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
|
// +build ignore
package main
import (
. "github.com/mmcloughlin/avo/build"
. "github.com/mmcloughlin/avo/operand"
. "github.com/mmcloughlin/avo/reg"
. "github.com/segmentio/asm/build/internal/x86"
"fmt"
"github.com/segmentio/asm/cpu"
)
func init() {
ConstraintExpr("!purego")
}
func main() {
TEXT("EqualFoldString", NOSPLIT, "func(a, b string) bool")
Doc(
"EqualFoldString is a version of strings.EqualFold designed to work on ASCII",
"input instead of UTF-8.",
"",
"When the program has guarantees that the input is composed of ASCII",
"characters only, it allows for greater optimizations.",
)
// Use index for byte position. We have plenty of registers, and it saves an
// ADD operation as the memory index is the same for both a and b.
i := GP64()
a := Mem{Base: Load(Param("a").Base(), GP64()), Index: i, Scale: 1}
n := Load(Param("a").Len(), GP64())
b := Mem{Base: Load(Param("b").Base(), GP64()), Index: i, Scale: 1}
bn, _ := Param("b").Len().Resolve()
ret, _ := ReturnIndex(0).Resolve()
CMPQ(n, bn.Addr) // if len(a) != len(b):
JNE(LabelRef("done")) // return false
XORQ(i, i) // i = 0
CMPQ(n, U8(16)) // if n < 16:
JB(LabelRef("init_x86")) // goto init_x86
JumpIfFeature("init_avx", cpu.AVX2) // goto init_avx if supported
Label("init_x86")
cmp := GP32()
av := GP32()
bv := GP32()
// Map to convert ASCII upper characters to lower case.
lowerCase := Mem{Base: GP64(), Scale: 1}
LEAQ(NewDataAddr(Symbol{Name: "github·com∕segmentio∕asm∕ascii·lowerCase"}, 0), lowerCase.Base)
XORL(cmp, cmp)
Label("cmp8")
CMPQ(n, U8(8)) // if n < 0:
JB(LabelRef("cmp7")) // goto cmp7
for i := 0; i < 8; i++ {
MOVBLZX(a.Offset(i), av) // av = a[i]
MOVBLZX(b.Offset(i), bv) // bv = b[i]
MOVB(lowerCase.Idx(av, 1), av.As8()) // av = lowerCase[av]
XORB(lowerCase.Idx(bv, 1), av.As8()) // av = lowerCase[bv] ^ av
ORB(av.As8(), cmp.As8()) // cmp |= av
}
JNE(LabelRef("done")) // return false if ZF == 0
ADDQ(Imm(8), a.Index) // i += 8
SUBQ(Imm(8), n) // n -= 8
JMP(LabelRef("cmp8"))
for i := 6; i >= 0; i-- {
Label(fmt.Sprintf("cmp%d", i+1))
next := "success"
if i > 0 {
next = fmt.Sprintf("cmp%d", i)
}
CMPQ(n, U8(i+1)) // if n < i:
JB(LabelRef(next)) // goto cmp${i-1}
MOVBLZX(a.Offset(i), av) // av = a[i]
MOVBLZX(b.Offset(i), bv) // bv = b[i]
MOVB(lowerCase.Idx(av, 1), av.As8()) // av = lowerCase[av]
XORB(lowerCase.Idx(bv, 1), av.As8()) // av = lowerCase[bv] ^ av
ORB(av.As8(), cmp.As8()) // cmp |= av
}
Label("done")
SETEQ(ret.Addr) // return ZF
RET() // ...
Label("success")
MOVB(U8(1), ret.Addr) // return true
RET() // ...
Label("init_avx")
bit := VecBroadcast(U8(0x20), YMM()) // "case" bit
msk := VecBroadcast(U8(0x1F), YMM()) // 0b10000000 - 'a'
rng := VecBroadcast(U8(0x9A), YMM()) // 'z' - 'a' + 1 - 0x80 (overflowed 8-bits)
one := VecBroadcast(U8(0x01), YMM()) // 1-bit for ANDing with comparison
vec := NewVectorizer(12, func(l VectorLane) Register {
v0 := l.Read(a)
v1 := l.Read(b)
v2 := l.Alloc()
VXORPD(v0, v1, v1) // calculate difference between a and b
VPCMPEQB(bit, v1, v2) // check if above difference is the 6th bit
VORPD(bit, v0, v0) // set the 6th bit for a
VPADDB(msk, v0, v0) // add 0x1f to each byte to set top bit for letters
VPCMPGTB(v0, rng, v0) // compare if not letter: v - 'a' < 'z' - 'a' + 1
VPAND(v2, v0, v0) // combine 6th-bit difference with letter range
VPAND(one, v0, v0) // merge test mask
VPSLLW(Imm(5), v0, v0) // shift into case bit position
VPCMPEQB(v1, v0, v0) // compare original difference with case-only difference
return v0
}).Reduce(ReduceAnd) // merge all comparisons together
cmpAVX := func(spec Spec, lanes int, incr bool) {
sz := int(spec.Size())
tmp := GP32()
out := vec.Compile(spec, lanes)[0] // [compare sz*lanes bytes]
if incr {
ADDQ(U8(sz*lanes), a.Index) // i += sz*lanes
SUBQ(U8(sz*lanes), n) // n -= sz*lanes
}
VPMOVMSKB(out, tmp) // tmp[0,1,2,...] = y0[0,8,16,...]
XORL(U32(^uint32(0)>>(32-sz)), tmp) // ZF = (tmp == 0xFFFFFFFF)
}
Label("cmp128")
CMPQ(n, U8(128)) // if n < 128:
JB(LabelRef("cmp64")) // goto cmp64
cmpAVX(S256, 4, true) // ZF = [compare 128 bytes]
JNE(LabelRef("done")) // return if ZF == 0
JMP(LabelRef("cmp128")) // loop cmp128
Label("cmp64")
CMPQ(n, U8(64)) // if n < 64:
JB(LabelRef("cmp32")) // goto cmp32
cmpAVX(S256, 2, true) // ZF = [compare 64 bytes]
JNE(LabelRef("done")) // return if ZF == 0
Label("cmp32")
CMPQ(n, U8(32)) // if n < 32:
JB(LabelRef("cmp16")) // goto cmp16
cmpAVX(S256, 1, true) // ZF = [compare 32 bytes]
JNE(LabelRef("done")) // return if ZF == 0
Label("cmp16")
// Convert YMM masks to XMM
bit = bit.(Vec).AsX()
msk = msk.(Vec).AsX()
rng = rng.(Vec).AsX()
one = one.(Vec).AsX()
CMPQ(n, U8(16)) // if n <= 16:
JLE(LabelRef("cmp_tail")) // goto cmp_tail
cmpAVX(S128, 1, true) // ZF = [compare 16 bytes]
JNE(LabelRef("done")) // return if ZF == 0
Label("cmp_tail")
// At this point, we have <= 16 bytes to compare, but we know the total input
// is >= 16 bytes. Move the pointer to the *last* 16 bytes of the input so we
// can skip the fallback.
SUBQ(Imm(16), n) // n -= 16
ADDQ(n, a.Index) // i += n
cmpAVX(S128, 1, false) // ZF = [compare 16 bytes]
JMP(LabelRef("done")) // return ZF
Generate()
}
|
