crc32_amd64.s 5.97 KB
Newer Older
1 2 3 4
// Copyright 2011 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

5 6
// +build gc

7 8 9
#define NOSPLIT 4
#define RODATA 8

10 11
// castagnoliSSE42 updates the (non-inverted) crc with the given buffer.
//
12 13 14 15 16 17
// func castagnoliSSE42(crc uint32, p []byte) uint32
TEXT ·castagnoliSSE42(SB), NOSPLIT, $0
	MOVL crc+0(FP), AX    // CRC value
	MOVQ p+8(FP), SI      // data pointer
	MOVQ p_len+16(FP), CX // len(p)

18
	// If there are fewer than 8 bytes to process, skip alignment.
19
	CMPQ CX, $8
20
	JL   less_than_8
21 22 23 24 25

	MOVQ SI, BX
	ANDQ $7, BX
	JZ   aligned

26 27 28 29 30 31 32 33 34
	// Process the first few bytes to 8-byte align the input.

	// BX = 8 - BX. We need to process this many bytes to align.
	SUBQ $1, BX
	XORQ $7, BX

	BTQ $0, BX
	JNC align_2

35 36 37
	CRC32B (SI), AX
	DECQ   CX
	INCQ   SI
38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57

align_2:
	BTQ $1, BX
	JNC align_4

	// CRC32W (SI), AX
	BYTE $0x66; BYTE $0xf2; BYTE $0x0f; BYTE $0x38; BYTE $0xf1; BYTE $0x06

	SUBQ $2, CX
	ADDQ $2, SI

align_4:
	BTQ $2, BX
	JNC aligned

	// CRC32L (SI), AX
	BYTE $0xf2; BYTE $0x0f; BYTE $0x38; BYTE $0xf1; BYTE $0x06

	SUBQ $4, CX
	ADDQ $4, SI
58 59 60 61

aligned:
	// The input is now 8-byte aligned and we can process 8-byte chunks.
	CMPQ CX, $8
62
	JL   less_than_8
63 64 65 66 67 68

	CRC32Q (SI), AX
	ADDQ   $8, SI
	SUBQ   $8, CX
	JMP    aligned

69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88
less_than_8:
	// We may have some bytes left over; process 4 bytes, then 2, then 1.
	BTQ $2, CX
	JNC less_than_4

	// CRC32L (SI), AX
	BYTE $0xf2; BYTE $0x0f; BYTE $0x38; BYTE $0xf1; BYTE $0x06
	ADDQ $4, SI

less_than_4:
	BTQ $1, CX
	JNC less_than_2

	// CRC32W (SI), AX
	BYTE $0x66; BYTE $0xf2; BYTE $0x0f; BYTE $0x38; BYTE $0xf1; BYTE $0x06
	ADDQ $2, SI

less_than_2:
	BTQ $0, CX
	JNC done
89 90 91 92 93 94 95

	CRC32B (SI), AX

done:
	MOVL AX, ret+32(FP)
	RET

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
// castagnoliSSE42Triple updates three (non-inverted) crcs with (24*rounds)
// bytes from each buffer.
//
// func castagnoliSSE42Triple(
//     crc1, crc2, crc3 uint32,
//     a, b, c []byte,
//     rounds uint32,
// ) (retA uint32, retB uint32, retC uint32)
TEXT ·castagnoliSSE42Triple(SB), NOSPLIT, $0
	MOVL crcA+0(FP), AX
	MOVL crcB+4(FP), CX
	MOVL crcC+8(FP), DX

	MOVQ a+16(FP), R8  // data pointer
	MOVQ b+40(FP), R9  // data pointer
	MOVQ c+64(FP), R10 // data pointer

	MOVL rounds+88(FP), R11

loop:
	CRC32Q (R8), AX
	CRC32Q (R9), CX
	CRC32Q (R10), DX

	CRC32Q 8(R8), AX
	CRC32Q 8(R9), CX
	CRC32Q 8(R10), DX

	CRC32Q 16(R8), AX
	CRC32Q 16(R9), CX
	CRC32Q 16(R10), DX

	ADDQ $24, R8
	ADDQ $24, R9
	ADDQ $24, R10

	DECQ R11
	JNZ  loop

	MOVL AX, retA+96(FP)
	MOVL CX, retB+100(FP)
	MOVL DX, retC+104(FP)
	RET

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
// func haveSSE42() bool
TEXT ·haveSSE42(SB), NOSPLIT, $0
	XORQ AX, AX
	INCL AX
	CPUID
	SHRQ $20, CX
	ANDQ $1, CX
	MOVB CX, ret+0(FP)
	RET

// func haveCLMUL() bool
TEXT ·haveCLMUL(SB), NOSPLIT, $0
	XORQ AX, AX
	INCL AX
	CPUID
	SHRQ $1, CX
	ANDQ $1, CX
	MOVB CX, ret+0(FP)
	RET

// func haveSSE41() bool
TEXT ·haveSSE41(SB), NOSPLIT, $0
	XORQ AX, AX
	INCL AX
	CPUID
	SHRQ $19, CX
	ANDQ $1, CX
	MOVB CX, ret+0(FP)
	RET

