md5block_386.s

     1  // Original source:
     2  //	http://www.zorinaq.com/papers/md5-amd64.html
     3  //	http://www.zorinaq.com/papers/md5-amd64.tar.bz2
     4  //
     5  // Translated from Perl generating GNU assembly into
     6  // #defines generating 8a assembly, and adjusted for 386,
     7  // by the Go Authors.
     8  
     9  #include "textflag.h"
    10  
    11  // MD5 optimized for AMD64.
    12  //
    13  // Author: Marc Bevand <bevand_m (at) epita.fr>
    14  // Licence: I hereby disclaim the copyright on this code and place it
    15  // in the public domain.
    16  
    17  #define ROUND1(a, b, c, d, index, const, shift) \
    18  	XORL	c, BP; \
    19  	LEAL	const(a)(DI*1), a; \
    20  	ANDL	b, BP; \
    21  	XORL d, BP; \
    22  	MOVL (index*4)(SI), DI; \
    23  	ADDL BP, a; \
    24  	ROLL $shift, a; \
    25  	MOVL c, BP; \
    26  	ADDL b, a
    27  
    28  #define ROUND2(a, b, c, d, index, const, shift) \
    29  	LEAL	const(a)(DI*1),a; \
    30  	MOVL	d,		DI; \
    31  	ANDL	b,		DI; \
    32  	MOVL	d,		BP; \
    33  	NOTL	BP; \
    34  	ANDL	c,		BP; \
    35  	ORL	DI,		BP; \
    36  	MOVL	(index*4)(SI),DI; \
    37  	ADDL	BP,		a; \
    38  	ROLL	$shift,	a; \
    39  	ADDL	b,		a
    40  
    41  #define ROUND3(a, b, c, d, index, const, shift) \
    42  	LEAL	const(a)(DI*1),a; \
    43  	MOVL	(index*4)(SI),DI; \
    44  	XORL	d,		BP; \
    45  	XORL	b,		BP; \
    46  	ADDL	BP,		a; \
    47  	ROLL	$shift,		a; \
    48  	MOVL	b,		BP; \
    49  	ADDL	b,		a
    50  
    51  #define ROUND4(a, b, c, d, index, const, shift) \
    52  	LEAL	const(a)(DI*1),a; \
    53  	ORL	b,		BP; \
    54  	XORL	c,		BP; \
    55  	ADDL	BP,		a; \
    56  	MOVL	(index*4)(SI),DI; \
    57  	MOVL	$0xffffffff,	BP; \
    58  	ROLL	$shift,		a; \
    59  	XORL	c,		BP; \
    60  	ADDL	b,		a
    61  
    62  TEXT	·block(SB),NOSPLIT,$24-16
    63  	MOVL	dig+0(FP),	BP
    64  	MOVL	p+4(FP),	SI
    65  	MOVL	p_len+8(FP), DX
    66  	SHRL	$6,		DX
    67  	SHLL	$6,		DX
    68  
    69  	LEAL	(SI)(DX*1),	DI
    70  	MOVL	(0*4)(BP),	AX
    71  	MOVL	(1*4)(BP),	BX
    72  	MOVL	(2*4)(BP),	CX
    73  	MOVL	(3*4)(BP),	DX
    74  
    75  	CMPL	SI,		DI
    76  	JEQ	end
    77  
    78  	MOVL	DI,		16(SP)
    79  
    80  loop:
    81  	MOVL	AX,		0(SP)
    82  	MOVL	BX,		4(SP)
    83  	MOVL	CX,		8(SP)
    84  	MOVL	DX,		12(SP)
    85  
    86  	MOVL	(0*4)(SI),	DI
    87  	MOVL	DX,		BP
    88  
    89  	ROUND1(AX,BX,CX,DX, 1,0xd76aa478, 7);
    90  	ROUND1(DX,AX,BX,CX, 2,0xe8c7b756,12);
    91  	ROUND1(CX,DX,AX,BX, 3,0x242070db,17);
    92  	ROUND1(BX,CX,DX,AX, 4,0xc1bdceee,22);
    93  	ROUND1(AX,BX,CX,DX, 5,0xf57c0faf, 7);
    94  	ROUND1(DX,AX,BX,CX, 6,0x4787c62a,12);
    95  	ROUND1(CX,DX,AX,BX, 7,0xa8304613,17);
    96  	ROUND1(BX,CX,DX,AX, 8,0xfd469501,22);
    97  	ROUND1(AX,BX,CX,DX, 9,0x698098d8, 7);
    98  	ROUND1(DX,AX,BX,CX,10,0x8b44f7af,12);
    99  	ROUND1(CX,DX,AX,BX,11,0xffff5bb1,17);
   100  	ROUND1(BX,CX,DX,AX,12,0x895cd7be,22);
   101  	ROUND1(AX,BX,CX,DX,13,0x6b901122, 7);
