// Copyright 2017 The Wuffs Authors.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//    https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

pub status "#bad Huffman code (over-subscribed)"
pub status "#bad Huffman code (under-subscribed)"
pub status "#bad Huffman code length count"
pub status "#bad Huffman code length repetition"
pub status "#bad Huffman code"
pub status "#bad Huffman minimum code length"
pub status "#bad block"
pub status "#bad distance"
pub status "#bad distance code count"
pub status "#bad literal/length code count"
pub status "#inconsistent stored block length"
pub status "#missing end-of-block code"
pub status "#no Huffman codes"

pri status "#internal error: inconsistent Huffman decoder state"
pri status "#internal error: inconsistent I/O"
pri status "#internal error: inconsistent distance"
pri status "#internal error: inconsistent n_bits"

// TODO: replace the placeholder 1 value with either 0 or (32768 + 512),
// depending on whether we'll move decoder.history into the workbuf.
pub const decoder_workbuf_len_max_incl_worst_case base.u64 = 1

// The next two tables were created by script/print-deflate-magic-numbers.go.
//
// The u32 values' meanings are the same as the decoder.huffs u32 values. In
// particular, bit 30 indicates a base number + extra bits, bits 23-8 are the
// base number and bits 7-4 are the number of those extra bits.
//
// Some trailing elements are 0x08000000. Bit 27 indicates an invalid value.

pri const lcode_magic_numbers array[32] base.u32 = [
	0x40000000, 0x40000100, 0x40000200, 0x40000300, 0x40000400, 0x40000500, 0x40000600, 0x40000700,
	0x40000810, 0x40000A10, 0x40000C10, 0x40000E10, 0x40001020, 0x40001420, 0x40001820, 0x40001C20,
	0x40002030, 0x40002830, 0x40003030, 0x40003830, 0x40004040, 0x40005040, 0x40006040, 0x40007040,
	0x40008050, 0x4000A050, 0x4000C050, 0x4000E050, 0x4000FF00, 0x08000000, 0x08000000, 0x08000000,
]

pri const dcode_magic_numbers array[32] base.u32 = [
	0x40000000, 0x40000100, 0x40000200, 0x40000300, 0x40000410, 0x40000610, 0x40000820, 0x40000C20,
	0x40001030, 0x40001830, 0x40002040, 0x40003040, 0x40004050, 0x40006050, 0x40008060, 0x4000C060,
	0x40010070, 0x40018070, 0x40020080, 0x40030080, 0x40040090, 0x40060090, 0x400800A0, 0x400C00A0,
	0x401000B0, 0x401800B0, 0x402000C0, 0x403000C0, 0x404000D0, 0x406000D0, 0x08000000, 0x08000000,
]

// huffs_table_size is the smallest power of 2 that is greater than or equal to
// the worst-case size of the Huffman tables. See
// script/print-deflate-huff-table-size.go which calculates that, for a 9-bit
// primary table, the worst-case size is 852 for the Lit/Len table and 592 for
// the Distance table.
pri const huffs_table_size base.u32 = 1024
pri const huffs_table_mask base.u32 = 1023

pub struct decoder? implements base.io_transformer(
	// These fields yield src's bits in Least Significant Bits order.
	bits   : base.u32,
	n_bits : base.u32,

	// history_index indexes the history array, defined below.
	history_index : base.u32,

	// n_huffs_bits is discussed in the huffs field comment.
	n_huffs_bits : array[2] base.u32[..= 9],

