* about dealing with the unaligned case. */
#define DECODE_TABLE_ALIGNMENT 16
-/* Maximum supported symbol count for make_huffman_decode_table().
- *
- * Reason: In direct mapping entries, we store the symbol in 11 bits. */
-#define DECODE_TABLE_MAX_SYMBOLS 2048
+#define DECODE_TABLE_SYMBOL_SHIFT 4
+#define DECODE_TABLE_LENGTH_MASK DECODE_TABLE_MAX_LENGTH
-/* Maximum supported table bits for make_huffman_decode_table().
- *
- * Reason: In internal binary tree nodes, offsets are encoded in 14 bits.
- * But the real limit is 13, because we allocate entries past the end of
- * the direct lookup part of the table for binary tree nodes. (Note: if
- * needed this limit could be removed by encoding the offsets relative to
- * &decode_table[1 << table_bits].) */
-#define DECODE_TABLE_MAX_TABLE_BITS 13
-
-/* Maximum supported codeword length for make_huffman_decode_table().
+#define DECODE_TABLE_MAX_NUM_SYMS ((1 << (16 - DECODE_TABLE_SYMBOL_SHIFT)) - 1)
+#define DECODE_TABLE_MAX_LENGTH ((1 << DECODE_TABLE_SYMBOL_SHIFT) - 1)
+
+/*
+ * Reads and returns the next Huffman-encoded symbol from a bitstream.
*
- * Reason: In direct mapping entries, we encode the codeword length in 5
- * bits, and the top 2 bits can't both be set because that has special
- * meaning. */
-#define DECODE_TABLE_MAX_CODEWORD_LEN 23
-
-/* Reads and returns the next Huffman-encoded symbol from a bitstream. If the
- * input data is exhausted, the Huffman symbol is decoded as if the missing bits
- * are all zeroes.
+ * If the input data is exhausted, the Huffman symbol is decoded as if the
+ * missing bits are all zeroes.
*
* XXX: This is mostly duplicated in lzms_decode_huffman_symbol() in
- * lzms_decompress.c. */
+ * lzms_decompress.c.
+ */
static inline unsigned
-read_huffsym(struct input_bitstream *istream, const u16 decode_table[],
+read_huffsym(struct input_bitstream *is, const u16 decode_table[],
unsigned table_bits, unsigned max_codeword_len)
{
unsigned entry;
- unsigned key_bits;
-
- bitstream_ensure_bits(istream, max_codeword_len);
-
- /* Index the decode table by the next table_bits bits of the input. */
- key_bits = bitstream_peek_bits(istream, table_bits);
- entry = decode_table[key_bits];
- if (likely(entry < 0xC000)) {
- /* Fast case: The decode table directly provided the
- * symbol and codeword length. The low 11 bits are the
- * symbol, and the high 5 bits are the codeword length. */
- bitstream_remove_bits(istream, entry >> 11);
- return entry & 0x7FF;
- } else {
- /* Slow case: The codeword for the symbol is longer than
- * table_bits, so the symbol does not have an entry
- * directly in the first (1 << table_bits) entries of the
- * decode table. Traverse the appropriate binary tree
- * bit-by-bit to decode the symbol. */
- bitstream_remove_bits(istream, table_bits);
- do {
- key_bits = (entry & 0x3FFF) + bitstream_pop_bits(istream, 1);
- } while ((entry = decode_table[key_bits]) >= 0xC000);
- return entry;
+ unsigned sym;
+ unsigned len;
+
+ /* If the bitbuffer contains fewer bits than might be required by a
+ * single codeword, then refill it. */
+ bitstream_ensure_bits(is, max_codeword_len);
+
+ /* Index the root table by the next 'table_bits' bits of input. */
+ entry = decode_table[bitstream_peek_bits(is, table_bits)];
+
+ /* Extract the symbol and length from the entry. */
+ sym = entry >> DECODE_TABLE_SYMBOL_SHIFT;
+ len = entry & DECODE_TABLE_LENGTH_MASK;
+
