Squeeze a bit more speed out of MD5 assembler.

CHANGES

@@ -5,6 +5,10 @@
 
  Changes between 0.9.1c and 0.9.2
 
+  *) Squeeze another 7% of speed out of MD5 assembler, at least on a P2. I'd
+     like to hear about it if this slows down other processors.
+     [Ben Laurie]
+
   *) Add CygWin32 platform information to Configure script.
      [Alan Batie <batie@aahz.jf.intel.com>]
 

crypto/md5/Makefile.ssl

@@ -61,7 +61,7 @@ asm/mx86-out.o: asm/mx86unix.cpp
 asm/mx86bsdi.o: asm/mx86unix.cpp
 	$(CPP) -DBSDI asm/mx86unix.cpp | sed 's/ :/:/' | as -o asm/mx86bsdi.o
 
-asm/mx86unix.cpp:
+asm/mx86unix.cpp: asm/md5-586.pl
 	(cd asm; perl md5-586.pl cpp >mx86unix.cpp)
 
 files:

crypto/md5/asm/md5-586.pl

@@ -44,7 +44,7 @@ sub R0
 	local($pos,$a,$b,$c,$d,$K,$ki,$s,$t)=@_;
 
 	&mov($tmp1,$C)  if $pos < 0;
-	 &mov($tmp2,&DWP($xo[$ki]*4,$K,"",0)) if $pos < 0; # very first one 
+	&mov($tmp2,&DWP($xo[$ki]*4,$K,"",0)) if $pos < 0; # very first one 
 
 	# body proper
 
@@ -54,7 +54,6 @@ sub R0
 	&and($tmp1,$b); # F function - part 3
 	&lea($a,&DWP($t,$a,$tmp2,1));
 
-	&mov($tmp2,&DWP($xo[$ki+1]*4,$K,"",0)) if ($pos != 2);
 	&xor($tmp1,$d); # F function - part 4
 
 	&add($a,$tmp1);
@@ -62,8 +61,10 @@ sub R0
 	&mov($tmp1,&Np($c)) if $pos == 1;	# next tmp1 for R1
 
 	&rotl($a,$s);
-	&add($a,$b);
 
+	&mov($tmp2,&DWP($xo[$ki+1]*4,$K,"",0)) if ($pos != 2);
+
+	&add($a,$b);
 	}
 
 sub R1
@@ -100,16 +101,16 @@ if (($n & 1) == 0)
 	# make sure to do 'D' first, not 'B', else we clash with
 	# the last add from the previous round.
 
-	 &xor($tmp1,$d); # H function - part 2
+	&xor($tmp1,$d); # H function - part 2
 
 	&xor($tmp1,$b); # H function - part 3
-	 &lea($a,&DWP($t,$a,$tmp2,1));
+	&lea($a,&DWP($t,$a,$tmp2,1));
 
 	&add($a,$tmp1);
-	 &mov($tmp2,&DWP($xo[$ki+1]*4,$K,"",0));
 
 	&rotl($a,$s);
 
+	&mov($tmp2,&DWP($xo[$ki+1]*4,$K,"",0));
 	&mov($tmp1,&Np($c));
 	}
 else
@@ -118,17 +119,17 @@ else
 	# make sure to do 'D' first, not 'B', else we clash with
 	# the last add from the previous round.
 
-	 &lea($a,&DWP($t,$a,$tmp2,1));
+	&lea($a,&DWP($t,$a,$tmp2,1));
 
 	&add($b,$c);			# MOVED FORWARD
-	 &xor($tmp1,$d); # H function - part 2
+	&xor($tmp1,$d); # H function - part 2
 
 	&xor($tmp1,$b); # H function - part 3
-	 &mov($tmp2,&DWP($xo[$ki+1]*4,$K,"",0)) if ($pos != 2);
+	&mov($tmp2,&DWP($xo[$ki+1]*4,$K,"",0)) if ($pos != 2);
 
 	&add($a,$tmp1);
-	 &mov($tmp1,&Np($c)) if $pos < 1;	# H function - part 1
-	 &mov($tmp1,-1) if $pos == 1;		# I function - part 1
+	&mov($tmp1,&Np($c)) if $pos < 1;	# H function - part 1
+	&mov($tmp1,-1) if $pos == 1;		# I function - part 1
 