	// end_of_block is whether decode_huffman_xxx saw an end-of-block code.
	//
	// TODO: can decode_huffman_xxx signal this in band instead of out of band?
	end_of_block : base.bool,

	util : base.utility,
)(
	// huffs and n_huffs_bits are the lookup tables for Huffman decodings.
	//
	// There are up to 2 Huffman decoders active at any one time. As per this
	// package's README.md:
	//  - huffs[0] is used for clcode and lcode.
	//  - huffs[1] is used for dcode.
	//
	// The initial table key is the low n_huffs_bits of the decoder.bits field.
	// Keys longer than 9 bits require a two step lookup, the first step
	// examines the low 9 bits, the second step examines the remaining bits.
	// Two steps are required at most, as keys are at most 15 bits long.
	//
	// Using decoder.bits's low n_huffs_bits as a table key is valid even if
	// decoder.n_bits is less than n_huffs_bits, because the immediate next
	// step after indexing the table by the key is to compare decoder.n_bits to
	// the table value's number of decoder.bits to consume. If it compares
	// less, then more source bytes are read and the table lookup re-tried.
	//
	// The table value's bits:
	//  - bit         31 indicates a literal.
	//  - bit         30 indicates a base number + extra bits.
	//  - bit         29 indicates end-of-block.
	//  - bit         28 indicates a redirect to another part of the table.
	//  - bit         27 indicates an invalid value.
	//  - bits 24 ..= 26 are zero.
	//  - bits  8 ..= 23 are the redirect offset, literal (in bits [8 ..= 15])
	//                   or base number.
	//  - bits  4 ..=  7 are the redirected table's size in bits or the number
	//                   of extra bits.
	//  - bits  0 ..=  3 are the number of decoder.bits to consume.
	//
	// Exactly one of the eight bits [24 ..= 31] should be set.
	huffs : array[2] array[huffs_table_size] base.u32,

	// history holds up to the last 32KiB of decoded output, if the decoding
	// was incomplete (e.g. due to a short read or write). RFC 1951 (DEFLATE)
	// gives the maximum distance in a length-distance back-reference as 32768,
	// or 0x8000.
	//
	// It is a ringbuffer, so that the most distant byte in the decoding isn't
	// necessarily history[0]. The ringbuffer is full (i.e. it holds 32KiB of
	// history) if and only if history_index >= 0x8000.
	//
	// history[history_index & 0x7FFF] is where the next byte of decoded output
	// will be written.
	history : array[0x8000] base.u8,  // 32 KiB.

	// code_lengths is used to pass out-of-band data to init_huff.
	//
	// code_lengths[args.n_codes0 + i] holds the number of bits in the i'th
	// code.
	code_lengths : array[320] base.u8,
)

pub func decoder.add_history!(hist: slice base.u8) {
	var s            : slice base.u8
	var n_copied     : base.u64
	var already_full : base.u32[..= 0x8000]

	s = args.hist
	if s.length() >= 0x8000 {
		// If s is longer than the ringbuffer, we can ignore the previous value
		// of history_index, as we will overwrite the whole ringbuffer.
		s = s.suffix(up_to: 0x8000)
		this.history[..].copy_from_slice!(s: s)
		this.history_index = 0x8000
	} else {
		// Otherwise, append s to the history ringbuffer starting at the
		// previous history_index (modulo 0x8000).
		n_copied = this.history[this.history_index & 0x7FFF ..].copy_from_slice!(s: s)
		if n_copied < s.length() {
			// a_slice.copy_from(s:b_slice) returns the minimum of the two
			// slice lengths. If that value is less than b_slice.length(), then
			// not all of b_slice was copied.
			//
			// In terms of the history ringbuffer, that means that we have to
			// wrap around and copy the remainder of s over the start of the
			// history ringbuffer.
			s = s[n_copied ..]
			n_copied = this.history[..].copy_from_slice!(s: s)
			// Set history_index (modulo 0x8000) to the length of this
			// remainder. The &0x7FFF is redundant, but proves to the compiler
			// that the conversion to u32 will not overflow. The +0x8000 is to
			// maintain that the history ringbuffer is full if and only if
			// history_index >= 0x8000.
			this.history_index = ((n_copied & 0x7FFF) as base.u32) + 0x8000
		} else {
			// We didn't need to wrap around.
			already_full = 0
			if this.history_index >= 0x8000 {
				already_full = 0x8000
			}
			this.history_index = (this.history_index & 0x7FFF) + ((n_copied & 0x7FFF) as base.u32) + already_full
		}
	}
}

