RC4 tune-up.
[openssl.git] / crypto / rc4 / rc4_enc.c
index de57a97..93a75cd 100644 (file)
@@ -1,5 +1,5 @@
-/* crypto/rc4/rc4_enc.org */
-/* Copyright (C) 1995-1997 Eric Young (eay@cryptsoft.com)
+/* crypto/rc4/rc4_enc.c */
+/* Copyright (C) 1995-1998 Eric Young (eay@cryptsoft.com)
  * All rights reserved.
  *
  * This package is an SSL implementation written
  * [including the GNU Public Licence.]
  */
 
-/* WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING
- *
- * Always modify rc4_enc.org since rc4_enc.c is automatically generated from
- * it during SSLeay configuration.
- * WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING 
- */
-
-#include "rc4.h"
-
-/* if this is defined data[i] is used instead of *data, this is a %20
- * speedup on x86 */
-#undef RC4_INDEX
-
-char *RC4_version="RC4 part of SSLeay 0.8.1a 24-Jun-1998";
-
-char *RC4_options()
-       {
-#ifdef RC4_INDEX
-       if (sizeof(RC4_INT) == 1)
-               return("rc4(idx,char)");
-       else
-               return("rc4(idx,int)");
-#else
-       if (sizeof(RC4_INT) == 1)
-               return("rc4(ptr,char)");
-       else
-               return("rc4(ptr,int)");
-#endif
-       }
+#include <openssl/rc4.h>
+#include "rc4_locl.h"
 
 /* RC4 as implemented from a posting from
  * Newsgroups: sci.crypt
@@ -94,44 +67,8 @@ char *RC4_options()
  * Date: Wed, 14 Sep 1994 06:35:31 GMT
  */
 
-void RC4_set_key(key, len, data)
-RC4_KEY *key;
-int len;
-register unsigned char *data;
-       {
-        register RC4_INT tmp;
-        register int id1,id2;
-        register RC4_INT *d;
-        unsigned int i;
-        
-        d= &(key->data[0]);
-       for (i=0; i<256; i++)
-               d[i]=i;
-        key->x = 0;     
-        key->y = 0;     
-        id1=id2=0;     
-
-#define SK_LOOP(n) { \
-               tmp=d[(n)]; \
-               id2 = (data[id1] + tmp + id2) & 0xff; \
-               if (++id1 == len) id1=0; \
-               d[(n)]=d[id2]; \
-               d[id2]=tmp; }
-
-       for (i=0; i < 256; i+=4)
-               {
-               SK_LOOP(i+0);
-               SK_LOOP(i+1);
-               SK_LOOP(i+2);
-               SK_LOOP(i+3);
-               }
-       }
-    
-void RC4(key, len, indata, outdata)
-RC4_KEY *key;
-unsigned long len;
-unsigned char *indata;
-unsigned char *outdata;
+void RC4(RC4_KEY *key, unsigned long len, unsigned char *indata,
+            unsigned char *outdata)
        {
         register RC4_INT *d;
         register RC4_INT x,y,tx,ty;
@@ -141,6 +78,103 @@ unsigned char *outdata;
         y=key->y;     
         d=key->data; 
 
