Text file
src/crypto/md5/md5block_arm64.s
1// Copyright 2018 The Go Authors. All rights reserved.
2// Use of this source code is governed by a BSD-style
3// license that can be found in the LICENSE file.
4//
5// ARM64 version of md5block.go
6// derived from crypto/md5/md5block_amd64.s
7
8//go:build !purego
9
10#include "textflag.h"
11
12TEXT ·block(SB),NOSPLIT,$0-32
13 MOVD dig+0(FP), R0
14 MOVD p+8(FP), R1
15 MOVD p_len+16(FP), R2
16 AND $~63, R2
17 CBZ R2, zero
18
19 ADD R1, R2, R21
20 LDPW (0*8)(R0), (R4, R5)
21 LDPW (1*8)(R0), (R6, R7)
22
23loop:
24 MOVW R4, R12
25 MOVW R5, R13
26 MOVW R6, R14
27 MOVW R7, R15
28
29 MOVW (0*4)(R1), R8
30 MOVW R7, R9
31
32#define ROUND1(a, b, c, d, index, const, shift) \
33 ADDW $const, a; \
34 ADDW R8, a; \
35 MOVW (index*4)(R1), R8; \
36 EORW c, R9; \
37 ANDW b, R9; \
38 EORW d, R9; \
39 ADDW R9, a; \
40 RORW $(32-shift), a; \
41 MOVW c, R9; \
42 ADDW b, a
43
44 ROUND1(R4,R5,R6,R7, 1,0xd76aa478, 7);
45 ROUND1(R7,R4,R5,R6, 2,0xe8c7b756,12);
46 ROUND1(R6,R7,R4,R5, 3,0x242070db,17);
47 ROUND1(R5,R6,R7,R4, 4,0xc1bdceee,22);
48 ROUND1(R4,R5,R6,R7, 5,0xf57c0faf, 7);
49 ROUND1(R7,R4,R5,R6, 6,0x4787c62a,12);
50 ROUND1(R6,R7,R4,R5, 7,0xa8304613,17);
51 ROUND1(R5,R6,R7,R4, 8,0xfd469501,22);
52 ROUND1(R4,R5,R6,R7, 9,0x698098d8, 7);
53 ROUND1(R7,R4,R5,R6,10,0x8b44f7af,12);
54 ROUND1(R6,R7,R4,R5,11,0xffff5bb1,17);
55 ROUND1(R5,R6,R7,R4,12,0x895cd7be,22);
56 ROUND1(R4,R5,R6,R7,13,0x6b901122, 7);
57 ROUND1(R7,R4,R5,R6,14,0xfd987193,12);
58 ROUND1(R6,R7,R4,R5,15,0xa679438e,17);
59 ROUND1(R5,R6,R7,R4, 0,0x49b40821,22);
60
61 MOVW (1*4)(R1), R8
62 MOVW R7, R9
63 MOVW R7, R10
64
65#define ROUND2(a, b, c, d, index, const, shift) \
66 ADDW $const, a; \
67 ADDW R8, a; \
68 MOVW (index*4)(R1), R8; \
69 ANDW b, R10; \
70 BICW R9, c, R9; \
71 ORRW R9, R10; \
72 MOVW c, R9; \
73 ADDW R10, a; \
74 MOVW c, R10; \
75 RORW $(32-shift), a; \
76 ADDW b, a
77
78 ROUND2(R4,R5,R6,R7, 6,0xf61e2562, 5);
79 ROUND2(R7,R4,R5,R6,11,0xc040b340, 9);
80 ROUND2(R6,R7,R4,R5, 0,0x265e5a51,14);
81 ROUND2(R5,R6,R7,R4, 5,0xe9b6c7aa,20);
82 ROUND2(R4,R5,R6,R7,10,0xd62f105d, 5);
83 ROUND2(R7,R4,R5,R6,15, 0x2441453, 9);
84 ROUND2(R6,R7,R4,R5, 4,0xd8a1e681,14);
85 ROUND2(R5,R6,R7,R4, 9,0xe7d3fbc8,20);
86 ROUND2(R4,R5,R6,R7,14,0x21e1cde6, 5);
87 ROUND2(R7,R4,R5,R6, 3,0xc33707d6, 9);
88 ROUND2(R6,R7,R4,R5, 8,0xf4d50d87,14);