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

pub func decoder.transform_io?(dst: base.io_writer, src: base.io_reader, workbuf: slice base.u8) {
	var mark   : base.u64
	var status : base.status

	while true {
		mark = args.dst.mark()
		status =? this.decode_blocks?(dst: args.dst, src: args.src)
		if not status.is_suspension() {
			return status
		}
		// TODO: should "since" be "since!", as the return value lets you
		// modify the state of args.dst, so future mutations (via the slice)
		// can change the veracity of any args.dst assertions?
		this.add_history!(hist: args.dst.since(mark: mark))
		yield? status
	}
}

pri func decoder.decode_blocks?(dst: base.io_writer, src: base.io_reader) {
	var final  : base.u32
	var b0     : base.u32[..= 255]
	var type   : base.u32
	var status : base.status

	while.outer final == 0 {
		while this.n_bits < 3,
			post this.n_bits >= 3,
		{
			b0 = args.src.read_u8_as_u32?()
			this.bits |= b0 << this.n_bits
			this.n_bits += 8
		}
		final = this.bits & 0x01
		type = (this.bits >> 1) & 0x03
		this.bits >>= 3
		this.n_bits -= 3

		if type == 0 {
			this.decode_uncompressed?(dst: args.dst, src: args.src)
			continue
		} else if type == 1 {
			status = this.init_fixed_huffman!()
			// TODO: "if status.is_error()" is probably more idiomatic, but for
			// some mysterious, idiosyncratic reason, performs noticeably worse
			// for gcc (but not for clang).
			//
			// See git commit 3bf9573 "Work around strange status.is_error
			// performance".
			if not status.is_ok() {
				return status
			}
		} else if type == 2 {
			this.init_dynamic_huffman?(src: args.src)
		} else {
			return "#bad block"
		}

		this.end_of_block = false
		while true {
			status = this.decode_huffman_fast!(dst: args.dst, src: args.src)
			if status.is_error() {
				return status
			}
			if this.end_of_block {
				continue.outer
			}
			this.decode_huffman_slow?(dst: args.dst, src: args.src)
			if this.end_of_block {
				continue.outer
			}
		}
	}
}

// decode_uncompressed decodes an uncompresed block as per the RFC section
// 3.2.4.
pri func decoder.decode_uncompressed?(dst: base.io_writer, src: base.io_reader) {
	var length   : base.u32
	var n_copied : base.u32

	// TODO: make this "if" into a function invariant?
	//
	// Ditto for decode_huffman_slow and decode_huffman_fast.
	if (this.n_bits >= 8) or ((this.bits >> (this.n_bits & 7)) <> 0) {
		return "#internal error: inconsistent n_bits"
	}
	this.n_bits = 0
	this.bits = 0

	length = args.src.read_u32le?()
	if (length.low_bits(n: 16) + length.high_bits(n: 16)) <> 0xFFFF {
		return "#inconsistent stored block length"
	}
	length = length.low_bits(n: 16)
	while true {
		n_copied = args.dst.copy_n32_from_reader!(n: length, r: args.src)
		if length <= n_copied {
			return ok
		}
		length -= n_copied
		if args.dst.available() == 0 {
			yield? base."$short write"
		} else {
			yield? base."$short read"
		}
	}
}

// init_fixed_huffman initializes this.huffs as per the RFC section 3.2.6.
pri func decoder.init_fixed_huffman!() base.status {
	var i      : base.u32
	var status : base.status

	while i < 144 {
		this.code_lengths[i] = 8
		i += 1
	}
	while i < 256 {
		this.code_lengths[i] = 9
		i += 1
	}
	while i < 280 {
		this.code_lengths[i] = 7
		i += 1
	}
	while i < 288 {
		this.code_lengths[i] = 8
		i += 1
	}
	while i < 320 {
		this.code_lengths[i] = 5
		i += 1
	}

	status = this.init_huff!(which: 0, n_codes0: 0, n_codes1: 288, base_symbol: 257)
	if status.is_error() {
		return status
	}
	status = this.init_huff!(which: 1, n_codes0: 288, n_codes1: 320, base_symbol: 0)
	if status.is_error() {
		return status
	}
	return ok
}

