json-as 1.4.0 → 1.5.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +50 -29
- package/README.md +84 -33
- package/assembly/custom/chars.ts +39 -78
- package/assembly/deserialize/index/arbitrary.ts +26 -8
- package/assembly/deserialize/index/float.ts +2 -4
- package/assembly/deserialize/index/integer.ts +2 -4
- package/assembly/deserialize/index/object.ts +6 -1
- package/assembly/deserialize/index/string.ts +2 -7
- package/assembly/deserialize/index/unsigned.ts +2 -4
- package/assembly/deserialize/naive/array/integer.ts +1 -1
- package/assembly/deserialize/naive/array/map.ts +1 -1
- package/assembly/deserialize/naive/array/object.ts +1 -1
- package/assembly/deserialize/naive/array/struct.ts +19 -1
- package/assembly/deserialize/naive/bool.ts +1 -5
- package/assembly/deserialize/naive/date.ts +1 -2
- package/assembly/deserialize/naive/float.ts +2 -7
- package/assembly/deserialize/naive/integer.ts +1 -2
- package/assembly/deserialize/naive/map.ts +5 -6
- package/assembly/deserialize/naive/object.ts +151 -13
- package/assembly/deserialize/naive/raw.ts +1 -5
- package/assembly/deserialize/naive/set.ts +2 -4
- package/assembly/deserialize/naive/staticarray.ts +1 -2
- package/assembly/deserialize/naive/string.ts +6 -9
- package/assembly/deserialize/naive/unsigned.ts +1 -2
- package/assembly/deserialize/simd/array/integer.ts +2 -7
- package/assembly/deserialize/simd/float.ts +3 -5
- package/assembly/deserialize/simd/integer.ts +2 -7
- package/assembly/deserialize/simd/string.ts +5 -22
- package/assembly/deserialize/swar/array/arbitrary.ts +1 -2
- package/assembly/deserialize/swar/array/array.ts +1 -2
- package/assembly/deserialize/swar/array/bool.ts +1 -2
- package/assembly/deserialize/swar/array/box.ts +1 -2
- package/assembly/deserialize/swar/array/float.ts +6 -18
- package/assembly/deserialize/swar/array/generic.ts +1 -2
- package/assembly/deserialize/swar/array/integer.ts +7 -16
- package/assembly/deserialize/swar/array/map.ts +1 -2
- package/assembly/deserialize/swar/array/object.ts +1 -2
- package/assembly/deserialize/swar/array/raw.ts +1 -2
- package/assembly/deserialize/swar/array/shared.ts +6 -13
- package/assembly/deserialize/swar/array/string.ts +3 -8
- package/assembly/deserialize/swar/array/struct.ts +2 -8
- package/assembly/deserialize/swar/array.ts +1 -3
- package/assembly/deserialize/swar/float.ts +4 -9
- package/assembly/deserialize/swar/integer.ts +2 -7
- package/assembly/deserialize/swar/string.ts +13 -15
- package/assembly/deserialize/swar/typedarray.ts +4 -4
- package/assembly/index.d.ts +29 -24
- package/assembly/index.ts +1362 -246
- package/assembly/serialize/index/arbitrary.ts +70 -4
- package/assembly/serialize/index/jsonarray.ts +51 -0
- package/assembly/serialize/index/object.ts +25 -3
- package/assembly/serialize/index/string.ts +1 -2
- package/assembly/serialize/index/typedarray.ts +1 -2
- package/assembly/serialize/index.ts +1 -0
- package/assembly/serialize/naive/array.ts +23 -34
- package/assembly/serialize/naive/bool.ts +0 -1
- package/assembly/serialize/naive/float.ts +16 -25
- package/assembly/serialize/naive/integer.ts +1 -5
- package/assembly/serialize/naive/raw.ts +1 -2
- package/assembly/serialize/naive/set.ts +0 -4
- package/assembly/serialize/naive/staticarray.ts +0 -5
- package/assembly/serialize/naive/string.ts +2 -5
- package/assembly/serialize/naive/typedarray.ts +0 -6
- package/assembly/serialize/simd/string.ts +1 -3
- package/assembly/serialize/swar/string.ts +1 -2
- package/assembly/util/atoi-fast.ts +4 -14
- package/assembly/util/bytes.ts +1 -2
- package/assembly/util/idofd.ts +1 -2
- package/assembly/util/isSpace.ts +1 -2
- package/assembly/util/itoa-fast.ts +6 -9
- package/assembly/util/nextPowerOf2.ts +1 -2
- package/assembly/util/parsefloat-fast.ts +3 -5
- package/assembly/util/ptrToStr.ts +1 -2
- package/assembly/util/scanValueEndSimd.ts +54 -16
- package/assembly/util/scanValueEndSwar.ts +67 -25
- package/assembly/util/scientific.ts +5 -8
- package/assembly/util/snp.ts +1 -2
- package/assembly/util/swar-int.ts +5 -10
- package/assembly/util/swar.ts +2 -4
- package/lib/as-bs.ts +23 -45
- package/package.json +14 -7
- package/transform/lib/index.js +108 -64
- package/transform/lib/types.d.ts +2 -1
- package/transform/lib/types.js +3 -0
- package/assembly/util/dragonbox-cache.ts +0 -445
- package/assembly/util/dragonbox.ts +0 -652
|
@@ -6,8 +6,8 @@
|
|
|
6
6
|
// below. Two reasons:
|
|
7
7
|
//
|
|
8
8
|
// 1. V8/wasm lowers `v / 100` (and other `/ <const>`s) to a single
|
|
9
|
-
// multiply-shift, so jeaiii's main selling point
|
|
10
|
-
// hardware
|
|
9
|
+
// multiply-shift, so jeaiii's main selling point - avoiding division
|
|
10
|
+
// hardware - gives no win on this target. The op counts come out
|
|
11
11
|
// roughly equal.