+ /* If the root table is indexed by the full 'max_codeword_len' bits,
+ * then there cannot be any subtables. This will be known at compile
+ * time. Otherwise, we must check whether the decoded symbol is really
+ * a subtable pointer. If so, we must discard the bits with which the
+ * root table was indexed, then index the subtable by the next 'len'
+ * bits of input to get the real entry. */
+ if (max_codeword_len > table_bits &&
+ entry >= (1U << (table_bits + DECODE_TABLE_SYMBOL_SHIFT)))
+ {
+ /* Subtable required */
+ bitstream_remove_bits(is, table_bits);
+ entry = decode_table[sym + bitstream_peek_bits(is, len)];
+ sym = entry >> DECODE_TABLE_SYMBOL_SHIFT;;
+ len = entry & DECODE_TABLE_LENGTH_MASK;
}
+
+ /* Discard the bits (or the remaining bits, if a subtable was required)
+ * of the codeword. */
+ bitstream_remove_bits(is, len);
+
+ /* Return the decoded symbol. */
+ return sym;
}
+/*
+ * The ENOUGH() macro returns the maximum number of decode table entries,
+ * including all subtable entries, that may be required for decoding a given
+ * Huffman code. This depends on three parameters:
+ *
+ * num_syms: the maximum number of symbols in the code
+ * table_bits: the number of bits with which the root table will be indexed
+ * max_codeword_len: the maximum allowed codeword length
+ *
+ * Given these parameters, the utility program 'enough' from zlib, when run as
+ * './enough num_syms table_bits max_codeword_len', will compute the maximum
+ * number of entries required. This has already been done for the combinations
+ * we need (or may need) and incorporated into the macro below so that the
+ * mapping can be done at compilation time. If an unknown combination is used,
+ * then a compilation error will result. To fix this, use 'enough' to find the
+ * missing value and add it below.
+ */
+#define ENOUGH(num_syms, table_bits, max_codeword_len) ( \
+ ((num_syms) == 8 && (table_bits) == 7 && (max_codeword_len) == 15) ? 128 : \
+ ((num_syms) == 8 && (table_bits) == 5 && (max_codeword_len) == 7) ? 36 : \
+ ((num_syms) == 8 && (table_bits) == 6 && (max_codeword_len) == 7) ? 66 : \
+ ((num_syms) == 8 && (table_bits) == 7 && (max_codeword_len) == 7) ? 128 : \
+ ((num_syms) == 20 && (table_bits) == 5 && (max_codeword_len) == 15) ? 1062 : \
+ ((num_syms) == 20 && (table_bits) == 6 && (max_codeword_len) == 15) ? 582 : \
+ ((num_syms) == 20 && (table_bits) == 7 && (max_codeword_len) == 15) ? 390 : \
+ ((num_syms) == 54 && (table_bits) == 9 && (max_codeword_len) == 15) ? 618 : \
+ ((num_syms) == 54 && (table_bits) == 10 && (max_codeword_len) == 15) ? 1098 : \
+ ((num_syms) == 249 && (table_bits) == 9 && (max_codeword_len) == 16) ? 878 : \
+ ((num_syms) == 249 && (table_bits) == 10 && (max_codeword_len) == 16) ? 1326 : \
+ ((num_syms) == 249 && (table_bits) == 11 && (max_codeword_len) == 16) ? 2318 : \
+ ((num_syms) == 256 && (table_bits) == 9 && (max_codeword_len) == 15) ? 822 : \
+ ((num_syms) == 256 && (table_bits) == 10 && (max_codeword_len) == 15) ? 1302 : \
+ ((num_syms) == 256 && (table_bits) == 11 && (max_codeword_len) == 15) ? 2310 : \
+ ((num_syms) == 512 && (table_bits) == 10 && (max_codeword_len) == 15) ? 1558 : \
+ ((num_syms) == 512 && (table_bits) == 11 && (max_codeword_len) == 15) ? 2566 : \
+ ((num_syms) == 512 && (table_bits) == 12 && (max_codeword_len) == 15) ? 4606 : \
+ ((num_syms) == 656 && (table_bits) == 10 && (max_codeword_len) == 16) ? 1734 : \
+ ((num_syms) == 656 && (table_bits) == 11 && (max_codeword_len) == 16) ? 2726 : \
