1/* SPDX-License-Identifier: GPL-2.0-only */
2/*
3 * Copyright 2012 Xyratex Technology Limited
4 *
5 * Using hardware provided PCLMULQDQ instruction to accelerate the CRC32
6 * calculation.
7 * CRC32 polynomial:0x04c11db7(BE)/0xEDB88320(LE)
8 * PCLMULQDQ is a new instruction in Intel SSE4.2, the reference can be found
9 * at:
10 * http://www.intel.com/products/processor/manuals/
11 * Intel(R) 64 and IA-32 Architectures Software Developer's Manual
12 * Volume 2B: Instruction Set Reference, N-Z
13 *
14 * Authors:   Gregory Prestas <Gregory_Prestas@us.xyratex.com>
15 *	      Alexander Boyko <Alexander_Boyko@xyratex.com>
16 */
17
18#include <linux/linkage.h>
19
20
21.section .rodata
22.align 16
23/*
24 * [x4*128+32 mod P(x) << 32)]'  << 1   = 0x154442bd4
25 * #define CONSTANT_R1  0x154442bd4LL
26 *
27 * [(x4*128-32 mod P(x) << 32)]' << 1   = 0x1c6e41596
28 * #define CONSTANT_R2  0x1c6e41596LL
29 */
30.Lconstant_R2R1:
31	.octa 0x00000001c6e415960000000154442bd4
32/*
33 * [(x128+32 mod P(x) << 32)]'   << 1   = 0x1751997d0
34 * #define CONSTANT_R3  0x1751997d0LL
35 *
36 * [(x128-32 mod P(x) << 32)]'   << 1   = 0x0ccaa009e
37 * #define CONSTANT_R4  0x0ccaa009eLL
38 */
39.Lconstant_R4R3:
40	.octa 0x00000000ccaa009e00000001751997d0
41/*
42 * [(x64 mod P(x) << 32)]'       << 1   = 0x163cd6124
43 * #define CONSTANT_R5  0x163cd6124LL
44 */
45.Lconstant_R5:
46	.octa 0x00000000000000000000000163cd6124
47.Lconstant_mask32:
48	.octa 0x000000000000000000000000FFFFFFFF
49/*
50 * #define CRCPOLY_TRUE_LE_FULL 0x1DB710641LL
51 *
52 * Barrett Reduction constant (u64`) = u` = (x**64 / P(x))` = 0x1F7011641LL
53 * #define CONSTANT_RU  0x1F7011641LL
54 */
55.Lconstant_RUpoly:
56	.octa 0x00000001F701164100000001DB710641
57
58#define CONSTANT %xmm0
59
60#ifdef __x86_64__
61#define BUF     %rdi
62#define LEN     %rsi
63#define CRC     %edx
64#else
65#define BUF     %eax
66#define LEN     %edx
67#define CRC     %ecx
68#endif
69
70
71
72.text
73/**
74 *      Calculate crc32
75 *      BUF - buffer (16 bytes aligned)
76 *      LEN - sizeof buffer (16 bytes aligned), LEN should be grater than 63
77 *      CRC - initial crc32
78 *      return %eax crc32
79 *      uint crc32_pclmul_le_16(unsigned char const *buffer,
80 *	                     size_t len, uint crc32)
81 */
82
83SYM_FUNC_START(crc32_pclmul_le_16) /* buffer and buffer size are 16 bytes aligned */
84	movdqa  (BUF), %xmm1
85	movdqa  0x10(BUF), %xmm2
86	movdqa  0x20(BUF), %xmm3
87	movdqa  0x30(BUF), %xmm4
88	movd    CRC, CONSTANT
89	pxor    CONSTANT, %xmm1
90	sub     $0x40, LEN
91	add     $0x40, BUF
92	cmp     $0x40, LEN
93	jb      .Lless_64
94
95#ifdef __x86_64__
96	movdqa .Lconstant_R2R1(%rip), CONSTANT
97#else
98	movdqa .Lconstant_R2R1, CONSTANT
99#endif
100
101.Lloop_64:/*  64 bytes Full cache line folding */
102	prefetchnta    0x40(BUF)
103	movdqa  %xmm1, %xmm5
104	movdqa  %xmm2, %xmm6
105	movdqa  %xmm3, %xmm7
106#ifdef __x86_64__
107	movdqa  %xmm4, %xmm8
