Home | History | Annotate | Download | only in md5
      1 // Original source:
      2 //	http://www.zorinaq.com/papers/md5-amd64.html
      3 //	http://www.zorinaq.com/papers/md5-amd64.tar.bz2
      4 //
      5 // MD5 adapted for s390x using Go's assembler for
      6 // s390x, based on md5block_amd64.s implementation by
      7 // the Go authors.
      8 //
      9 // Author: Marc Bevand <bevand_m (at) epita.fr>
     10 // Licence: I hereby disclaim the copyright on this code and place it
     11 // in the public domain.
     12 
     13 #include "textflag.h"
     14 
     15 // func block(dig *digest, p []byte)
     16 TEXT block(SB),NOSPLIT,$16-32
     17 	MOVD	dig+0(FP), R1
     18 	MOVD	p+8(FP), R6
     19 	MOVD	p_len+16(FP), R5
     20 	AND	$-64, R5
     21 	LAY	(R6)(R5*1), R7
     22 
     23 	LMY	0(R1), R2, R5
     24 	CMPBEQ	R6, R7, end
     25 
     26 loop:
     27 	STMY	R2, R5, tmp-16(SP)
     28 
     29 	MOVWBR	0(R6), R8
     30 	MOVWZ	R5, R9
     31 
     32 #define ROUND1(a, b, c, d, index, const, shift) \
     33 	XOR	c, R9; \
     34 	ADD	$const, a; \
     35 	ADD	R8, a; \
     36 	MOVWBR	(index*4)(R6), R8; \
     37 	AND	b, R9; \
     38 	XOR	d, R9; \
     39 	ADD	R9, a; \
     40 	RLL	$shift, a; \
     41 	MOVWZ	c, R9; \
     42 	ADD	b, a
     43 
     44 	ROUND1(R2,R3,R4,R5, 1,0xd76aa478, 7);
     45 	ROUND1(R5,R2,R3,R4, 2,0xe8c7b756,12);
     46 	ROUND1(R4,R5,R2,R3, 3,0x242070db,17);
     47 	ROUND1(R3,R4,R5,R2, 4,0xc1bdceee,22);
     48 	ROUND1(R2,R3,R4,R5, 5,0xf57c0faf, 7);
     49 	ROUND1(R5,R2,R3,R4, 6,0x4787c62a,12);
     50 	ROUND1(R4,R5,R2,R3, 7,0xa8304613,17);
     51 	ROUND1(R3,R4,R5,R2, 8,0xfd469501,22);
     52 	ROUND1(R2,R3,R4,R5, 9,0x698098d8, 7);
     53 	ROUND1(R5,R2,R3,R4,10,0x8b44f7af,12);
     54 	ROUND1(R4,R5,R2,R3,11,0xffff5bb1,17);
     55 	ROUND1(R3,R4,R5,R2,12,0x895cd7be,22);
     56 	ROUND1(R2,R3,R4,R5,13,0x6b901122, 7);
     57 	ROUND1(R5,R2,R3,R4,14,0xfd987193,12);
     58 	ROUND1(R4,R5,R2,R3,15,0xa679438e,17);
     59 	ROUND1(R3,R4,R5,R2, 0,0x49b40821,22);
     60 
     61 	MOVWBR	(1*4)(R6), R8
     62 	MOVWZ	R5, R9
     63 	MOVWZ	R5, R1
     64 
     65 #define ROUND2(a, b, c, d, index, const, shift) \
     66 	XOR	$0xffffffff, R9; \ // NOTW R9
     67 	ADD	$const, a; \
     68 	ADD	R8, a; \
     69 	MOVWBR	(index*4)(R6), R8; \
     70 	AND	b, R1; \
     71 	AND	c, R9; \
     72 	OR	R9, R1; \
     73 	MOVWZ	c, R9; \
     74 	ADD	R1, a; \
     75 	MOVWZ	c, R1; \
     76 	RLL	$shift,	a; \
     77 	ADD	b, a
     78 
     79 	ROUND2(R2,R3,R4,R5, 6,0xf61e2562, 5);
     80 	ROUND2(R5,R2,R3,R4,11,0xc040b340, 9);
     81 	ROUND2(R4,R5,R2,R3, 0,0x265e5a51,14);
     82 	ROUND2(R3,R4,R5,R2, 5,0xe9b6c7aa,20);
     83 	ROUND2(R2,R3,R4,R5,10,0xd62f105d, 5);
     84 	ROUND2(R5,R2,R3,R4,15, 0x2441453, 9);
     85 	ROUND2(R4,R5,R2,R3, 4,0xd8a1e681,14);
     86 	ROUND2(R3,R4,R5,R2, 9,0xe7d3fbc8,20);
     87 	ROUND2(R2,R3,R4,R5,14,0x21e1cde6, 5);
