RC4 tune-up for Intel P4 core, both 32- and 64-bit ones. As it's

apparently impossible to compose blended code with would perform
satisfactory on all x86 and x86_64 cores, an extra RC4_CHAR
code-path is introduced and P4 core is detected at run-time. This
way we keep original performance on non-P4 implementations and
turbo-charge P4 performance by factor of 2.8x (on 32-bit core).

4 files changed, 117 insertions, 19 deletions

diff --git a/crypto/rc4/asm/rc4-586.pl b/crypto/rc4/asm/rc4-586.pl
index 7ef889e5a1..dbe3803f55 100644
--- a/crypto/rc4/asm/rc4-586.pl
+++ b/crypto/rc4/asm/rc4-586.pl
@@ -7,10 +7,10 @@ require "x86asm.pl";
 
 &asm_init($ARGV[0],"rc4-586.pl");
 
-$tx="eax";
-$ty="ebx";
-$x="ecx";
-$y="edx";
+$x="eax";
+$y="ebx";
+$tx="ecx";
+$ty="edx";
 $in="esi";
 $out="edi";
 $d="ebp";
@@ -31,7 +31,7 @@ sub RC4_loop
 			{
 			 &mov($ty,	&swtmp(2));
 			&cmp($ty,	$in);
-			 &jle(&label("finished"));
+			 &jbe(&label("finished"));
 			&inc($in);
 			}
 		else
@@ -39,7 +39,7 @@ sub RC4_loop
 			&add($ty,	8);
 			 &inc($in);
 			&cmp($ty,	$in);
-			 &jl(&label("finished"));
+			 &jb(&label("finished"));
 			&mov(&swtmp(2),	$ty);
 			}
 		}
@@ -88,35 +88,44 @@ sub RC4
 
 	&function_begin_B($name,"");
 
+	&mov($ty,&wparam(1));		# len
+	&cmp($ty,0);
+	&jne(&label("proceed"));
+	&ret();
+	&set_label("proceed");
+
 	&comment("");
 
 	&push("ebp");
 	 &push("ebx");
-	&mov(	$d,	&wparam(0));	# key
-	 &mov(	$ty,	&wparam(1));	# num
 	&push("esi");
 	 &push("edi");
+	&mov(	$d,	&wparam(0));	# key
+	 &mov(	$in,	&wparam(2));
 
 	&mov(	$x,	&DWP(0,$d,"",1));
 	 &mov(	$y,	&DWP(4,$d,"",1));
 
-	&mov(	$in,	&wparam(2));
+	&mov(	$out,	&wparam(3));
 	 &inc(	$x);
 
 	&stack_push(3);	# 3 temp variables
 	 &add(	$d,	8);
 	&and(	$x,		0xff);
 
+	# detect compressed schedule, see commentary section in rc4_skey.c...
+	&cmp(&DWP(256,$d),-1);
+	&je(&label("RC4_CHAR"));
+
 	 &lea(	$ty,	&DWP(-8,$ty,$in));
 
 	# check for 0 length input
 
-	&mov(	$out,	&wparam(3));
 	 &mov(	&swtmp(2),	$ty);	# this is now address to exit at
 	&mov(	$tx,	&DWP(0,$d,$x,4));
 
 	 &cmp(	$ty,	$in);
-	&jl(	&label("end")); # less than 8 bytes
+	&jb(	&label("end")); # less than 8 bytes
 
 	&set_label("start");
 
@@ -148,7 +157,7 @@ sub RC4
 	&mov(	&DWP(-4,$out,"",0),	$tx);
 	 &mov(	$tx,		&DWP(0,$d,$x,4));
 	&cmp($in,	$ty);
-	 &jle(&label("start"));
+	 &jbe(&label("start"));
 
 	&set_label("end");
 
@@ -162,6 +171,32 @@ sub RC4
 	&RC4_loop(5,0,1);
 	&RC4_loop(6,1,1);
 