108#endif
109	pclmulqdq $0x00, CONSTANT, %xmm1
110	pclmulqdq $0x00, CONSTANT, %xmm2
111	pclmulqdq $0x00, CONSTANT, %xmm3
112#ifdef __x86_64__
113	pclmulqdq $0x00, CONSTANT, %xmm4
114#endif
115	pclmulqdq $0x11, CONSTANT, %xmm5
116	pclmulqdq $0x11, CONSTANT, %xmm6
117	pclmulqdq $0x11, CONSTANT, %xmm7
118#ifdef __x86_64__
119	pclmulqdq $0x11, CONSTANT, %xmm8
120#endif
121	pxor    %xmm5, %xmm1
122	pxor    %xmm6, %xmm2
123	pxor    %xmm7, %xmm3
124#ifdef __x86_64__
125	pxor    %xmm8, %xmm4
126#else
127	/* xmm8 unsupported for x32 */
128	movdqa  %xmm4, %xmm5
129	pclmulqdq $0x00, CONSTANT, %xmm4
130	pclmulqdq $0x11, CONSTANT, %xmm5
131	pxor    %xmm5, %xmm4
132#endif
133
134	pxor    (BUF), %xmm1
135	pxor    0x10(BUF), %xmm2
136	pxor    0x20(BUF), %xmm3
137	pxor    0x30(BUF), %xmm4
138
139	sub     $0x40, LEN
140	add     $0x40, BUF
141	cmp     $0x40, LEN
142	jge     .Lloop_64
143.Lless_64:/*  Folding cache line into 128bit */
144#ifdef __x86_64__
145	movdqa  .Lconstant_R4R3(%rip), CONSTANT
146#else
147	movdqa  .Lconstant_R4R3, CONSTANT
148#endif
149	prefetchnta     (BUF)
150
151	movdqa  %xmm1, %xmm5
152	pclmulqdq $0x00, CONSTANT, %xmm1
153	pclmulqdq $0x11, CONSTANT, %xmm5
154	pxor    %xmm5, %xmm1
155	pxor    %xmm2, %xmm1
156
157	movdqa  %xmm1, %xmm5
158	pclmulqdq $0x00, CONSTANT, %xmm1
159	pclmulqdq $0x11, CONSTANT, %xmm5
160	pxor    %xmm5, %xmm1
161	pxor    %xmm3, %xmm1
162
163	movdqa  %xmm1, %xmm5
164	pclmulqdq $0x00, CONSTANT, %xmm1
165	pclmulqdq $0x11, CONSTANT, %xmm5
166	pxor    %xmm5, %xmm1
167	pxor    %xmm4, %xmm1
168
169	cmp     $0x10, LEN
170	jb      .Lfold_64
171.Lloop_16:/* Folding rest buffer into 128bit */
172	movdqa  %xmm1, %xmm5
173	pclmulqdq $0x00, CONSTANT, %xmm1
174	pclmulqdq $0x11, CONSTANT, %xmm5
175	pxor    %xmm5, %xmm1
176	pxor    (BUF), %xmm1
177	sub     $0x10, LEN
178	add     $0x10, BUF
179	cmp     $0x10, LEN
180	jge     .Lloop_16
181
182.Lfold_64:
183	/* perform the last 64 bit fold, also adds 32 zeroes
184	 * to the input stream */
185	pclmulqdq $0x01, %xmm1, CONSTANT /* R4 * xmm1.low */
186	psrldq  $0x08, %xmm1
187	pxor    CONSTANT, %xmm1
188
189	/* final 32-bit fold */
190	movdqa  %xmm1, %xmm2
191#ifdef __x86_64__
192	movdqa  .Lconstant_R5(%rip), CONSTANT
193	movdqa  .Lconstant_mask32(%rip), %xmm3
194#else
195	movdqa  .Lconstant_R5, CONSTANT
196	movdqa  .Lconstant_mask32, %xmm3
197#endif
198	psrldq  $0x04, %xmm2
199	pand    %xmm3, %xmm1
200	pclmulqdq $0x00, CONSTANT, %xmm1
201	pxor    %xmm2, %xmm1
202
203	/* Finish up with the bit-reversed barrett reduction 64 ==> 32 bits */
204#ifdef __x86_64__
205	movdqa  .Lconstant_RUpoly(%rip), CONSTANT
206#else
207	movdqa  .Lconstant_RUpoly, CONSTANT
208#endif
209	movdqa  %xmm1, %xmm2
210	pand    %xmm3, %xmm1
211	pclmulqdq $0x10, CONSTANT, %xmm1
212	pand    %xmm3, %xmm1
213	pclmulqdq $0x00, CONSTANT, %xmm1
214	pxor    %xmm2, %xmm1
215	pextrd  $0x01, %xmm1, %eax
216
217	RET
218SYM_FUNC_END(crc32_pclmul_le_16)
219