| 1 |
// +build !noasm |
| 2 |
|
| 3 |
#include "go_asm.h" |
| 4 |
#include "textflag.h" |
| 5 |
|
| 6 |
// Register allocation. |
| 7 |
#define p R0 |
| 8 |
#define n R1 |
| 9 |
#define h R2 |
| 10 |
#define v1 R2 // Alias for h. |
| 11 |
#define v2 R3 |
| 12 |
#define v3 R4 |
| 13 |
#define v4 R5 |
| 14 |
#define x1 R6 |
| 15 |
#define x2 R7 |
| 16 |
#define x3 R8 |
| 17 |
#define x4 R9 |
| 18 |
|
| 19 |
// We need the primes in registers. The 16-byte loop only uses prime{1,2}. |
| 20 |
#define prime1r R11 |
| 21 |
#define prime2r R12 |
| 22 |
#define prime3r R3 // The rest can alias v{2-4}. |
| 23 |
#define prime4r R4 |
| 24 |
#define prime5r R5 |
| 25 |
|
| 26 |
// Update round macros. These read from and increment p. |
| 27 |
|
| 28 |
#define round16aligned \ |
| 29 |
MOVM.IA.W (p), [x1, x2, x3, x4] \ |
| 30 |
\ |
| 31 |
MULA x1, prime2r, v1, v1 \ |
| 32 |
MULA x2, prime2r, v2, v2 \ |
| 33 |
MULA x3, prime2r, v3, v3 \ |
| 34 |
MULA x4, prime2r, v4, v4 \ |
| 35 |
\ |
| 36 |
MOVW v1 @> 19, v1 \ |
| 37 |
MOVW v2 @> 19, v2 \ |
| 38 |
MOVW v3 @> 19, v3 \ |
| 39 |
MOVW v4 @> 19, v4 \ |
| 40 |
\ |
| 41 |
MUL prime1r, v1 \ |
| 42 |
MUL prime1r, v2 \ |
| 43 |
MUL prime1r, v3 \ |
| 44 |
MUL prime1r, v4 \ |
| 45 |
|
| 46 |
#define round16unaligned \ |
| 47 |
MOVBU.P 16(p), x1 \ |
| 48 |
MOVBU -15(p), x2 \ |
| 49 |
ORR x2 << 8, x1 \ |
| 50 |
MOVBU -14(p), x3 \ |
| 51 |
MOVBU -13(p), x4 \ |
| 52 |
ORR x4 << 8, x3 \ |
| 53 |
ORR x3 << 16, x1 \ |
| 54 |
\ |
| 55 |
MULA x1, prime2r, v1, v1 \ |
| 56 |
MOVW v1 @> 19, v1 \ |
| 57 |
MUL prime1r, v1 \ |
| 58 |
\ |
| 59 |
MOVBU -12(p), x1 \ |
| 60 |
MOVBU -11(p), x2 \ |
| 61 |
ORR x2 << 8, x1 \ |
| 62 |
MOVBU -10(p), x3 \ |
| 63 |
MOVBU -9(p), x4 \ |
| 64 |
ORR x4 << 8, x3 \ |
| 65 |
ORR x3 << 16, x1 \ |
| 66 |
\ |
| 67 |
MULA x1, prime2r, v2, v2 \ |
| 68 |
MOVW v2 @> 19, v2 \ |
| 69 |
MUL prime1r, v2 \ |
| 70 |
\ |
| 71 |
MOVBU -8(p), x1 \ |
| 72 |
MOVBU -7(p), x2 \ |
| 73 |
ORR x2 << 8, x1 \ |
| 74 |
MOVBU -6(p), x3 \ |
| 75 |
MOVBU -5(p), x4 \ |
| 76 |
ORR x4 << 8, x3 \ |
| 77 |
ORR x3 << 16, x1 \ |
| 78 |
\ |
| 79 |
MULA x1, prime2r, v3, v3 \ |
| 80 |
MOVW v3 @> 19, v3 \ |
| 81 |
MUL prime1r, v3 \ |
| 82 |
\ |
| 83 |
MOVBU -4(p), x1 \ |
| 84 |
MOVBU -3(p), x2 \ |
| 85 |
ORR x2 << 8, x1 \ |
| 86 |
MOVBU -2(p), x3 \ |
| 87 |
MOVBU -1(p), x4 \ |
| 88 |
ORR x4 << 8, x3 \ |
| 89 |
ORR x3 << 16, x1 \ |
| 90 |
\ |
| 91 |
MULA x1, prime2r, v4, v4 \ |
| 92 |
MOVW v4 @> 19, v4 \ |
| 93 |
MUL prime1r, v4 \ |
| 94 |
|
| 95 |
|
| 96 |
// func ChecksumZero([]byte) uint32 |
| 97 |
TEXT ·ChecksumZero(SB), NOFRAME|NOSPLIT, $-4-16 |
| 98 |
MOVW input_base+0(FP), p |
| 99 |
MOVW input_len+4(FP), n |
| 100 |
|
| 101 |
MOVW $const_prime1, prime1r |
| 102 |
MOVW $const_prime2, prime2r |
| 103 |
|
| 104 |
// Set up h for n < 16. It's tempting to say {ADD prime5, n, h} |
| 105 |
// here, but that's a pseudo-op that generates a load through R11. |
| 106 |
MOVW $const_prime5, prime5r |
| 107 |
ADD prime5r, n, h |
| 108 |
CMP $0, n |
| 109 |
BEQ end |
| 110 |
|
| 111 |
// We let n go negative so we can do comparisons with SUB.S |
| 112 |
// instead of separate CMP. |
| 113 |
SUB.S $16, n |
| 114 |
BMI loop16done |