 	&rotl($a,$s);
 
@@ -146,21 +147,21 @@ sub R3
 	&xor($tmp1,$d) if $pos < 0; 	# I function - part 2
 
 	&or($tmp1,$b);				# I function - part 3
-	 &lea($a,&DWP($t,$a,$tmp2,1));
+	&lea($a,&DWP($t,$a,$tmp2,1));
 
 	&xor($tmp1,$c); 			# I function - part 4
-	 &mov($tmp2,&DWP($xo[$ki+1]*4,$K,"",0))	if $pos != 2; # load X/k value
-	 &mov($tmp2,&wparam(0)) if $pos == 2;
+	&mov($tmp2,&DWP($xo[$ki+1]*4,$K,"",0))	if $pos != 2; # load X/k value
+	&mov($tmp2,&wparam(0)) if $pos == 2;
 
 	&add($a,$tmp1);
-	 &mov($tmp1,-1) if $pos < 1;	# H function - part 1
-	 &add($K,64) if $pos >=1 && !$normal;
+	&mov($tmp1,-1) if $pos < 1;	# H function - part 1
+	&add($K,64) if $pos >=1 && !$normal;
 
 	&rotl($a,$s);
 
 	&xor($tmp1,&Np($d)) if $pos <= 0; 	# I function - part = first time
 	&mov($tmp1,&DWP( 0,$tmp2,"",0)) if $pos > 0;
-	 &add($a,$b);
+	&add($a,$b);
 	}
 
 

