4 files changed, 35 insertions, 20 deletions

diff --git a/xdelta3/examples/Makefile b/xdelta3/examples/Makefile
index d171a80..d6d9cc9 100755
--- a/xdelta3/examples/Makefile
+++ b/xdelta3/examples/Makefile
@@ -1,6 +1,5 @@
-#CFLAGS = -g -Wall -I..
-# -DXD3_DEBUG=1
-CFLAGS = -O3 -pg -Wall -I.. -DXD3_DEBUG=0
+#CFLAGS = -g -Wall -I.. -DXD3_DEBUG=1
+CFLAGS = -O3 -Wall -I.. -DXD3_DEBUG=0
 # -pg
 
 SOURCES = small_page_test.c encode_decode_test.c speed_test.c
diff --git a/xdelta3/xdelta3-decode.h b/xdelta3/xdelta3-decode.h
index 5ddf823..0b49fb9 100644
--- a/xdelta3/xdelta3-decode.h
+++ b/xdelta3/xdelta3-decode.h
@@ -1021,8 +1021,9 @@ xd3_decode_input (xd3_stream *stream)
               return XD3_INVALID_INPUT;
             }
 
-          src->cpyoff_blocks = stream->dec_cpyoff / src->blksize;
-          src->cpyoff_blkoff = stream->dec_cpyoff % src->blksize;
+          xd3_blksize_div(stream->dec_cpyoff, src,
+                          &src->cpyoff_blocks,
+                          &src->cpyoff_blkoff);
         }
 
       /* xd3_decode_emit returns XD3_OUTPUT on every success. */
diff --git a/xdelta3/xdelta3.c b/xdelta3/xdelta3.c
index 58fb8b5..88de13a 100644
--- a/xdelta3/xdelta3.c
+++ b/xdelta3/xdelta3.c
@@ -2629,15 +2629,14 @@ xd3_set_source (xd3_stream *stream,
                 xd3_source *src)
 {
   xoff_t blk_num;
-  xoff_t tail_size;
+  usize_t tail_size;
 
   IF_DEBUG1 (DP(RINT "[set source] size %"Q"u\n", src->size));
 
   if (src == NULL || src->size < stream->smatcher.large_look) { return 0; }
 
   stream->src  = src;
-  blk_num      = src->size / src->blksize;
-  tail_size    = src->size - (blk_num * src->blksize);
+  xd3_blksize_div (src->size, src, &blk_num, &tail_size);
   src->blocks  = blk_num + (tail_size > 0);
   src->onlastblk = xd3_bytes_on_srcblk (src, src->blocks - 1);
   src->srclen  = 0;
@@ -4416,8 +4415,7 @@ xd3_source_extend_match (xd3_stream *stream)
        * match_fwd/match_back and direction.  Consolidate? */
       matchoff  = stream->match_srcpos - stream->match_back;
       streamoff = stream->input_position - stream->match_back;
-      tryblk    = matchoff / src->blksize;
-      tryoff    = matchoff - (tryblk * src->blksize);
+      xd3_blksize_div (matchoff, src, &tryblk, &tryoff);
 
       /* this loops backward over source blocks */
       while (stream->match_back < stream->match_maxback)
@@ -4465,8 +4463,7 @@ xd3_source_extend_match (xd3_stream *stream)
 
   matchoff  = stream->match_srcpos + stream->match_fwd;
   streamoff = stream->input_position + stream->match_fwd;
-  tryblk    = matchoff / src->blksize;
-  tryoff    = matchoff - (tryblk * src->blksize);
+  xd3_blksize_div (matchoff, src, & tryblk, & tryoff);
 
   /* Note: practically the same code as backwards case above: same comments */
   while (stream->match_fwd < stream->match_maxfwd)
@@ -4839,11 +4836,16 @@ xd3_srcwin_move_point (xd3_stream *stream, usize_t *next_move_point)
   while (stream->srcwin_cksum_pos < logical_input_cksum_pos &&
          stream->srcwin_cksum_pos < stream->src->size)
     {
-      xoff_t  blkno = stream->srcwin_cksum_pos / stream->src->blksize;
-      xoff_t  blkbaseoffset = blkno * stream->src->blksize;
-      ssize_t oldpos = stream->srcwin_cksum_pos - blkbaseoffset;
-      ssize_t blkpos = xd3_bytes_on_srcblk_fast (stream->src, blkno);
+      xoff_t  blkno;
+      xoff_t  blkbaseoffset;
+      usize_t blkrem;
+      ssize_t oldpos;
+      ssize_t blkpos;
       int ret;
+      xd3_blksize_div (stream->srcwin_cksum_pos,
+                       stream->src, &blkno, &blkrem);
+      oldpos = blkrem;
+      blkpos = xd3_bytes_on_srcblk_fast (stream->src, blkno);
 
       if (oldpos + stream->smatcher.large_look > blkpos)
         {
diff --git a/xdelta3/xdelta3.h b/xdelta3/xdelta3.h
index af79839..9abddbf 100644
--- a/xdelta3/xdelta3.h
+++ b/xdelta3/xdelta3.h
@@ -677,8 +677,7 @@ struct _xd3_source
   xoff_t              blocks;        /* the total number of blocks in
                                         this source */
   usize_t             onlastblk;     /* cached size info, avoid __udivdi3 */
-  usize_t             cpyoff_blocks; /* offset of copy window in
-                                        blocks */
+  xoff_t              cpyoff_blocks; /* offset of dec_cpyoff in blocks */
   usize_t             cpyoff_blkoff; /* offset of copy window in
                                         blocks, remainder */
   xoff_t              getblkno;      /* request block number: xd3 sets
@@ -1164,6 +1163,18 @@ const char* xd3_errstring (xd3_stream  *stream)
   return stream->msg ? stream->msg : "";
 }
 
+
+/* 64-bit divisions are expensive.  on a 32bit platform, these show in
+ * a profile as __udivdi3().  these are all the xoff_t divisions: */
+static inline
+void xd3_blksize_div (const xoff_t offset,
+                      const xd3_source *source,
+                      xoff_t *blkno,
+                      usize_t *blkoff) {
+  *blkno = offset / source->blksize;
+  *blkoff = offset - (*blkno * source->blksize);
+}
+
 /* This function tells the number of bytes expected to be set in
  * source->onblk after a getblk request.  This is for convenience of
  * handling a partial last block.  Note that this is a relatively
@@ -1173,14 +1184,16 @@ const char* xd3_errstring (xd3_stream  *stream)
 static inline
 usize_t xd3_bytes_on_srcblk (xd3_source *source, xoff_t blkno)
 {
+  xoff_t s_1_div;
+  usize_t s_1_rem;
   XD3_ASSERT (blkno < source->blocks);
 
   if (blkno != source->blocks - 1)
     {
       return source->blksize;
     }
-
-  return (usize_t)((source->size - 1) % source->blksize) + 1;
+  xd3_blksize_div(source->size - 1, source, &s_1_div, &s_1_rem);
+  return s_1_rem + 1;
 }
 
 static inline