                    continue.resume
                                }
                                run_length = (args.src.peek_u8_as_u32() & 0x7F) + 1
                                args.src.skip_u32_fast!(actual: 1, worst_case: 1)
                                this.scratch[0] = args.src.peek_u8()
                                args.src.skip_u32_fast!(actual: 1, worst_case: 1)
                                this.scratch[1] = args.src.peek_u8()
                                args.src.skip_u32_fast!(actual: 1, worst_case: 1)
                                this.scratch[2] = args.src.peek_u8()
                                args.src.skip_u32_fast!(actual: 1, worst_case: 1)
                            } else {  // this.src_bytes_per_pixel == 4
                                if args.src.length() < 5 {
                                    yield? base."$short read"
                                    continue.resume
                                }
                                run_length = (args.src.peek_u8_as_u32() & 0x7F) + 1
                                args.src.skip_u32_fast!(actual: 1, worst_case: 1)
                                this.scratch[0] = args.src.peek_u8()
                                args.src.skip_u32_fast!(actual: 1, worst_case: 1)
                                this.scratch[1] = args.src.peek_u8()
                                args.src.skip_u32_fast!(actual: 1, worst_case: 1)
                                this.scratch[2] = args.src.peek_u8()
                                args.src.skip_u32_fast!(actual: 1, worst_case: 1)
                                this.scratch[3] = args.src.peek_u8()
                                args.src.skip_u32_fast!(actual: 1, worst_case: 1)
                            }

                            if (run_length + dst_x) > this.width {
                                return "#bad run length encoding"
                            }
                        }
                    }

                } else {
                    // Wuffs' base.pixel_swizzler doesn't support BGRX5551, so
                    // we manually convert to BGRX8888, one pixel at a time.

                    if lit_length > 0 {
                        if args.src.length() < 2 {
                            yield? base."$short read"
                            continue.resume
                        }
                        c32 = args.src.peek_u16le_as_u32()
                        args.src.skip_u32_fast!(actual: 2, worst_case: 2)
                        c5 = 0x1F & (c32 >> 0)
                        this.scratch[0] = ((c5 << 3) | (c5 >> 2)) as base.u8
                        c5 = 0x1F & (c32 >> 5)
                        this.scratch[1] = ((c5 << 3) | (c5 >> 2)) as base.u8
                        c5 = 0x1F & (c32 >> 10)
                        this.scratch[2] = ((c5 << 3) | (c5 >> 2)) as base.u8
                        // TODO: can the alpha value be zero (BGRA5551 not BGRX5551)?
                        this.scratch[3] = 0xFF
                        this.swizzler.swizzle_interleaved_from_slice!(
                                dst: dst,
                                dst_palette: dst_palette,
                                src: this.scratch[.. 4])
                        if dst_bytes_per_pixel <= dst.length() {
                            dst = dst[dst_bytes_per_pixel ..]
                        }
                        dst_x += 1
                        lit_length -= 1

                    } else if run_length > 0 {
                        run_length -= 1
                        this.swizzler.swizzle_interleaved_from_slice!(
                                dst: dst,
                                dst_palette: dst_palette,
                                src: this.scratch[.. this.scratch_bytes_per_pixel])
                        if dst_bytes_per_pixel <= dst.length() {
                            dst = dst[dst_bytes_per_pixel ..]
                        }
                        dst_x += 1

                    } else {
                        // Handle Raw vs RLE packets.
                        if args.src.length() <= 0 {
                            yield? base."$short read"
                            continue.resume
                        }
                        if args.src.peek_u8_as_u32() < 0x80 {
                            lit_length = args.src.peek_u8_as_u32() + 1
                            args.src.skip_u32_fast!(actual: 1, worst_case: 1)
                            if (lit_length + dst_x) > this.width {
                                return "#bad run length encoding"
                            }

                        } else {
                            if args.src.length() < 3 {
                                yield? base."$short read"
                                continue.resume
                            }
                            run_length = (args.src.peek_u8_as_u32() & 0x7F) + 1
                            args.src.skip_u32_fast!(actual: 1, worst_case: 1)
                            c32 = args.src.peek_u16le_as_u32()
                            args.src.skip_u32_fast!(actual: 2, worst_case: 2)
                            c5 = 0x1F & (c32 >> 0)
                            this.scratch[0] = ((c5 << 3) | (c5 >> 2)) as base.u8
                            c5 = 0x1F & (c32 >> 5)
                            this.scratch[1] = ((c5 << 3) | (c5 >> 2)) as base.u8
                            c5 = 0x1F & (c32 >> 10)
                            this.scratch[2] = ((c5 << 3) | (c5 >> 2)) as base.u8
                            // TODO: can the alpha value be zero (BGRA5551 not BGRX5551)?
                            this.scratch[3] = 0xFF

                            if (run_length + dst_x) > this.width {
                                return "#bad run length encoding"
                            }
                        }
                    }
                }
            }
            dst_x = 0

            if (this.header_image_descriptor & 0x20) == 0 {  // Bottom-to-top.
                dst_y ~mod-= 1
            } else {  // Top-to-bottom.
                dst_y ~mod+= 1
            }
            if (this.header_image_type & 8) == 0 {
                // No RLE (run length encoding) means that the entire row is
                // effectively literals.
                lit_length = this.width
            }
        }
        break.resume
    }.resume

    this.call_sequence = 0x60
}

pub func decoder.frame_dirty_rect() base.rect_ie_u32 {
    return this.util.make_rect_ie_u32(
            min_incl_x: 0,
            min_incl_y: 0,
            max_excl_x: this.width,
            max_excl_y: this.height)
}

pub func decoder.num_animation_loops() base.u32 {
    return 0
}

pub func decoder.num_decoded_frame_configs() base.u64 {
    if this.call_sequence > 0x20 {
        return 1
    }
    return 0
}

pub func decoder.num_decoded_frames() base.u64 {
    if this.call_sequence > 0x40 {
        return 1
    }
    return 0
}

pub func decoder.restart_frame!(index: base.u64, io_position: base.u64) base.status {
    if this.call_sequence < 0x20 {
        return base."#bad call sequence"
    }
    if args.index <> 0 {
        return base."#bad argument"
    }
    this.call_sequence = 0x28
    this.frame_config_io_position = args.io_position
    return ok
}

pub func decoder.set_report_metadata!(fourcc: base.u32, report: base.bool) {
    // TODO: implement.
}

pub func decoder.tell_me_more?(dst: base.io_writer, minfo: nptr base.more_information, src: base.io_reader) {
    return base."#no more information"
}

pub func decoder.workbuf_len() base.range_ii_u64 {
    return this.util.make_range_ii_u64(min_incl: 0, max_incl: 0)
}