X-Git-Url: https://git.openssl.org/gitweb/?p=openssl.git;a=blobdiff_plain;f=crypto%2Frc4%2Frc4_enc.c;h=8c4fc6c7a3da38451da27f8b429ef340d2684a40;hp=26da6d520cc534547b40418d551d2c2070704715;hb=6bcb68da419e9804d3b2cc88023d25546e4a7d60;hpb=dfeab0689f69c0b4bd3480ffd37a9cacc2f17d9c

diff --git a/crypto/rc4/rc4_enc.c b/crypto/rc4/rc4_enc.c
index 26da6d520c..8c4fc6c7a3 100644
--- a/crypto/rc4/rc4_enc.c
+++ b/crypto/rc4/rc4_enc.c
@@ -56,7 +56,7 @@
  * [including the GNU Public Licence.]
  */
 
-#include "rc4.h"
+#include <openssl/rc4.h>
 #include "rc4_locl.h"
 
 /* RC4 as implemented from a posting from
@@ -67,20 +67,201 @@
  * Date: Wed, 14 Sep 1994 06:35:31 GMT
  */
 
-void RC4(key, len, indata, outdata)
-RC4_KEY *key;
-unsigned long len;
-unsigned char *indata;
-unsigned char *outdata;
+void RC4(RC4_KEY *key, size_t len, const unsigned char *indata,
+	     unsigned char *outdata)
 	{
         register RC4_INT *d;
         register RC4_INT x,y,tx,ty;
-	int i;
+	size_t i;
         
         x=key->x;     
         y=key->y;     
         d=key->data; 
 
