5 files changed, 302 insertions, 60 deletions

diff --git a/src/lib/libcrypto/rc4/asm/rc4-586.pl b/src/lib/libcrypto/rc4/asm/rc4-586.pl
index d6e98f0811..ef7eee766c 100644
--- a/src/lib/libcrypto/rc4/asm/rc4-586.pl
+++ b/src/lib/libcrypto/rc4/asm/rc4-586.pl
@@ -200,22 +200,23 @@ sub RC4
 
         &lea    ($ty,&DWP(0,$in,$ty));
         &mov    (&swtmp(2),$ty);
+        &movz   ($tx,&BP(0,$d,$x));
 
         # strangely enough unrolled loop performs over 20% slower...
         &set_label("RC4_CHAR_loop");
-                &movz   ($tx,&BP(0,$d,$x));
                 &add    (&LB($y),&LB($tx));
                 &movz   ($ty,&BP(0,$d,$y));
                 &movb   (&BP(0,$d,$y),&LB($tx));
                 &movb   (&BP(0,$d,$x),&LB($ty));
                 &add    (&LB($ty),&LB($tx));
                 &movz   ($ty,&BP(0,$d,$ty));
+                &add    (&LB($x),1);
                 &xorb   (&LB($ty),&BP(0,$in));
-                &movb   (&BP(0,$out),&LB($ty));
-                &inc    (&LB($x));
-                &inc    ($in);
-                &inc    ($out);
+                &lea    ($in,&DWP(1,$in));
+                &movz   ($tx,&BP(0,$d,$x));
                 &cmp    ($in,&swtmp(2));
+                &movb   (&BP(0,$out),&LB($ty));
+                &lea    ($out,&DWP(1,$out));
         &jb     (&label("RC4_CHAR_loop"));
 
         &set_label("finished");
diff --git a/src/lib/libcrypto/rc4/asm/rc4-x86_64.pl b/src/lib/libcrypto/rc4/asm/rc4-x86_64.pl
index b628daca70..2d47320485 100755
--- a/src/lib/libcrypto/rc4/asm/rc4-x86_64.pl
+++ b/src/lib/libcrypto/rc4/asm/rc4-x86_64.pl
@@ -2,29 +2,70 @@
 #
 # ====================================================================
 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
-# project. Rights for redistribution and usage in source and binary
-# forms are granted according to the OpenSSL license.
+# project. The module is, however, dual licensed under OpenSSL and
+# CRYPTOGAMS licenses depending on where you obtain it. For further
+# details see http://www.openssl.org/~appro/cryptogams/.
 # ====================================================================
 #
-# Unlike 0.9.7f this code expects RC4_CHAR back in config line! See
-# commentary section in corresponding script in development branch
-# for background information about this option carousel. For those
-# who don't have energy to figure out these gory details, here is
-# basis in form of performance matrix relative to the original
-# 0.9.7e C code-base:
-#
-#               0.9.7e  0.9.7f  this
-# AMD64         1x      3.3x    2.4x
-# EM64T         1x      0.8x    1.5x
-#
-# In other words idea is to trade -25% AMD64 performance to compensate
-# for deterioration and gain +90% on EM64T core. Development branch
-# maintains best performance for either target, i.e. 3.3x for AMD64
-# and 1.5x for EM64T.
+# 2.22x RC4 tune-up:-) It should be noted though that my hand [as in
+# "hand-coded assembler"] doesn't stand for the whole improvement
+# coefficient. It turned out that eliminating RC4_CHAR from config
+# line results in ~40% improvement (yes, even for C implementation).
+# Presumably it has everything to do with AMD cache architecture and
+# RAW or whatever penalties. Once again! The module *requires* config
+# line *without* RC4_CHAR! As for coding "secret," I bet on partial
+# register arithmetics. For example instead of 'inc %r8; and $255,%r8'
+# I simply 'inc %r8b'. Even though optimization manual discourages
+# to operate on partial registers, it turned out to be the best bet.
+# At least for AMD... How IA32E would perform remains to be seen...
+
+# As was shown by Marc Bevand reordering of couple of load operations
+# results in even higher performance gain of 3.3x:-) At least on
+# Opteron... For reference, 1x in this case is RC4_CHAR C-code
+# compiled with gcc 3.3.2, which performs at ~54MBps per 1GHz clock.
+# Latter means that if you want to *estimate* what to expect from
+# *your* Opteron, then multiply 54 by 3.3 and clock frequency in GHz.
+
+# Intel P4 EM64T core was found to run the AMD64 code really slow...
+# The only way to achieve comparable performance on P4 was to keep
+# RC4_CHAR. Kind of ironic, huh? As it's apparently impossible to
+# compose blended code, which would perform even within 30% marginal
+# on either AMD and Intel platforms, I implement both cases. See
+# rc4_skey.c for further details...
+
+# P4 EM64T core appears to be "allergic" to 64-bit inc/dec. Replacing 
+# those with add/sub results in 50% performance improvement of folded
+# loop...
+
+# As was shown by Zou Nanhai loop unrolling can improve Intel EM64T
+# performance by >30% [unlike P4 32-bit case that is]. But this is
+# provided that loads are reordered even more aggressively! Both code
+# pathes, AMD64 and EM64T, reorder loads in essentially same manner
+# as my IA-64 implementation. On Opteron this resulted in modest 5%
+# improvement [I had to test it], while final Intel P4 performance
+# achieves respectful 432MBps on 2.8GHz processor now. For reference.
+# If executed on Xeon, current RC4_CHAR code-path is 2.7x faster than
+# RC4_INT code-path. While if executed on Opteron, it's only 25%
+# slower than the RC4_INT one [meaning that if CPU µ-arch detection
+# is not implemented, then this final RC4_CHAR code-path should be
+# preferred, as it provides better *all-round* performance].
+
+# Intel Core2 was observed to perform poorly on both code paths:-( It
+# apparently suffers from some kind of partial register stall, which
+# occurs in 64-bit mode only [as virtually identical 32-bit loop was
+# observed to outperform 64-bit one by almost 50%]. Adding two movzb to
+# cloop1 boosts its performance by 80%! This loop appears to be optimal
+# fit for Core2 and therefore the code was modified to skip cloop8 on
+# this CPU.
 