89 ROUND2(R5,R6,R7,R4,13,0x455a14ed,20);
90 ROUND2(R4,R5,R6,R7, 2,0xa9e3e905, 5);
91 ROUND2(R7,R4,R5,R6, 7,0xfcefa3f8, 9);
92 ROUND2(R6,R7,R4,R5,12,0x676f02d9,14);
93 ROUND2(R5,R6,R7,R4, 0,0x8d2a4c8a,20);
94
95 MOVW (5*4)(R1), R8
96 MOVW R6, R9
97
98#define ROUND3(a, b, c, d, index, const, shift) \
99 ADDW $const, a; \
100 ADDW R8, a; \
101 MOVW (index*4)(R1), R8; \
102 EORW d, R9; \
103 EORW b, R9; \
104 ADDW R9, a; \
105 RORW $(32-shift), a; \
106 MOVW b, R9; \
107 ADDW b, a
108
109 ROUND3(R4,R5,R6,R7, 8,0xfffa3942, 4);
110 ROUND3(R7,R4,R5,R6,11,0x8771f681,11);
111 ROUND3(R6,R7,R4,R5,14,0x6d9d6122,16);
112 ROUND3(R5,R6,R7,R4, 1,0xfde5380c,23);
113 ROUND3(R4,R5,R6,R7, 4,0xa4beea44, 4);
114 ROUND3(R7,R4,R5,R6, 7,0x4bdecfa9,11);
115 ROUND3(R6,R7,R4,R5,10,0xf6bb4b60,16);
116 ROUND3(R5,R6,R7,R4,13,0xbebfbc70,23);
117 ROUND3(R4,R5,R6,R7, 0,0x289b7ec6, 4);
118 ROUND3(R7,R4,R5,R6, 3,0xeaa127fa,11);
119 ROUND3(R6,R7,R4,R5, 6,0xd4ef3085,16);
120 ROUND3(R5,R6,R7,R4, 9, 0x4881d05,23);
121 ROUND3(R4,R5,R6,R7,12,0xd9d4d039, 4);
122 ROUND3(R7,R4,R5,R6,15,0xe6db99e5,11);
123 ROUND3(R6,R7,R4,R5, 2,0x1fa27cf8,16);
124 ROUND3(R5,R6,R7,R4, 0,0xc4ac5665,23);
125
126 MOVW (0*4)(R1), R8
127 MVNW R7, R9
128
129#define ROUND4(a, b, c, d, index, const, shift) \
130 ADDW $const, a; \
131 ADDW R8, a; \
132 MOVW (index*4)(R1), R8; \
133 ORRW b, R9; \
134 EORW c, R9; \
135 ADDW R9, a; \
136 RORW $(32-shift), a; \
137 MVNW c, R9; \
138 ADDW b, a
139
140 ROUND4(R4,R5,R6,R7, 7,0xf4292244, 6);
141 ROUND4(R7,R4,R5,R6,14,0x432aff97,10);
142 ROUND4(R6,R7,R4,R5, 5,0xab9423a7,15);
143 ROUND4(R5,R6,R7,R4,12,0xfc93a039,21);
144 ROUND4(R4,R5,R6,R7, 3,0x655b59c3, 6);
145 ROUND4(R7,R4,R5,R6,10,0x8f0ccc92,10);
146 ROUND4(R6,R7,R4,R5, 1,0xffeff47d,15);
147 ROUND4(R5,R6,R7,R4, 8,0x85845dd1,21);
148 ROUND4(R4,R5,R6,R7,15,0x6fa87e4f, 6);
149 ROUND4(R7,R4,R5,R6, 6,0xfe2ce6e0,10);
150 ROUND4(R6,R7,R4,R5,13,0xa3014314,15);
151 ROUND4(R5,R6,R7,R4, 4,0x4e0811a1,21);
152 ROUND4(R4,R5,R6,R7,11,0xf7537e82, 6);
153 ROUND4(R7,R4,R5,R6, 2,0xbd3af235,10);
154 ROUND4(R6,R7,R4,R5, 9,0x2ad7d2bb,15);
155 ROUND4(R5,R6,R7,R4, 0,0xeb86d391,21);
156
157 ADDW R12, R4
158 ADDW R13, R5
159 ADDW R14, R6
160 ADDW R15, R7
161
162 ADD $64, R1
163 CMP R1, R21
164 BNE loop
165
166 STPW (R4, R5), (0*8)(R0)
167 STPW (R6, R7), (1*8)(R0)
168zero:
169 RET
View as plain text