// init_dynamic_huffman initializes this.huffs as per the RFC section 3.2.7.
pri func decoder.init_dynamic_huffman?(src: base.io_reader) {
	var bits               : base.u32
	var n_bits             : base.u32
	var b0                 : base.u32[..= 255]
	var n_lit              : base.u32[..= 288]
	var n_dist             : base.u32[..= 32]
	var n_clen             : base.u32[..= 19]
	var i                  : base.u32
	var b1                 : base.u32[..= 255]
	var status             : base.status
	var mask               : base.u32[..= 511]
	var table_entry        : base.u32
	var table_entry_n_bits : base.u32[..= 15]
	var b2                 : base.u32[..= 255]
	var n_extra_bits       : base.u32[..= 7]
	var rep_symbol         : base.u8[..= 15]
	var rep_count          : base.u32
	var b3                 : base.u32[..= 255]

	bits = this.bits
	n_bits = this.n_bits
	while n_bits < 14,
		post n_bits >= 14,
	{
		b0 = args.src.read_u8_as_u32?()
		bits |= b0 << n_bits
		n_bits += 8
	}
	n_lit = bits.low_bits(n: 5) + 257
	if n_lit > 286 {
		return "#bad literal/length code count"
	}
	bits >>= 5
	n_dist = bits.low_bits(n: 5) + 1
	if n_dist > 30 {
		return "#bad distance code count"
	}
	bits >>= 5
	n_clen = bits.low_bits(n: 4) + 4
	bits >>= 4
	n_bits -= 14

	// Read the clcode Huffman table: H-CL.
	i = 0
	while i < n_clen {
		while n_bits < 3,
			inv i < n_clen,
			post n_bits >= 3,
		{
			b1 = args.src.read_u8_as_u32?()
			bits |= b1 << n_bits
			n_bits += 8
		}
		assert i < 19 via "a < b: a < c; c <= b"(c: n_clen)
		this.code_lengths[code_order[i]] = (bits & 0x07) as base.u8
		bits >>= 3
		n_bits -= 3
		i += 1
	}
	while i < 19 {
		this.code_lengths[code_order[i]] = 0
		i += 1
	}
	status = this.init_huff!(which: 0, n_codes0: 0, n_codes1: 19, base_symbol: 0xFFF)
	if status.is_error() {
		return status
	}

	// Decode the code lengths for the next two Huffman tables.
	mask = ((1 as base.u32) << (this.n_huffs_bits[0])) - 1
	i = 0
	while i < (n_lit + n_dist) {
		assert i < (288 + 32) via "a < (b + c): a < (b0 + c0); b0 <= b; c0 <= c"(b0: n_lit, c0: n_dist)

		// Decode a clcode symbol from H-CL.
		while true,
			inv i < 320,
		{
			table_entry = this.huffs[0][bits & mask]
			table_entry_n_bits = table_entry & 15
			if n_bits >= table_entry_n_bits {
				bits >>= table_entry_n_bits
				n_bits -= table_entry_n_bits
				break
			}
			assert n_bits < 15 via "a < b: a < c; c <= b"(c: table_entry_n_bits)
			b2 = args.src.read_u8_as_u32?()
			bits |= b2 << n_bits
			n_bits += 8
		}
		// For H-CL, there should be no redirections and all symbols should be
		// literals.
		if (table_entry >> 24) <> 0x80 {
			return "#internal error: inconsistent Huffman decoder state"
		}
		table_entry = (table_entry >> 8) & 0xFF

		// Write a literal code length.
		if table_entry < 16 {
			this.code_lengths[i] = table_entry as base.u8
			i += 1
			continue
		}