 $output=shift;
 
-open STDOUT,">$output" || die "can't open $output: $!";
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
+( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
+( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
+die "can't locate x86_64-xlate.pl";
+
+open STDOUT,"| $^X $xlate $output";
 
 $dat="%rdi";        # arg1
 $len="%rsi";        # arg2
@@ -36,29 +77,101 @@ $out="%rcx";	    # arg4
 $YY="%r12";
 $TY="%r13";
 
-$code=<<___;;
+$code=<<___;
 .text
 
 .globl  RC4
-.type   RC4,\@function
+.type   RC4,\@function,4
 .align  16
 RC4:    or      $len,$len
         jne     .Lentry
-        repret
+        ret
 .Lentry:
         push    %r12
         push    %r13
 
-        add     \$2,$dat
-        movzb   -2($dat),$XX[0]#d
-        movzb   -1($dat),$YY#d
+        add     \$8,$dat
+        movl    -8($dat),$XX[0]#d
+        movl    -4($dat),$YY#d
+        cmpl    \$-1,256($dat)
+        je      .LRC4_CHAR
+        inc     $XX[0]#b
+        movl    ($dat,$XX[0],4),$TX[0]#d
+        test    \$-8,$len
+        jz      .Lloop1
+        jmp     .Lloop8
+.align  16
+.Lloop8:
+___
+for ($i=0;$i<8;$i++) {
+$code.=<<___;
+        add     $TX[0]#b,$YY#b
+        mov     $XX[0],$XX[1]
+        movl    ($dat,$YY,4),$TY#d
+        ror     \$8,%rax                        # ror is redundant when $i=0
+        inc     $XX[1]#b
+        movl    ($dat,$XX[1],4),$TX[1]#d
+        cmp     $XX[1],$YY
+        movl    $TX[0]#d,($dat,$YY,4)
+        cmove   $TX[0],$TX[1]
+        movl    $TY#d,($dat,$XX[0],4)
+        add     $TX[0]#b,$TY#b
+        movb    ($dat,$TY,4),%al
+___
+push(@TX,shift(@TX)); push(@XX,shift(@XX));     # "rotate" registers
+}
+$code.=<<___;
+        ror     \$8,%rax
+        sub     \$8,$len
+
+        xor     ($inp),%rax
+        add     \$8,$inp
+        mov     %rax,($out)
+        add     \$8,$out
 
