unsigned table_bits;
};
-struct lzms_probabilites {
- struct lzms_probability_entry main[LZMS_NUM_MAIN_PROBS];
- struct lzms_probability_entry match[LZMS_NUM_MATCH_PROBS];
- struct lzms_probability_entry lz[LZMS_NUM_LZ_PROBS];
- struct lzms_probability_entry delta[LZMS_NUM_DELTA_PROBS];
- struct lzms_probability_entry lz_rep[LZMS_NUM_LZ_REP_DECISIONS]
- [LZMS_NUM_LZ_REP_PROBS];
- struct lzms_probability_entry delta_rep[LZMS_NUM_DELTA_REP_DECISIONS]
- [LZMS_NUM_DELTA_REP_PROBS];
-};
-
struct lzms_decompressor {
/* 'last_target_usages' is in union with everything else because it is
/* Load the probability entry corresponding to the current state. */
prob_entry = &probs[*state_p];
+ /* Update the state early. We'll still need to OR the state with 1
+ * later if the decoded bit is a 1. */
+ *state_p = (*state_p << 1) & (num_states - 1);
+
+ /* Get the probability (out of LZMS_PROBABILITY_DENOMINATOR) that the
+ * next bit is 0. */
+ prob = lzms_get_probability(prob_entry);
+
/* Normalize if needed. */
- if (rd->range <= 0xffff) {
+ if (!(rd->range & 0xFFFF0000)) {
rd->range <<= 16;
rd->code <<= 16;
if (likely(rd->next != rd->end))
rd->code |= le16_to_cpu(*rd->next++);
}
- /* Get the probability (out of LZMS_PROBABILITY_DENOMINATOR) that the
- * next bit is 0. */
- prob = lzms_get_probability(prob_entry);
-
/* Based on the probability, calculate the bound between the 0-bit
* region and the 1-bit region of the range. */
bound = (rd->range >> LZMS_PROBABILITY_BITS) * prob;
rd->range = bound;
/* Update the state and probability entry based on the decoded bit. */
- *state_p = ((*state_p << 1) | 0) & (num_states - 1);
lzms_update_probability_entry(prob_entry, 0);
return 0;
} else {
rd->code -= bound;
/* Update the state and probability entry based on the decoded bit. */
- *state_p = ((*state_p << 1) | 1) & (num_states - 1);
lzms_update_probability_entry(prob_entry, 1);
+ *state_p |= 1;
return 1;
}
}
/* LRU queues for match sources */
u32 recent_lz_offsets[LZMS_NUM_LZ_REPS + 1];
u64 recent_delta_pairs[LZMS_NUM_DELTA_REPS + 1];
- u32 pending_lz_offset = 0;
- u64 pending_delta_pair = 0;
- const u8 *lz_offset_still_pending;
- const u8 *delta_pair_still_pending;
+
+ /* Previous item type: 0 = literal, 1 = LZ match, 2 = delta match.
+ * This is used to handle delayed updates of the LRU queues. Instead of
+ * actually delaying the updates, we can check when decoding each rep
+ * match whether a delayed update needs to be taken into account, and if
+ * so get the match source from slot 'rep_idx + 1' instead of from slot
+ * 'rep_idx'. */
+ unsigned prev_item_type = 0;
/* States and probability entries for item type disambiguation */
u32 main_state = 0;
lzms_input_bitstream_init(&is, in, in_nbytes / sizeof(le16));
- lzms_init_probability_entries((struct lzms_probability_entry *)&d->probs,
- sizeof(d->probs) /
- sizeof(struct lzms_probability_entry));
+ lzms_init_probabilities(&d->probs);
lzms_init_huffman_codes(d, lzms_get_num_offset_slots(out_nbytes));
{
/* Literal */
*out_next++ = lzms_decode_literal(d, &is);
+ prev_item_type = 0;
} else if (!lzms_decode_bit(&rd, &match_state,
LZMS_NUM_MATCH_PROBS,
u32 offset;
u32 length;
+ STATIC_ASSERT(LZMS_NUM_LZ_REPS == 3);
+
if (!lzms_decode_bit(&rd, &lz_state,
LZMS_NUM_LZ_PROBS, d->probs.lz))
{
/* Explicit offset */
offset = lzms_decode_lz_offset(d, &is);
+
+ recent_lz_offsets[3] = recent_lz_offsets[2];
+ recent_lz_offsets[2] = recent_lz_offsets[1];
+ recent_lz_offsets[1] = recent_lz_offsets[0];
} else {
/* Repeat offset */
- if (pending_lz_offset != 0 &&
- out_next != lz_offset_still_pending)
- {
- BUILD_BUG_ON(LZMS_NUM_LZ_REPS != 3);
- recent_lz_offsets[3] = recent_lz_offsets[2];
- recent_lz_offsets[2] = recent_lz_offsets[1];
- recent_lz_offsets[1] = recent_lz_offsets[0];
- recent_lz_offsets[0] = pending_lz_offset;
- pending_lz_offset = 0;
- }
-
- BUILD_BUG_ON(LZMS_NUM_LZ_REPS != 3);
if (!lzms_decode_bit(&rd, &lz_rep_states[0],
LZMS_NUM_LZ_REP_PROBS,
d->probs.lz_rep[0]))
{
- offset = recent_lz_offsets[0];
- recent_lz_offsets[0] = recent_lz_offsets[1];
- recent_lz_offsets[1] = recent_lz_offsets[2];