|
|
12
12
|
//
|
|
13
13
|
// 2. The div-by-const variant computes each digit pair independently
|
|
@@ -24,7 +24,7 @@
|
|
|
24
24
|
// - A 100-entry digit-pair LUT keyed on `value % 100`. One `store<u32>`
|
|
25
25
|
// emits a UTF-16 pair.
|
|
26
26
|
//
|
|
27
|
-
// - Forward write in one pass
|
|
27
|
+
// - Forward write in one pass - no `decimalCount32` precomputation, no
|
|
28
28
|
// backward write.
|
|
29
29
|
//
|
|
30
30
|
// Reference H2H bench: `__benches__/custom/itoa-h2h.bench.ts`.
|
|
@@ -43,13 +43,11 @@ function initPairs(): void {
|
|
|
43
43
|
_pairsInited = true;
|
|
44
44
|
}
|
|
45
45
|
|
|
46
|
-
|
|
47
|
-
@inline export function ensureItoaPairs(): void {
|
|
46
|
+
export function ensureItoaPairs(): void {
|
|
48
47
|
if (!_pairsInited) initPairs();
|
|
49
48
|
}
|
|
50
49
|
|
|
51
|
-
|
|
52
|
-
@inline function pair(i: u32): u32 {
|
|
50
|
+
function pair(i: u32): u32 {
|
|
53
51
|
return load<u32>(DIGIT_PAIRS_UTF16 + ((<usize>i) << 2));
|
|
54
52
|
}
|
|
55
53
|
|
|
@@ -213,8 +211,7 @@ export function itoaU64(buf: usize, v: u64): u32 {
|
|
|
213
211
|
*
|
|
214
212
|
* Returns the number of UTF-16 chars written.
|
|
215
213
|
*/
|
|
216
|
-
|
|
217
|
-
@inline export function itoaFast<T extends number>(buf: usize, value: T): u32 {
|
|
214
|
+
export function itoaFast<T extends number>(buf: usize, value: T): u32 {
|
|
218
215
|
if (sizeof<T>() <= 4) {
|
|
219
216
|
if (isSigned<T>()) {
|
|
220
217
|
let v = <i32>value;
|
|
@@ -3,7 +3,7 @@ import { ptrToStr } from "./ptrToStr";
|
|
|
3
3
|
// Lemire-style fast float parser.
|
|
4
4
|
//
|
|
5
5
|
// Reference: Daniel Lemire, "Number parsing at a gigabyte per second"
|
|
6
|
-
// (2021). https://arxiv.org/abs/2101.11408
|
|
6
|
+
// (2021). https://arxiv.org/abs/2101.11408 - implemented in
|
|
7
7
|
// https://github.com/fastfloat/fast_float.
|
|
8
8
|
//
|
|
9
9
|
// The "fast path" applies when:
|
|
@@ -35,13 +35,11 @@ const MAX_EXACT_POW10: i32 = 22;
|
|
|
35
35
|
// 2^53 = 9_007_199_254_740_992. Any u64 <= this is exact in f64.