crypto/md5/asm/mx86unix.cpp

-/* Run the C pre-processor over this file with one of the following defined
- * ELF - elf object files,
- * OUT - a.out object files,
- * BSDI - BSDI style a.out object files
- * SOL - Solaris style elf
- */
-
-#define TYPE(a,b)       .type   a,b
-#define SIZE(a,b)       .size   a,b
-
-#if defined(OUT) || defined(BSDI)
-#define md5_block_x86 _md5_block_x86
-
-#endif
-
-#ifdef OUT
-#define OK	1
-#define ALIGN	4
-#endif
-
-#ifdef BSDI
-#define OK              1
-#define ALIGN           4
-#undef SIZE
-#undef TYPE
-#define SIZE(a,b)
-#define TYPE(a,b)
-#endif
-
-#if defined(ELF) || defined(SOL)
-#define OK              1
-#define ALIGN           16
-#endif
-
-#ifndef OK
-You need to define one of
-ELF - elf systems - linux-elf, NetBSD and DG-UX
-OUT - a.out systems - linux-a.out and FreeBSD
-SOL - solaris systems, which are elf with strange comment lines
-BSDI - a.out with a very primative version of as.
-#endif
-
-/* Let the Assembler begin :-) */
-	/* Don't even think of reading this code */
-	/* It was automatically generated by md5-586.pl */
-	/* Which is a perl program used to generate the x86 assember for */
-	/* any of elf, a.out, BSDI,Win32, or Solaris */
-	/* eric <eay@cryptsoft.com> */
-
-	.file	"md5-586.s"
-	.version	"01.01"
-gcc2_compiled.:
-.text
-	.align ALIGN
-.globl md5_block_x86
-	TYPE(md5_block_x86,@function)
-md5_block_x86:
-	pushl	%esi
-	pushl	%edi
-	movl	12(%esp),	%edi
-	movl	16(%esp),	%esi
-	movl	20(%esp),	%ecx
-	pushl	%ebp
-	pushl	%ebx
-	addl	%esi,		%ecx
-	subl	$64,		%ecx
-	movl	(%edi),		%eax
-	pushl	%ecx
-	movl	4(%edi),	%ebx
-	movl	8(%edi),	%ecx
-	movl	12(%edi),	%edx
-.L000start:
-
-	/* R0 section */
-	movl	%ecx,		%edi
-	movl	(%esi),		%ebp
-	/* R0 0 */
-	xorl	%edx,		%edi
-	andl	%ebx,		%edi
-	leal	3614090360(%eax,%ebp,1),%eax
-	movl	4(%esi),	%ebp
-	xorl	%edx,		%edi
-	addl	%edi,		%eax
-	movl	%ebx,		%edi
-	roll	$7,		%eax
-	addl	%ebx,		%eax
-	/* R0 1 */
-	xorl	%ecx,		%edi
-	andl	%eax,		%edi
-	leal	3905402710(%edx,%ebp,1),%edx
-	movl	8(%esi),	%ebp
-	xorl	%ecx,		%edi
-	addl	%edi,		%edx
-	movl	%eax,		%edi
-	roll	$12,		%edx
-	addl	%eax,		%edx
-	/* R0 2 */
-	xorl	%ebx,		%edi
-	andl	%edx,		%edi
-	leal	606105819(%ecx,%ebp,1),%ecx
-	movl	12(%esi),	%ebp
-	xorl	%ebx,		%edi
-	addl	%edi,		%ecx
-	movl	%edx,		%edi
-	roll	$17,		%ecx
-	addl	%edx,		%ecx
-	/* R0 3 */
-	xorl	%eax,		%edi
-	andl	%ecx,		%edi
-	leal	3250441966(%ebx,%ebp,1),%ebx
-	movl	16(%esi),	%ebp
-	xorl	%eax,		%edi
-	addl	%edi,		%ebx
-	movl	%ecx,		%edi