		// Write a repeated code length.
		n_extra_bits = 0
		rep_symbol = 0
		rep_count = 0
		if table_entry == 16 {
			n_extra_bits = 2
			if i <= 0 {
				return "#bad Huffman code length repetition"
			}
			rep_symbol = this.code_lengths[i - 1] & 15
			rep_count = 3
			assert rep_count <= 11
		} else if table_entry == 17 {
			n_extra_bits = 3
			rep_symbol = 0
			rep_count = 3
			assert rep_count <= 11
		} else if table_entry == 18 {
			n_extra_bits = 7
			rep_symbol = 0
			rep_count = 11
			assert rep_count <= 11
		} else {
			return "#internal error: inconsistent Huffman decoder state"
		}
		while n_bits < n_extra_bits,
			inv i < 320,
			inv rep_count <= 11,
			post n_bits >= n_extra_bits,
		{
			assert n_bits < 7 via "a < b: a < c; c <= b"(c: n_extra_bits)
			b3 = args.src.read_u8_as_u32?()
			bits |= b3 << n_bits
			n_bits += 8
		}
		rep_count += bits.low_bits(n: n_extra_bits)
		bits >>= n_extra_bits
		n_bits -= n_extra_bits

		while rep_count > 0 {
			// TODO: hoist this check up one level?
			if i >= (n_lit + n_dist) {
				return "#bad Huffman code length count"
			}
			assert i < (288 + 32) via "a < (b + c): a < (b0 + c0); b0 <= b; c0 <= c"(b0: n_lit, c0: n_dist)
			this.code_lengths[i] = rep_symbol
			i += 1
			rep_count -= 1
		}
	}

	if i <> (n_lit + n_dist) {
		return "#bad Huffman code length count"
	}
	if this.code_lengths[256] == 0 {
		return "#missing end-of-block code"
	}

	status = this.init_huff!(which: 0, n_codes0: 0, n_codes1: n_lit, base_symbol: 257)
	if status.is_error() {
		return status
	}
	status = this.init_huff!(which: 1, n_codes0: n_lit, n_codes1: n_lit + n_dist, base_symbol: 0)
	if status.is_error() {
		return status
	}

	this.bits = bits
	this.n_bits = n_bits
}

// TODO: make named constants for 15, 19, 319, etc.

pri func decoder.init_huff!(which: base.u32[..= 1], n_codes0: base.u32[..= 288], n_codes1: base.u32[..= 320], base_symbol: base.u32) base.status {
	var counts            : array[16] base.u16[..= 320]
	var i                 : base.u32
	var remaining         : base.u32
	var offsets           : array[16] base.u16[..= 320]
	var n_symbols         : base.u32[..= 320]
	var count             : base.u32[..= 320]
	var symbols           : array[320] base.u16[..= 319]
	var min_cl            : base.u32[..= 9]
	var max_cl            : base.u32[..= 15]
	var initial_high_bits : base.u32
	var prev_cl           : base.u32[..= 15]
	var prev_redirect_key : base.u32
	var top               : base.u32[..= huffs_table_size]
	var next_top          : base.u32[..= huffs_table_size]
	var code              : base.u32
	var key               : base.u32
	var value             : base.u32
	var cl                : base.u32[..= 15]
	var redirect_key      : base.u32[..= 511]
	var j                 : base.u32[..= 16]
	var reversed_key      : base.u32[..= 511]
	var symbol            : base.u32[..= 319]
	var high_bits         : base.u32
	var delta             : base.u32

	// For the clcode example in this package's README.md:
	//  - n_codes0 = 0
	//  - n_codes1 = 19
	//  - code_lengths[ 0] = 3
	//  - code_lengths[ 1] = 0
	//  - code_lengths[ 2] = 0
	//  - code_lengths[ 3] = 5
	//  - code_lengths[ 4] = 3
	//  - code_lengths[ 5] = 3
	//  - code_lengths[ 6] = 3
	//  - code_lengths[ 7] = 3
	//  - code_lengths[ 8] = 3
	//  - code_lengths[ 9] = 3
	//  - code_lengths[10] = 0
	//  - code_lengths[11] = 0
	//  - code_lengths[12] = 0
	//  - code_lengths[13] = 0
	//  - code_lengths[14] = 0
	//  - code_lengths[15] = 0
	//  - code_lengths[16] = 0
	//  - code_lengths[17] = 4
	//  - code_lengths[18] = 5