|
|
36
36
|
const MAX_EXACT_MANTISSA: u64 = 1 << 53;
|
|
37
37
|
|
|
38
|
-
|
|
39
|
-
@inline function loadPow10(exp: u32): f64 {
|
|
38
|
+
function loadPow10(exp: u32): f64 {
|
|
40
39
|
return load<f64>(POW10_F64_POS + ((<usize>exp) << 3));
|
|
41
40
|
}
|
|
42
41
|
|
|
43
|
-
|
|
44
|
-
@inline function fallback<T>(srcStart: usize, srcEnd: usize): T {
|
|
42
|
+
function fallback<T>(srcStart: usize, srcEnd: usize): T {
|
|
45
43
|
const s = ptrToStr(srcStart, srcEnd);
|
|
46
44
|
// @ts-ignore: type
|
|
47
45
|
const type: T = 0;
|
|
@@ -4,6 +4,7 @@ import {
|
|
|
4
4
|
BRACE_RIGHT,
|
|
5
5
|
BRACKET_LEFT,
|
|
6
6
|
BRACKET_RIGHT,
|
|
7
|
+
COLON,
|
|
7
8
|
COMMA,
|
|
8
9
|
QUOTE,
|
|
9
10
|
} from "../custom/chars";
|
|
@@ -22,23 +23,40 @@ import { isSpace } from "./isSpace";
|
|
|
22
23
|
// @ts-expect-error: @lazy is a valid decorator
|
|
23
24
|
@lazy const SPLAT_SPACE = i16x8.splat(0x20);
|
|
24
25
|
// JSON whitespace besides space is the contiguous range 0x09..0x0d
|
|
25
|
-
// (tab/LF/VT/FF/CR), matched as `(c - 9) u<= 4`
|
|
26
|
+
// (tab/LF/VT/FF/CR), matched as `(c - 9) u<= 4` - one sub + one unsigned
|
|
26
27
|
// compare instead of five equality tests. Exact: matches `isSpace` with no
|
|
27
28
|
// false positives.
|
|
28
29
|
// @ts-expect-error: @lazy is a valid decorator
|
|
29
30
|
@lazy const SPLAT_WS_LO = i16x8.splat(0x09);
|
|
30
31
|
// @ts-expect-error: @lazy is a valid decorator
|
|
31
32
|
@lazy const SPLAT_WS_SPAN = i16x8.splat(0x04);
|
|
33
|
+
// @ts-expect-error: @lazy is a valid decorator
|
|
34
|
+
@lazy const SPLAT_BRACKET_LEFT = i16x8.splat(<i16>BRACKET_LEFT);
|
|
35
|
+
// Clears bit 5 (0x20), folding `{`/`}` onto `[`/`]` so one pair of compares
|
|
36
|
+
// matches either bracket flavor. No ASCII char besides `[{`/`]}` folds onto
|
|
37
|
+
// `[`/`]`, so the structural mask stays exact.
|
|
38
|
+
// @ts-expect-error: @lazy is a valid decorator
|
|
39
|
+
@lazy const SPLAT_BRACKET_FOLD = i16x8.splat(<i16>0xffdf);
|
|
32
40
|
|
|
33
|
-
|
|
34
|
-
@inline function quoteOrBackslashMask(block: v128): i32 {
|
|
41
|
+
function quoteOrBackslashMask(block: v128): i32 {
|
|
35
42
|
return i16x8.bitmask(
|
|
36
43
|
v128.or(i16x8.eq(block, SPLAT_QUOTE), i16x8.eq(block, SPLAT_BACK_SLASH)),
|
|
37
44
|
);
|
|
38
45
|
}
|
|
39
46
|
|
|
40
|
-
//
|
|
41
|
-
|
|
47
|
+
// Lanes equal to `"`, `{`, `}`, `[`, or `]` - the only bytes that, outside a
|
|
48
|
+
// string, change depth or open a string. Everything else (digits, `:`, `,`,
|
|
49
|
+
// whitespace, true/false/null) can be bulk-skipped between them.
