]> pd.if.org Git - zpackage/blobdiff - lzma/lzma/lzma_encoder_optimum_fast.c
integrate lzma
[zpackage] / lzma / lzma / lzma_encoder_optimum_fast.c
diff --git a/lzma/lzma/lzma_encoder_optimum_fast.c b/lzma/lzma/lzma_encoder_optimum_fast.c
new file mode 100644 (file)
index 0000000..9b30347
--- /dev/null
@@ -0,0 +1,169 @@
+///////////////////////////////////////////////////////////////////////////////
+//
+/// \file       lzma_encoder_optimum_fast.c
+//
+//  Author:     Igor Pavlov
+//
+//  This file has been put into the public domain.
+//  You can do whatever you want with this file.
+//
+///////////////////////////////////////////////////////////////////////////////
+
+#include "lzma_encoder_private.h"
+#include "memcmplen.h"
+
+
+#define change_pair(small_dist, big_dist) \
+       (((big_dist) >> 7) > (small_dist))
+
+
+extern void
+lzma_lzma_optimum_fast(lzma_coder *restrict coder, lzma_mf *restrict mf,
+               uint32_t *restrict back_res, uint32_t *restrict len_res)
+{
+       const uint32_t nice_len = mf->nice_len;
+
+       uint32_t len_main;
+       uint32_t matches_count;
+       if (mf->read_ahead == 0) {
+               len_main = mf_find(mf, &matches_count, coder->matches);
+       } else {
+               assert(mf->read_ahead == 1);
+               len_main = coder->longest_match_length;
+               matches_count = coder->matches_count;
+       }
+
+       const uint8_t *buf = mf_ptr(mf) - 1;
+       const uint32_t buf_avail = my_min(mf_avail(mf) + 1, MATCH_LEN_MAX);
+
+       if (buf_avail < 2) {
+               // There's not enough input left to encode a match.
+               *back_res = UINT32_MAX;
+               *len_res = 1;
+               return;
+       }
+
+       // Look for repeated matches; scan the previous four match distances
+       uint32_t rep_len = 0;
+       uint32_t rep_index = 0;
+
+       for (uint32_t i = 0; i < REPS; ++i) {
+               // Pointer to the beginning of the match candidate
+               const uint8_t *const buf_back = buf - coder->reps[i] - 1;
+
+               // If the first two bytes (2 == MATCH_LEN_MIN) do not match,
+               // this rep is not useful.
+               if (not_equal_16(buf, buf_back))
+                       continue;
+
+               // The first two bytes matched.
+               // Calculate the length of the match.
+               const uint32_t len = lzma_memcmplen(
+                               buf, buf_back, 2, buf_avail);
+
+               // If we have found a repeated match that is at least
+               // nice_len long, return it immediately.
+               if (len >= nice_len) {
+                       *back_res = i;
+                       *len_res = len;
+                       mf_skip(mf, len - 1);
+                       return;
+               }
+
+               if (len > rep_len) {
+                       rep_index = i;
+                       rep_len = len;
+               }
+       }
+
+       // We didn't find a long enough repeated match. Encode it as a normal
+       // match if the match length is at least nice_len.
+       if (len_main >= nice_len) {
+               *back_res = coder->matches[matches_count - 1].dist + REPS;
+               *len_res = len_main;
+               mf_skip(mf, len_main - 1);
+               return;
+       }
+
+       uint32_t back_main = 0;
+       if (len_main >= 2) {
+               back_main = coder->matches[matches_count - 1].dist;
+
+               while (matches_count > 1 && len_main ==
+                               coder->matches[matches_count - 2].len + 1) {
+                       if (!change_pair(coder->matches[
+                                               matches_count - 2].dist,
+                                       back_main))
+                               break;
+
+                       --matches_count;
+                       len_main = coder->matches[matches_count - 1].len;
+                       back_main = coder->matches[matches_count - 1].dist;
+               }
+
+               if (len_main == 2 && back_main >= 0x80)
+                       len_main = 1;
+       }
+
+       if (rep_len >= 2) {
+               if (rep_len + 1 >= len_main
+                               || (rep_len + 2 >= len_main
+                                       && back_main > (UINT32_C(1) << 9))
+                               || (rep_len + 3 >= len_main
+                                       && back_main > (UINT32_C(1) << 15))) {
+                       *back_res = rep_index;
+                       *len_res = rep_len;
+                       mf_skip(mf, rep_len - 1);
+                       return;
+               }
+       }
+
+       if (len_main < 2 || buf_avail <= 2) {
+               *back_res = UINT32_MAX;
+               *len_res = 1;
+               return;
+       }
+
+       // Get the matches for the next byte. If we find a better match,
+       // the current byte is encoded as a literal.
+       coder->longest_match_length = mf_find(mf,
+                       &coder->matches_count, coder->matches);
+
+       if (coder->longest_match_length >= 2) {
+               const uint32_t new_dist = coder->matches[
+                               coder->matches_count - 1].dist;
+
+               if ((coder->longest_match_length >= len_main
+                                       && new_dist < back_main)
+                               || (coder->longest_match_length == len_main + 1
+                                       && !change_pair(back_main, new_dist))
+                               || (coder->longest_match_length > len_main + 1)
+                               || (coder->longest_match_length + 1 >= len_main
+                                       && len_main >= 3
+                                       && change_pair(new_dist, back_main))) {
+                       *back_res = UINT32_MAX;
+                       *len_res = 1;
+                       return;
+               }
+       }
+
+       // In contrast to LZMA SDK, dictionary could not have been moved
+       // between mf_find() calls, thus it is safe to just increment
+       // the old buf pointer instead of recalculating it with mf_ptr().
+       ++buf;
+
+       const uint32_t limit = my_max(2, len_main - 1);
+
+       for (uint32_t i = 0; i < REPS; ++i) {
+               if (memcmp(buf, buf - coder->reps[i] - 1, limit) == 0) {
+                       *back_res = UINT32_MAX;
+                       *len_res = 1;
+                       return;
+               }
+       }
+
+       *back_res = back_main + REPS;
+       *len_res = len_main;
+       mf_skip(mf, len_main - 2);
+       return;
+}