1// Copyright 2025 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//go:build gc && !purego
6
7// func update(state *macState, msg []byte)
8TEXT ·update(SB), $0-32
9 MOVV state+0(FP), R4
10 MOVV msg_base+8(FP), R5
11 MOVV msg_len+16(FP), R6
12
13 MOVV $0x10, R7
14
15 MOVV (R4), R8 // h0
16 MOVV 8(R4), R9 // h1
17 MOVV 16(R4), R10 // h2
18 MOVV 24(R4), R11 // r0
19 MOVV 32(R4), R12 // r1
20
21 BLT R6, R7, bytes_between_0_and_15
22
23loop:
24 MOVV (R5), R14 // msg[0:8]
25 MOVV 8(R5), R16 // msg[8:16]
26 ADDV R14, R8, R8 // h0 (x1 + y1 = z1', if z1' < x1 then z1' overflow)
27 ADDV R16, R9, R27
28 SGTU R14, R8, R24 // h0.carry
29 SGTU R9, R27, R28
30 ADDV R27, R24, R9 // h1
31 SGTU R27, R9, R24
32 OR R24, R28, R24 // h1.carry
33 ADDV $0x01, R24, R24
34 ADDV R10, R24, R10 // h2
35
36 ADDV $16, R5, R5 // msg = msg[16:]
37
38multiply:
39 MULV R8, R11, R14 // h0r0.lo
40 MULHVU R8, R11, R15 // h0r0.hi
41 MULV R9, R11, R13 // h1r0.lo
42 MULHVU R9, R11, R16 // h1r0.hi
43 ADDV R13, R15, R15
44 SGTU R13, R15, R24
45 ADDV R24, R16, R16
46 MULV R10, R11, R25
47 ADDV R16, R25, R25
48 MULV R8, R12, R13 // h0r1.lo
49 MULHVU R8, R12, R16 // h0r1.hi
50 ADDV R13, R15, R15
51 SGTU R13, R15, R24
52 ADDV R24, R16, R16
53 MOVV R16, R8
54 MULV R10, R12, R26 // h2r1
55 MULV R9, R12, R13 // h1r1.lo
56 MULHVU R9, R12, R16 // h1r1.hi
57 ADDV R13, R25, R25
58 ADDV R16, R26, R27
59 SGTU R13, R25, R24
60 ADDV R27, R24, R26
61 ADDV R8, R25, R25
62 SGTU R8, R25, R24
63 ADDV R24, R26, R26
64 AND $3, R25, R10
65 AND $-4, R25, R17
66 ADDV R17, R14, R8
67 ADDV R26, R15, R27
68 SGTU R17, R8, R24
69 SGTU R26, R27, R28
70 ADDV R27, R24, R9
71 SGTU R27, R9, R24
72 OR R24, R28, R24
73 ADDV R24, R10, R10
74 SLLV $62, R26, R27
75 SRLV $2, R25, R28
76 SRLV $2, R26, R26
77 OR R27, R28, R25
78 ADDV R25, R8, R8
79 ADDV R26, R9, R27
80 SGTU R25, R8, R24
81 SGTU R26, R27, R28
82 ADDV R27, R24, R9
83 SGTU R27, R9, R24
84 OR R24, R28, R24
85 ADDV R24, R10, R10
86
87 SUBV $16, R6, R6
88 BGE R6, R7, loop
89
90bytes_between_0_and_15:
91 BEQ R6, R0, done
92 MOVV $1, R14
93 XOR R15, R15
94 ADDV R6, R5, R5
95
96flush_buffer:
97 MOVBU -1(R5), R25
98 SRLV $56, R14, R24
99 SLLV $8, R15, R28
100 SLLV $8, R14, R14
101 OR R24, R28, R15
102 XOR R25, R14, R14
103 SUBV $1, R6, R6
104 SUBV $1, R5, R5
105 BNE R6, R0, flush_buffer
106
107 ADDV R14, R8, R8
108 SGTU R14, R8, R24
109 ADDV R15, R9, R27
110 SGTU R15, R27, R28
111 ADDV R27, R24, R9
112 SGTU R27, R9, R24
113 OR R24, R28, R24
114 ADDV R10, R24, R10
115
116 MOVV $16, R6
117 JMP multiply
118
119done:
120 MOVV R8, (R4)
121 MOVV R9, 8(R4)
122 MOVV R10, 16(R4)
123 RET