feat: switch grisu2 float-to-string algorithm to hybrid of xjb & zmij algorithms by JairusSW · Pull Request #3025 · AssemblyScript/assemblyscript

JairusSW · 2026-06-09T05:20:07Z

Changes proposed in this pull request:
⯈ Switch to xjb-as which is an improvement over zmij-as
⯈ Comply to the ECMA262 Specification with the exception of a trailing .0

Note: I wrote xjb-as with claude assisting me for an initial port. I then optimized by hand. Everything is carefully checked and passes over a trillion fuzz cases compared directly against Number::parse() in V8. It also reaches 100% code coverage and tests every edge case both I and claude could think of.

It's good quality stuff though, and I'm quite confident in that.

Here's some performance notes. xjb-as/README.md has more extensive notes and stuff. These benches are taken on an AMD 7800x3D.

program uses	grisu2	xjb	Δ
f64 only	4.56	6.84	+2.28
f32 only	4.56	4.61	+0.05
both	5.17	8.97	+3.80

Lookup-table footprint vs the old grisu2 implementation (omiting shared tables):

program uses	grisu2	xjb	Δ
f64 (`dtoa`) only	910 B	672 B	−238 B
f32 (`ftoa`) only	910 B	616 B	−294 B
both	910 B	1288 B	+378 B

I've read the contributing guidelines
I've added my name and email to the NOTICE file

P.S. Sorry about the diff. Most of it is just the wasm files.

Signed-off-by: Jairus Tanaka <me@jairus.dev>

PaperPrototype · 2026-06-17T23:01:49Z

Just curious here, why the exception of a trailing .0?

JairusSW · 2026-06-17T23:04:04Z

Just curious here, why the exception of a trailing .0?

Max and Dan made that decision so that floats are easily identifiable when converting them to a string. For example, you know that the string "1.0" is a result of f64.toString() while the string "1" is likely a result of i32.toString(). Slight deviation from ECMA262-spec, but it's not awful

MaxGraey · 2026-06-17T23:06:22Z

Can you rename xjb.ts to dtoa.ts and in comments mentioned about implementation based on xjb and zmij? xjb filename is very exotic and doesn't explains exactly what this file does

Co-authored-by: Max Graey <maxgraey@gmail.com>

MaxGraey · 2026-06-17T23:25:34Z

+let gBcd: u64 = 0;
+let gBcdLen: i32 = 0;


gBcd -> gBcdVal
?

MaxGraey · 2026-06-17T23:28:11Z

+export let gDigHi: u64 = 0;
+export let gDigLo: u64 = 0;
+export let gDigNum: i32 = 0;


You have a lot of global vars. Does we need all of them? And can we declare all of necessary vars in one place?

