diff --git a/src/matchfinder.c b/src/matchfinder.c
index df4d828..6c33385 100644
--- a/src/matchfinder.c
+++ b/src/matchfinder.c
@@ -195,10 +195,11 @@ int lzsa_build_suffix_array(lzsa_compressor *pCompressor, const unsigned char *p
  * @param nOffset offset to find matches at, in the input window
  * @param pMatches pointer to returned matches
  * @param nMaxMatches maximum number of matches to return (0 for none)
+ * @param nInWindowSize total input size in bytes (previously compressed bytes + bytes to compress)
  *
  * @return number of matches
  */
-int lzsa_find_matches_at(lzsa_compressor *pCompressor, const int nOffset, lzsa_match *pMatches, const int nMaxMatches) {
+int lzsa_find_matches_at(lzsa_compressor *pCompressor, const int nOffset, lzsa_match *pMatches, const int nMaxMatches, const int nInWindowSize) {
    unsigned int *intervals = pCompressor->intervals;
    unsigned int *pos_data = pCompressor->pos_data;
    unsigned int ref;
@@ -238,6 +239,24 @@ int lzsa_find_matches_at(lzsa_compressor *pCompressor, const int nOffset, lzsa_m
    /* Ascend indirectly via pos_data[] links.  */
    match_pos = super_ref & EXCL_VISITED_MASK;
    matchptr = pMatches;
+
+   if (pCompressor->format_version >= 2 && nInWindowSize < 65536) {
+      if ((matchptr - pMatches) < nMaxMatches) {
+         int nMatchOffset = (int)(nOffset - match_pos);
+
+         if (nMatchOffset <= MAX_OFFSET) {
+            if (pCompressor->format_version >= 2) {
+               matchptr->length = (unsigned short)(ref >> (LCP_SHIFT + TAG_BITS));
+            }
+            else {
+               matchptr->length = (unsigned short)(ref >> LCP_SHIFT);
+            }
+            matchptr->offset = (unsigned short)nMatchOffset;
+            matchptr++;
+         }
+      }
+   }
+
    for (;;) {
       while ((super_ref = pos_data[match_pos]) > ref)
          match_pos = intervals[super_ref & POS_MASK] & EXCL_VISITED_MASK;
@@ -282,7 +301,7 @@ void lzsa_skip_matches(lzsa_compressor *pCompressor, const int nStartOffset, con
    /* Skipping still requires scanning for matches, as this also performs a lazy update of the intervals. However,
     * we don't store the matches. */
    for (i = nStartOffset; i < nEndOffset; i++) {
-      lzsa_find_matches_at(pCompressor, i, &match, 0);
+      lzsa_find_matches_at(pCompressor, i, &match, 0, 0);
    }
 }
 
@@ -299,7 +318,7 @@ void lzsa_find_all_matches(lzsa_compressor *pCompressor, const int nMatchesPerOf
    int i;
 
    for (i = nStartOffset; i < nEndOffset; i++) {
-      int nMatches = lzsa_find_matches_at(pCompressor, i, pMatch, nMatchesPerOffset);
+      int nMatches = lzsa_find_matches_at(pCompressor, i, pMatch, nMatchesPerOffset, nEndOffset - nStartOffset);
 
       while (nMatches < nMatchesPerOffset) {
          pMatch[nMatches].length = 0;
diff --git a/src/matchfinder.h b/src/matchfinder.h
index 3526948..cc3b5ab 100644
--- a/src/matchfinder.h
+++ b/src/matchfinder.h
@@ -59,10 +59,11 @@ int lzsa_build_suffix_array(lzsa_compressor *pCompressor, const unsigned char *p
  * @param nOffset offset to find matches at, in the input window
  * @param pMatches pointer to returned matches
  * @param nMaxMatches maximum number of matches to return (0 for none)
+ * @param nInWindowSize total input size in bytes (previously compressed bytes + bytes to compress)
  *
  * @return number of matches
  */
-int lzsa_find_matches_at(lzsa_compressor *pCompressor, const int nOffset, lzsa_match *pMatches, const int nMaxMatches);
+int lzsa_find_matches_at(lzsa_compressor *pCompressor, const int nOffset, lzsa_match *pMatches, const int nMaxMatches, const int nInWindowSize);
 
 /**
  * Skip previously compressed bytes
diff --git a/src/shrink_block_v2.c b/src/shrink_block_v2.c
index 543eee5..9c2c28f 100644
--- a/src/shrink_block_v2.c
+++ b/src/shrink_block_v2.c
@@ -303,9 +303,16 @@ static void lzsa_optimize_forward_v2(lzsa_compressor *pCompressor, const unsigne
                      (nCodingChoiceCost == pDestArrival->cost && nScore < (pDestArrival->score + nDisableScore))) {
 
                      if (pDestArrival->from_slot) {
+                        int z;
+
+                        for (z = n; z < nMatchesPerArrival - 1; z++) {
+                           if (pDestSlots[z].rep_offset == arrival[(i << MATCHES_PER_ARRIVAL_SHIFT) + j].rep_offset)
+                              break;
+                        }
+
                         memmove(&pDestSlots[n + 1],
                            &pDestSlots[n],
-                           sizeof(lzsa_arrival) * (nMatchesPerArrival - n - 1));
+                           sizeof(lzsa_arrival) * (z - n));
                      }
 
                      pDestArrival->cost = nCodingChoiceCost;
@@ -400,15 +407,25 @@ static void lzsa_optimize_forward_v2(lzsa_compressor *pCompressor, const unsigne
                         }
 