	// Calculate counts.
	//
	// For the clcode example in this package's README.md:
	//  - counts[0] = 9
	//  - counts[1] = 0
	//  - counts[2] = 0
	//  - counts[3] = 7
	//  - counts[4] = 1
	//  - counts[5] = 2
	//  - all other counts elements are 0.
	i = args.n_codes0
	while i < args.n_codes1 {
		assert i < 320 via "a < b: a < c; c <= b"(c: args.n_codes1)
		// TODO: this if should be unnecessary. Have some way to assert that,
		// for all j, counts[j] <= i, and thus counts[j]++ will not overflow.
		if counts[this.code_lengths[i] & 15] >= 320 {
			return "#internal error: inconsistent Huffman decoder state"
		}
		counts[this.code_lengths[i] & 15] += 1
		i += 1
	}
	if ((counts[0] as base.u32) + args.n_codes0) == args.n_codes1 {
		return "#no Huffman codes"
	}

	// Check that the Huffman code completely covers all possible input bits.
	remaining = 1  // There is 1 possible 0-bit code.
	i = 1
	while i <= 15 {
		if remaining > (1 << 30) {
			return "#internal error: inconsistent Huffman decoder state"
		}
		// Each iteration doubles the number of possible remaining codes.
		remaining <<= 1
		if remaining < (counts[i] as base.u32) {
			return "#bad Huffman code (over-subscribed)"
		}
		remaining -= counts[i] as base.u32
		i += 1
	}
	if remaining <> 0 {
		// As a special case, allow a degenerate H-D Huffman table, with only
		// one 1-bit code, for the smallest possible distance.
		if (args.which == 1) and (counts[1] == 1) and
			(this.code_lengths[args.n_codes0] == 1) and
			(((counts[0] as base.u32) + args.n_codes0 + 1) == args.n_codes1) {

			this.n_huffs_bits[1] = 1
			this.huffs[1][0] = dcode_magic_numbers[0] | 1
			this.huffs[1][1] = dcode_magic_numbers[31] | 1
			return ok
		}

		return "#bad Huffman code (under-subscribed)"
	}

	// Calculate offsets and n_symbols.
	//
	// For the clcode example in this package's README.md:
	//  - offsets[0] =  0
	//  - offsets[1] =  0
	//  - offsets[2] =  0
	//  - offsets[3] =  0
	//  - offsets[4] =  7
	//  - offsets[5] =  8
	//  - offsets[6] = 10
	//  - all other offsets elements are 10.
	//  - n_symbols = 10
	i = 1
	while i <= 15 {
		offsets[i] = n_symbols as base.u16
		count = counts[i] as base.u32
		if n_symbols > (320 - count) {
			return "#internal error: inconsistent Huffman decoder state"
		}
		assert (n_symbols + count) <= 320 via "(a + b) <= c: a <= (c - b)"()
		// TODO: change this to n_symbols += count, once the proof engine's
		// bounds checking can handle it.
		n_symbols = n_symbols + count
		i += 1
	}
	if n_symbols > 288 {
		return "#internal error: inconsistent Huffman decoder state"
	}

