1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
|
/**
* Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
* SPDX-License-Identifier: Apache-2.0.
*/
#include <aws/checksums/private/crc_priv.h>
#include <aws/common/cpuid.h>
/* this implementation is only for the x86_64 intel architecture */
#if defined(__x86_64__)
# if defined(__clang__)
# pragma clang diagnostic push
# pragma clang diagnostic ignored "-Wdollar-in-identifier-extension"
# endif
/*
* Factored out common inline asm for folding crc0,crc1,crc2 stripes in rcx, r11, r10 using
* the specified Magic Constants K1 and K2.
* Assumes rcx, r11, r10 contain crc0, crc1, crc2 that need folding
* Utilizes xmm1, xmm2, xmm3, xmm4 as well as clobbering r8, r9, r11
* Result is placed in ecx
*/
# define FOLD_K1K2(NAME, K1, K2) \
"fold_k1k2_" #NAME "_%=: \n" \
"movl " #K1 ", %%r8d # Magic K1 constant \n" \
"movl " #K2 ", %%r9d # Magic K2 constant \n" \
"movq %%rcx, %%xmm1 # crc0 into lower dword of xmm1 \n" \
"movq %%r8, %%xmm3 # K1 into lower dword of xmm3 \n" \
"movq %%r11, %%xmm2 # crc1 into lower dword of xmm2 \n" \
"movq %%r9, %%xmm4 # K2 into lower dword of xmm4 \n" \
"pclmulqdq $0x00, %%xmm3, %%xmm1 # Multiply crc0 by K1 \n" \
"pclmulqdq $0x00, %%xmm4, %%xmm2 # Multiply crc1 by K2 \n" \
"xor %%rcx, %%rcx # \n" \
"xor %%r11, %%r11 # \n" \
"movq %%xmm1, %%r8 # \n" \
"movq %%xmm2, %%r9 # \n" \
"crc32q %%r8, %%rcx # folding crc0 \n" \
"crc32q %%r9, %%r11 # folding crc1 \n" \
"xor %%r10d, %%ecx # combine crc2 and crc0 \n" \
"xor %%r11d, %%ecx # combine crc1 and crc0 \n"
/**
* Private (static) function.
* Computes the Castagnoli CRC32c (iSCSI) of the specified data buffer using the Intel CRC32Q (quad word) machine
* instruction by operating on 24-byte stripes in parallel. The results are folded together using CLMUL. This function
* is optimized for exactly 256 byte blocks that are best aligned on 8-byte memory addresses. It MUST be passed a
* pointer to input data that is exactly 256 bytes in length. Note: this function does NOT invert bits of the input crc
* or return value.
*/
static inline uint32_t s_crc32c_sse42_clmul_256(const uint8_t *input, uint32_t crc) {
__asm__ __volatile__(
"enter_256_%=:"
"xor %%r11, %%r11 # zero all 64 bits in r11, will track crc1 \n"
"xor %%r10, %%r10 # zero all 64 bits in r10, will track crc2 \n"
"crc32q 0(%[in]), %%rcx # crc0 \n"
"crc32q 88(%[in]), %%r11 # crc1 \n"
"crc32q 176(%[in]), %%r10 # crc2 \n"
"crc32q 8(%[in]), %%rcx # crc0 \n"
"crc32q 96(%[in]), %%r11 # crc1 \n"
"crc32q 184(%[in]), %%r10 # crc2 \n"
"crc32q 16(%[in]), %%rcx # crc0 \n"
"crc32q 104(%[in]), %%r11 # crc1 \n"
"crc32q 192(%[in]), %%r10 # crc2 \n"
"crc32q 24(%[in]), %%rcx # crc0 \n"
"crc32q 112(%[in]), %%r11 # crc1 \n"
"crc32q 200(%[in]), %%r10 # crc2 \n"
"crc32q 32(%[in]), %%rcx # crc0 \n"
"crc32q 120(%[in]), %%r11 # crc1 \n"
"crc32q 208(%[in]), %%r10 # crc2 \n"
"crc32q 40(%[in]), %%rcx # crc0 \n"
"crc32q 128(%[in]), %%r11 # crc1 \n"
"crc32q 216(%[in]), %%r10 # crc2 \n"
"crc32q 48(%[in]), %%rcx # crc0 \n"
"crc32q 136(%[in]), %%r11 # crc1 \n"
"crc32q 224(%[in]), %%r10 # crc2 \n"
"crc32q 56(%[in]), %%rcx # crc0 \n"
"crc32q 144(%[in]), %%r11 # crc1 \n"
"crc32q 232(%[in]), %%r10 # crc2 \n"
"crc32q 64(%[in]), %%rcx # crc0 \n"
"crc32q 152(%[in]), %%r11 # crc1 \n"
"crc32q 240(%[in]), %%r10 # crc2 \n"
"crc32q 72(%[in]), %%rcx # crc0 \n"
"crc32q 160(%[in]), %%r11 # crc1 \n"
"crc32q 248(%[in]), %%r10 # crc2 \n"
"crc32q 80(%[in]), %%rcx # crc0 \n"
"crc32q 168(%[in]), %%r11 # crc2 \n"
FOLD_K1K2(256, $0x1b3d8f29, $0x39d3b296) /* Magic Constants used to fold crc stripes into ecx */
/* output registers
[crc] is an input and and output so it is marked read/write (i.e. "+c")*/
: [ crc ] "+c"(crc)
/* input registers */
: [ in ] "d"(input)
/* additional clobbered registers */
: "%r8", "%r9", "%r11", "%r10", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "cc");
return crc;
}
/**
* Private (static) function.
