(memxor3): Optimized.

(memxor3_common_alignment): New function.
(memxor3_different_alignment_b): New function.
(memxor3_different_alignment_ab): New function.
(memxor3_different_alignment_all): New function.

Rev: nettle/memxor.c:1.3

memxor.c

@@ -23,7 +23,8 @@
  * MA 02111-1307, USA.
  */
 
-/* Implementation inspired by memcmp in glibc, contributed to the FSF by Torbjorn Granlund.
+/* Implementation inspired by memcmp in glibc, contributed to the FSF
+   by Torbjorn Granlund.
  */
 
 #if HAVE_CONFIG_H
@@ -34,7 +35,6 @@
 
 #include "memxor.h"
 
-#if 1
 typedef unsigned long int word_t;
 
 #if SIZEOF_LONG & (SIZEOF_LONG - 1)
@@ -55,16 +55,19 @@ typedef unsigned long int word_t;
 static void
 memxor_common_alignment (word_t *dst, const word_t *src, size_t n)
 {
-  size_t i;
-  /* FIXME: Unroll four times, like memcmp? */
-  i = n & 1;
-  if (i)
-    dst[0] ^= src[0];
+  /* FIXME: Require n > 0? */
+  /* FIXME: Unroll four times, like memcmp? Probably not worth the
+     effort. */
 
-  for (; i < n; i += 2)
+  if (n & 1)
     {
-      dst[i] ^= src[i];
-      dst[i+1] ^= src[i+1];
+      *dst++ ^= *src++;
+      n--;
+    }
+  for (; n >= 2; dst += 2, src += 2, n -= 2)
+    {
+      dst[0] ^= src[0];
+      dst[1] ^= src[1];
     }
 }
 
@@ -83,7 +86,7 @@ memxor_different_alignment (word_t *dst, const uint8_t *src, size_t n)
   shl = CHAR_BIT * offset;
   shr = CHAR_BIT * (sizeof(word_t) - offset);
 
-  src_word = (const word_t *) (src - offset);
+  src_word = (const word_t *) ((uintptr_t) src & -SIZEOF_LONG);
 
   /* FIXME: Unroll four times, like memcmp? */
   i = n & 1;
@@ -91,7 +94,6 @@ memxor_different_alignment (word_t *dst, const uint8_t *src, size_t n)
   if (i)
     {
       s1 = src_word[0];
-      s0 = src_word[1];
       dst[0] ^= MERGE (s1, shl, s0, shr);
     }
 
@@ -104,6 +106,9 @@ memxor_different_alignment (word_t *dst, const uint8_t *src, size_t n)
     }
 }
 
+/* Performance, Intel SU1400 (x86_64): 0.25 cycles/byte aligned, 0.45
+   cycles/byte unaligned. */
+
 /* XOR LEN bytes starting at SRCADDR onto DESTADDR. Result undefined
    if the source overlaps with the destination. Return DESTADDR. */
 uint8_t *
@@ -113,8 +118,6 @@ memxor(uint8_t *dst, const uint8_t *src, size_t n)
 
   if (n >= WORD_T_THRESH)
     {
-      size_t left_over;
-
       /* There are at least some bytes to compare.  No need to test
 	 for N == 0 in this alignment loop.  */
       while (ALIGN_OFFSET (dst))
@@ -127,35 +130,195 @@ memxor(uint8_t *dst, const uint8_t *src, size_t n)
       else
 	memxor_common_alignment ((word_t *) dst, (const word_t *) src, n / sizeof(word_t));
 
-      left_over = n % sizeof(word_t);
-      dst += n - left_over;
-      src += n - left_over;
-      n = left_over;
+      dst += n & -SIZEOF_LONG;
+      src += n & -SIZEOF_LONG;
+      n = n & (SIZEOF_LONG - 1);
     }
   for (; n > 0; n--)
     *dst++ ^= *src++;
 
   return orig_dst;
 }
-#else
-uint8_t *
-memxor(uint8_t *dst, const uint8_t *src, size_t n)
+
+
+/* XOR word-aligned areas. n is the number of words, not bytes. */
+static void
+memxor3_common_alignment (word_t *dst,
+			  const word_t *a, const word_t *b, size_t n)
 {
-  size_t i;
-  for (i = 0; i<n; i++)
-    dst[i] ^= src[i];
+  /* FIXME: Require n > 0? */
+  while (n-- > 0)
+    dst[n] = a[n] ^ b[n];
+}
 
-  return dst;
+static void
+memxor3_different_alignment_b (word_t *dst,
+			       const word_t *a, const uint8_t *b, unsigned offset, size_t n)
+{
+  int shl, shr;
+  const word_t *b_word;
+
+  word_t s0, s1;
+
+  shl = CHAR_BIT * offset;
+  shr = CHAR_BIT * (sizeof(word_t) - offset);
+
+  b_word = (const word_t *) ((uintptr_t) b & -SIZEOF_LONG);
+
+  if (n & 1)
+    {
+      n--;
+      s1 = b_word[n];
+      s0 = b_word[n+1];
+      dst[n] = a[n] ^ MERGE (s1, shl, s0, shr);
     }
-#endif
+  else
+    s1 = b_word[n];
+  
+  while (n > 0)
+    {
+      n -= 2;
+      s0 = b_word[n+1]; 
+      dst[n+1] = a[n+1] ^ MERGE(s0, shl, s1, shr);
+      s1 = b_word[n];
+      dst[n] = a[n] ^ MERGE(s1, shl, s0, shr);
+    }
+}
+
+static void
+memxor3_different_alignment_ab (word_t *dst,
+				const uint8_t *a, const uint8_t *b,
+				unsigned offset, size_t n)
+{
+  int shl, shr;
+  const word_t *a_word;
+  const word_t *b_word;
   