+        test    \$-8,$len
+        jnz     .Lloop8
+        cmp     \$0,$len
+        jne     .Lloop1
+___
+$code.=<<___;
+.Lexit:
+        sub     \$1,$XX[0]#b
+        movl    $XX[0]#d,-8($dat)
+        movl    $YY#d,-4($dat)
+
+        pop     %r13
+        pop     %r12
+        ret
+.align  16
+.Lloop1:
+        add     $TX[0]#b,$YY#b
+        movl    ($dat,$YY,4),$TY#d
+        movl    $TX[0]#d,($dat,$YY,4)
+        movl    $TY#d,($dat,$XX[0],4)
+        add     $TY#b,$TX[0]#b
+        inc     $XX[0]#b
+        movl    ($dat,$TX[0],4),$TY#d
+        movl    ($dat,$XX[0],4),$TX[0]#d
+        xorb    ($inp),$TY#b
+        inc     $inp
+        movb    $TY#b,($out)
+        inc     $out
+        dec     $len
+        jnz     .Lloop1
+        jmp     .Lexit
+
+.align  16
+.LRC4_CHAR:
         add     \$1,$XX[0]#b
         movzb   ($dat,$XX[0]),$TX[0]#d
         test    \$-8,$len
         jz      .Lcloop1
+        cmp     \$0,260($dat)
+        jnz     .Lcloop1
         push    %rbx
-.align  16      # incidentally aligned already
+        jmp     .Lcloop8
+.align  16
 .Lcloop8:
         mov     ($inp),%eax
         mov     4($inp),%ebx
@@ -114,15 +227,9 @@ $code.=<<___;
         pop     %rbx
         cmp     \$0,$len
         jne     .Lcloop1
-.Lexit:
-        sub     \$1,$XX[0]#b
-        movb    $XX[0]#b,-2($dat)
-        movb    $YY#b,-1($dat)
-
-        pop     %r13
-        pop     %r12
-        repret
-
+        jmp     .Lexit
+___
+$code.=<<___;
 .align  16
 .Lcloop1:
         add     $TX[0]#b,$YY#b
@@ -131,6 +238,8 @@ $code.=<<___;
         movb    $TY#b,($dat,$XX[0])
         add     $TX[0]#b,$TY#b
         add     \$1,$XX[0]#b
+        movzb   $TY#b,$TY#d
+        movzb   $XX[0]#b,$XX[0]#d
         movzb   ($dat,$TY),$TY#d
         movzb   ($dat,$XX[0]),$TX[0]#d
         xorb    ($inp),$TY#b
@@ -143,8 +252,113 @@ $code.=<<___;
 .size   RC4,.-RC4
 ___
 
-$code =~ s/#([bwd])/$1/gm;
+$idx="%r8";
+$ido="%r9";
+
+$code.=<<___;
+.extern OPENSSL_ia32cap_P
+.globl  RC4_set_key
+.type   RC4_set_key,\@function,3
+.align  16
+RC4_set_key:
+        lea     8($dat),$dat
+        lea     ($inp,$len),$inp
+        neg     $len
+        mov     $len,%rcx
+        xor     %eax,%eax
+        xor     $ido,$ido
+        xor     %r10,%r10
+        xor     %r11,%r11
 