* Computes the Castagnoli CRC32c (iSCSI) of the specified data buffer using the Intel CRC32Q (quad word) machine
* instruction by operating on 3 24-byte stripes in parallel. The results are folded together using CLMUL. This function
* is optimized for exactly 1024 byte blocks that are best aligned on 8-byte memory addresses. It MUST be passed a
* pointer to input data that is exactly 1024 bytes in length. Note: this function does NOT invert bits of the input crc
* or return value.
*/
static inline uint32_t s_crc32c_sse42_clmul_1024(const uint8_t *input, uint32_t crc) {
__asm__ __volatile__(
"enter_1024_%=:"
"xor %%r11, %%r11 # zero all 64 bits in r11, will track crc1 \n"
"xor %%r10, %%r10 # zero all 64 bits in r10, will track crc2 \n"
"movl $5, %%r8d # Loop 5 times through 64 byte chunks in 3 parallel stripes \n"
"loop_1024_%=:"
"prefetcht0 128(%[in]) # \n"
"prefetcht0 472(%[in]) # \n"
"prefetcht0 808(%[in]) # \n"
"crc32q 0(%[in]), %%rcx # crc0: stripe0 \n"
"crc32q 344(%[in]), %%r11 # crc1: stripe1 \n"
"crc32q 680(%[in]), %%r10 # crc2: stripe2 \n"
"crc32q 8(%[in]), %%rcx # crc0 \n"
"crc32q 352(%[in]), %%r11 # crc1 \n"
"crc32q 688(%[in]), %%r10 # crc2 \n"
"crc32q 16(%[in]), %%rcx # crc0 \n"
"crc32q 360(%[in]), %%r11 # crc1 \n"
"crc32q 696(%[in]), %%r10 # crc2 \n"
"crc32q 24(%[in]), %%rcx # crc0 \n"
"crc32q 368(%[in]), %%r11 # crc1 \n"
"crc32q 704(%[in]), %%r10 # crc2 \n"
"crc32q 32(%[in]), %%rcx # crc0 \n"
"crc32q 376(%[in]), %%r11 # crc1 \n"
"crc32q 712(%[in]), %%r10 # crc2 \n"
"crc32q 40(%[in]), %%rcx # crc0 \n"
"crc32q 384(%[in]), %%r11 # crc1 \n"
"crc32q 720(%[in]), %%r10 # crc2 \n"
"crc32q 48(%[in]), %%rcx # crc0 \n"
"crc32q 392(%[in]), %%r11 # crc1 \n"
"crc32q 728(%[in]), %%r10 # crc2 \n"
"crc32q 56(%[in]), %%rcx # crc0 \n"
"crc32q 400(%[in]), %%r11 # crc1 \n"
"crc32q 736(%[in]), %%r10 # crc2 \n"
"add $64, %[in] # \n"
"sub $1, %%r8d # \n"
"jnz loop_1024_%= # \n"
"crc32q 0(%[in]), %%rcx # crc0 \n"
"crc32q 344(%[in]), %%r11 # crc1 \n"
"crc32q 680(%[in]), %%r10 # crc2 \n"
"crc32q 8(%[in]), %%rcx # crc0 \n"
"crc32q 352(%[in]), %%r11 # crc1 \n"
"crc32q 688(%[in]), %%r10 # crc2 \n"
"crc32q 16(%[in]), %%rcx # crc0 \n"
"crc32q 696(%[in]), %%r10 # crc2 \n"
FOLD_K1K2(1024, $0xe417f38a, $0x8f158014) /* Magic Constants used to fold crc stripes into ecx
output registers
[crc] is an input and and output so it is marked read/write (i.e. "+c")
we clobber the register for [input] (via add instruction) so we must also
tag it read/write (i.e. "+d") in the list of outputs to tell gcc about the clobber */
: [ crc ] "+c"(crc), [ in ] "+d"(input)
:
/* additional clobbered registers */
/* "cc" is the flags - we add and sub, so the flags are also clobbered */
: "%r8", "%r9", "%r11", "%r10", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "cc");
return crc;
}
/**
* Private (static) function.