   102  	ROUND1(DX,AX,BX,CX,14,0xfd987193,12);
   103  	ROUND1(CX,DX,AX,BX,15,0xa679438e,17);
   104  	ROUND1(BX,CX,DX,AX, 0,0x49b40821,22);
   105  
   106  	MOVL	(1*4)(SI),	DI
   107  	MOVL	DX,		BP
   108  
   109  	ROUND2(AX,BX,CX,DX, 6,0xf61e2562, 5);
   110  	ROUND2(DX,AX,BX,CX,11,0xc040b340, 9);
   111  	ROUND2(CX,DX,AX,BX, 0,0x265e5a51,14);
   112  	ROUND2(BX,CX,DX,AX, 5,0xe9b6c7aa,20);
   113  	ROUND2(AX,BX,CX,DX,10,0xd62f105d, 5);
   114  	ROUND2(DX,AX,BX,CX,15, 0x2441453, 9);
   115  	ROUND2(CX,DX,AX,BX, 4,0xd8a1e681,14);
   116  	ROUND2(BX,CX,DX,AX, 9,0xe7d3fbc8,20);
   117  	ROUND2(AX,BX,CX,DX,14,0x21e1cde6, 5);
   118  	ROUND2(DX,AX,BX,CX, 3,0xc33707d6, 9);
   119  	ROUND2(CX,DX,AX,BX, 8,0xf4d50d87,14);
   120  	ROUND2(BX,CX,DX,AX,13,0x455a14ed,20);
   121  	ROUND2(AX,BX,CX,DX, 2,0xa9e3e905, 5);
   122  	ROUND2(DX,AX,BX,CX, 7,0xfcefa3f8, 9);
   123  	ROUND2(CX,DX,AX,BX,12,0x676f02d9,14);
   124  	ROUND2(BX,CX,DX,AX, 0,0x8d2a4c8a,20);
   125  
   126  	MOVL	(5*4)(SI),	DI
   127  	MOVL	CX,		BP
   128  
   129  	ROUND3(AX,BX,CX,DX, 8,0xfffa3942, 4);
   130  	ROUND3(DX,AX,BX,CX,11,0x8771f681,11);
   131  	ROUND3(CX,DX,AX,BX,14,0x6d9d6122,16);
   132  	ROUND3(BX,CX,DX,AX, 1,0xfde5380c,23);
   133  	ROUND3(AX,BX,CX,DX, 4,0xa4beea44, 4);
   134  	ROUND3(DX,AX,BX,CX, 7,0x4bdecfa9,11);
   135  	ROUND3(CX,DX,AX,BX,10,0xf6bb4b60,16);
   136  	ROUND3(BX,CX,DX,AX,13,0xbebfbc70,23);
   137  	ROUND3(AX,BX,CX,DX, 0,0x289b7ec6, 4);
   138  	ROUND3(DX,AX,BX,CX, 3,0xeaa127fa,11);
   139  	ROUND3(CX,DX,AX,BX, 6,0xd4ef3085,16);
   140  	ROUND3(BX,CX,DX,AX, 9, 0x4881d05,23);
   141  	ROUND3(AX,BX,CX,DX,12,0xd9d4d039, 4);
   142  	ROUND3(DX,AX,BX,CX,15,0xe6db99e5,11);
   143  	ROUND3(CX,DX,AX,BX, 2,0x1fa27cf8,16);
   144  	ROUND3(BX,CX,DX,AX, 0,0xc4ac5665,23);
   145  
   146  	MOVL	(0*4)(SI),	DI
   147  	MOVL	$0xffffffff,	BP
   148  	XORL	DX,		BP
   149  
   150  	ROUND4(AX,BX,CX,DX, 7,0xf4292244, 6);
   151  	ROUND4(DX,AX,BX,CX,14,0x432aff97,10);
   152  	ROUND4(CX,DX,AX,BX, 5,0xab9423a7,15);
   153  	ROUND4(BX,CX,DX,AX,12,0xfc93a039,21);
   154  	ROUND4(AX,BX,CX,DX, 3,0x655b59c3, 6);
   155  	ROUND4(DX,AX,BX,CX,10,0x8f0ccc92,10);
   156  	ROUND4(CX,DX,AX,BX, 1,0xffeff47d,15);
   157  	ROUND4(BX,CX,DX,AX, 8,0x85845dd1,21);
   158  	ROUND4(AX,BX,CX,DX,15,0x6fa87e4f, 6);
   159  	ROUND4(DX,AX,BX,CX, 6,0xfe2ce6e0,10);
   160  	ROUND4(CX,DX,AX,BX,13,0xa3014314,15);
   161  	ROUND4(BX,CX,DX,AX, 4,0x4e0811a1,21);
   162  	ROUND4(AX,BX,CX,DX,11,0xf7537e82, 6);
   163  	ROUND4(DX,AX,BX,CX, 2,0xbd3af235,10);
   164  	ROUND4(CX,DX,AX,BX, 9,0x2ad7d2bb,15);
   165  	ROUND4(BX,CX,DX,AX, 0,0xeb86d391,21);
   166  
   167  	ADDL	0(SP),	AX
   168  	ADDL	4(SP),	BX
   169  	ADDL	8(SP),	CX
   170  	ADDL	12(SP),	DX
   171  
   172  	ADDL	$64,		SI
   173  	CMPL	SI,		16(SP)
   174  	JB	loop
   175  
   176  end:
   177  	MOVL	dig+0(FP),	BP
   178  	MOVL	AX,		(0*4)(BP)
   179  	MOVL	BX,		(1*4)(BP)
   180  	MOVL	CX,		(2*4)(BP)
   181  	MOVL	DX,		(3*4)(BP)
   182  	RET
   183
View as plain text