     88 	ROUND2(R5,R2,R3,R4, 3,0xc33707d6, 9);
     89 	ROUND2(R4,R5,R2,R3, 8,0xf4d50d87,14);
     90 	ROUND2(R3,R4,R5,R2,13,0x455a14ed,20);
     91 	ROUND2(R2,R3,R4,R5, 2,0xa9e3e905, 5);
     92 	ROUND2(R5,R2,R3,R4, 7,0xfcefa3f8, 9);
     93 	ROUND2(R4,R5,R2,R3,12,0x676f02d9,14);
     94 	ROUND2(R3,R4,R5,R2, 0,0x8d2a4c8a,20);
     95 
     96 	MOVWBR	(5*4)(R6), R8
     97 	MOVWZ	R4, R9
     98 
     99 #define ROUND3(a, b, c, d, index, const, shift) \
    100 	ADD	$const, a; \
    101 	ADD	R8, a; \
    102 	MOVWBR	(index*4)(R6), R8; \
    103 	XOR	d, R9; \
    104 	XOR	b, R9; \
    105 	ADD	R9, a; \
    106 	RLL	$shift, a; \
    107 	MOVWZ	b, R9; \
    108 	ADD	b, a
    109 
    110 	ROUND3(R2,R3,R4,R5, 8,0xfffa3942, 4);
    111 	ROUND3(R5,R2,R3,R4,11,0x8771f681,11);
    112 	ROUND3(R4,R5,R2,R3,14,0x6d9d6122,16);
    113 	ROUND3(R3,R4,R5,R2, 1,0xfde5380c,23);
    114 	ROUND3(R2,R3,R4,R5, 4,0xa4beea44, 4);
    115 	ROUND3(R5,R2,R3,R4, 7,0x4bdecfa9,11);
    116 	ROUND3(R4,R5,R2,R3,10,0xf6bb4b60,16);
    117 	ROUND3(R3,R4,R5,R2,13,0xbebfbc70,23);
    118 	ROUND3(R2,R3,R4,R5, 0,0x289b7ec6, 4);
    119 	ROUND3(R5,R2,R3,R4, 3,0xeaa127fa,11);
    120 	ROUND3(R4,R5,R2,R3, 6,0xd4ef3085,16);
    121 	ROUND3(R3,R4,R5,R2, 9, 0x4881d05,23);
    122 	ROUND3(R2,R3,R4,R5,12,0xd9d4d039, 4);
    123 	ROUND3(R5,R2,R3,R4,15,0xe6db99e5,11);
    124 	ROUND3(R4,R5,R2,R3, 2,0x1fa27cf8,16);
    125 	ROUND3(R3,R4,R5,R2, 0,0xc4ac5665,23);
    126 
    127 	MOVWBR	(0*4)(R6), R8
    128 	MOVWZ	$0xffffffff, R9
    129 	XOR	R5, R9
    130 
    131 #define ROUND4(a, b, c, d, index, const, shift) \
    132 	ADD	$const, a; \
    133 	ADD	R8, a; \
    134 	MOVWBR	(index*4)(R6), R8; \
    135 	OR	b, R9; \
    136 	XOR	c, R9; \
    137 	ADD	R9, a; \
    138 	MOVWZ	$0xffffffff, R9; \
    139 	RLL	$shift,	a; \
    140 	XOR	c, R9; \
    141 	ADD	b, a
    142 
    143 	ROUND4(R2,R3,R4,R5, 7,0xf4292244, 6);
    144 	ROUND4(R5,R2,R3,R4,14,0x432aff97,10);
    145 	ROUND4(R4,R5,R2,R3, 5,0xab9423a7,15);
    146 	ROUND4(R3,R4,R5,R2,12,0xfc93a039,21);
    147 	ROUND4(R2,R3,R4,R5, 3,0x655b59c3, 6);
    148 	ROUND4(R5,R2,R3,R4,10,0x8f0ccc92,10);
    149 	ROUND4(R4,R5,R2,R3, 1,0xffeff47d,15);
    150 	ROUND4(R3,R4,R5,R2, 8,0x85845dd1,21);
    151 	ROUND4(R2,R3,R4,R5,15,0x6fa87e4f, 6);
    152 	ROUND4(R5,R2,R3,R4, 6,0xfe2ce6e0,10);
    153 	ROUND4(R4,R5,R2,R3,13,0xa3014314,15);
    154 	ROUND4(R3,R4,R5,R2, 4,0x4e0811a1,21);
    155 	ROUND4(R2,R3,R4,R5,11,0xf7537e82, 6);
    156 	ROUND4(R5,R2,R3,R4, 2,0xbd3af235,10);
    157 	ROUND4(R4,R5,R2,R3, 9,0x2ad7d2bb,15);
    158 	ROUND4(R3,R4,R5,R2, 0,0xeb86d391,21);
    159 
    160 	MOVWZ	tmp-16(SP), R1
    161 	ADD	R1, R2
    162 	MOVWZ	tmp-12(SP), R1
    163 	ADD	R1, R3
    164 	MOVWZ	tmp-8(SP), R1
    165 	ADD	R1, R4
    166 	MOVWZ	tmp-4(SP), R1
    167 	ADD	R1, R5
    168 
    169 	LA	64(R6), R6
    170 	CMPBLT	R6, R7, loop
    171 
    172 end:
    173 	MOVD	dig+0(FP), R1
    174 	STMY	R2, R5, 0(R1)
    175 	RET
    176