std/jpeg/decode_mcu_progressive_dc_high_bits.wuffs

// Copyright 2023 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

pri func decoder.decode_mcu_progressive_dc_high_bits!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], my: base.u32[..= 0x1FFF]) base.u32 {
    var ret : base.u32

    var bits   : base.u64
    var n_bits : base.u32

    var csel : base.u8[..= 3]

    var r   : base.io_reader
    var pos : base.u32

    var dc_h       : base.u8[..= 3]
    var dc_symbol  : base.u32[..= 0x0F]
    var dc_ht_fast : base.u32
    var dc_bl      : base.u32
    var dc_code    : base.u32
    var dc_blm1    : base.u32[..= 15]
    var dc_ht_slow : base.u32
    var dc_value   : base.u16
    var dc_extend  : base.u16

    bits = this.bitstream_bits
    n_bits = this.bitstream_n_bits

    if this.bitstream_ri > this.bitstream_wi {
        return 2  // Internal error.
    }
    io_bind (io: r, data: this.bitstream_buffer[this.bitstream_ri .. this.bitstream_wi], history_position: this.bitstream_ri as base.u64) {
        while.goto_done true {{

        while.block this.mcu_current_block < this.mcu_num_blocks {
            assert this.mcu_current_block < 10 via "a < b: a < c; c <= b"(c: this.mcu_num_blocks)

            // Ensure that we have enough bits for this iteration of the
            // while.block loop body. Worst case, there are 64 components and
            // each one needs (16 + 15) bits (round that up to 4 bytes), so we
            // need (64 * 4) = 256 bytes available. 8 more bytes of slack means
            // that we can always call peek_u64be.
            if r.length() < 264 {
                ret = 1  // Request another fill_bitstream call.
                break.goto_done
            }

            while.dc_component true,
                    inv this.mcu_current_block < 10,
            {
                // Load at least 56 bits.
                if r.length() < 8 {
                    ret = 2  // Internal error.
                    break.goto_done
                }
                bits |= r.peek_u64be() >> (n_bits & 63)
                r.skip_u32_fast!(actual: (63 - (n_bits & 63)) >> 3, worst_case: 8)
                n_bits |= 56

                // Read the Huffman-encoded dc_symbol, up to 16 bits long.
                dc_h = this.mcu_blocks_dc_hselector[this.mcu_current_block]
                dc_ht_fast = this.huff_tables_fast[dc_h][bits >> 56] as base.u32
                dc_bl = dc_ht_fast >> 8
                if n_bits >= dc_bl {
                    dc_symbol = 0x0F & dc_ht_fast
                    dc_extend = EXTEND[dc_symbol]
                    bits ~mod<<= (dc_bl & 63)
                    n_bits -= dc_bl
                } else {
                    dc_code = (bits >> 55) as base.u32
                    dc_blm1 = 8
                    bits ~mod<<= 9
                    n_bits ~mod-= 9
                    while true,
                            inv this.mcu_current_block < 10,
                    {
                        dc_ht_slow = this.huff_tables_slow[dc_h][dc_blm1]
                        if dc_code < (dc_ht_slow >> 8) {
                            dc_symbol = 0x0F & (this.huff_tables_symbols[dc_h][0xFF & (dc_code ~mod+ dc_ht_slow)] as base.u32)
                            dc_extend = EXTEND[dc_symbol]
                            break
                        }
                        dc_code = (dc_code ~mod<< 1) | ((bits >> 63) as base.u32)
                        bits ~mod<<= 1
                        n_bits ~mod-= 1
                        dc_blm1 = (dc_blm1 + 1) & 15
                        if dc_blm1 == 0 {
                            dc_symbol = 0
                            dc_extend = EXTEND[dc_symbol]
                            break
                        }
                    } endwhile
                }

                // Process the dc_value in the next dc_symbol (up to 15) bits.
                //
                // The dc_value is shifted by (64 - dc_symbol) in two steps,
                // because we want to shift by 64 (not 0) when dc_symbol is 0.
                dc_value = (((bits >> 32) >> (32 - dc_symbol)) & 0xFFFF) as base.u16
                dc_value ~mod+= dc_extend &
                        (((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF) as base.u16) ^ 0xFFFF)
                bits ~mod<<= dc_symbol
                n_bits ~mod-= dc_symbol
                csel = this.scan_comps_cselector[this.mcu_blocks_sselector[this.mcu_current_block]]
                this.mcu_previous_dc_values[csel] ~mod+= dc_value
                this.mcu_blocks[this.mcu_current_block][0] =
                        this.mcu_previous_dc_values[csel] ~mod<< this.scan_al

                break.dc_component
            } endwhile.dc_component

            assert this.mcu_current_block < 10
            this.mcu_current_block += 1
        } endwhile.block
        this.mcu_current_block = 0

        break.goto_done
        }} endwhile.goto_done

        pos = (r.position() & 0xFFFF_FFFF) as base.u32
        if pos > this.bitstream_wi {
            ret = 2  // Internal error.
        } else {
            assert pos <= 0x800 via "a <= b: a <= c; c <= b"(c: this.bitstream_wi)
            this.bitstream_ri = pos
        }
    }

    this.bitstream_bits = bits
    this.bitstream_n_bits = n_bits
    return ret
}