- recent_lz_offsets[2] = recent_lz_offsets[3];
+ offset = recent_lz_offsets[0 + (prev_item_type & 1)];
+ recent_lz_offsets[0 + (prev_item_type & 1)] = recent_lz_offsets[0];
} else if (!lzms_decode_bit(&rd, &lz_rep_states[1],
LZMS_NUM_LZ_REP_PROBS,
d->probs.lz_rep[1]))
{
- offset = recent_lz_offsets[1];
- recent_lz_offsets[1] = recent_lz_offsets[2];
- recent_lz_offsets[2] = recent_lz_offsets[3];
+ offset = recent_lz_offsets[1 + (prev_item_type & 1)];
+ recent_lz_offsets[1 + (prev_item_type & 1)] = recent_lz_offsets[1];
+ recent_lz_offsets[1] = recent_lz_offsets[0];
} else {
- offset = recent_lz_offsets[2];
- recent_lz_offsets[2] = recent_lz_offsets[3];
+ offset = recent_lz_offsets[2 + (prev_item_type & 1)];
+ recent_lz_offsets[2 + (prev_item_type & 1)] = recent_lz_offsets[2];
+ recent_lz_offsets[2] = recent_lz_offsets[1];
+ recent_lz_offsets[1] = recent_lz_offsets[0];
}
}
-
- if (pending_lz_offset != 0) {
- BUILD_BUG_ON(LZMS_NUM_LZ_REPS != 3);
- recent_lz_offsets[3] = recent_lz_offsets[2];
- recent_lz_offsets[2] = recent_lz_offsets[1];
- recent_lz_offsets[1] = recent_lz_offsets[0];
- recent_lz_offsets[0] = pending_lz_offset;
- }
- pending_lz_offset = offset;
+ recent_lz_offsets[0] = offset;
+ prev_item_type = 1;
length = lzms_decode_length(d, &is);
lz_copy(out_next, length, offset, out_end, LZMS_MIN_MATCH_LENGTH);
out_next += length;
-
- lz_offset_still_pending = out_next;
} else {
/* Delta match */
u32 offset;
const u8 *matchptr;
u32 length;
+ u64 pair;
+
+ STATIC_ASSERT(LZMS_NUM_DELTA_REPS == 3);
if (!lzms_decode_bit(&rd, &delta_state,
LZMS_NUM_DELTA_PROBS,
/* Explicit offset */
power = lzms_decode_delta_power(d, &is);
raw_offset = lzms_decode_delta_offset(d, &is);
- } else {
- /* Repeat offset */
- u64 val;
-
- if (pending_delta_pair != 0 &&
- out_next != delta_pair_still_pending)
- {
- BUILD_BUG_ON(LZMS_NUM_DELTA_REPS != 3);
- recent_delta_pairs[3] = recent_delta_pairs[2];
- recent_delta_pairs[2] = recent_delta_pairs[1];
- recent_delta_pairs[1] = recent_delta_pairs[0];
- recent_delta_pairs[0] = pending_delta_pair;
- pending_delta_pair = 0;
- }
- BUILD_BUG_ON(LZMS_NUM_DELTA_REPS != 3);
+ pair = ((u64)power << 32) | raw_offset;
+ recent_delta_pairs[3] = recent_delta_pairs[2];
+ recent_delta_pairs[2] = recent_delta_pairs[1];
+ recent_delta_pairs[1] = recent_delta_pairs[0];
+ } else {
if (!lzms_decode_bit(&rd, &delta_rep_states[0],
LZMS_NUM_DELTA_REP_PROBS,
d->probs.delta_rep[0]))
{
- val = recent_delta_pairs[0];
- recent_delta_pairs[0] = recent_delta_pairs[1];
- recent_delta_pairs[1] = recent_delta_pairs[2];
- recent_delta_pairs[2] = recent_delta_pairs[3];
+ pair = recent_delta_pairs[0 + (prev_item_type >> 1)];
+ recent_delta_pairs[0 + (prev_item_type >> 1)] = recent_delta_pairs[0];
} else if (!lzms_decode_bit(&rd, &delta_rep_states[1],
LZMS_NUM_DELTA_REP_PROBS,
d->probs.delta_rep[1]))
{
- val = recent_delta_pairs[1];
- recent_delta_pairs[1] = recent_delta_pairs[2];
- recent_delta_pairs[2] = recent_delta_pairs[3];
+ pair = recent_delta_pairs[1 + (prev_item_type >> 1)];
+ recent_delta_pairs[1 + (prev_item_type >> 1)] = recent_delta_pairs[1];
+ recent_delta_pairs[1] = recent_delta_pairs[0];
} else {
- val = recent_delta_pairs[2];
- recent_delta_pairs[2] = recent_delta_pairs[3];
+ pair = recent_delta_pairs[2 + (prev_item_type >> 1)];
+ recent_delta_pairs[2 + (prev_item_type >> 1)] = recent_delta_pairs[2];
+ recent_delta_pairs[2] = recent_delta_pairs[1];
+ recent_delta_pairs[1] = recent_delta_pairs[0];
}
- power = val >> 32;
- raw_offset = (u32)val;
- }
- if (pending_delta_pair != 0) {
- BUILD_BUG_ON(LZMS_NUM_DELTA_REPS != 3);
- recent_delta_pairs[3] = recent_delta_pairs[2];
- recent_delta_pairs[2] = recent_delta_pairs[1];
- recent_delta_pairs[1] = recent_delta_pairs[0];
- recent_delta_pairs[0] = pending_delta_pair;
+ power = pair >> 32;
+ raw_offset = (u32)pair;
}
- pending_delta_pair = raw_offset | ((u64)power << 32);
+ recent_delta_pairs[0] = pair;
+ prev_item_type = 2;
length = lzms_decode_length(d, &is);
out_next++;
matchptr++;
} while (--length);
-
- delta_pair_still_pending = out_next;
}
}
git://wimlib.net / wimlib / blobdiff