github.com/zuoyebang/bitalosdb@v1.1.1-0.20240516111551-79a8c4d8ce20/internal/hash/md5block_arm64.s (about)

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