+  word_t s0, s1;
+
+  shl = CHAR_BIT * offset;
+  shr = CHAR_BIT * (sizeof(word_t) - offset);
+
+  a_word = (const word_t *) ((uintptr_t) a & -SIZEOF_LONG);
+  b_word = (const word_t *) ((uintptr_t) b & -SIZEOF_LONG);
+
+  if (n & 1)
+    {
+      n--;
+      s1 = a_word[n] ^ b_word[n];
+      s0 = a_word[n+1] ^ b_word[n+1];
+      dst[n] = MERGE (s1, shl, s0, shr);
+    }
+  else    
+    s1 = a_word[n] ^ b_word[n];
+  
+  while (n > 0)
+    {
+      n -= 2;
+      s0 = a_word[n+1] ^ b_word[n+1]; 
+      dst[n+1] = MERGE(s0, shl, s1, shr);
+      s1 = a_word[n] ^ b_word[n];
+      dst[n] = MERGE(s1, shl, s0, shr);
+    }
+}
+
+static void
+memxor3_different_alignment_all (word_t *dst,
+				 const uint8_t *a, const uint8_t *b,
+				 unsigned a_offset, unsigned b_offset,
+				 size_t n)
+{
+  int al, ar, bl, br;
+  const word_t *a_word;
+  const word_t *b_word;
+  
+  word_t a0, a1, b0, b1;
+
+  al = CHAR_BIT * a_offset;
+  ar = CHAR_BIT * (sizeof(word_t) - a_offset);
+  bl = CHAR_BIT * b_offset;
+  br = CHAR_BIT * (sizeof(word_t) - b_offset);
+
+  a_word = (const word_t *) ((uintptr_t) a & -SIZEOF_LONG);
+  b_word = (const word_t *) ((uintptr_t) b & -SIZEOF_LONG);
+
+  if (n & 1)
+    {
+      n--;
+      a1 = a_word[n]; a0 = a_word[n+1];
+      b1 = b_word[n]; b0 = b_word[n+1];
+      
+      dst[n] = MERGE (a1, al, a0, ar) ^ MERGE (b1, bl, b0, br);
+    }
+  else    
+    {
+      a1 = a_word[n];
+      b1 = b_word[n];
+    }
+  
+  while (n > 0)
+    {
+      n -= 2;
+      a0 = a_word[n+1]; b0 = b_word[n+1]; 
+      dst[n+1] = MERGE(a0, al, a1, ar) ^ MERGE(b0, bl, b1, br);
+      a1 = a_word[n]; b1 = b_word[n];
+      dst[n] = MERGE(a1, al, a0, ar) ^ MERGE(b1, bl, b0, br);
+    }
+}
+
+/* Current implementation processes data in descending order, to
+   support overlapping operation with one of the sources overlapping
+   the start of the destination area. This feature is used only
+   internally by cbc decrypt, and it is not advertised or documented
+   to nettle users. */
 uint8_t *
 memxor3(uint8_t *dst, const uint8_t *a, const uint8_t *b, size_t n)
 {
-  size_t i;
-  for (i = 0; i<n; i++)
-    dst[i] = a[i] ^ b[i];
+  if (n >= WORD_T_THRESH)
+    {
+      unsigned i;
+      unsigned a_offset;
+      unsigned b_offset;
+      size_t nwords;
 
-  return dst;
+      for (i = ALIGN_OFFSET(dst + n); i > 0; i--)
+	{
+	  n--;
+	  dst[n] = a[n] ^ b[n];
 	}
 
+      a_offset = ALIGN_OFFSET(a + n);
+      b_offset = ALIGN_OFFSET(b + n);
+
+      nwords = n / sizeof (word_t);
+      n %= sizeof (word_t);
+
+      if (a_offset == b_offset)
+	{
+	  if (!a_offset)
+	    memxor3_common_alignment((word_t *) (dst + n),
+				     (const word_t *) (a + n),
+				     (const word_t *) (b + n), nwords);
+	  else
+	    memxor3_different_alignment_ab((word_t *) (dst + n),
+					   a + n, b + n, a_offset,
+					   nwords);
+	}
+      else if (!a_offset)
+	memxor3_different_alignment_b((word_t *) (dst + n),
+				      (const word_t *) (a + n), b + n,
+				      b_offset, nwords);
+      else if (!b_offset)
+	memxor3_different_alignment_b((word_t *) (dst + n),
+				      (const word_t *) (b + n), a + n,
+				      a_offset, nwords);
+      else
+	memxor3_different_alignment_all((word_t *) (dst + n), a + n, b + n,
+					a_offset, b_offset, nwords);
+					
+    }
+  while (n-- > 0)
+    dst[n] = a[n] ^ b[n];
+
+  return dst;
+}