* Computes the Castagnoli CRC32c (iSCSI) of the specified data buffer using the Intel CRC32Q (quad word) machine
* instruction by operating on 24-byte stripes in parallel. The results are folded together using CLMUL. This function
* is optimized for exactly 3072 byte blocks that are best aligned on 8-byte memory addresses. It MUST be passed a
* pointer to input data that is exactly 3072 bytes in length. Note: this function does NOT invert bits of the input crc
* or return value.
*/
static inline uint32_t s_crc32c_sse42_clmul_3072(const uint8_t *input, uint32_t crc) {
__asm__ __volatile__(
"enter_3072_%=:"
"xor %%r11, %%r11 # zero all 64 bits in r11, will track crc1 \n"
"xor %%r10, %%r10 # zero all 64 bits in r10, will track crc2 \n"
"movl $16, %%r8d # Loop 16 times through 64 byte chunks in 3 parallel stripes \n"
"loop_3072_%=:"
"prefetcht0 128(%[in]) # \n"
"prefetcht0 1152(%[in]) # \n"
"prefetcht0 2176(%[in]) # \n"
"crc32q 0(%[in]), %%rcx # crc0: stripe0 \n"
"crc32q 1024(%[in]), %%r11 # crc1: stripe1 \n"
"crc32q 2048(%[in]), %%r10 # crc2: stripe2 \n"
"crc32q 8(%[in]), %%rcx # crc0: stripe0 \n"
"crc32q 1032(%[in]), %%r11 # crc1: stripe1 \n"
"crc32q 2056(%[in]), %%r10 # crc2: stripe2 \n"
"crc32q 16(%[in]), %%rcx # crc0: stripe0 \n"
"crc32q 1040(%[in]), %%r11 # crc1: stripe1 \n"
"crc32q 2064(%[in]), %%r10 # crc2: stripe2 \n"
"crc32q 24(%[in]), %%rcx # crc0: stripe0 \n"
"crc32q 1048(%[in]), %%r11 # crc1: stripe1 \n"
"crc32q 2072(%[in]), %%r10 # crc2: stripe2 \n"
"crc32q 32(%[in]), %%rcx # crc0: stripe0 \n"
"crc32q 1056(%[in]), %%r11 # crc1: stripe1 \n"
"crc32q 2080(%[in]), %%r10 # crc2: stripe2 \n"
"crc32q 40(%[in]), %%rcx # crc0: stripe0 \n"
"crc32q 1064(%[in]), %%r11 # crc1: stripe1 \n"
"crc32q 2088(%[in]), %%r10 # crc2: stripe2 \n"
"crc32q 48(%[in]), %%rcx # crc0: stripe0 \n"
"crc32q 1072(%[in]), %%r11 # crc1: stripe1 \n"
"crc32q 2096(%[in]), %%r10 # crc2: stripe2 \n"
"crc32q 56(%[in]), %%rcx # crc0: stripe0 \n"
"crc32q 1080(%[in]), %%r11 # crc1: stripe1 \n"
"crc32q 2104(%[in]), %%r10 # crc2: stripe2 \n"
"add $64, %[in] # \n"
"sub $1, %%r8d # \n"
"jnz loop_3072_%= # \n"
FOLD_K1K2(
3072,
$0xa51b6135,
$0x170076fa) /* Magic Constants used to fold crc stripes into ecx
output registers
[crc] is an input and and output so it is marked read/write (i.e. "+c")
we clobber the register for [input] (via add instruction) so we must also
tag it read/write (i.e. "+d") in the list of outputs to tell gcc about the clobber*/
: [ crc ] "+c"(crc), [ in ] "+d"(input)
:
/* additional clobbered registers
"cc" is the flags - we add and sub, so the flags are also clobbered */
: "%r8", "%r9", "%r11", "%r10", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "cc");
return crc;
}
static bool detection_performed = false;
static bool detected_clmul = false;
/*
* Computes the Castagnoli CRC32c (iSCSI) of the specified data buffer using the Intel CRC32Q (64-bit quad word) and
* PCLMULQDQ machine instructions (if present).