	// Calculate symbols.
	//
	// For the clcode example in this package's README.md:
	//  - symbols[0] =  0
	//  - symbols[1] =  4
	//  - symbols[2] =  5
	//  - symbols[3] =  6
	//  - symbols[4] =  7
	//  - symbols[5] =  8
	//  - symbols[6] =  9
	//  - symbols[7] = 17
	//  - symbols[8] =  3
	//  - symbols[9] = 18
	//
	// As a (local variable) side effect, offsets' values will be updated:
	//  - offsets[3] =  7, formerly 0
	//  - offsets[4] =  8, formerly 7
	//  - offsets[5] = 10, formerly 8
	i = args.n_codes0
	while i < args.n_codes1,
		inv n_symbols <= 288,
	{
		assert i < 320 via "a < b: a < c; c <= b"(c: args.n_codes1)
		// TODO: this if check should be unnecessary.
		if i < args.n_codes0 {
			return "#internal error: inconsistent Huffman decoder state"
		}
		if this.code_lengths[i] <> 0 {
			if offsets[this.code_lengths[i] & 15] >= 320 {
				return "#internal error: inconsistent Huffman decoder state"
			}
			symbols[offsets[this.code_lengths[i] & 15]] = (i - args.n_codes0) as base.u16
			offsets[this.code_lengths[i] & 15] += 1
		}
		i += 1
	}

	// Calculate min_cl and max_cl.
	//
	// For the clcode example in this package's README.md:
	//  - min_cl = 3
	//  - max_cl = 5
	min_cl = 1
	while true,
		inv n_symbols <= 288,
	{
		if counts[min_cl] <> 0 {
			break
		}
		if min_cl >= 9 {
			return "#bad Huffman minimum code length"
		}
		min_cl += 1
	}
	max_cl = 15
	while true,
		inv n_symbols <= 288,
	{
		if counts[max_cl] <> 0 {
			break
		}
		if max_cl <= 1 {
			return "#no Huffman codes"
		}
		max_cl -= 1
	}
	if max_cl <= 9 {
		this.n_huffs_bits[args.which] = max_cl
	} else {
		this.n_huffs_bits[args.which] = 9
	}

	// Calculate this.huffs[args.which].
	//
	// For the clcode example in this package's README.md:
	//  - this.huffs[0][0b..000] = 0x80000003 (literal, symbols[0]=0x00, code_length=3)
	//  - this.huffs[0][0b..100] = 0x80000403 (literal, symbols[1]=0x04, code_length=3)
	//  - this.huffs[0][0b..010] = 0x80000503 (literal, symbols[2]=0x05, code_length=3)
	//  - this.huffs[0][0b..110] = 0x80000603 (literal, symbols[3]=0x06, code_length=3)
	//  - this.huffs[0][0b..001] = 0x80000703 (literal, symbols[4]=0x07, code_length=3)
	//  - this.huffs[0][0b..101] = 0x80000803 (literal, symbols[5]=0x08, code_length=3)
	//  - this.huffs[0][0b..011] = 0x80000903 (literal, symbols[6]=0x09, code_length=3)
	//  - this.huffs[0][0b.0111] = 0x80001104 (literal, symbols[7]=0x11, code_length=4)
	//  - this.huffs[0][0b01111] = 0x80000305 (literal, symbols[8]=0x03, code_length=5)
	//  - this.huffs[0][0b11111] = 0x80001805 (literal, symbols[9]=0x18, code_length=5)
	i = 0
	if (n_symbols <> (offsets[max_cl] as base.u32)) or (n_symbols <> (offsets[15] as base.u32)) {
		return "#internal error: inconsistent Huffman decoder state"
	}
	if (args.n_codes0 + (symbols[0] as base.u32)) >= 320 {
		return "#internal error: inconsistent Huffman decoder state"
	}

	initial_high_bits = 1 << 9
	if max_cl < 9 {
		initial_high_bits = (1 as base.u32) << max_cl
	}
	prev_cl = (this.code_lengths[args.n_codes0 + (symbols[0] as base.u32)] & 15) as base.u32
	prev_redirect_key = 0xFFFFFFFF
	top = 0
	next_top = 512
	code = 0
	key = 0
	value = 0
	while true,
		pre code < (1 << 15),
		pre i < 288,
		inv n_symbols <= 288,
	{
		if (args.n_codes0 + (symbols[i] as base.u32)) >= 320 {
			return "#internal error: inconsistent Huffman decoder state"
		}
		cl = (this.code_lengths[args.n_codes0 + (symbols[i] as base.u32)] & 15) as base.u32
		if cl > prev_cl {
			code <<= cl - prev_cl
			if code >= (1 << 15) {
				return "#internal error: inconsistent Huffman decoder state"
			}
		}
		// For the remainder of this loop body, prev_cl is the original code
		// length, cl is possibly clipped by 9, if in the 2nd-level table.
		prev_cl = cl

		key = code
		if cl > 9 {
			cl -= 9
			assert cl <= 9

			redirect_key = (key >> cl) & 511
			key = key.low_bits(n: cl)
			if prev_redirect_key <> redirect_key {
				prev_redirect_key = redirect_key