-$code =~ s/repret/.byte\t0xF3,0xC3/gm;
+        mov     OPENSSL_ia32cap_P(%rip),$idx#d
+        bt      \$20,$idx#d
+        jnc     .Lw1stloop
+        bt      \$30,$idx#d
+        setc    $ido#b
+        mov     $ido#d,260($dat)
+        jmp     .Lc1stloop
+
+.align  16
+.Lw1stloop:
+        mov     %eax,($dat,%rax,4)
+        add     \$1,%al
+        jnc     .Lw1stloop
+
+        xor     $ido,$ido
+        xor     $idx,$idx
+.align  16
+.Lw2ndloop:
+        mov     ($dat,$ido,4),%r10d
+        add     ($inp,$len,1),$idx#b
+        add     %r10b,$idx#b
+        add     \$1,$len
+        mov     ($dat,$idx,4),%r11d
+        cmovz   %rcx,$len
+        mov     %r10d,($dat,$idx,4)
+        mov     %r11d,($dat,$ido,4)
+        add     \$1,$ido#b
+        jnc     .Lw2ndloop
+        jmp     .Lexit_key
+
+.align  16
+.Lc1stloop:
+        mov     %al,($dat,%rax)
+        add     \$1,%al
+        jnc     .Lc1stloop
+
+        xor     $ido,$ido
+        xor     $idx,$idx
+.align  16
+.Lc2ndloop:
+        mov     ($dat,$ido),%r10b
+        add     ($inp,$len),$idx#b
+        add     %r10b,$idx#b
+        add     \$1,$len
+        mov     ($dat,$idx),%r11b
+        jnz     .Lcnowrap
+        mov     %rcx,$len
+.Lcnowrap:
+        mov     %r10b,($dat,$idx)
+        mov     %r11b,($dat,$ido)
+        add     \$1,$ido#b
+        jnc     .Lc2ndloop
+        movl    \$-1,256($dat)
+
+.align  16
+.Lexit_key:
+        xor     %eax,%eax
+        mov     %eax,-8($dat)
+        mov     %eax,-4($dat)
+        ret
+.size   RC4_set_key,.-RC4_set_key
+
+.globl  RC4_options
+.type   RC4_options,\@function,0
+.align  16
+RC4_options:
+        .picmeup %rax
+        lea     .Lopts-.(%rax),%rax
+        mov     OPENSSL_ia32cap_P(%rip),%edx
+        bt      \$20,%edx
+        jnc     .Ldone
+        add     \$12,%rax
+        bt      \$30,%edx
+        jnc     .Ldone
+        add     \$13,%rax
+.Ldone:
+        ret
+.align  64
+.Lopts:
+.asciz  "rc4(8x,int)"
+.asciz  "rc4(8x,char)"
+.asciz  "rc4(1x,char)"
+.asciz  "RC4 for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
+.align  64
+.size   RC4_options,.-RC4_options
+___
+
+$code =~ s/#([bwd])/$1/gm;
 
 print $code;
+
+close STDOUT;
diff --git a/src/lib/libcrypto/rc4/rc4.h b/src/lib/libcrypto/rc4/rc4.h
index ae0cea75b8..7aec04fe93 100644
--- a/src/lib/libcrypto/rc4/rc4.h
+++ b/src/lib/libcrypto/rc4/rc4.h
@@ -59,12 +59,11 @@
 #ifndef HEADER_RC4_H
 #define HEADER_RC4_H
 
+#include <openssl/opensslconf.h> /* OPENSSL_NO_RC4, RC4_INT */
 #ifdef OPENSSL_NO_RC4
 #error RC4 is disabled.
 #endif
 
-#include <openssl/opensslconf.h> /* RC4_INT */
-
 #ifdef  __cplusplus
 extern "C" {
 #endif
@@ -77,9 +76,6 @@ typedef struct rc4_key_st
 
  
 const char *RC4_options(void);
-#ifdef OPENSSL_FIPS
-void private_RC4_set_key(RC4_KEY *key, int len, const unsigned char *data);
-#endif
 void RC4_set_key(RC4_KEY *key, int len, const unsigned char *data);
 void RC4(RC4_KEY *key, unsigned long len, const unsigned char *indata,
                 unsigned char *outdata);
diff --git a/src/lib/libcrypto/rc4/rc4_enc.c b/src/lib/libcrypto/rc4/rc4_enc.c
index d5f18a3a70..0660ea60a2 100644
--- a/src/lib/libcrypto/rc4/rc4_enc.c
+++ b/src/lib/libcrypto/rc4/rc4_enc.c
@@ -157,7 +157,7 @@ void RC4(RC4_KEY *key, unsigned long len, const unsigned char *indata,
                 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))
+                        for (;len&~(sizeof(RC4_CHUNK)-1);len-=sizeof(RC4_CHUNK))
                                 {
                                 ichunk  = *(RC4_CHUNK *)indata;
                                 otp  = RC4_STEP<<BESHFT(0);
@@ -210,7 +210,7 @@ void RC4(RC4_KEY *key, unsigned long len, const unsigned char *indata,
                 else
                         {       /* LITTLE-ENDIAN CASE */
 # define LESHFT(c)      (((c)*8)&(sizeof(RC4_CHUNK)*8-1))
-                        for (;len&-sizeof(RC4_CHUNK);len-=sizeof(RC4_CHUNK))
+                        for (;len&~(sizeof(RC4_CHUNK)-1);len-=sizeof(RC4_CHUNK))
                                 {
                                 ichunk  = *(RC4_CHUNK *)indata;
                                 otp  = RC4_STEP;
diff --git a/src/lib/libcrypto/rc4/rc4_skey.c b/src/lib/libcrypto/rc4/rc4_skey.c
index 60510624fd..46b77ec321 100644
--- a/src/lib/libcrypto/rc4/rc4_skey.c
+++ b/src/lib/libcrypto/rc4/rc4_skey.c
@@ -57,12 +57,10 @@
  */
 