-	roll	$22,		%ebx
-	addl	%ecx,		%ebx
-	/* R0 4 */
-	xorl	%edx,		%edi
-	andl	%ebx,		%edi
-	leal	4118548399(%eax,%ebp,1),%eax
-	movl	20(%esi),	%ebp
-	xorl	%edx,		%edi
-	addl	%edi,		%eax
-	movl	%ebx,		%edi
-	roll	$7,		%eax
-	addl	%ebx,		%eax
-	/* R0 5 */
-	xorl	%ecx,		%edi
-	andl	%eax,		%edi
-	leal	1200080426(%edx,%ebp,1),%edx
-	movl	24(%esi),	%ebp
-	xorl	%ecx,		%edi
-	addl	%edi,		%edx
-	movl	%eax,		%edi
-	roll	$12,		%edx
-	addl	%eax,		%edx
-	/* R0 6 */
-	xorl	%ebx,		%edi
-	andl	%edx,		%edi
-	leal	2821735955(%ecx,%ebp,1),%ecx
-	movl	28(%esi),	%ebp
-	xorl	%ebx,		%edi
-	addl	%edi,		%ecx
-	movl	%edx,		%edi
-	roll	$17,		%ecx
-	addl	%edx,		%ecx
-	/* R0 7 */
-	xorl	%eax,		%edi
-	andl	%ecx,		%edi
-	leal	4249261313(%ebx,%ebp,1),%ebx
-	movl	32(%esi),	%ebp
-	xorl	%eax,		%edi
-	addl	%edi,		%ebx
-	movl	%ecx,		%edi
-	roll	$22,		%ebx
-	addl	%ecx,		%ebx
-	/* R0 8 */
-	xorl	%edx,		%edi
-	andl	%ebx,		%edi
-	leal	1770035416(%eax,%ebp,1),%eax
-	movl	36(%esi),	%ebp
-	xorl	%edx,		%edi
-	addl	%edi,		%eax
-	movl	%ebx,		%edi
-	roll	$7,		%eax
-	addl	%ebx,		%eax
-	/* R0 9 */
-	xorl	%ecx,		%edi
-	andl	%eax,		%edi
-	leal	2336552879(%edx,%ebp,1),%edx
-	movl	40(%esi),	%ebp
-	xorl	%ecx,		%edi
-	addl	%edi,		%edx
-	movl	%eax,		%edi
-	roll	$12,		%edx
-	addl	%eax,		%edx
-	/* R0 10 */
-	xorl	%ebx,		%edi
-	andl	%edx,		%edi
-	leal	4294925233(%ecx,%ebp,1),%ecx
-	movl	44(%esi),	%ebp
-	xorl	%ebx,		%edi
-	addl	%edi,		%ecx
-	movl	%edx,		%edi
-	roll	$17,		%ecx
-	addl	%edx,		%ecx
-	/* R0 11 */
-	xorl	%eax,		%edi
-	andl	%ecx,		%edi
-	leal	2304563134(%ebx,%ebp,1),%ebx
-	movl	48(%esi),	%ebp
-	xorl	%eax,		%edi
-	addl	%edi,		%ebx
-	movl	%ecx,		%edi
-	roll	$22,		%ebx
-	addl	%ecx,		%ebx
-	/* R0 12 */
-	xorl	%edx,		%edi
-	andl	%ebx,		%edi
-	leal	1804603682(%eax,%ebp,1),%eax
-	movl	52(%esi),	%ebp
-	xorl	%edx,		%edi
-	addl	%edi,		%eax
-	movl	%ebx,		%edi
-	roll	$7,		%eax
-	addl	%ebx,		%eax
-	/* R0 13 */
-	xorl	%ecx,		%edi
-	andl	%eax,		%edi
-	leal	4254626195(%edx,%ebp,1),%edx
-	movl	56(%esi),	%ebp
-	xorl	%ecx,		%edi
-	addl	%edi,		%edx
-	movl	%eax,		%edi
-	roll	$12,		%edx
-	addl	%eax,		%edx
-	/* R0 14 */
-	xorl	%ebx,		%edi
-	andl	%edx,		%edi
-	leal	2792965006(%ecx,%ebp,1),%ecx
-	movl	60(%esi),	%ebp
-	xorl	%ebx,		%edi
-	addl	%edi,		%ecx
-	movl	%edx,		%edi
-	roll	$17,		%ecx
-	addl	%edx,		%ecx
-	/* R0 15 */
-	xorl	%eax,		%edi
-	andl	%ecx,		%edi
-	leal	1236535329(%ebx,%ebp,1),%ebx
-	movl	4(%esi),	%ebp
-	xorl	%eax,		%edi
-	addl	%edi,		%ebx
-	movl	%ecx,		%edi
-	roll	$22,		%ebx
-	addl	%ecx,		%ebx
-