|
|
50
|
+
function structuralOrQuoteMask(block: v128): i32 {
|
|
51
|
+
const folded = v128.and(block, SPLAT_BRACKET_FOLD);
|
|
52
|
+
const brackets = v128.or(
|
|
53
|
+
i16x8.eq(folded, SPLAT_BRACKET_LEFT),
|
|
54
|
+
i16x8.eq(folded, SPLAT_BRACKET_RIGHT),
|
|
55
|
+
);
|
|
56
|
+
return i16x8.bitmask(v128.or(brackets, i16x8.eq(block, SPLAT_QUOTE)));
|
|
57
|
+
}
|
|
58
|
+
|
|
59
|
+
function scalarTerminatorMask(block: v128): i32 {
|
|
42
60
|
const structural = v128.or(
|
|
43
61
|
v128.or(i16x8.eq(block, SPLAT_COMMA), i16x8.eq(block, SPLAT_BRACE_RIGHT)),
|
|
44
62
|
i16x8.eq(block, SPLAT_BRACKET_RIGHT),
|
|
@@ -86,12 +104,14 @@ function scanQuotedValueEnd_SIMD(srcStart: usize, srcEnd: usize): usize {
|
|
|
86
104
|
}
|
|
87
105
|
|
|
88
106
|
function scanCompositeValueEnd_SIMD(srcStart: usize, srcEnd: usize): usize {
|
|
89
|
-
//
|
|
90
|
-
//
|
|
91
|
-
//
|
|
92
|
-
//
|
|
107
|
+
// Process structural tokens scalar-side (cheap, and token-dense regions stay
|
|
108
|
+
// in a tight loop), but bulk-skip the bytes between them: nested string VALUES
|
|
109
|
+
// via the vectorized quoted scan (URLs, base64, prose), and runs of digits /
|
|
110
|
+
// punctuation / whitespace (numeric arrays like coordinate lists) via a
|
|
111
|
+
// vectorized hunt for the next `"`/`{`/`}`/`[`/`]`.
|
|
93
112
|
let depth: i32 = 1;
|
|
94
113
|
let ptr = srcStart + 2;
|
|
114
|
+
const srcEnd16 = srcEnd >= 16 ? srcEnd - 16 : 0;
|
|
95
115
|
while (ptr < srcEnd) {
|
|
96
116
|
const code = load<u16>(ptr);
|
|
97
117
|
if (code == QUOTE) {
|
|
@@ -99,12 +119,34 @@ function scanCompositeValueEnd_SIMD(srcStart: usize, srcEnd: usize): usize {
|
|
|
99
119
|
if (!ptr) return 0;
|
|
100
120
|
continue;
|
|
101
121
|
}
|
|
102
|
-
|
|
122
|
+
const folded = code & 0xffdf;
|
|
123
|
+
if (folded == BRACKET_LEFT) {
|
|
124
|
+
// `[` or `{`
|
|
103
125
|
depth++;
|
|
104
|
-
|
|
126
|
+
ptr += 2;
|
|
127
|
+
continue;
|
|
128
|
+
}
|
|
129
|
+
if (folded == BRACKET_RIGHT) {
|
|
130
|
+
// `]` or `}`
|
|
105
131
|
if (--depth == 0) return ptr + 2;
|
|
132
|
+
ptr += 2;
|
|
133
|
+
continue;
|
|
106
134
|
}
|
|
107
135
|
ptr += 2;
|
|
136
|
+
// `,` and `:` sit one byte from the next token, so vectorizing them only
|
|
137
|
+
// adds SIMD setup on string-dense objects - stay scalar. Other fillers
|
|
138
|
+
// (number digits, whitespace, true/false/null) can run long; vectorize past
|
|
139
|
+
// them to the next `"`/`{`/`}`/`[`/`]`.
|
|
140
|
+
if (code == COMMA || code == COLON) continue;
|
|
141
|
+
while (ptr <= srcEnd16) {
|
|
142
|
+
const mask = structuralOrQuoteMask(load<v128>(ptr));
|
|
143
|
+
if (mask == 0) {
|
|
144
|
+
ptr += 16;
|
|
145
|
+
continue;
|
|
146
|
+
}
|
|
147
|
+
ptr += usize(ctz(mask) << 1);
|
|
148
|
+
break;
|
|
149
|
+
}
|
|
108
150
|
}
|
|
109
151
|
return 0;
|
|
110
152
|
}
|
|
@@ -135,11 +177,7 @@ function scanScalarValueEnd_SIMD(srcStart: usize, srcEnd: usize): usize {
|
|
|
135
177
|
return srcStart;
|
|
136
178
|
}
|
|
137
179
|
|
|
138
|
-
|
|
139
|
-
@inline export function scanValueEnd_SIMD<T>(
|
|
140
|
-
srcStart: usize,
|
|
141
|
-
srcEnd: usize,
|
|
142
|
-
): usize {
|
|
180
|
+
export function scanValueEnd_SIMD<T>(srcStart: usize, srcEnd: usize): usize {
|
|
143
181
|
if (srcStart >= srcEnd) return 0;
|
|
144
182
|
const first = load<u16>(srcStart);
|
|
145
183
|
|
|
@@ -4,6 +4,7 @@ import {
|
|
|
4
4
|
BRACE_RIGHT,
|
|
5
5
|
BRACKET_LEFT,
|
|
6
6
|
BRACKET_RIGHT,
|
|
7
|
+
COLON,
|
|
7
8
|
COMMA,
|
|
8
9
|
QUOTE,
|
|
9
10
|
} from "../custom/chars";
|
|
@@ -11,33 +12,45 @@ import { isSpace } from "./isSpace";
|
|
|
11
12
|
|
|
12
13
|
// SWAR analogue of `scanValueEndSimd.ts`, processing four UTF-16 lanes per
|
|
13
14
|
// 64-bit word for the SWAR build mode (no SIMD feature). Each mask is a fast
|
|
14
|
-
// FILTER
|
|
15
|
+
// FILTER - a matched lane is re-checked with a real `load<u16>` before acting -
|
|
15
16
|
// so the masks may over-match non-ASCII lanes whose low byte equals a target
|
|
16
17
|
// (the verify rejects them). Lane byte offset within a hit word is
|
|
17
18
|
// `ctz(mask) >> 3` (detection bit sits at lane*16 + 7).