Co-authored-by: Max Graey <maxgraey@gmail.com>

dcodeIO · 2026-06-17T23:31:34Z

+ (data $5 (i32.const 1239) "\80\00\00\00\00\00\00\00\a0\00\00\00\00\00\00\00\c8\00\00\00\00\00\00\00\fa\00\00\00\00\00\00@\9c\00\00\00\00\00\00P\c3\00\00\00\00\00\00$\f4\00\00\00\00\00\80\96\98\00\00\00\00\00 \bc\be\00\00\00\00\00(k\ee\00\00\00\00\00\f9\02\95\00\00\00\00@\b7C\ba\00\00\00\00\10\a5\d4\e8\00\00\00\00*\e7\84\91\00\00\00\80\f4 \e6\b5\00\00\00\a01\a9_\e3\00\00\00\04\bf\c9\1b\8e\00\00\00\c5.\bc\a2\b1\00\00@v:k\0b\de\00\00\e8\89\04#\c7\8a\00\00b\ac\c5\ebx\ad\00\80z\17\b7&\d7\d8\00\90\acn2x\86\87\00\b4W\n?\16h\a9\00\a1\ed\cc\ce\1b\c2\d3\a0\84\14@aQY\84\c8\a5\19\90\b9\a5o\a5:\0f \f4\'\8f\cb\ce")
+ (data $6 (i32.const 1456) "o\1b\8e(\10T\8e\af\daM\e4^\ae\f0\ec\07J\fb\9f\f4\98\'D\b1\9dwA\df\cf\11\cd\99\07\ef\99\85\0b?\fe\b2\15\aa\b4\dc\e6\a7\1f\86c\beZ\06\0b\a5\bc\b4\aaSkuz\07\ed\0f\08\bf,)Ud\7f\b6C\d5\b1\17L\c8;\1a\fb;\efi\c2\87F\b8B\a7\ee@OQ]=\eb\dd\e4PF\1a\12\ba\13\e4labM\f3\92\ea\af(\b6\ef&\e2\bb\8c6U\n\f7\89\04\89\0f`\cb\05\e9\b8\b6\bd!\c9\c1\bb\87\e9\00T\96_\9a\84x\db\8f\bf4\d0\bdr\04R\98\de\'\8a\92\95\00\9am\c1\94\82\17\0f<\05\b7u\00\00\00\00\00\00P\c3\00\00\00\00\00\00\00\00\05\e3L6\12\197\c5\00\00\00\00\00\00(l\d6\aa\80\9d\ef\f0\"\c7\f6~\b9\b7\d2:MBL\c8q\d5m\93\13\c9\ea8\1e\cd\19:\bc\03\1cU\ab\01\80\0c\t\cb\c6,\07\d3\bf\f5\ad\\\a1\90\08\137h\03\cd\10\8cz\c3\87\a8\db6.\ef\07\12\c2\b2\02\cf\bc\f4\03^\e4g\f9\94\c7\85\d7in\f8\06\d1R\ba\be\01\d763\e1|\a0\1c4\a8E\10\d3Q\a0\t\12\11H\de\1e1Vx\85\fa\a6\1e\d5f\a5>\7f\"t*U3\f1\ca\ba\0f)2\d7\96@\adGy\17|\a9t\088\c7\b1\d8J\d9\bc\"x\ae\81R7\18")
+ (data $7 (i32.const 1824) "?6N\n@\18\00\00\00d\00\00@\00 $\00\00\00\00\00\00\00\0c\80\13\c8\82\1f\e0L^\0f\f60\d7\1b\00\00\00\00\00\00\00\fc\ff\f7\cd\d8\01\82n\d1?\cd@\01%d\db\r\r\00\00\00$\04\14@8qS\b4\1dx\11")
+ (data $9 (i32.const 2032) "p\\\ea{\ce2~\8f\1a\c7C\c6\b0\b7\96\e5\ae\05\03\05\'\c6\ab\b7\bf7\cf\d0\b8\d1\ef\92\fe%\e5\1a\8eO\19\eb2\ebP\e2\a4?\14\bc\f5\88\r\b5P\99v\96!\dbH\bb\1a\c2\bd\f0\b4\15\07\c9{\ce\97\c0]\11l:\96\0b\13\9a\c7\1b\e0\c3V\df\84\f6\06\e3L6\12\197\c5\9e\b5p+\a8\ad\c5\9d\97\"\81E@|o\fc\dfNg\04\cd\c9\f2\c9\e6\0b\b96\d7\07\8f\a1\85\t\94\f8x9?\81:\0f \f4\'\8f\cb\ce\c8\a5\19\90\b9\a5o\a5\a0\84\14@aQY\84\00\a1\ed\cc\ce\1b\c2\d3\00\b4W\n?\16h\a9\00\90\acn2x\86\87\00\80z\17\b7&\d7\d8\00\00b\ac\c5\ebx\ad\00\00\e8\89\04#\c7\8a\00\00@v:k\0b\de\00\00\00\c5.\bc\a2\b1\00\00\00\04\bf\c9\1b\8e\00\00\00\a01\a9_\e3\00\00\00\80\f4 \e6\b5\00\00\00\00*\e7\84\91\00\00\00\00\10\a5\d4\e8\00\00\00\00@\b7C\ba\00\00\00\00\00\f9\02\95\00\00\00\00\00(k\ee\00\00\00\00\00 \bc\be\00\00\00\00\00\80\96\98\00\00\00\00\00\00$\f4\00\00\00\00\00\00P\c3\00\00\00\00\00\00@\9c\00\00\00\00\00\00\00\fa\00\00\00\00\00\00\00\c8\00\00\00\00\00\00\00\a0\00\00\00\00\00\00\00\80\cd\cc\cc\cc\cc\cc\cc\cc\0b\d7\a3p=\n\d7\a3<\dfO\8d\97n\12\83,e\19\e2X\17\b7\d1$\84G\1bG\ac\c5\a7\b6il\af\05\bd7\86\bdBz\e5\d5\94\bf\d6\fd\cea\84\11w\cc\ab\98\a5\b46A_p\89\bf\d5\ed\bd\ce\fe\e6\db\ff\aa$\cb\0b\ff\eb\af\cc\88Po\t\cc\bc\8c\14\0e\b4KB\13.\e1\10\d8\\\t5\dc$\b4\da\ac\b0:\f7|\1d\90\\\e1M\c4\be\94\95\e6J\b4\a462\aaw\b8\08]\1d\92\8e\ee\92\93\a6a\95\b6}J\1e\ec\eb\1a\11\92d\08\e5\bc\ef{\datP\a0\1d\97\b2,\f7\ba\80\00\c9\f1(\8a\92\95\00\9am\c1S;uD\cd\14\be\9aR\c5\ee\d3\ae\87\96\f7\db\9dXv%\06\12\c6I~\e0\91\b7\d1t\9e\0e\ca\00\83\f2\b5\87\fd?;\9a5\f5\f7\d2\ca2\fc\14^\f7_B\a2\f5\fcCK,\b3\ce\81\bb\949E\ad\1e\b1\cf")