 #include <openssl/rc4.h>
-#include <openssl/crypto.h>
-#include <openssl/fips.h>
 #include "rc4_locl.h"
 #include <openssl/opensslv.h>
 
-const char *RC4_version="RC4" OPENSSL_VERSION_PTEXT;
+const char RC4_version[]="RC4" OPENSSL_VERSION_PTEXT;
 
 const char *RC4_options(void)
         {
@@ -87,7 +85,7 @@ const char *RC4_options(void)
  * Date: Wed, 14 Sep 1994 06:35:31 GMT
  */
 
-FIPS_NON_FIPS_VCIPHER_Init(RC4)
+void RC4_set_key(RC4_KEY *key, int len, const unsigned char *data)
         {
         register RC4_INT tmp;
         register int id1,id2;
@@ -95,26 +93,59 @@ FIPS_NON_FIPS_VCIPHER_Init(RC4)
         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) { \
+#define SK_LOOP(d,n) { \
                 tmp=d[(n)]; \
                 id2 = (data[id1] + tmp + id2) & 0xff; \
                 if (++id1 == len) id1=0; \
                 d[(n)]=d[id2]; \
                 d[id2]=tmp; }
 
+#if defined(OPENSSL_CPUID_OBJ) && !defined(OPENSSL_NO_ASM)
+# if    defined(__i386)   || defined(__i386__)   || defined(_M_IX86) || \
+        defined(__INTEL__) || \
+        defined(__x86_64) || defined(__x86_64__) || defined(_M_AMD64)
+        if (sizeof(RC4_INT) > 1) {
+                /*
+                 * Unlike all other x86 [and x86_64] implementations,
+                 * Intel P4 core [including EM64T] was found to perform
+                 * poorly with wider RC4_INT. Performance improvement
+                 * for IA-32 hand-coded assembler turned out to be 2.8x
+                 * if re-coded for RC4_CHAR! It's however inappropriate
+                 * to just switch to RC4_CHAR for x86[_64], as non-P4
+                 * implementations suffer from significant performance
+                 * losses then, e.g. PIII exhibits >2x deterioration,
+                 * and so does Opteron. In order to assure optimal
+                 * all-round performance, we detect P4 at run-time by
+                 * checking upon reserved bit 20 in CPU capability
+                 * vector and set up compressed key schedule, which is
+                 * recognized by correspondingly updated assembler
+                 * module... Bit 20 is set up by OPENSSL_ia32_cpuid.
+                 *
+                 *                              <appro@fy.chalmers.se>
+                 */
+                if (OPENSSL_ia32cap_P & (1<<20)) {
+                        unsigned char *cp=(unsigned char *)d;
+
+                        for (i=0;i<256;i++) cp[i]=i;
+                        for (i=0;i<256;i++) SK_LOOP(cp,i);
+                        /* mark schedule as compressed! */
+                        d[256/sizeof(RC4_INT)]=-1;
+                        return;
+                }
+        }
+# endif
+#endif
+        for (i=0; i < 256; i++) d[i]=i;
         for (i=0; i < 256; i+=4)
                 {
-                SK_LOOP(i+0);
-                SK_LOOP(i+1);
-                SK_LOOP(i+2);
-                SK_LOOP(i+3);
+                SK_LOOP(d,i+0);
+                SK_LOOP(d,i+1);
+                SK_LOOP(d,i+2);
+                SK_LOOP(d,i+3);
                 }
         }