-	/* R1 section */
-	/* R1 16 */
-	leal	4129170786(%eax,%ebp,1),%eax
-	xorl	%ebx,		%edi
-	andl	%edx,		%edi
-	movl	24(%esi),	%ebp
-	xorl	%ecx,		%edi
-	addl	%edi,		%eax
-	movl	%ebx,		%edi
-	roll	$5,		%eax
-	addl	%ebx,		%eax
-	/* R1 17 */
-	leal	3225465664(%edx,%ebp,1),%edx
-	xorl	%eax,		%edi
-	andl	%ecx,		%edi
-	movl	44(%esi),	%ebp
-	xorl	%ebx,		%edi
-	addl	%edi,		%edx
-	movl	%eax,		%edi
-	roll	$9,		%edx
-	addl	%eax,		%edx
-	/* R1 18 */
-	leal	643717713(%ecx,%ebp,1),%ecx
-	xorl	%edx,		%edi
-	andl	%ebx,		%edi
-	movl	(%esi),		%ebp
-	xorl	%eax,		%edi
-	addl	%edi,		%ecx
-	movl	%edx,		%edi
-	roll	$14,		%ecx
-	addl	%edx,		%ecx
-	/* R1 19 */
-	leal	3921069994(%ebx,%ebp,1),%ebx
-	xorl	%ecx,		%edi
-	andl	%eax,		%edi
-	movl	20(%esi),	%ebp
-	xorl	%edx,		%edi
-	addl	%edi,		%ebx
-	movl	%ecx,		%edi
-	roll	$20,		%ebx
-	addl	%ecx,		%ebx
-	/* R1 20 */
-	leal	3593408605(%eax,%ebp,1),%eax
-	xorl	%ebx,		%edi
-	andl	%edx,		%edi
-	movl	40(%esi),	%ebp
-	xorl	%ecx,		%edi
-	addl	%edi,		%eax
-	movl	%ebx,		%edi
-	roll	$5,		%eax
-	addl	%ebx,		%eax
-	/* R1 21 */
-	leal	38016083(%edx,%ebp,1),%edx
-	xorl	%eax,		%edi
-	andl	%ecx,		%edi
-	movl	60(%esi),	%ebp
-	xorl	%ebx,		%edi
-	addl	%edi,		%edx
-	movl	%eax,		%edi
-	roll	$9,		%edx
-	addl	%eax,		%edx
-	/* R1 22 */
-	leal	3634488961(%ecx,%ebp,1),%ecx
-	xorl	%edx,		%edi
-	andl	%ebx,		%edi
-	movl	16(%esi),	%ebp
-	xorl	%eax,		%edi
-	addl	%edi,		%ecx
-	movl	%edx,		%edi
-	roll	$14,		%ecx
-	addl	%edx,		%ecx
-	/* R1 23 */
-	leal	3889429448(%ebx,%ebp,1),%ebx
-	xorl	%ecx,		%edi
-	andl	%eax,		%edi
-	movl	36(%esi),	%ebp
-	xorl	%edx,		%edi
-	addl	%edi,		%ebx
-	movl	%ecx,		%edi
-	roll	$20,		%ebx
-	addl	%ecx,		%ebx
-	/* R1 24 */
-	leal	568446438(%eax,%ebp,1),%eax
-	xorl	%ebx,		%edi
-	andl	%edx,		%edi
-	movl	56(%esi),	%ebp
-	xorl	%ecx,		%edi
-	addl	%edi,		%eax
-	movl	%ebx,		%edi
-	roll	$5,		%eax
-	addl	%ebx,		%eax
-	/* R1 25 */
-	leal	3275163606(%edx,%ebp,1),%edx
-	xorl	%eax,		%edi
-	andl	%ecx,		%edi
-	movl	12(%esi),	%ebp
-	xorl	%ebx,		%edi
-	addl	%edi,		%edx
-	movl	%eax,		%edi
-	roll	$9,		%edx
-	addl	%eax,		%edx
-	/* R1 26 */
-	leal	4107603335(%ecx,%ebp,1),%ecx
-	xorl	%edx,		%edi
-	andl	%ebx,		%edi
-	movl	32(%esi),	%ebp
-	xorl	%eax,		%edi
-	addl	%edi,		%ecx
-	movl	%edx,		%edi
-	roll	$14,		%ecx
-	addl	%edx,		%ecx
-	/* R1 27 */
-	leal	1163531501(%ebx,%ebp,1),%ebx
-	xorl	%ecx,		%edi
-	andl	%eax,		%edi
-	movl	52(%esi),	%ebp
-	xorl	%edx,		%edi
-	addl	%edi,		%ebx
-	movl	%ecx,		%edi
-	roll	$20,		%ebx
-	addl	%ecx,		%ebx
-	/* R1 28 */
-	leal	2850285829(%eax,%ebp,1),%eax
