Added Curt McDowell's platform-independent optimized checksum routine.

src/core/inet.c

@@ -80,7 +80,7 @@ lwip_standard_chksum(void *dataptr, u16_t len)
   octetptr = (u8_t*)dataptr;
   while (len > 1)
   {
-    /* delacre first octet as most significant
+    /* declare first octet as most significant
        thus assume network order, ignoring host order */
     src = (*octetptr) << 8;
     octetptr++;
@@ -109,6 +109,133 @@ lwip_standard_chksum(void *dataptr, u16_t len)
 
 #endif
 
+Curt McDowell
+Broadcom Corp.
+csm@broadcom.com
+
+/*
+ * IP checksum two bytes at a time with support for
+ * unaligned buffer.
+ * Works for len up to and including 0x20000.
+ * by Curt McDowell, Broadcom Corp. 12/08/2005
+ */
+
+static u16_t
+lwip_standard_chksum2(void *dataptr, int len)
+{
+  u8_t *pb = dataptr;
+  u16_t *ps, t = 0;
+  u32_t sum = 0;
+  int odd = ((u32_t)pb & 1);
+
+  /* Get aligned to u16_t */
+  if (odd && len > 0) {
+    ((u8_t *)&t)[1] = *pb++;
+    len--;
+  }
+
+  /* Add the bulk of the data */
+  ps = (u16_t *)pb;
+  while (len > 1) {
+    sum += *ps++;
+    len -= 2;
+  }
+
+  /* Consume left-over byte, if any */
+  if (len > 0)
+    ((u8_t *)&t)[0] = *(u8_t *)ps;;
+
+  /* Add end bytes */
+  sum += t;
+
+  /*  Fold 32-bit sum to 16 bits */
+  while (sum >> 16)
+    sum = (sum & 0xffff) + (sum >> 16);
+
+  /* Swap if alignment was odd */
+  if (odd)
+    sum = ((sum & 0xff) << 8) | ((sum & 0xff00) >> 8);
+
+  return sum;
+}
+
+#if 0 
+
+/**
+ * An optimized checksum routine. Basically, it uses loop-unrolling on
+ * the checksum loop, treating the head and tail bytes specially, whereas
+ * the inner loop acts on 8 bytes at a time. 
+ *
+ * @arg start of buffer to be checksummed. May be an odd byte address.
+ * @len number of bytes in the buffer to be checksummed.
+ * 
+ * @todo First argument type conflicts with generic checksum routine.
+ * 
+ * by Curt McDowell, Broadcom Corp. December 8th, 2005
+ */
+
+static u16_t
+lwip_standard_chksum4(u8_t *pb, int len)
+{
+  u16_t *ps, t = 0;
+  u32_t *pl;
+  u32_t sum = 0, tmp;
+  /* starts at odd byte address? */
+  int odd = ((u32_t)pb & 1);
+
+  if (odd && len > 0) {
+    ((u8_t *)&t)[1] = *pb++;
+    len--;
+  }
+
+  ps = (u16_t *)pb;
+
+  if (((u32_t)ps & 3) && len > 1) {
+    sum += *ps++;
+    len -= 2;
+  }
+
+  pl = (u32_t *)ps;
+
+  while (len > 7)  {
+    tmp = sum + *pl++;          /* ping */
+    if (tmp < sum)
+      tmp++;                    /* add back carry */
+
+    sum = tmp + *pl++;          /* pong */
+    if (sum < tmp)
+      sum++;                    /* add back carry */
+
+    len -= 8;
+  }
+
+  /* make room in upper bits */
+  sum = (sum >> 16) + (sum & 0xffff);
+
+  ps = (u16_t *)pl;
+
+  /* 16-bit aligned word remaining? */
+  while (len > 1) {
+    sum += *ps++;
+    len -= 2;
+  }
+
+  /* dangling tail byte remaining? */
+  if (len > 0)                  /* include odd byte */
+    ((u8_t *)&t)[0] = *(u8_t *)ps;
+
+  sum += t;                     /* add end bytes */
+
+  while (sum >> 16)             /* combine halves */
+    sum = (sum >> 16) + (sum & 0xffff);
+
+  if (odd)
+    sum = ((sum & 0xff) << 8) | ((sum & 0xff00) >> 8);
+
+  return sum;
+}
+#endif
+
 /* inet_chksum_pseudo:
  *
  * Calculates the pseudo Internet checksum used by TCP and UDP for a pbuf chain.