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
|
// +build ignore
//
// This code is a go assembly implementation of:
//
// Muła, Wojciech, & Lemire, Daniel (Thu, 14 Jun 2018).
// Faster Base64 Encoding and Decoding Using AVX2 Instructions.
// [arXiv:1704.00605](https://arxiv.org/abs/1704.00605)
//
// ...with changes to support multiple encodings.
package main
import (
. "github.com/mmcloughlin/avo/build"
. "github.com/mmcloughlin/avo/gotypes"
. "github.com/mmcloughlin/avo/operand"
. "github.com/mmcloughlin/avo/reg"
. "github.com/segmentio/asm/build/internal/asm"
. "github.com/segmentio/asm/build/internal/x86"
)
var lutHi = ConstBytes("b64_dec_lut_hi", []byte{
16, 16, 1, 2, 4, 8, 4, 8, 16, 16, 16, 16, 16, 16, 16, 16,
16, 16, 1, 2, 4, 8, 4, 8, 16, 16, 16, 16, 16, 16, 16, 16,
})
var madd1 = ConstBytes("b64_dec_madd1", []byte{
64, 1, 64, 1, 64, 1, 64, 1, 64, 1, 64, 1, 64, 1, 64, 1,
64, 1, 64, 1, 64, 1, 64, 1, 64, 1, 64, 1, 64, 1, 64, 1,
})
var madd2 = ConstArray16("b64_dec_madd2",
4096, 1, 4096, 1, 4096, 1, 4096, 1,
4096, 1, 4096, 1, 4096, 1, 4096, 1,
)
var shufLo = ConstBytes("b64_dec_shuf_lo", []byte{
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 1, 0, 6,
})
var shuf = ConstBytes("b64_dec_shuf", []byte{
2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, 0, 0, 0, 0,
5, 4, 10, 9, 8, 14, 13, 12, 0, 0, 0, 0, 0, 0, 0, 0,
})
func init() {
ConstraintExpr("!purego")
}
func main() {
TEXT("decodeAVX2", NOSPLIT, "func(dst, src []byte, lut *int8) (int, int)")
createDecode(Param("dst"), Param("src"), Param("lut"), func(m Mem, r VecVirtual) {
VMOVDQU(m, r)
})
TEXT("decodeAVX2URI", NOSPLIT, "func(dst, src []byte, lut *int8) (int, int)")
slash := VecBroadcast(U8('/'), YMM())
underscore := VecBroadcast(U8('_'), YMM())
createDecode(Param("dst"), Param("src"), Param("lut"), func(m Mem, r VecVirtual) {
eq := YMM()
VMOVDQU(m, r)
VPCMPEQB(r, underscore, eq)
VPBLENDVB(eq, slash, r, r)
})
Generate()
}
func createDecode(pdst, psrc, plut Component, load func(m Mem, r VecVirtual)) {
dst := Mem{Base: Load(pdst.Base(), GP64()), Index: GP64(), Scale: 1}
src := Mem{Base: Load(psrc.Base(), GP64()), Index: GP64(), Scale: 1}
lut := Mem{Base: Load(plut, GP64())}
rem := Load(psrc.Len(), GP64())
rsrc := YMM()
rdst := YMM()
nibh := YMM()
nibl := YMM()
emsk := YMM()
roll := YMM()
shfl := YMM()
lutl := YMM()
luth := YMM()
lutr := YMM()
zero := YMM()
lo := YMM()
hi := YMM()
mask := VecBroadcast(U8(0x2f), YMM())
XORQ(dst.Index, dst.Index)
XORQ(src.Index, src.Index)
VPXOR(zero, zero, zero)
VPERMQ(Imm(1<<6|1<<2), lut, lutr)
VPERMQ(Imm(1<<6|1<<2), lut.Offset(16), lutl)
VMOVDQA(lutHi, luth)
Label("loop")
load(src, rsrc)
VPSRLD(Imm(4), rsrc, nibh)
VPAND(mask, rsrc, nibl)
VPSHUFB(nibl, lutl, lo)
VPAND(mask, nibh, nibh)
VPSHUFB(nibh, luth, hi)
VPTEST(hi, lo)
JNE(LabelRef("done"))
VPCMPEQB(mask, rsrc, emsk)
VPADDB(emsk, nibh, roll)
VPSHUFB(roll, lutr, roll)
VPADDB(rsrc, roll, shfl)
VPMADDUBSW(madd1, shfl, shfl)
VPMADDWD(madd2, shfl, shfl)
VEXTRACTI128(Imm(1), shfl, rdst.AsX())
VPSHUFB(shufLo, rdst.AsX(), rdst.AsX())
VPSHUFB(shuf, shfl, shfl)
VPBLENDD(Imm(8), rdst, shfl, rdst)
VPBLENDD(Imm(192), zero, rdst, rdst)
VMOVDQU(rdst, dst)
ADDQ(Imm(24), dst.Index)
ADDQ(Imm(32), src.Index)
SUBQ(Imm(32), rem)
CMPQ(rem, Imm(45))
JB(LabelRef("done"))
JMP(LabelRef("loop"))
Label("done")
Store(dst.Index, ReturnIndex(0))
Store(src.Index, ReturnIndex(1))
VZEROUPPER()
RET()
}
|