-	xorl	%ebx,		%edi
-	andl	%edx,		%edi
-	movl	8(%esi),	%ebp
-	xorl	%ecx,		%edi
-	addl	%edi,		%eax
-	movl	%ebx,		%edi
-	roll	$5,		%eax
-	addl	%ebx,		%eax
-	/* R1 29 */
-	leal	4243563512(%edx,%ebp,1),%edx
-	xorl	%eax,		%edi
-	andl	%ecx,		%edi
-	movl	28(%esi),	%ebp
-	xorl	%ebx,		%edi
-	addl	%edi,		%edx
-	movl	%eax,		%edi
-	roll	$9,		%edx
-	addl	%eax,		%edx
-	/* R1 30 */
-	leal	1735328473(%ecx,%ebp,1),%ecx
-	xorl	%edx,		%edi
-	andl	%ebx,		%edi
-	movl	48(%esi),	%ebp
-	xorl	%eax,		%edi
-	addl	%edi,		%ecx
-	movl	%edx,		%edi
-	roll	$14,		%ecx
-	addl	%edx,		%ecx
-	/* R1 31 */
-	leal	2368359562(%ebx,%ebp,1),%ebx
-	xorl	%ecx,		%edi
-	andl	%eax,		%edi
-	movl	20(%esi),	%ebp
-	xorl	%edx,		%edi
-	addl	%edi,		%ebx
-	movl	%ecx,		%edi
-	roll	$20,		%ebx
-	addl	%ecx,		%ebx
-
-	/* R2 section */
-	/* R2 32 */
-	xorl	%edx,		%edi
-	xorl	%ebx,		%edi
-	leal	4294588738(%eax,%ebp,1),%eax
-	addl	%edi,		%eax
-	movl	32(%esi),	%ebp
-	roll	$4,		%eax
-	movl	%ebx,		%edi
-	/* R2 33 */
-	leal	2272392833(%edx,%ebp,1),%edx
-	addl	%ebx,		%eax
-	xorl	%ecx,		%edi
-	xorl	%eax,		%edi
-	movl	44(%esi),	%ebp
-	addl	%edi,		%edx
-	movl	%eax,		%edi
-	roll	$11,		%edx
-	addl	%eax,		%edx
-	/* R2 34 */
-	xorl	%ebx,		%edi
-	xorl	%edx,		%edi
-	leal	1839030562(%ecx,%ebp,1),%ecx
-	addl	%edi,		%ecx
-	movl	56(%esi),	%ebp
-	roll	$16,		%ecx
-	movl	%edx,		%edi
-	/* R2 35 */
-	leal	4259657740(%ebx,%ebp,1),%ebx
-	addl	%edx,		%ecx
-	xorl	%eax,		%edi
-	xorl	%ecx,		%edi
-	movl	4(%esi),	%ebp
-	addl	%edi,		%ebx
-	movl	%ecx,		%edi
-	roll	$23,		%ebx
-	addl	%ecx,		%ebx
-	/* R2 36 */
-	xorl	%edx,		%edi
-	xorl	%ebx,		%edi
-	leal	2763975236(%eax,%ebp,1),%eax
-	addl	%edi,		%eax
-	movl	16(%esi),	%ebp
-	roll	$4,		%eax
-	movl	%ebx,		%edi
-	/* R2 37 */
-	leal	1272893353(%edx,%ebp,1),%edx
-	addl	%ebx,		%eax
-	xorl	%ecx,		%edi
-	xorl	%eax,		%edi
-	movl	28(%esi),	%ebp
-	addl	%edi,		%edx
-	movl	%eax,		%edi
-	roll	$11,		%edx
-	addl	%eax,		%edx
-	/* R2 38 */
-	xorl	%ebx,		%edi
-	xorl	%edx,		%edi
-	leal	4139469664(%ecx,%ebp,1),%ecx
-	addl	%edi,		%ecx
-	movl	40(%esi),	%ebp
-	roll	$16,		%ecx
-	movl	%edx,		%edi
-	/* R2 39 */
-	leal	3200236656(%ebx,%ebp,1),%ebx
-	addl	%edx,		%ecx
-	xorl	%eax,		%edi
-	xorl	%ecx,		%edi
-	movl	52(%esi),	%ebp
-	addl	%edi,		%ebx
-	movl	%ecx,		%edi
-	roll	$23,		%ebx
-	addl	%ecx,		%ebx
-	/* R2 40 */
-	xorl	%edx,		%edi
-	xorl	%ebx,		%edi
-	leal	681279174(%eax,%ebp,1),%eax
-	addl	%edi,		%eax
-	movl	(%esi),		%ebp
-	roll	$4,		%eax
-	movl	%ebx,		%edi
-	/* R2 41 */
-	leal	3936430074(%edx,%ebp,1),%edx