				// Calculate the number of bits needed for the 2nd level table.
				// This computation is similar to "check that the Huffman code
				// completely covers all possible input bits" above.
				remaining = (1 as base.u32) << cl
				j = prev_cl
				while j <= 15,
					inv cl <= 9,
					inv code < (1 << 15),
					inv i < 288,
					inv n_symbols <= 288,
				{
					if remaining <= (counts[j] as base.u32) {
						break
					}
					remaining -= counts[j] as base.u32
					if remaining > (1 << 30) {
						return "#internal error: inconsistent Huffman decoder state"
					}
					remaining <<= 1
					j += 1
				}
				if (j <= 9) or (15 < j) {
					return "#internal error: inconsistent Huffman decoder state"
				}
				j -= 9
				initial_high_bits = (1 as base.u32) << j

				top = next_top
				if (top + ((1 as base.u32) << j)) > huffs_table_size {
					return "#internal error: inconsistent Huffman decoder state"
				}
				assert (top + ((1 as base.u32) << j)) <= 1024 via "a <= b: a <= c; c <= b"(c: huffs_table_size)
				next_top = top + ((1 as base.u32) << j)

				redirect_key = (reverse8[redirect_key >> 1] as base.u32) | ((redirect_key & 1) << 8)
				this.huffs[args.which][redirect_key] = 0x10000009 | (top << 8) | (j << 4)
			}
		}
		if (key >= (1 << 9)) or (counts[prev_cl] <= 0) {
			return "#internal error: inconsistent Huffman decoder state"
		}
		counts[prev_cl] -= 1

		reversed_key = (reverse8[key >> 1] as base.u32) | ((key & 1) << 8)
		reversed_key >>= 9 - cl

		symbol = symbols[i] as base.u32
		if symbol == 256 {
			// End-of-block.
			value = 0x20000000 | cl
		} else if (symbol < 256) and (args.which == 0) {
			// Literal.
			value = 0x80000000 | (symbol << 8) | cl
		} else if symbol >= args.base_symbol {
			// Base number + extra bits.
			symbol -= args.base_symbol
			if args.which == 0 {
				value = lcode_magic_numbers[symbol & 31] | cl
			} else {
				value = dcode_magic_numbers[symbol & 31] | cl
			}
		} else {
			return "#internal error: inconsistent Huffman decoder state"
		}

		// The table uses log2(initial_high_bits) bits, but reversed_key only
		// has cl bits. We duplicate the key-value pair across all possible
		// values of the high (log2(initial_high_bits) - cl) bits.
		high_bits = initial_high_bits
		delta = (1 as base.u32) << cl
		while high_bits >= delta,
			inv code < (1 << 15),
			inv i < 288,
			inv n_symbols <= 288,
		{
			high_bits -= delta
			if (top + ((high_bits | reversed_key) & 511)) >= huffs_table_size {
				return "#internal error: inconsistent Huffman decoder state"
			}
			assert (top + ((high_bits | reversed_key) & 511)) < 1024 via "a < b: a < c; c <= b"(c: huffs_table_size)
			this.huffs[args.which][top + ((high_bits | reversed_key) & 511)] = value
		}

		i += 1
		if i >= n_symbols {
			break
		}
		assert i < 288 via "a < b: a < c; c <= b"(c: n_symbols)
		code += 1
		if code >= (1 << 15) {
			return "#internal error: inconsistent Huffman decoder state"
		}
	}
	return ok
}