+#if defined(RC4_CHUNK) && (defined(L_ENDIAN) || defined(B_ENDIAN))
+       /*
+        * The original reason for implementing this(*) was the fact that
+        * pre-21164a Alpha CPUs don't have byte load/store instructions
+        * and e.g. a byte store has to be done with 64-bit load, shift,
+        * and, or and finally 64-bit store. Peaking data and operating
+        * at natural word size made it possible to reduce amount of
+        * instructions as well as to perform early read-ahead without
+        * suffering from RAW (read-after-write) hazard. This resulted
+        * in >40%(**) performance improvement (on 21064 box with gcc).
+        * But it's not only Alpha users who win here:-) Thanks to the
+        * early-n-wide read-ahead this implementation also exhibits
+        * >40% speed-up on SPARC and almost 20% on MIPS.
+        *
+        * (*)  "this" means code which recognizes the case when input
+        *      and output pointers appear to be aligned at natural CPU
+        *      word boundary.
+        * (**) i.e. according to 'apps/openssl speed rc4' benchmark,
+        *      crypto/rc4/rc4speed.c exhibits almost 70% speed-up.
+        *
+        *                                      <appro@fy.chalmers.se>
+        */
+
+#define RC4_STEP       ( \
+                       x=(x+1) &0xff,  \
+                       tx=d[x],        \
+                       y=(tx+y)&0xff,  \
+                       ty=d[y],        \
+                       d[y]=tx,        \
+                       d[x]=ty,        \
+                       (RC4_CHUNK)d[(tx+ty)&0xff]\
+                       )
+
+#if defined(L_ENDIAN)
+#  define SHFT(c)      ((c)*8)
+#  define MASK(i)      (((RC4_CHUNK)-1)>>((sizeof(RC4_CHUNK)-(i))<<3))
+#  define SHINC                8
+#elif defined(B_ENDIAN)
+#  define SHFT(c)      ((sizeof(RC4_CHUNK)-(c)-1)*8)
+#  define MASK(i)      (((RC4_CHUNK)-1)<<((sizeof(RC4_CHUNK)-(i))<<3))
+#  define SHINC                -8
+#else
+#  error "L_ENDIAN or B_ENDIAN *must* be defined!"
+#endif
+
+       if ( ( ((unsigned long)indata  & (sizeof(RC4_CHUNK)-1)) | 
+              ((unsigned long)outdata & (sizeof(RC4_CHUNK)-1)) ) == 0
+          ) {
+               RC4_CHUNK ichunk,cipher;
+
+               for (;len&-sizeof(RC4_CHUNK);len-=sizeof(RC4_CHUNK)) {
+                       ichunk  = *(RC4_CHUNK *)indata;
+                       cipher  = RC4_STEP<<SHFT(0);
+                       cipher |= RC4_STEP<<SHFT(1);
+                       cipher |= RC4_STEP<<SHFT(2);
+                       cipher |= RC4_STEP<<SHFT(3);
+#ifdef RC4_CHUNK_IS_64_BIT
+                       cipher |= RC4_STEP<<SHFT(4);
+                       cipher |= RC4_STEP<<SHFT(5);
+                       cipher |= RC4_STEP<<SHFT(6);
+                       cipher |= RC4_STEP<<SHFT(7);
+#endif
+                       *(RC4_CHUNK *)outdata = cipher^ichunk;
+                       indata  += sizeof(RC4_CHUNK);
+                       outdata += sizeof(RC4_CHUNK);
+               }
+               if (len) {
+                       RC4_CHUNK mask,ochunk;
+
+                       ichunk = *(RC4_CHUNK *)indata;
+                       ochunk = *(RC4_CHUNK *)outdata;
+                       cipher = 0;
+                       i = SHFT(0);
+                       mask = MASK(len);
+                       switch (len) {
+#ifdef RC4_CHUNK_IS_64_BIT
+                               case 7: cipher  = RC4_STEP<<SHFT(0), i+=SHINC;
+                               case 6: cipher |= RC4_STEP<<i,  i+=SHINC;
+                               case 5: cipher |= RC4_STEP<<i,  i+=SHINC;
+                               case 4: cipher |= RC4_STEP<<i,  i+=SHINC;
+                               case 3: cipher |= RC4_STEP<<i,  i+=SHINC;
+#else
+                               case 3: cipher  = RC4_STEP<<SHFT(0), i+=SHINC;
+#endif
+                               case 2: cipher |= RC4_STEP<<i,  i+=SHINC;
+                               case 1: cipher |= RC4_STEP<<i,  i+=SHINC;
+                               case 0: ; /* it's never the case, but it
+                                            has to be here for ultrix? */
+                       }
+                       ochunk &= ~mask;
+                       ochunk |= (cipher^ichunk) & mask;
+                       *(RC4_CHUNK *)outdata = ochunk;
+               }
+       }
+       else
+#endif
+       {
 #define LOOP(in,out) \
                x=((x+1)&0xff); \
                tx=d[x]; \
@@ -155,7 +189,6 @@ unsigned char *outdata;
 #define RC4_LOOP(a,b,i)        LOOP(a[i],b[i])
 #endif
 
-       i= -(int)len;
        i=(int)(len>>3L);
        if (i)
                {
@@ -190,6 +223,7 @@ unsigned char *outdata;
                        RC4_LOOP(indata,outdata,6); if (--i == 0) break;
                        }
                }               
+       }
        key->x=x;     
        key->y=y;
        }