                         if (!exists) {
-                           for (n = 0; n < nMatchesPerArrival; n++) {
+                           for (n = 0; n < nMatchesPerArrival - 1; n++) {
                               lzsa_arrival *pDestArrival = &pDestSlots[n];
 
                               if (nCodingChoiceCost < pDestArrival->cost ||
                                  (nCodingChoiceCost == pDestArrival->cost && nScore < (pDestArrival->score + nDisableScore))) {
                                  if (pDestArrival->from_slot) {
+                                    int z;
+
+                                    for (z = n; z < nMatchesPerArrival - 1; z++) {
+                                       if (pDestSlots[z].rep_offset == nMatchOffset)
+                                          break;
+                                    }
+
+                                    if (z == (nMatchesPerArrival - 1) && pDestSlots[z].from_slot && pDestSlots[z].match_len < MIN_MATCH_SIZE_V2)
+                                       z--;
+
                                     memmove(&pDestSlots[n + 1],
                                        &pDestSlots[n],
-                                       sizeof(lzsa_arrival) * (nMatchesPerArrival - n - 1));
+                                       sizeof(lzsa_arrival) * (z - n));
                                  }
 
                                  pDestArrival->cost = nCodingChoiceCost;
@@ -456,9 +473,16 @@ static void lzsa_optimize_forward_v2(lzsa_compressor *pCompressor, const unsigne
                            if (nRepCodingChoiceCost < pDestArrival->cost ||
                               (nRepCodingChoiceCost == pDestArrival->cost && nScore < (pDestArrival->score + nDisableScore))) {
                               if (pDestArrival->from_slot) {
+                                 int z;
+
+                                 for (z = n; z < nMatchesPerArrival - 1; z++) {
+                                    if (pDestSlots[z].rep_offset == nRepOffset)
+                                       break;
+                                 }
+
                                  memmove(&pDestSlots[n + 1],
                                     &pDestSlots[n],
-                                    sizeof(lzsa_arrival) * (nMatchesPerArrival - n - 1));
+                                    sizeof(lzsa_arrival) * (z - n));
                               }
 
                               pDestArrival->cost = nRepCodingChoiceCost;
@@ -1078,8 +1102,6 @@ int lzsa_optimize_and_write_block_v2(lzsa_compressor *pCompressor, const unsigne
    
    memset(pCompressor->best_match, 0, BLOCK_SIZE * sizeof(lzsa_match));
    lzsa_optimize_forward_v2(pCompressor, pInWindow, pCompressor->best_match - nPreviousBlockSize, nPreviousBlockSize, nPreviousBlockSize + nInDataSize, 0 /* reduce */, (nInDataSize < 65536) ? 1 : 0 /* insert forward reps */, nMatchesPerArrival);
-   if (nInDataSize < 65536)
-      lzsa_optimize_forward_v2(pCompressor, pInWindow, pCompressor->best_match - nPreviousBlockSize, nPreviousBlockSize, nPreviousBlockSize + nInDataSize, 0 /* reduce */, 1 /* insert forward reps */, nMatchesPerArrival);
 
    int nDidReduce;
    int nPasses = 0;
diff --git a/src/shrink_context.c b/src/shrink_context.c
index ff187c1..c1e7ab3 100644
--- a/src/shrink_context.c
+++ b/src/shrink_context.c
@@ -89,7 +89,7 @@ int lzsa_compressor_init(lzsa_compressor *pCompressor, const int nMaxWindowSize,
             pCompressor->open_intervals = (unsigned int *)malloc((LCP_AND_TAG_MAX + 1) * sizeof(unsigned int));
 
             if (pCompressor->open_intervals) {
-               pCompressor->arrival = (lzsa_arrival *)malloc((BLOCK_SIZE + 1) * NMATCHES_PER_ARRIVAL_BIG * sizeof(lzsa_arrival));
+               pCompressor->arrival = (lzsa_arrival *)malloc(((BLOCK_SIZE + 1) << MATCHES_PER_ARRIVAL_SHIFT) * sizeof(lzsa_arrival));
    
                if (pCompressor->arrival) {
                   pCompressor->best_match = (lzsa_match *)malloc(BLOCK_SIZE * sizeof(lzsa_match));
diff --git a/src/shrink_context.h b/src/shrink_context.h
index 05e86fe..f418992 100644
--- a/src/shrink_context.h
+++ b/src/shrink_context.h
@@ -50,8 +50,8 @@ extern "C" {
 #define EXCL_VISITED_MASK  0x7fffffff
 
 #define NMATCHES_PER_ARRIVAL_SMALL 8
-#define NMATCHES_PER_ARRIVAL_BIG 16
-#define MATCHES_PER_ARRIVAL_SHIFT 4
+#define NMATCHES_PER_ARRIVAL_BIG 23
+#define MATCHES_PER_ARRIVAL_SHIFT 5
 
 #define NMATCHES_PER_INDEX_V1 8
 #define MATCHES_PER_INDEX_SHIFT_V1 3
@@ -59,7 +59,7 @@ extern "C" {
 #define NMATCHES_PER_INDEX_V2 64
 #define MATCHES_PER_INDEX_SHIFT_V2 6
 
-#define LEAVE_ALONE_MATCH_SIZE 1000
+#define LEAVE_ALONE_MATCH_SIZE 300
 
 #define LAST_LITERALS 0