* Handles data that isn't 8-byte aligned as well as any trailing data with the CRC32B (byte) instruction.
* Pass 0 in the previousCrc32 parameter as an initial value unless continuing to update a running CRC in a subsequent
* call.
*/
uint32_t aws_checksums_crc32c_hw(const uint8_t *input, int length, uint32_t previousCrc32) {
if (AWS_UNLIKELY(!detection_performed)) {
detected_clmul = aws_cpu_has_feature(AWS_CPU_FEATURE_CLMUL);
/* Simply setting the flag true to skip HW detection next time
Not using memory barriers since the worst that can
happen is a fallback to the non HW accelerated code. */
detection_performed = true;
}
uint32_t crc = ~previousCrc32;
/* For small input, forget about alignment checks - simply compute the CRC32c one byte at a time */
if (AWS_UNLIKELY(length < 8)) {
while (length-- > 0) {
__asm__("loop_small_%=: CRC32B (%[in]), %[crc]" : [ crc ] "+c"(crc) : [ in ] "r"(input));
input++;
}
return ~crc;
}
/* Get the 8-byte memory alignment of our input buffer by looking at the least significant 3 bits */
int input_alignment = (unsigned long int)input & 0x7;
/* Compute the number of unaligned bytes before the first aligned 8-byte chunk (will be in the range 0-7) */
int leading = (8 - input_alignment) & 0x7;
/* reduce the length by the leading unaligned bytes we are about to process */
length -= leading;
/* spin through the leading unaligned input bytes (if any) one-by-one */
while (leading-- > 0) {
__asm__("loop_leading_%=: CRC32B (%[in]), %[crc]" : [ crc ] "+c"(crc) : [ in ] "r"(input));
input++;
}
/* Using likely to keep this code inlined */
if (AWS_LIKELY(detected_clmul)) {
while (AWS_LIKELY(length >= 3072)) {
/* Compute crc32c on each block, chaining each crc result */
crc = s_crc32c_sse42_clmul_3072(input, crc);
input += 3072;
length -= 3072;
}
while (AWS_LIKELY(length >= 1024)) {
/* Compute crc32c on each block, chaining each crc result */
crc = s_crc32c_sse42_clmul_1024(input, crc);
input += 1024;
length -= 1024;
}
while (AWS_LIKELY(length >= 256)) {
/* Compute crc32c on each block, chaining each crc result */
crc = s_crc32c_sse42_clmul_256(input, crc);
input += 256;
length -= 256;
}
}
/* Spin through remaining (aligned) 8-byte chunks using the CRC32Q quad word instruction */
while (AWS_LIKELY(length >= 8)) {
/* Hardcoding %rcx register (i.e. "+c") to allow use of qword instruction */
__asm__ __volatile__("loop_8_%=: CRC32Q (%[in]), %%rcx" : [ crc ] "+c"(crc) : [ in ] "r"(input));
input += 8;
length -= 8;
}
/* Finish up with any trailing bytes using the CRC32B single byte instruction one-by-one */
while (length-- > 0) {
__asm__ __volatile__("loop_trailing_%=: CRC32B (%[in]), %[crc]" : [ crc ] "+c"(crc) : [ in ] "r"(input));
input++;
}
return ~crc;
}
uint32_t aws_checksums_crc32_hw(const uint8_t *input, int length, uint32_t previousCrc32) {
return aws_checksums_crc32_sw(input, length, previousCrc32);
}
# if defined(__clang__)
# pragma clang diagnostic pop
# endif
#else
uint32_t aws_checksums_crc32_hw(const uint8_t *input, int length, uint32_t previousCrc32) {
return aws_checksums_crc32_sw(input, length, previousCrc32);
}
uint32_t aws_checksums_crc32c_hw(const uint8_t *input, int length, uint32_t previousCrc32) {
return aws_checksums_crc32c_sw(input, length, previousCrc32);
}
#endif
|
