diff --git a/delta2/delta2.c b/delta2/delta2.c
index d93afc9..1a019b7 100644
--- a/delta2/delta2.c
+++ b/delta2/delta2.c
@@ -27,7 +27,7 @@
  * (3, 5, 7, 8) are tried to find the one that gives the maximum
  * reduction. A span length threshold in bytes is used. Byte spans
  * less than this threshold are ignored.
- * Bytes are packed into integers in big-endian format.
+ * Bytes are packed into integers in little-endian format.
  *
  * After an optimal stride length has been identified the encoder
  * performs a delta run length encoding on the spans. Two types of
@@ -71,7 +71,7 @@
 #define	MSB_SHIFT	(56)
 
 /*
- * Delta2 algorithm processes data in chunks. The 4K size below is somewhat
+ * Delta2 algorithm processes data in blocks. The 4K size below is somewhat
  * adhoc but a couple of considerations were looked at:
  * 1) Balance between number of headers and delta runs. Too small chunks
  *    will increase header counts for long delta runs spanning chunks.
@@ -86,8 +86,8 @@
  * and interpreted in little-endian order.
  */
 #if BYTE_ORDER == BIG_ENDIAN
-#define	HTONLL	__bswap_64(x)
-#define	NTOHLL	__bswap_64(x)
+#define	HTONLL	htonll
+#define	NTOHLL	ntohll
 #else
 #define	HTONLL
 #define	NTOHLL
@@ -249,14 +249,14 @@ delta2_encode_real(uchar_t *src, uint64_t srclen, uchar_t *dst, uint64_t *dstlen
 			gtot2 += DELTA_HDR;
 			/*
 			 * If this ended into another table reset next scan
-			 * point to before the table.
+			 * point to beginning of the table.
 			 */
 			val = cnt - snum;
 		} else {
 			gtot2 += snum;
 			/*
 			 * If this ended into another table reset next scan
-			 * point to before the table.
+			 * point to beginning of the table.
 			 */
 			if (snum >= (MIN_THRESH>>1))
 				val = cnt - snum;
diff --git a/main.c b/main.c
index 7a4c3e1..b62aea2 100644
--- a/main.c
+++ b/main.c
@@ -247,6 +247,7 @@ preproc_compress(compress_func_ptr cmp_func, void *src, uint64_t srclen, void *d
 		*dstlen = _dstlen + 9;
 		DEBUG_STAT_EN(fprintf(stderr, "Chunk compression speed %.3f MB/s\n", get_mb_s(srclen, strt, en)));
 	} else {
+		DEBUG_STAT_EN(fprintf(stderr, "Chunk did not compress.\n"));
 		memcpy(dest+1, src, srclen);
 		*dstlen = srclen + 1;
 		result = 0;
@@ -274,10 +275,13 @@ preproc_decompress(compress_func_ptr dec_func, void *src, uint64_t srclen, void
 		DEBUG_STAT_EN(strt = get_wtime_millis());
 		result = dec_func(sorc, srclen, dst, dstlen, level, chdr, data);
 		DEBUG_STAT_EN(en = get_wtime_millis());
+
 		if (result < 0) return (result);
 		DEBUG_STAT_EN(fprintf(stderr, "Chunk decompression speed %.3f MB/s\n", get_mb_s(srclen, strt, en)));
 		memcpy(src, dst, *dstlen);
 		srclen = *dstlen;
+	} else {
+		src = sorc;
 	}
 
 	if (type & PREPROC_TYPE_DELTA2) {
@@ -1253,8 +1257,14 @@ plain_compress:
 			    compressed_chunk, &_chunksize, tdat->level, 0, tdat->data,
 			    tdat->props);
 		} else {
+			DEBUG_STAT_EN(double strt, en);
+
+			DEBUG_STAT_EN(strt = get_wtime_millis());
 			rv = tdat->compress(tdat->uncompressed_chunk, tdat->rbytes,
 			    compressed_chunk, &_chunksize, tdat->level, 0, tdat->data);
+			DEBUG_STAT_EN(en = get_wtime_millis());
+			DEBUG_STAT_EN(fprintf(stderr, "Chunk compression speed %.3f MB/s\n",
+					      get_mb_s(_chunksize, strt, en)));
 		}
 	}