|
|
18
19
|
|
|
19
|
-
|
|
20
|
-
|
|
21
|
-
// @ts-expect-error: @inline is a valid decorator
|
|
22
|
-
@inline const HI: u64 = 0x0080_0080_0080_0080;
|
|
20
|
+
const ONES: u64 = 0x0001_0001_0001_0001;
|
|
21
|
+
const HI: u64 = 0x0080_0080_0080_0080;
|
|
23
22
|
|
|
24
23
|
// 16-bit-lane "equals" partials (pre-`& HI`); OR several, then `& HI` once.
|
|
25
|
-
|
|
26
|
-
@inline function eqPart(block: u64, splat: u64): u64 {
|
|
24
|
+
function eqPart(block: u64, splat: u64): u64 {
|
|
27
25
|
const t = block ^ splat;
|
|
28
26
|
return (t - ONES) & ~t;
|
|
29
27
|
}
|
|
30
28
|
|
|
31
|
-
|
|
32
|
-
|
|
33
|
-
|
|
34
|
-
|
|
29
|
+
const S_QUOTE: u64 = 0x0022_0022_0022_0022;
|
|
30
|
+
const S_BACK_SLASH: u64 = 0x005c_005c_005c_005c;
|
|
31
|
+
const S_BRACKET_LEFT: u64 = 0x005b_005b_005b_005b;
|
|
32
|
+
const S_BRACKET_RIGHT: u64 = 0x005d_005d_005d_005d;
|
|
33
|
+
// Clears bit 5 (0x20) of each lane, folding `{`/`}` onto `[`/`]`.
|
|
34
|
+
const FOLD: u64 = 0xffdf_ffdf_ffdf_ffdf;
|
|
35
35
|
|
|
36
|
-
|
|
37
|
-
@inline function quoteOrBackslashMask(block: u64): u64 {
|
|
36
|
+
function quoteOrBackslashMask(block: u64): u64 {
|
|
38
37
|
return (eqPart(block, S_QUOTE) | eqPart(block, S_BACK_SLASH)) & HI;
|
|
39
38
|
}
|
|
40
39
|
|
|
40
|
+
// Filter for lanes equal to `"`, `{`, `}`, `[`, or `]` - the only bytes that,
|
|
41
|
+
// outside a string, change depth or open a string. As with the other SWAR
|
|
42
|
+
// masks, a hit is a candidate to verify with a real load (it may over-match a
|
|
43
|
+
// non-ASCII lane whose low byte collides).
|
|
44
|
+
function structuralOrQuoteMask(block: u64): u64 {
|
|
45
|
+
const folded = block & FOLD;
|
|
46
|
+
return (
|
|
47
|
+
(eqPart(folded, S_BRACKET_LEFT) |
|
|
48
|
+
eqPart(folded, S_BRACKET_RIGHT) |
|
|
49
|
+
eqPart(block, S_QUOTE)) &
|
|
50
|
+
HI
|
|
51
|
+
);
|
|
52
|
+
}
|
|
53
|
+
|
|
41
54
|
function scanQuotedValueEnd_SWAR(srcStart: usize, srcEnd: usize): usize {
|
|
42
55
|
srcStart += 2;
|
|
43
56
|
const srcEnd8 = srcEnd >= 8 ? srcEnd - 8 : 0;
|
|
@@ -63,21 +76,30 @@ function scanQuotedValueEnd_SWAR(srcStart: usize, srcEnd: usize): usize {
|
|
|
63
76
|
break;
|
|
64
77
|
}
|
|
65
78
|
|
|
79
|
+
// Resolve escapes by consuming a backslash *and the char it escapes* together,
|
|
80
|
+
// so escape parity is tracked exactly. A look-back `prev != BACK_SLASH` test is
|
|
81
|
+
// wrong for an escaped backslash: in `"x\\"` the closing quote follows a `\`
|
|
82
|
+
// (the second of the pair) yet still closes the string.