+ ((num_syms) == 656 && (table_bits) == 12 && (max_codeword_len) == 16) ? 4758 : \
+ ((num_syms) == 799 && (table_bits) == 9 && (max_codeword_len) == 15) ? 1366 : \
+ ((num_syms) == 799 && (table_bits) == 10 && (max_codeword_len) == 15) ? 1846 : \
+ ((num_syms) == 799 && (table_bits) == 11 && (max_codeword_len) == 15) ? 2854 : \
+ -1)
+
+/* Wrapper around ENOUGH() that does additional compile-time validation. */
+#define DECODE_TABLE_LENGTH(num_syms, table_bits, max_codeword_len) ( \
+ \
+ /* Every possible symbol value must fit into the symbol portion \
+ * of a decode table entry. */ \
+ STATIC_ASSERT_ZERO((num_syms) <= DECODE_TABLE_MAX_NUM_SYMS) + \
+ \
+ /* There cannot be more symbols than possible codewords. */ \
+ STATIC_ASSERT_ZERO((num_syms) <= 1U << (max_codeword_len)) + \
+ \
+ /* It doesn't make sense to use a table_bits more than the \
+ * maximum codeword length. */ \
+ STATIC_ASSERT_ZERO((max_codeword_len) >= (table_bits)) + \
+ \
+ /* The maximum length in the root table must fit into the \
+ * length portion of a decode table entry. */ \
+ STATIC_ASSERT_ZERO((table_bits) <= DECODE_TABLE_MAX_LENGTH) + \
+ \
+ /* The maximum length in a subtable must fit into the length
+ * portion of a decode table entry. */ \
+ STATIC_ASSERT_ZERO((max_codeword_len) - (table_bits) <= \
+ DECODE_TABLE_MAX_LENGTH) + \
+ \
+ /* The needed 'enough' value must have been defined. */ \
+ STATIC_ASSERT_ZERO(ENOUGH((num_syms), (table_bits), \
+ (max_codeword_len)) >= 0) + \
+ \
+ /* The maximum subtable index must fit in the field which would \
+ * normally hold a symbol value. */ \
+ STATIC_ASSERT_ZERO(ENOUGH((num_syms), (table_bits), \
+ (max_codeword_len)) <= \
+ DECODE_TABLE_MAX_NUM_SYMS) + \
+ \
+ /* The minimum subtable index must be greater than the greatest \
+ * possible symbol value. */ \
+ STATIC_ASSERT_ZERO((1U << table_bits) >= num_syms) + \
+ \
+ ENOUGH(num_syms, table_bits, max_codeword_len) \
+)
+
+/*
+ * Declare the decode table for a Huffman code, given several compile-time
+ * constants that describe that code (see ENOUGH() for details).
+ *
+ * Decode tables must be aligned to a DECODE_TABLE_ALIGNMENT-boundary. This
+ * implies that if a decode table is nested a dynamically allocated structure,
+ * then the outer structure must be allocated on a DECODE_TABLE_ALIGNMENT-byte
+ * boundary as well.
+ */
+#define DECODE_TABLE(name, num_syms, table_bits, max_codeword_len) \
+ u16 name[DECODE_TABLE_LENGTH((num_syms), (table_bits), \
+ (max_codeword_len))] \
+ _aligned_attribute(DECODE_TABLE_ALIGNMENT)
+
extern int
make_huffman_decode_table(u16 decode_table[], unsigned num_syms,
- unsigned num_bits, const u8 lens[],
+ unsigned table_bits, const u8 lens[],
unsigned max_codeword_len);
static inline void
}
static inline machine_word_t
-repeat_byte(u8 b)
+repeat_u16(u16 b)
{
- machine_word_t v;
+ machine_word_t v = b;
STATIC_ASSERT(WORDBITS == 32 || WORDBITS == 64);
-
- v = b;
- v |= v << 8;
v |= v << 16;
v |= v << ((WORDBITS == 64) ? 32 : 0);
return v;
}
+static inline machine_word_t
+repeat_u8(u8 b)
+{
+ return repeat_u16(((u16)b << 8) | b);
+}
+
/*
* Copy an LZ77 match at (dst - offset) to dst.
*
* encoding of the previous byte. This case is common
* if the data contains many repeated bytes. */
- machine_word_t v = repeat_byte(*(dst - 1));
+ machine_word_t v = repeat_u8(*(dst - 1));
do {
store_word_unaligned(v, dst);
src += WORDBYTES;