Text file
src/crypto/md5/md5block_ppc64x.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 // MD5 optimized for ppc64le using Go's assembler for
6 // ppc64le, 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 //go:build (ppc64 || ppc64le) && !purego
14
15 #include "textflag.h"
16
17 // ENDIAN_MOVE generates the appropriate
18 // 4 byte load for big or little endian.
19 // The 4 bytes at ptr+off is loaded into dst.
20 // The idx reg is only needed for big endian
21 // and is clobbered when used.
22 #ifdef GOARCH_ppc64le
23 #define ENDIAN_MOVE(off, ptr, dst, idx) \
24 MOVWZ off(ptr),dst
25 #else
26 #define ENDIAN_MOVE(off, ptr, dst, idx) \
27 MOVD $off,idx; \
28 MOVWBR (idx)(ptr), dst
29 #endif
30
31 #define M00 R18
32 #define M01 R19
33 #define M02 R20
34 #define M03 R24
35 #define M04 R25
36 #define M05 R26
37 #define M06 R27
38 #define M07 R28
39 #define M08 R29
40 #define M09 R21
41 #define M10 R11
42 #define M11 R8
43 #define M12 R7
44 #define M13 R12
45 #define M14 R23
46 #define M15 R10
47
48 #define ROUND1(a, b, c, d, index, const, shift) \
49 ADD $const, index, R9; \
50 ADD R9, a; \
51 AND b, c, R9; \
52 ANDN b, d, R31; \
53 OR R9, R31, R9; \
54 ADD R9, a; \
55 ROTLW $shift, a; \
56 ADD b, a;
57
58 #define ROUND2(a, b, c, d, index, const, shift) \
59 ADD $const, index, R9; \
60 ADD R9, a; \
61 AND b, d, R31; \
62 ANDN d, c, R9; \
63 OR R9, R31; \
64 ADD R31, a; \
65 ROTLW $shift, a; \
66 ADD b, a;
67
68 #define ROUND3(a, b, c, d, index, const, shift) \
69 ADD $const, index, R9; \
70 ADD R9, a; \
71 XOR d, c, R31; \
72 XOR b, R31; \
73 ADD R31, a; \
74 ROTLW $shift, a; \
75 ADD b, a;
76
77 #define ROUND4(a, b, c, d, index, const, shift) \
78 ADD $const, index, R9; \
79 ADD R9, a; \
80 ORN d, b, R31; \
81 XOR c, R31; \
82 ADD R31, a; \
83 ROTLW $shift, a; \
84 ADD b, a;
85
86
87 TEXT ·block(SB),NOSPLIT,$0-32
88 MOVD dig+0(FP), R10
89 MOVD p+8(FP), R6
90 MOVD p_len+16(FP), R5
91
92 // We assume p_len >= 64
93 SRD $6, R5
94 MOVD R5, CTR
95
96 MOVWZ 0(R10), R22
97 MOVWZ 4(R10), R3
98 MOVWZ 8(R10), R4
99 MOVWZ 12(R10), R5
100
101 loop:
102 MOVD R22, R14
103 MOVD R3, R15
104 MOVD R4, R16
105 MOVD R5, R17
106
107 ENDIAN_MOVE( 0,R6,M00,M15)
108 ENDIAN_MOVE( 4,R6,M01,M15)
109 ENDIAN_MOVE( 8,R6,M02,M15)
110 ENDIAN_MOVE(12,R6,M03,M15)
111
112 ROUND1(R22,R3,R4,R5,M00,0xd76aa478, 7);
113 ROUND1(R5,R22,R3,R4,M01,0xe8c7b756,12);
114 ROUND1(R4,R5,R22,R3,M02,0x242070db,17);
115 ROUND1(R3,R4,R5,R22,M03,0xc1bdceee,22);
116
117 ENDIAN_MOVE(16,R6,M04,M15)
118 ENDIAN_MOVE(20,R6,M05,M15)
119 ENDIAN_MOVE(24,R6,M06,M15)
120 ENDIAN_MOVE(28,R6,M07,M15)
121
122 ROUND1(R22,R3,R4,R5,M04,0xf57c0faf, 7);
123 ROUND1(R5,R22,R3,R4,M05,0x4787c62a,12);
124 ROUND1(R4,R5,R22,R3,M06,0xa8304613,17);
125 ROUND1(R3,R4,R5,R22,M07,0xfd469501,22);
126
127 ENDIAN_MOVE(32,R6,M08,M15)
128 ENDIAN_MOVE(36,R6,M09,M15)
129 ENDIAN_MOVE(40,R6,M10,M15)
130 ENDIAN_MOVE(44,R6,M11,M15)