+#if defined(RC4_CHUNK)
+	/*
+	 * 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 20-30% on 64-bit MIPS (depending
+	 * on sizeof(RC4_INT)).
+	 *
+	 * (*)	"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...
+	 *
+	 * Cavets.
+	 *
+	 * - RC4_CHUNK="unsigned long long" should be a #1 choice for
+	 *   UltraSPARC. Unfortunately gcc generates very slow code
+	 *   (2.5-3 times slower than one generated by Sun's WorkShop
+	 *   C) and therefore gcc (at least 2.95 and earlier) should
+	 *   always be told that RC4_CHUNK="unsigned long".
+	 *
+	 *					<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 ( ( ((size_t)indata  & (sizeof(RC4_CHUNK)-1)) | 
+	       ((size_t)outdata & (sizeof(RC4_CHUNK)-1)) ) == 0 )
+		{
+		RC4_CHUNK ichunk,otp;
+		const union { long one; char little; } is_endian = {1};
+
+		/*
+		 * I reckon we can afford to implement both endian
+		 * cases and to decide which way to take at run-time
+		 * because the machine code appears to be very compact
+		 * and redundant 1-2KB is perfectly tolerable (i.e.
+		 * in case the compiler fails to eliminate it:-). By
+		 * suggestion from Terrel Larson <terr@terralogic.net>
+		 * who also stands for the is_endian union:-)
+		 *
+		 * Special notes.
+		 *
+		 * - is_endian is declared automatic as doing otherwise
+		 *   (declaring static) prevents gcc from eliminating
+		 *   the redundant code;
+		 * - compilers (those I've tried) don't seem to have
+		 *   problems eliminating either the operators guarded
+		 *   by "if (sizeof(RC4_CHUNK)==8)" or the condition
+		 *   expressions themselves so I've got 'em to replace
+		 *   corresponding #ifdefs from the previous version;
+		 * - I chose to let the redundant switch cases when
+		 *   sizeof(RC4_CHUNK)!=8 be (were also #ifdefed
+		 *   before);
+		 * - in case you wonder "&(sizeof(RC4_CHUNK)*8-1)" in
+		 *   [LB]ESHFT guards against "shift is out of range"
+		 *   warnings when sizeof(RC4_CHUNK)!=8 
+		 *
+		 *			<appro@fy.chalmers.se>
+		 */
+		if (!is_endian.little)
+			{	/* BIG-ENDIAN CASE */
+# define BESHFT(c)	(((sizeof(RC4_CHUNK)-(c)-1)*8)&(sizeof(RC4_CHUNK)*8-1))
+			for (;len&(0-sizeof(RC4_CHUNK));len-=sizeof(RC4_CHUNK))
+				{
+				ichunk  = *(RC4_CHUNK *)indata;
+				otp  = RC4_STEP<<BESHFT(0);
+				otp |= RC4_STEP<<BESHFT(1);
+				otp |= RC4_STEP<<BESHFT(2);
+				otp |= RC4_STEP<<BESHFT(3);
+				if (sizeof(RC4_CHUNK)==8)
+					{
+					otp |= RC4_STEP<<BESHFT(4);
+					otp |= RC4_STEP<<BESHFT(5);
+					otp |= RC4_STEP<<BESHFT(6);
+					otp |= RC4_STEP<<BESHFT(7);
+					}
+				*(RC4_CHUNK *)outdata = otp^ichunk;
+				indata  += sizeof(RC4_CHUNK);
+				outdata += sizeof(RC4_CHUNK);
+				}
+			if (len)
+				{
+				RC4_CHUNK mask=(RC4_CHUNK)-1, ochunk;
+
+				ichunk = *(RC4_CHUNK *)indata;
+				ochunk = *(RC4_CHUNK *)outdata;
+				otp = 0;
+				i = BESHFT(0);
+				mask <<= (sizeof(RC4_CHUNK)-len)<<3;
+				switch (len&(sizeof(RC4_CHUNK)-1))
+					{
+					case 7:	otp  = RC4_STEP<<i, i-=8;
+					case 6:	otp |= RC4_STEP<<i, i-=8;
+					case 5:	otp |= RC4_STEP<<i, i-=8;
+					case 4:	otp |= RC4_STEP<<i, i-=8;
+					case 3:	otp |= RC4_STEP<<i, i-=8;
+					case 2:	otp |= RC4_STEP<<i, i-=8;
+					case 1:	otp |= RC4_STEP<<i, i-=8;
+					case 0: ; /*
+						   * it's never the case,
+						   * but it has to be here
+						   * for ultrix?
+						   */
+					}
+				ochunk &= ~mask;
+				ochunk |= (otp^ichunk) & mask;
+				*(RC4_CHUNK *)outdata = ochunk;
+				}
+			key->x=x;     
+			key->y=y;
+			return;
+			}
+		else
+			{	/* LITTLE-ENDIAN CASE */
+# define LESHFT(c)	(((c)*8)&(sizeof(RC4_CHUNK)*8-1))
+			for (;len&(0-sizeof(RC4_CHUNK));len-=sizeof(RC4_CHUNK))
+				{
+				ichunk  = *(RC4_CHUNK *)indata;
+				otp  = RC4_STEP;
+				otp |= RC4_STEP<<8;
+				otp |= RC4_STEP<<16;
+				otp |= RC4_STEP<<24;
+				if (sizeof(RC4_CHUNK)==8)
+					{
+					otp |= RC4_STEP<<LESHFT(4);
+					otp |= RC4_STEP<<LESHFT(5);
+					otp |= RC4_STEP<<LESHFT(6);
+					otp |= RC4_STEP<<LESHFT(7);
+					}
+				*(RC4_CHUNK *)outdata = otp^ichunk;
+				indata  += sizeof(RC4_CHUNK);
+				outdata += sizeof(RC4_CHUNK);
+				}
+			if (len)
+				{
+				RC4_CHUNK mask=(RC4_CHUNK)-1, ochunk;
+
+				ichunk = *(RC4_CHUNK *)indata;
+				ochunk = *(RC4_CHUNK *)outdata;
+				otp = 0;
+				i   = 0;
+				mask >>= (sizeof(RC4_CHUNK)-len)<<3;
+				switch (len&(sizeof(RC4_CHUNK)-1))
+					{
+					case 7:	otp  = RC4_STEP,    i+=8;
+					case 6:	otp |= RC4_STEP<<i, i+=8;
+					case 5:	otp |= RC4_STEP<<i, i+=8;
+					case 4:	otp |= RC4_STEP<<i, i+=8;
+					case 3:	otp |= RC4_STEP<<i, i+=8;
+					case 2:	otp |= RC4_STEP<<i, i+=8;
+					case 1:	otp |= RC4_STEP<<i, i+=8;
+					case 0: ; /*
+						   * it's never the case,
+						   * but it has to be here
+						   * for ultrix?
+						   */
+					}
+				ochunk &= ~mask;
+				ochunk |= (otp^ichunk) & mask;
+				*(RC4_CHUNK *)outdata = ochunk;
+				}
+			key->x=x;     
+			key->y=y;
+			return;
+			}
+		}
+#endif
 #define LOOP(in,out) \
 		x=((x+1)&0xff); \
 		tx=d[x]; \
@@ -95,7 +276,7 @@ unsigned char *outdata;
 #define RC4_LOOP(a,b,i)	LOOP(a[i],b[i])
 #endif
 
-	i=(int)(len>>3L);
+	i=len>>3;
 	if (i)
 		{
 		for (;;)
@@ -115,7 +296,7 @@ unsigned char *outdata;
 			if (--i == 0) break;
 			}
 		}
-	i=(int)len&0x07;
+	i=len&0x07;
 	if (i)
 		{
 		for (;;)