-	addl	%ebx,		%eax
-	xorl	%ecx,		%edi
-	xorl	%eax,		%edi
-	movl	12(%esi),	%ebp
-	addl	%edi,		%edx
-	movl	%eax,		%edi
-	roll	$11,		%edx
-	addl	%eax,		%edx
-	/* R2 42 */
-	xorl	%ebx,		%edi
-	xorl	%edx,		%edi
-	leal	3572445317(%ecx,%ebp,1),%ecx
-	addl	%edi,		%ecx
-	movl	24(%esi),	%ebp
-	roll	$16,		%ecx
-	movl	%edx,		%edi
-	/* R2 43 */
-	leal	76029189(%ebx,%ebp,1),%ebx
-	addl	%edx,		%ecx
-	xorl	%eax,		%edi
-	xorl	%ecx,		%edi
-	movl	36(%esi),	%ebp
-	addl	%edi,		%ebx
-	movl	%ecx,		%edi
-	roll	$23,		%ebx
-	addl	%ecx,		%ebx
-	/* R2 44 */
-	xorl	%edx,		%edi
-	xorl	%ebx,		%edi
-	leal	3654602809(%eax,%ebp,1),%eax
-	addl	%edi,		%eax
-	movl	48(%esi),	%ebp
-	roll	$4,		%eax
-	movl	%ebx,		%edi
-	/* R2 45 */
-	leal	3873151461(%edx,%ebp,1),%edx
-	addl	%ebx,		%eax
-	xorl	%ecx,		%edi
-	xorl	%eax,		%edi
-	movl	60(%esi),	%ebp
-	addl	%edi,		%edx
-	movl	%eax,		%edi
-	roll	$11,		%edx
-	addl	%eax,		%edx
-	/* R2 46 */
-	xorl	%ebx,		%edi
-	xorl	%edx,		%edi
-	leal	530742520(%ecx,%ebp,1),%ecx
-	addl	%edi,		%ecx
-	movl	8(%esi),	%ebp
-	roll	$16,		%ecx
-	movl	%edx,		%edi
-	/* R2 47 */
-	leal	3299628645(%ebx,%ebp,1),%ebx
-	addl	%edx,		%ecx
-	xorl	%eax,		%edi
-	xorl	%ecx,		%edi
-	movl	(%esi),		%ebp
-	addl	%edi,		%ebx
-	movl	$-1,		%edi
-	roll	$23,		%ebx
-	addl	%ecx,		%ebx
-
-	/* R3 section */
-	/* R3 48 */
-	xorl	%edx,		%edi
-	orl	%ebx,		%edi
-	leal	4096336452(%eax,%ebp,1),%eax
-	xorl	%ecx,		%edi
-	movl	28(%esi),	%ebp
-	addl	%edi,		%eax
-	movl	$-1,		%edi
-	roll	$6,		%eax
-	xorl	%ecx,		%edi
-	addl	%ebx,		%eax
-	/* R3 49 */
-	orl	%eax,		%edi
-	leal	1126891415(%edx,%ebp,1),%edx
-	xorl	%ebx,		%edi
-	movl	56(%esi),	%ebp
-	addl	%edi,		%edx
-	movl	$-1,		%edi
-	roll	$10,		%edx
-	xorl	%ebx,		%edi
-	addl	%eax,		%edx
-	/* R3 50 */
-	orl	%edx,		%edi
-	leal	2878612391(%ecx,%ebp,1),%ecx
-	xorl	%eax,		%edi
-	movl	20(%esi),	%ebp
-	addl	%edi,		%ecx
-	movl	$-1,		%edi
-	roll	$15,		%ecx
-	xorl	%eax,		%edi
-	addl	%edx,		%ecx
-	/* R3 51 */
-	orl	%ecx,		%edi
-	leal	4237533241(%ebx,%ebp,1),%ebx
-	xorl	%edx,		%edi
-	movl	48(%esi),	%ebp
-	addl	%edi,		%ebx
-	movl	$-1,		%edi
-	roll	$21,		%ebx
-	xorl	%edx,		%edi
-	addl	%ecx,		%ebx
-	/* R3 52 */
-	orl	%ebx,		%edi
-	leal	1700485571(%eax,%ebp,1),%eax
-	xorl	%ecx,		%edi
-	movl	12(%esi),	%ebp
-	addl	%edi,		%eax
-	movl	$-1,		%edi
-	roll	$6,		%eax
-	xorl	%ecx,		%edi
-	addl	%ebx,		%eax
-	/* R3 53 */
-	orl	%eax,		%edi
-	leal	2399980690(%edx,%ebp,1),%edx
-	xorl	%ebx,		%edi
-	movl	40(%esi),	%ebp
-	addl	%edi,		%edx
-	movl	$-1,		%edi
-	roll	$10,		%edx
-	xorl	%ebx,		%edi