|
|
66
83
|
while (srcStart < srcEnd) {
|
|
67
84
|
const code = load<u16>(srcStart);
|
|
68
|
-
if (code ==
|
|
69
|
-
|
|
85
|
+
if (code == BACK_SLASH) {
|
|
86
|
+
srcStart += 4;
|
|
87
|
+
continue;
|
|
88
|
+
}
|
|
89
|
+
if (code == QUOTE) return srcStart + 2;
|
|
70
90
|
srcStart += 2;
|
|
71
91
|
}
|
|
72
92
|
return 0;
|
|
73
93
|
}
|
|
74
94
|
|
|
75
95
|
function scanCompositeValueEnd_SWAR(srcStart: usize, srcEnd: usize): usize {
|
|
76
|
-
//
|
|
77
|
-
//
|
|
78
|
-
//
|
|
96
|
+
// Process structural tokens scalar-side, but bulk-skip the bytes between them:
|
|
97
|
+
// nested string VALUES via the SWAR quoted scan, and runs of digits /
|
|
98
|
+
// punctuation / whitespace (numeric arrays) via a SWAR hunt for the next
|
|
99
|
+
// `"`/`{`/`}`/`[`/`]`.
|
|
79
100
|
let depth: i32 = 1;
|
|
80
101
|
let ptr = srcStart + 2;
|
|
102
|
+
const srcEnd8 = srcEnd >= 8 ? srcEnd - 8 : 0;
|
|
81
103
|
while (ptr < srcEnd) {
|
|
82
104
|
const code = load<u16>(ptr);
|
|
83
105
|
if (code == QUOTE) {
|
|
@@ -85,12 +107,36 @@ function scanCompositeValueEnd_SWAR(srcStart: usize, srcEnd: usize): usize {
|
|
|
85
107
|
if (!ptr) return 0;
|
|
86
108
|
continue;
|
|
87
109
|
}
|
|
88
|
-
|
|
110
|
+
const folded = code & 0xffdf;
|
|
111
|
+
if (folded == BRACKET_LEFT) {
|
|
89
112
|
depth++;
|
|
90
|
-
|
|
113
|
+
ptr += 2;
|
|
114
|
+
continue;
|
|
115
|
+
}
|
|
116
|
+
if (folded == BRACKET_RIGHT) {
|
|
91
117
|
if (--depth == 0) return ptr + 2;
|
|
118
|
+
ptr += 2;
|
|
119
|
+
continue;
|
|
92
120
|
}
|
|
93
121
|
ptr += 2;
|
|
122
|
+
// `,`/`:` sit one byte from the next token - stay scalar (string-dense
|
|
123
|
+
// objects); other fillers can run long, so SWAR-skip past them.
|
|
124
|
+
if (code == COMMA || code == COLON) continue;
|
|
125
|
+
while (ptr <= srcEnd8) {
|
|
126
|
+
const mask = structuralOrQuoteMask(load<u64>(ptr));
|
|
127
|
+
if (mask == 0) {
|
|
128
|
+
ptr += 8;
|
|
129
|
+
continue;
|
|
130
|
+
}
|
|
131
|
+
const idx = ptr + (usize(ctz(mask)) >> 3);
|
|
132
|
+
const c = load<u16>(idx);
|
|
133
|
+
const f = c & 0xffdf;
|
|
134
|
+
if (c == QUOTE || f == BRACKET_LEFT || f == BRACKET_RIGHT) {
|
|
135
|
+
ptr = idx; // real token - the outer loop processes it
|
|
136
|
+
break;
|
|
137
|
+
}
|
|
138
|
+
ptr = idx + 2; // spurious lane match - keep scanning
|
|
139
|
+
}
|
|
94
140
|
}
|
|
95
141
|
return 0;
|
|
96
142
|
}
|
|
@@ -117,11 +163,7 @@ function scanScalarValueEnd_SWAR(srcStart: usize, srcEnd: usize): usize {
|
|
|
117
163
|
* objects/arrays use the SWAR token scans above; scalars use a short scalar
|
|
118
164
|
* loop. Returns 0 on empty input or an unterminated string/composite.