+	&jmp(&label("finished"));
+
+	&align(16);
+	# this is essentially Intel P4 specific codepath, see rc4_skey.c...
+	&set_label("RC4_CHAR");
+
+	&lea	($ty,&DWP(0,$in,$ty));
+	&mov	(&swtmp(2),$ty);
+
+	# 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,$ty));
+		&xorb	(&LB($ty),&BP(0,$in));
+		&movb	(&BP(0,$out),&LB($ty));
+		&inc	(&LB($x));
+		&inc	($in);
+		&inc	($out);
+		&cmp	($in,&swtmp(2));
+	&jb	(&label("RC4_CHAR_loop"));
+
 	&set_label("finished");
 	&dec(	$x);
 	 &stack_pop(3);
diff --git a/crypto/rc4/asm/rc4-amd64.pl b/crypto/rc4/asm/rc4-amd64.pl
index cc3f0c0231..35e426d561 100755
--- a/crypto/rc4/asm/rc4-amd64.pl
+++ b/crypto/rc4/asm/rc4-amd64.pl
@@ -25,6 +25,13 @@
 # Latter means that if you want to *estimate* what to expect from
 # *your* CPU, 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 is 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...
+
 $output=shift;
 
 $win64a=1 if ($output =~ /win64a.[s|asm]/);
@@ -90,6 +97,8 @@ $code.=<<___;
 	add	\$8,$dat
 	movl	`&PTR("DWORD:-8[$dat]")`,$XX#d
 	movl	`&PTR("DWORD:-4[$dat]")`,$YY#d
+	cmpl	\$-1,`&PTR("DWORD:256[$dat]")`
+	je	.LRC4_CHAR
 	test	\$-8,$len
 	jz	.Lloop1
 .align	16
@@ -167,6 +176,24 @@ $code.=<<___;
 	dec	$len
 	jnz	.Lloop1
 	jmp	.Lexit
+
+.align	16
+.LRC4_CHAR:
+	inc	$XX#b
+	movzb	`&PTR("BYTE:[$dat+$XX]")`,$TX#d
+	add	$TX#b,$YY#b
+	movzb	`&PTR("BYTE:[$dat+$YY]")`,$TY#d
+	movb	$TX#b,`&PTR("BYTE:[$dat+$YY]")`
+	movb	$TY#b,`&PTR("BYTE:[$dat+$XX]")`
+	add	$TX#b,$TY#b
+	movzb	`&PTR("BYTE:[$dat+$TY]")`,$TY#d
+	xorb	`&PTR("BYTE:[$inp]")`,$TY#b
+	movb	$TY#b,`&PTR("BYTE:[$out]")`
+	inc	$inp
+	inc	$out
+	dec	$len
+	jnz	.LRC4_CHAR
+	jmp	.Lexit
 ___
 $code.=<<___ if (defined($win64a));
 RC4	ENDP
@@ -189,6 +216,8 @@ if (defined($win64a)) {
     $code =~ s/mov[bwlq]/mov/gm;
     $code =~ s/movzb/movzx/gm;
     $code =~ s/repret/DB\t0F3h,0C3h/gm;
+    $code =~ s/cmpl/cmp/gm;
+    $code =~ s/xorb/xor/gm;
 } else {
     $code =~ s/([QD]*WORD|BYTE)://gm;
     $code =~ s/repret/.byte\t0xF3,0xC3/gm;
diff --git a/crypto/rc4/rc4_locl.h b/crypto/rc4/rc4_locl.h
index 3bb80b6ce9..c712e1632e 100644
--- a/crypto/rc4/rc4_locl.h
+++ b/crypto/rc4/rc4_locl.h
@@ -1,4 +1,5 @@
 #ifndef HEADER_RC4_LOCL_H
 #define HEADER_RC4_LOCL_H
 #include <openssl/opensslconf.h>
+#include <cryptlib.h>
 #endif
diff --git a/crypto/rc4/rc4_skey.c b/crypto/rc4/rc4_skey.c
index bb10c1ebe2..781ff2d8b9 100644
--- a/crypto/rc4/rc4_skey.c
+++ b/crypto/rc4/rc4_skey.c
@@ -93,25 +93,58 @@ void RC4_set_key(RC4_KEY *key, int len, const unsigned char *data)
         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, let us [try to] detect P4 at
+		 * run-time by checking upon HTT bit in CPU capability
+		 * vector and set up compressed key schedule, which is
+		 * recognized by correspondingly updated assembler
+		 * module...
+		 *				<appro@fy.chalmers.se>
+		 */
+		if (OPENSSL_ia32cap_P & (1<<28)) {
+			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);
 		}
 	}