-	addl	%eax,		%edx
-	/* R3 54 */
-	orl	%edx,		%edi
-	leal	4293915773(%ecx,%ebp,1),%ecx
-	xorl	%eax,		%edi
-	movl	4(%esi),	%ebp
-	addl	%edi,		%ecx
-	movl	$-1,		%edi
-	roll	$15,		%ecx
-	xorl	%eax,		%edi
-	addl	%edx,		%ecx
-	/* R3 55 */
-	orl	%ecx,		%edi
-	leal	2240044497(%ebx,%ebp,1),%ebx
-	xorl	%edx,		%edi
-	movl	32(%esi),	%ebp
-	addl	%edi,		%ebx
-	movl	$-1,		%edi
-	roll	$21,		%ebx
-	xorl	%edx,		%edi
-	addl	%ecx,		%ebx
-	/* R3 56 */
-	orl	%ebx,		%edi
-	leal	1873313359(%eax,%ebp,1),%eax
-	xorl	%ecx,		%edi
-	movl	60(%esi),	%ebp
-	addl	%edi,		%eax
-	movl	$-1,		%edi
-	roll	$6,		%eax
-	xorl	%ecx,		%edi
-	addl	%ebx,		%eax
-	/* R3 57 */
-	orl	%eax,		%edi
-	leal	4264355552(%edx,%ebp,1),%edx
-	xorl	%ebx,		%edi
-	movl	24(%esi),	%ebp
-	addl	%edi,		%edx
-	movl	$-1,		%edi
-	roll	$10,		%edx
-	xorl	%ebx,		%edi
-	addl	%eax,		%edx
-	/* R3 58 */
-	orl	%edx,		%edi
-	leal	2734768916(%ecx,%ebp,1),%ecx
-	xorl	%eax,		%edi
-	movl	52(%esi),	%ebp
-	addl	%edi,		%ecx
-	movl	$-1,		%edi
-	roll	$15,		%ecx
-	xorl	%eax,		%edi
-	addl	%edx,		%ecx
-	/* R3 59 */
-	orl	%ecx,		%edi
-	leal	1309151649(%ebx,%ebp,1),%ebx
-	xorl	%edx,		%edi
-	movl	16(%esi),	%ebp
-	addl	%edi,		%ebx
-	movl	$-1,		%edi
-	roll	$21,		%ebx
-	xorl	%edx,		%edi
-	addl	%ecx,		%ebx
-	/* R3 60 */
-	orl	%ebx,		%edi
-	leal	4149444226(%eax,%ebp,1),%eax
-	xorl	%ecx,		%edi
-	movl	44(%esi),	%ebp
-	addl	%edi,		%eax
-	movl	$-1,		%edi
-	roll	$6,		%eax
-	xorl	%ecx,		%edi
-	addl	%ebx,		%eax
-	/* R3 61 */
-	orl	%eax,		%edi
-	leal	3174756917(%edx,%ebp,1),%edx
-	xorl	%ebx,		%edi
-	movl	8(%esi),	%ebp
-	addl	%edi,		%edx
-	movl	$-1,		%edi
-	roll	$10,		%edx
-	xorl	%ebx,		%edi
-	addl	%eax,		%edx
-	/* R3 62 */
-	orl	%edx,		%edi
-	leal	718787259(%ecx,%ebp,1),%ecx
-	xorl	%eax,		%edi
-	movl	36(%esi),	%ebp
-	addl	%edi,		%ecx
-	movl	$-1,		%edi
-	roll	$15,		%ecx
-	xorl	%eax,		%edi
-	addl	%edx,		%ecx
-	/* R3 63 */
-	orl	%ecx,		%edi
-	leal	3951481745(%ebx,%ebp,1),%ebx
-	xorl	%edx,		%edi
-	movl	24(%esp),	%ebp
-	addl	%edi,		%ebx
-	addl	$64,		%esi
-	roll	$21,		%ebx
-	movl	(%ebp),		%edi
-	addl	%ecx,		%ebx
-	addl	%edi,		%eax
-	movl	4(%ebp),	%edi
-	addl	%edi,		%ebx
-	movl	8(%ebp),	%edi
-	addl	%edi,		%ecx
-	movl	12(%ebp),	%edi
-	addl	%edi,		%edx
-	movl	%eax,		(%ebp)
-	movl	%ebx,		4(%ebp)
-	movl	(%esp),		%edi
-	movl	%ecx,		8(%ebp)
-	movl	%edx,		12(%ebp)
-	cmpl	%esi,		%edi
-	jge	.L000start
-	popl	%eax
-	popl	%ebx
-	popl	%ebp
-	popl	%edi
-	popl	%esi
-	ret
-.md5_block_x86_end:
-	SIZE(md5_block_x86,.md5_block_x86_end-md5_block_x86)
-.ident	"desasm.pl"