From d07469a5fad47569984c70b22be5c89af8714f2f Mon Sep 17 00:00:00 2001
From: "Nicholas J. Kain" <nicholas@kain.us>
Date: Sat, 24 Oct 2020 17:14:20 -0400
Subject: [PATCH] Optimize internet checksum to use 16-bit fetches.

We could use 32-bit fetches with the same technique on 64-bit
architectures, or SIMD could be used to do very fast 128-bit
fetches, but this isn't a performance bottleneck and this
method is very simple and relatively fast.
---
 src/nk/net_checksum.h | 40 ++++++++++++++++++++--------------------
 1 file changed, 20 insertions(+), 20 deletions(-)

diff --git a/src/nk/net_checksum.h b/src/nk/net_checksum.h
index 46dc55f..e983ef6 100644
--- a/src/nk/net_checksum.h
+++ b/src/nk/net_checksum.h
@@ -1,6 +1,8 @@
 #ifndef NCMLIB_NET_CHECKSUM_H
 #define NCMLIB_NET_CHECKSUM_H
 
+// RFC 1071 is still a good reference.
+
 #include <stdint.h>
 
 // When summing ones-complement 16-bit values using a 32-bit unsigned
@@ -14,30 +16,28 @@ static inline uint16_t net_checksum161c_foldcarry(uint32_t v)
     return v;
 }
 
-// This function is not suitable for summing buffers that are greater than
-// 128k bytes in length: failure case will be incorrect checksums via
-// unsigned overflow, which is a defined operation and is safe.  This limit
-// should not be an issue for IPv4 or IPv6 packet, which are limited to
-// at most 64k bytes.
+// Produces the correct result on little endian in the sense that
+// the binary value returned, when stored to memory, will match
+// the result on big endian; if the numeric value returned
+// must match big endian results, then call ntohs() on the result.
 static uint16_t net_checksum161c(const void *buf, size_t size)
 {
-    uint32_t sum = 0;
-    int odd = size & 0x01;
-    size_t i;
-    size &= ~((size_t)0x01);
-    size >>= 1;
-    const uint8_t *b = (const uint8_t *)buf;
-    for (i = 0; i < size; ++i) {
-        uint16_t hi = b[i*2];
-        uint16_t lo = b[i*2+1];
-        sum += ntohs((lo + (hi << 8)));
+    const char *b = (const char *)buf;
+    const char *bend = b + size;
+    uint32_t sum = 0, sumo = 0;
+    if (size & 1) {
+        --bend;
+        uint8_t z[2] = { (uint8_t)*bend, 0 };
+        uint16_t t;
+        memcpy(&t, z, 2);
+        sumo = t;
     }
-    if (odd) {
-        uint16_t hi = b[i*2];
-        uint16_t lo = 0;
-        sum += ntohs((lo + (hi << 8)));
+    for (; b != bend; b += 2) {
+        uint16_t t;
+        memcpy(&t, b, 2);
+        sum += t;
     }
-    return ~net_checksum161c_foldcarry(sum);
+    return ~net_checksum161c_foldcarry(sum + sumo);
 }
 
 // For two sequences of bytes A and B that return checksums CS(A) and CS(B),