| 115 |
|
| 116 |
ADD prime1r, prime2r, v1 |
| 117 |
MOVW prime2r, v2 |
| 118 |
MOVW $0, v3 |
| 119 |
RSB $0, prime1r, v4 |
| 120 |
|
| 121 |
TST $3, p |
| 122 |
BNE loop16unaligned |
| 123 |
|
| 124 |
loop16aligned: |
| 125 |
SUB.S $16, n |
| 126 |
round16aligned |
| 127 |
BPL loop16aligned |
| 128 |
B loop16finish |
| 129 |
|
| 130 |
loop16unaligned: |
| 131 |
SUB.S $16, n |
| 132 |
round16unaligned |
| 133 |
BPL loop16unaligned |
| 134 |
|
| 135 |
loop16finish: |
| 136 |
MOVW v1 @> 31, h |
| 137 |
ADD v2 @> 25, h |
| 138 |
ADD v3 @> 20, h |
| 139 |
ADD v4 @> 14, h |
| 140 |
|
| 141 |
// h += len(input) with v2 as temporary. |
| 142 |
MOVW input_len+4(FP), v2 |
| 143 |
ADD v2, h |
| 144 |
|
| 145 |
loop16done: |
| 146 |
ADD $16, n // Restore number of bytes left. |
| 147 |
|
| 148 |
SUB.S $4, n |
| 149 |
MOVW $const_prime3, prime3r |
| 150 |
BMI loop4done |
| 151 |
MOVW $const_prime4, prime4r |
| 152 |
|
| 153 |
TST $3, p |
| 154 |
BNE loop4unaligned |
| 155 |
|
| 156 |
loop4aligned: |
| 157 |
SUB.S $4, n |
| 158 |
|
| 159 |
MOVW.P 4(p), x1 |
| 160 |
MULA prime3r, x1, h, h |
| 161 |
MOVW h @> 15, h |
| 162 |
MUL prime4r, h |
| 163 |
|
| 164 |
BPL loop4aligned |
| 165 |
B loop4done |
| 166 |
|
| 167 |
loop4unaligned: |
| 168 |
SUB.S $4, n |
| 169 |
|
| 170 |
MOVBU.P 4(p), x1 |
| 171 |
MOVBU -3(p), x2 |
| 172 |
ORR x2 << 8, x1 |
| 173 |
MOVBU -2(p), x3 |
| 174 |
ORR x3 << 16, x1 |
| 175 |
MOVBU -1(p), x4 |
| 176 |
ORR x4 << 24, x1 |
| 177 |
|
| 178 |
MULA prime3r, x1, h, h |
| 179 |
MOVW h @> 15, h |
| 180 |
MUL prime4r, h |
| 181 |
|
| 182 |
BPL loop4unaligned |
| 183 |
|
| 184 |
loop4done: |
| 185 |
ADD.S $4, n // Restore number of bytes left. |
| 186 |
BEQ end |
| 187 |
|
| 188 |
MOVW $const_prime5, prime5r |
| 189 |
|
| 190 |
loop1: |
| 191 |
SUB.S $1, n |
| 192 |
|
| 193 |
MOVBU.P 1(p), x1 |
| 194 |
MULA prime5r, x1, h, h |
| 195 |
MOVW h @> 21, h |
| 196 |
MUL prime1r, h |
| 197 |
|
| 198 |
BNE loop1 |
| 199 |
|
| 200 |
end: |
| 201 |
MOVW $const_prime3, prime3r |
| 202 |
EOR h >> 15, h |
| 203 |
MUL prime2r, h |
| 204 |
EOR h >> 13, h |
| 205 |
MUL prime3r, h |
| 206 |
EOR h >> 16, h |
| 207 |
|
| 208 |
MOVW h, ret+12(FP) |
| 209 |
RET |
| 210 |
|
| 211 |
|
| 212 |
// func update(v *[4]uint64, buf *[16]byte, p []byte) |
| 213 |
TEXT ·update(SB), NOFRAME|NOSPLIT, $-4-20 |
| 214 |
MOVW v+0(FP), p |
| 215 |
MOVM.IA (p), [v1, v2, v3, v4] |
| 216 |
|
| 217 |
MOVW $const_prime1, prime1r |
| 218 |
MOVW $const_prime2, prime2r |
| 219 |
|
| 220 |
// Process buf, if not nil. |
| 221 |
MOVW buf+4(FP), p |
| 222 |
CMP $0, p |
| 223 |
BEQ noBuffered |
| 224 |
|
| 225 |
round16aligned |
| 226 |
|
| 227 |
noBuffered: |
| 228 |
MOVW input_base +8(FP), p |
| 229 |
MOVW input_len +12(FP), n |
| 230 |
|
| 231 |
SUB.S $16, n |
| 232 |
BMI end |
| 233 |
|
| 234 |
TST $3, p |
| 235 |
BNE loop16unaligned |
| 236 |
|
| 237 |
loop16aligned: |
| 238 |
SUB.S $16, n |
| 239 |
round16aligned |
| 240 |
BPL loop16aligned |
| 241 |
B end |
| 242 |
|
| 243 |
loop16unaligned: |
| 244 |
SUB.S $16, n |
| 245 |
round16unaligned |
| 246 |
BPL loop16unaligned |
| 247 |
|
| 248 |
end: |
| 249 |
MOVW v+0(FP), p |
| 250 |
MOVM.IA [v1, v2, v3, v4], (p) |
| 251 |
RET |