|
|
119
165
|
*/
|
|
120
|
-
|
|
121
|
-
@inline export function scanValueEnd_SWAR<T>(
|
|
122
|
-
srcStart: usize,
|
|
123
|
-
srcEnd: usize,
|
|
124
|
-
): usize {
|
|
166
|
+
export function scanValueEnd_SWAR<T>(srcStart: usize, srcEnd: usize): usize {
|
|
125
167
|
if (srcStart >= srcEnd) return 0;
|
|
126
168
|
const first = load<u16>(srcStart);
|
|
127
169
|
|
|
@@ -10,7 +10,7 @@
|
|
|
10
10
|
// `scientific` directly skips both costs.
|
|
11
11
|
//
|
|
12
12
|
// scientific() is correctly rounded for all u64 mantissas and decimal
|
|
13
|
-
// exponents that fit in IEEE-754 f64's range
|
|
13
|
+
// exponents that fit in IEEE-754 f64's range - including the [2^53, 2^64)
|
|
14
14
|
// mantissa range that breaks Lemire's single-fmul fast path.
|
|
15
15
|
|
|
16
16
|
const POWERS10: usize = memory.data<f64>([
|
|
@@ -25,13 +25,11 @@ const POWERS5: usize = memory.data<i32>([
|
|
|
25
25
|
244140625, 1220703125,
|
|
26
26
|
]);
|
|
27
27
|
|
|
28
|
-
|
|
29
|
-
@inline function pow10(n: i32): f64 {
|
|
28
|
+
function pow10(n: i32): f64 {
|
|
30
29
|
return load<f64>(POWERS10 + ((<usize>n) << alignof<f64>()));
|
|
31
30
|
}
|
|
32
31
|
|
|
33
|
-
|
|
34
|
-
@inline function pow5_32(n: i32): i32 {
|
|
32
|
+
function pow5_32(n: i32): i32 {
|
|
35
33
|
return load<i32>(POWERS5 + ((<usize>n) << alignof<i32>()));
|
|
36
34
|
}
|
|
37
35
|
|
|
@@ -40,8 +38,7 @@ const POWERS5: usize = memory.data<i32>([
|
|
|
40
38
|
// @ts-ignore: lazy decorator
|
|
41
39
|
@lazy let __fixmulShift: u64 = 0;
|
|
42
40
|
|
|
43
|
-
|
|
44
|
-
@inline function fixmul(a: u64, b: u32): u64 {
|
|
41
|
+
function fixmul(a: u64, b: u32): u64 {
|
|
45
42
|
const low = (a & 0xffffffff) * b;
|
|
46
43
|
const high = (a >> 32) * b + (low >> 32);
|
|
47
44
|
const overflow = <u32>(high >> 32);
|
|
@@ -98,7 +95,7 @@ function scaleup(significand: u64, exp: i32): f64 {
|
|
|
98
95
|
|
|
99
96
|
/**
|
|
100
97
|
* Construct an f64 from a u64 mantissa and decimal exponent. Result is
|
|
101
|
-
* correctly rounded
|
|
98
|
+
* correctly rounded - bit-identical to `f64.parse` for any input the SWAR
|
|
102
99
|
* float deserializer can pre-parse into this form.
|
|
103
100
|
*
|
|
104
101
|
* Caller guarantees the digit run that produced `significand` was already
|
package/assembly/util/snp.ts
CHANGED
|
@@ -58,8 +58,7 @@ export function snp<T extends number>(srcStart: usize, srcEnd: usize): T {
|
|
|
58
58
|
}
|
|
59
59
|
}
|
|
60
60
|
|
|
61
|
-
|
|
62
|
-
@inline function pow10<T extends number>(x: u16): T {
|
|
61
|
+
function pow10<T extends number>(x: u16): T {
|
|
63
62
|
if (sizeof<T>() == 8) {
|
|
64
63
|
return <T>load<u64>(POW_TEN_TABLE_64 + x);
|
|
65
64
|
} else {
|
|
@@ -26,8 +26,7 @@ const FINAL_4_MAGIC: u64 = 0x0000_0064_0000_0001;
|
|
|
26
26
|
* @param block Four UTF-16 code units packed into a `u64`.
|
|
27
27
|
* @returns The parsed 4-digit value, or `U32.MAX_VALUE` on invalid input.
|
|
28
28
|
*/
|
|
29
|
-
|
|
30
|
-
@inline export function parse4Digits_Baseline(block: u64): u32 {
|
|
29
|
+
export function parse4Digits_Baseline(block: u64): u32 {
|
|
31
30
|
const digits = (block & LANE_LO_4) - ZERO_4;
|
|
32
31
|
if (((digits | (digits + RANGE_ADD_4)) & RANGE_MASK_4) != 0) {
|
|
33
32
|
return U32.MAX_VALUE;
|
|
@@ -58,8 +57,7 @@ const FINAL_4_MAGIC: u64 = 0x0000_0064_0000_0001;
|
|
|
58
57
|
* @param block Four UTF-16 code units packed into a `u64`.