131
132 ROUND1(R22,R3,R4,R5,M08,0x698098d8, 7);
133 ROUND1(R5,R22,R3,R4,M09,0x8b44f7af,12);
134 ROUND1(R4,R5,R22,R3,M10,0xffff5bb1,17);
135 ROUND1(R3,R4,R5,R22,M11,0x895cd7be,22);
136
137 ENDIAN_MOVE(48,R6,M12,M15)
138 ENDIAN_MOVE(52,R6,M13,M15)
139 ENDIAN_MOVE(56,R6,M14,M15)
140 ENDIAN_MOVE(60,R6,M15,M15)
141
142 ROUND1(R22,R3,R4,R5,M12,0x6b901122, 7);
143 ROUND1(R5,R22,R3,R4,M13,0xfd987193,12);
144 ROUND1(R4,R5,R22,R3,M14,0xa679438e,17);
145 ROUND1(R3,R4,R5,R22,M15,0x49b40821,22);
146
147 ROUND2(R22,R3,R4,R5,M01,0xf61e2562, 5);
148 ROUND2(R5,R22,R3,R4,M06,0xc040b340, 9);
149 ROUND2(R4,R5,R22,R3,M11,0x265e5a51,14);
150 ROUND2(R3,R4,R5,R22,M00,0xe9b6c7aa,20);
151 ROUND2(R22,R3,R4,R5,M05,0xd62f105d, 5);
152 ROUND2(R5,R22,R3,R4,M10, 0x2441453, 9);
153 ROUND2(R4,R5,R22,R3,M15,0xd8a1e681,14);
154 ROUND2(R3,R4,R5,R22,M04,0xe7d3fbc8,20);
155 ROUND2(R22,R3,R4,R5,M09,0x21e1cde6, 5);
156 ROUND2(R5,R22,R3,R4,M14,0xc33707d6, 9);
157 ROUND2(R4,R5,R22,R3,M03,0xf4d50d87,14);
158 ROUND2(R3,R4,R5,R22,M08,0x455a14ed,20);
159 ROUND2(R22,R3,R4,R5,M13,0xa9e3e905, 5);
160 ROUND2(R5,R22,R3,R4,M02,0xfcefa3f8, 9);
161 ROUND2(R4,R5,R22,R3,M07,0x676f02d9,14);
162 ROUND2(R3,R4,R5,R22,M12,0x8d2a4c8a,20);
163
164 ROUND3(R22,R3,R4,R5,M05,0xfffa3942, 4);
165 ROUND3(R5,R22,R3,R4,M08,0x8771f681,11);
166 ROUND3(R4,R5,R22,R3,M11,0x6d9d6122,16);
167 ROUND3(R3,R4,R5,R22,M14,0xfde5380c,23);
168 ROUND3(R22,R3,R4,R5,M01,0xa4beea44, 4);
169 ROUND3(R5,R22,R3,R4,M04,0x4bdecfa9,11);
170 ROUND3(R4,R5,R22,R3,M07,0xf6bb4b60,16);
171 ROUND3(R3,R4,R5,R22,M10,0xbebfbc70,23);
172 ROUND3(R22,R3,R4,R5,M13,0x289b7ec6, 4);
173 ROUND3(R5,R22,R3,R4,M00,0xeaa127fa,11);
174 ROUND3(R4,R5,R22,R3,M03,0xd4ef3085,16);
175 ROUND3(R3,R4,R5,R22,M06, 0x4881d05,23);
176 ROUND3(R22,R3,R4,R5,M09,0xd9d4d039, 4);
177 ROUND3(R5,R22,R3,R4,M12,0xe6db99e5,11);
178 ROUND3(R4,R5,R22,R3,M15,0x1fa27cf8,16);
179 ROUND3(R3,R4,R5,R22,M02,0xc4ac5665,23);
180
181 ROUND4(R22,R3,R4,R5,M00,0xf4292244, 6);
182 ROUND4(R5,R22,R3,R4,M07,0x432aff97,10);
183 ROUND4(R4,R5,R22,R3,M14,0xab9423a7,15);
184 ROUND4(R3,R4,R5,R22,M05,0xfc93a039,21);
185 ROUND4(R22,R3,R4,R5,M12,0x655b59c3, 6);
186 ROUND4(R5,R22,R3,R4,M03,0x8f0ccc92,10);
187 ROUND4(R4,R5,R22,R3,M10,0xffeff47d,15);
188 ROUND4(R3,R4,R5,R22,M01,0x85845dd1,21);
189 ROUND4(R22,R3,R4,R5,M08,0x6fa87e4f, 6);
190 ROUND4(R5,R22,R3,R4,M15,0xfe2ce6e0,10);
191 ROUND4(R4,R5,R22,R3,M06,0xa3014314,15);
192 ROUND4(R3,R4,R5,R22,M13,0x4e0811a1,21);
193 ROUND4(R22,R3,R4,R5,M04,0xf7537e82, 6);
194 ROUND4(R5,R22,R3,R4,M11,0xbd3af235,10);
195 ROUND4(R4,R5,R22,R3,M02,0x2ad7d2bb,15);
196 ROUND4(R3,R4,R5,R22,M09,0xeb86d391,21);
197
198 ADD R14, R22
199 ADD R15, R3
200 ADD R16, R4
201 ADD R17, R5
202 ADD $64, R6
203 BDNZ loop
204
205 end:
206 MOVD dig+0(FP), R10
207 MOVWZ R22, 0(R10)
208 MOVWZ R3, 4(R10)
209 MOVWZ R4, 8(R10)
210 MOVWZ R5, 12(R10)
211
212 RET
213
View as plain text