// CRC32 polynomial data
//
// These constants are lifted from the
// Linux kernel, since they avoid the costly
// PSHUFB 16 byte reversal proposed in the
// original Intel paper.
DATA r2r1kp<>+0(SB)/8, $0x154442bd4
DATA r2r1kp<>+8(SB)/8, $0x1c6e41596
DATA r4r3kp<>+0(SB)/8, $0x1751997d0
DATA r4r3kp<>+8(SB)/8, $0x0ccaa009e
DATA rupolykp<>+0(SB)/8, $0x1db710641
DATA rupolykp<>+8(SB)/8, $0x1f7011641
DATA r5kp<>+0(SB)/8, $0x163cd6124

GLOBL r2r1kp<>(SB), RODATA, $16
GLOBL r4r3kp<>(SB), RODATA, $16
GLOBL rupolykp<>(SB), RODATA, $16
GLOBL r5kp<>(SB), RODATA, $8

// Based on http://www.intel.com/content/dam/www/public/us/en/documents/white-papers/fast-crc-computation-generic-polynomials-pclmulqdq-paper.pdf
// len(p) must be at least 64, and must be a multiple of 16.

// func ieeeCLMUL(crc uint32, p []byte) uint32
TEXT ·ieeeCLMUL(SB), NOSPLIT, $0
	MOVL crc+0(FP), X0    // Initial CRC value
	MOVQ p+8(FP), SI      // data pointer
	MOVQ p_len+16(FP), CX // len(p)

	MOVOU (SI), X1
	MOVOU 16(SI), X2
	MOVOU 32(SI), X3
	MOVOU 48(SI), X4
	PXOR  X0, X1
	ADDQ  $64, SI    // buf+=64
	SUBQ  $64, CX    // len-=64
	CMPQ  CX, $64    // Less than 64 bytes left
	JB    remain64

208
	MOVOA r2r1kp<>+0(SB), X0
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

loopback64:
	MOVOA X1, X5
	MOVOA X2, X6
	MOVOA X3, X7
	MOVOA X4, X8

	PCLMULQDQ $0, X0, X1
	PCLMULQDQ $0, X0, X2
	PCLMULQDQ $0, X0, X3
	PCLMULQDQ $0, X0, X4

	// Load next early
	MOVOU (SI), X11
	MOVOU 16(SI), X12
	MOVOU 32(SI), X13
	MOVOU 48(SI), X14

	PCLMULQDQ $0x11, X0, X5
	PCLMULQDQ $0x11, X0, X6
	PCLMULQDQ $0x11, X0, X7
	PCLMULQDQ $0x11, X0, X8

	PXOR X5, X1
	PXOR X6, X2
	PXOR X7, X3
	PXOR X8, X4

	PXOR X11, X1
	PXOR X12, X2
	PXOR X13, X3
	PXOR X14, X4

	ADDQ $0x40, DI
	ADDQ $64, SI    // buf+=64
	SUBQ $64, CX    // len-=64
	CMPQ CX, $64    // Less than 64 bytes left?
	JGE  loopback64

	// Fold result into a single register (X1)
remain64:
250
	MOVOA r4r3kp<>+0(SB), X0
251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

	MOVOA     X1, X5
	PCLMULQDQ $0, X0, X1
	PCLMULQDQ $0x11, X0, X5
	PXOR      X5, X1
	PXOR      X2, X1

	MOVOA     X1, X5
	PCLMULQDQ $0, X0, X1
	PCLMULQDQ $0x11, X0, X5
	PXOR      X5, X1
	PXOR      X3, X1

	MOVOA     X1, X5
	PCLMULQDQ $0, X0, X1
	PCLMULQDQ $0x11, X0, X5
	PXOR      X5, X1
	PXOR      X4, X1

270
	// If there is less than 16 bytes left we are done
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
	CMPQ CX, $16
	JB   finish

	// Encode 16 bytes
remain16:
	MOVOU     (SI), X10
	MOVOA     X1, X5
	PCLMULQDQ $0, X0, X1
	PCLMULQDQ $0x11, X0, X5
	PXOR      X5, X1
	PXOR      X10, X1
	SUBQ      $16, CX
	ADDQ      $16, SI
	CMPQ      CX, $16
	JGE       remain16

finish:
	// Fold final result into 32 bits and return it
	PCMPEQB   X3, X3
	PCLMULQDQ $1, X1, X0
	PSRLDQ    $8, X1
	PXOR      X0, X1

	MOVOA X1, X2
	MOVQ  r5kp<>+0(SB), X0

	// Creates 32 bit mask. Note that we don't care about upper half.
	PSRLQ $32, X3

	PSRLDQ    $4, X2
	PAND      X3, X1
	PCLMULQDQ $0, X0, X1
	PXOR      X2, X1

305
	MOVOA rupolykp<>+0(SB), X0
306 307 308 309 310 311 312 313 314 315 316 317 318 319

	MOVOA     X1, X2
	PAND      X3, X1
	PCLMULQDQ $0x10, X0, X1
	PAND      X3, X1
	PCLMULQDQ $0, X0, X1
	PXOR      X2, X1

	// PEXTRD   $1, X1, AX  (SSE 4.1)
	BYTE $0x66; BYTE $0x0f; BYTE $0x3a
	BYTE $0x16; BYTE $0xc8; BYTE $0x01
	MOVL AX, ret+32(FP)

	RET