X-Git-Url: https://git.librecmc.org/?a=blobdiff_plain;f=crypto%2Frc4%2Frc4_enc.c;h=d5f18a3a707040a30801dcc7344d001bb5b214fb;hb=c5e8580e7b6edb6626fca6c42e91f9e3992d1f88;hp=de57a970bf81013015bdab6b576fbe207b4e7a22;hpb=eda1f21f1af8b6f77327e7b37573af9c1ba73726;p=oweals%2Fopenssl.git diff --git a/crypto/rc4/rc4_enc.c b/crypto/rc4/rc4_enc.c index de57a970bf..d5f18a3a70 100644 --- a/crypto/rc4/rc4_enc.c +++ b/crypto/rc4/rc4_enc.c @@ -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 @@ -56,35 +56,8 @@ * [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 +#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, const unsigned char *indata, + unsigned char *outdata) { register RC4_INT *d; register RC4_INT x,y,tx,ty; @@ -141,6 +78,190 @@ unsigned char *outdata; 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". + * + * + */ + +# 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 ( ( ((unsigned long)indata & (sizeof(RC4_CHUNK)-1)) | + ((unsigned long)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 + * 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 + * + * + */ + if (!is_endian.little) + { /* BIG-ENDIAN CASE */ +# define BESHFT(c) (((sizeof(RC4_CHUNK)-(c)-1)*8)&(sizeof(RC4_CHUNK)*8-1)) + for (;len&-sizeof(RC4_CHUNK);len-=sizeof(RC4_CHUNK)) + { + ichunk = *(RC4_CHUNK *)indata; + otp = RC4_STEP<x=x; + key->y=y; + return; + } + else + { /* LITTLE-ENDIAN CASE */ +# define LESHFT(c) (((c)*8)&(sizeof(RC4_CHUNK)*8-1)) + for (;len&-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<>= (sizeof(RC4_CHUNK)-len)<<3; + switch (len&(sizeof(RC4_CHUNK)-1)) + { + case 7: otp = RC4_STEP, i+=8; + case 6: otp |= RC4_STEP<x=x; + key->y=y; + return; + } + } +#endif #define LOOP(in,out) \ x=((x+1)&0xff); \ tx=d[x]; \ @@ -155,7 +276,6 @@ unsigned char *outdata; #define RC4_LOOP(a,b,i) LOOP(a[i],b[i]) #endif - i= -(int)len; i=(int)(len>>3L); if (i) {