|
|
59
58
|
* @returns The parsed 4-digit value, or `U32.MAX_VALUE` on invalid input.
|
|
60
59
|
*/
|
|
61
|
-
|
|
62
|
-
@inline export function parse4Digits_PairMul(block: u64): u32 {
|
|
60
|
+
export function parse4Digits_PairMul(block: u64): u32 {
|
|
63
61
|
const digits = block - ZERO_4;
|
|
64
62
|
if (((digits | (digits + RANGE_ADD_4)) & RANGE_MASK_4) != 0) {
|
|
65
63
|
return U32.MAX_VALUE;
|
|
@@ -76,8 +74,7 @@ const FINAL_4_MAGIC: u64 = 0x0000_0064_0000_0001;
|
|
|
76
74
|
* @param block Four UTF-16 code units packed into a `u64`.
|
|
77
75
|
* @returns The parsed 4-digit value.
|
|
78
76
|
*/
|
|
79
|
-
|
|
80
|
-
@inline export function parse4Digits_PairMul_Unsafe(block: u64): u32 {
|
|
77
|
+
export function parse4Digits_PairMul_Unsafe(block: u64): u32 {
|
|
81
78
|
const digits = block - ZERO_4;
|
|
82
79
|
const pairs = (digits * 10 + (digits >> 16)) & U32_LO_PAIR;
|
|
83
80
|
return <u32>((pairs * FINAL_4_MAGIC) >> 32);
|
|
@@ -123,8 +120,7 @@ export function parse8Digits_PairMul(lo: u64, hi: u64): u32 {
|
|
|
123
120
|
* @param hi The second `u64`, four UTF-16 code units.
|
|
124
121
|
* @returns The parsed 8-digit value.
|
|
125
122
|
*/
|
|
126
|
-
|
|
127
|
-
@inline export function parse8Digits_PairMul_Unsafe(lo: u64, hi: u64): u32 {
|
|
123
|
+
export function parse8Digits_PairMul_Unsafe(lo: u64, hi: u64): u32 {
|
|
128
124
|
const loDigits = lo - ZERO_4;
|
|
129
125
|
const hiDigits = hi - ZERO_4;
|
|
130
126
|
const loPairs = (loDigits * 10 + (loDigits >> 16)) & U32_LO_PAIR;
|
|
@@ -149,8 +145,7 @@ export function parse8Digits_PairMul(lo: u64, hi: u64): u32 {
|
|
|
149
145
|
* @param block Four UTF-16 code units packed into a `u64`.
|
|
150
146
|
* @returns A mask with non-digit lanes flagged in their high bit, or 0.
|
|
151
147
|
*/
|
|
152
|
-
|
|
153
|
-
@inline export function nonDigitMask4(block: u64): u64 {
|
|
148
|
+
export function nonDigitMask4(block: u64): u64 {
|
|
154
149
|
const digits = (block & LANE_LO_4) - ZERO_4;
|
|
155
150
|
return (digits | (digits + RANGE_ADD_4)) & RANGE_MASK_4;
|
|
156
151
|
}
|
package/assembly/util/swar.ts
CHANGED
|
@@ -18,8 +18,7 @@
|
|
|
18
18
|
* @param block Packed UTF-16 ASCII hex digits.
|
|
19
19
|
* @returns The decoded 16-bit value.
|
|
20
20
|
*/
|
|
21
|
-
|
|
22
|
-
@inline export function hex4_to_u16_swar(block: u64): u16 {
|
|
21
|
+
export function hex4_to_u16_swar(block: u64): u16 {
|
|
23
22
|
// (c & 0xF) + 9 * (c >> 6)
|
|
24
23
|
block = (block & 0x0f000f000f000f) + ((block >> 6) & 0x03000300030003) * 9;
|
|
25
24
|
|
|
@@ -51,8 +50,7 @@
|
|
|
51
50
|
* @param code The 16-bit value to encode.
|
|
52
51
|
* @returns Four packed UTF-16 ASCII hex digits.
|
|
53
52
|
*/
|
|
54
|
-
|
|
55
|
-
@inline export function u16_to_hex4_swar(code: u16): u64 {
|
|
53
|
+
export function u16_to_hex4_swar(code: u16): u64 {
|
|
56
54
|
let block =
|
|
57
55
|
(<u64>((code >> 12) & 0xf)) |
|
|
58
56
|
((<u64>((code >> 8) & 0xf)) << 16) |
|