*** empty log message ***

1 files changed, 48 insertions, 9 deletions

diff --git a/lzf_d.c b/lzf_d.c
index 9e2cd82..076d1d9 100644
--- a/lzf_d.c
+++ b/lzf_d.c
@@ -36,6 +36,8 @@
 
 #include "lzfP.h"
 
+#include <string.h> /* for memcpy/memset */
+
 #if AVOID_ERRNO
 # define SET_ERRNO(n)
 #else
@@ -43,12 +45,14 @@
 # define SET_ERRNO(n) errno = (n)
 #endif
 
+#if USE_REP_MOVSB /* small win on amd, big loss on intel */
 #if (__i386 || __amd64) && __GNUC__ >= 3
 # define lzf_movsb(dst, src, len)                \
    asm ("rep movsb"                              \
         : "=D" (dst), "=S" (src), "=c" (len)     \
         :  "0" (dst),  "1" (src),  "2" (len));
 #endif
+#endif
 
 unsigned int 
 lzf_decompress (const void *const in_data,  unsigned int in_len,
@@ -84,9 +88,17 @@ lzf_decompress (const void *const in_data,  unsigned int in_len,
 #ifdef lzf_movsb
           lzf_movsb (op, ip, ctrl);
 #else
-          do
-            *op++ = *ip++;
-          while (--ctrl);
+          switch (ctrl)
+            {
+              case 32: *op++ = *ip++; case 31: *op++ = *ip++; case 30: *op++ = *ip++; case 29: *op++ = *ip++;
+              case 28: *op++ = *ip++; case 27: *op++ = *ip++; case 26: *op++ = *ip++; case 25: *op++ = *ip++;
+              case 24: *op++ = *ip++; case 23: *op++ = *ip++; case 22: *op++ = *ip++; case 21: *op++ = *ip++;
+              case 20: *op++ = *ip++; case 19: *op++ = *ip++; case 18: *op++ = *ip++; case 17: *op++ = *ip++;
+              case 16: *op++ = *ip++; case 15: *op++ = *ip++; case 14: *op++ = *ip++; case 13: *op++ = *ip++;
+              case 12: *op++ = *ip++; case 11: *op++ = *ip++; case 10: *op++ = *ip++; case  9: *op++ = *ip++;
+              case  8: *op++ = *ip++; case  7: *op++ = *ip++; case  6: *op++ = *ip++; case  5: *op++ = *ip++;
+              case  4: *op++ = *ip++; case  3: *op++ = *ip++; case  2: *op++ = *ip++; case  1: *op++ = *ip++;
+            }
 #endif
         }
       else /* back reference */
@@ -132,12 +144,39 @@ lzf_decompress (const void *const in_data,  unsigned int in_len,
           len += 2;
           lzf_movsb (op, ref, len);
 #else
-          *op++ = *ref++;
-          *op++ = *ref++;
-
-          do
-            *op++ = *ref++;
-          while (--len);
+          switch (len)
+            {
+              default:
+                len += 2;
+
+                if (op >= ref + len)
+                  {
+                    /* disjunct areas */
+                    memcpy (op, ref, len);
+                    op += len;
+                  }
+                else
+                  {
+                    /* overlapping, use octte by octte copying */
+                    do
+                      *op++ = *ref++;
+                    while (--len);
+                  }
+
+                break;
+
+              case 9: *op++ = *ref++;
+              case 8: *op++ = *ref++;
+              case 7: *op++ = *ref++;
+              case 6: *op++ = *ref++;
+              case 5: *op++ = *ref++;
+              case 4: *op++ = *ref++;
+              case 3: *op++ = *ref++;
+              case 2: *op++ = *ref++;
+              case 1: *op++ = *ref++;
+              case 0: *op++ = *ref++; /* two octets more */
+                      *op++ = *ref++;
+            }
 #endif
         }
     }