+ (data $10 (i32.const 2648) "\"\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$!\"#$\"#$\"#$\"#$!\"#")
+ (data $11 (i32.const 2908) ",")
+ (data $11.1 (i32.const 2920) "\02\00\00\00\1c\00\00\00I\00n\00v\00a\00l\00i\00d\00 \00l\00e\00n\00g\00t\00h")
+ (data $12 (i32.const 2956) "<")
+ (data $12.1 (i32.const 2968) "\02\00\00\00&\00\00\00~\00l\00i\00b\00/\00a\00r\00r\00a\00y\00b\00u\00f\00f\00e\00r\00.\00t\00s")
+ (data $13 (i32.const 3020) "<")
+ (data $13.1 (i32.const 3032) "\02\00\00\00(\00\00\00A\00l\00l\00o\00c\00a\00t\00i\00o\00n\00 \00t\00o\00o\00 \00l\00a\00r\00g\00e")
+ (data $14 (i32.const 3084) "<")
+ (data $14.1 (i32.const 3096) "\02\00\00\00 \00\00\00~\00l\00i\00b\00/\00r\00t\00/\00i\00t\00c\00m\00s\00.\00t\00s")
+ (data $17 (i32.const 3212) "<")
+ (data $17.1 (i32.const 3224) "\02\00\00\00$\00\00\00I\00n\00d\00e\00x\00 \00o\00u\00t\00 \00o\00f\00 \00r\00a\00n\00g\00e")
+ (data $18 (i32.const 3276) ",")
+ (data $18.1 (i32.const 3288) "\02\00\00\00\14\00\00\00~\00l\00i\00b\00/\00r\00t\00.\00t\00s")
+ (data $20 (i32.const 3356) "<")
+ (data $20.1 (i32.const 3368) "\02\00\00\00\1e\00\00\00~\00l\00i\00b\00/\00r\00t\00/\00t\00l\00s\00f\00.\00t\00s")
+ (data $21 (i32.const 3420) "\1c")
+ (data $21.1 (i32.const 3432) "\02")
+ (data $22 (i32.const 3452) "<")
+ (data $22.1 (i32.const 3464) "\02\00\00\00$\00\00\00~\00l\00i\00b\00/\00t\00y\00p\00e\00d\00a\00r\00r\00a\00y\00.\00t\00s")
+ (data $23 (i32.const 3516) "<")
+ (data $23.1 (i32.const 3528) "\02\00\00\00&\00\00\00~\00l\00i\00b\00/\00s\00t\00a\00t\00i\00c\00a\00r\00r\00a\00y\00.\00t\00s")
+ (data $24 (i32.const 3580) ",")
+ (data $24.1 (i32.const 3592) "\02\00\00\00\1a\00\00\00~\00l\00i\00b\00/\00a\00r\00r\00a\00y\00.\00t\00s")
+ (data $25 (i32.const 3628) "|")
+ (data $25.1 (i32.const 3640) "\02\00\00\00^\00\00\00E\00l\00e\00m\00e\00n\00t\00 \00t\00y\00p\00e\00 \00m\00u\00s\00t\00 \00b\00e\00 \00n\00u\00l\00l\00a\00b\00l\00e\00 \00i\00f\00 \00a\00r\00r\00a\00y\00 \00i\00s\00 \00h\00o\00l\00e\00y")
+ (data $26 (i32.const 3756) "<")
+ (data $26.1 (i32.const 3768) "\02\00\00\00*\00\00\00O\00b\00j\00e\00c\00t\00 \00a\00l\00r\00e\00a\00d\00y\00 \00p\00i\00n\00n\00e\00d")
+ (data $27 (i32.const 3820) "<")
+ (data $27.1 (i32.const 3832) "\02\00\00\00(\00\00\00O\00b\00j\00e\00c\00t\00 \00i\00s\00 \00n\00o\00t\00 \00p\00i\00n\00n\00e\00d")
+ (data $28 (i32.const 3888) "\10\00\00\00 \00\00\00 \00\00\00 ")
+ (data $28.1 (i32.const 3912) "\81\08\00\00\01\19\00\00\01\02\00\00$\t\00\00\a4\00\00\00$\n\00\00\02\t\00\00\02A\00\00\00\00\00\00A\00\00\00 ")


Wondering if the baseline amount of static memory can be reduced here by computing the parts of it that aren't strictly necessary to have as LUTs?

Co-authored-by: Max Graey <maxgraey@gmail.com>

MaxGraey · 2026-06-17T23:42:51Z

+  const shift = h + 1 + EXTRA_SHIFT;
+
+  loadPow10Xjb64(powExp);
+  const pHi = gPow10Hi, pLo = gPow10Lo;


Suggested change

const pHi = gPow10Hi, pLo = gPow10Lo;

let pHi = gPow10Hi;

let pLo = gPow10Lo;

MaxGraey · 2026-06-17T23:43:38Z

+  const plo64 = pHi * y;
+  const lo = plo64 + mulhi64(pLo, y);
+  const p_hi = a + u64(lo < plo64);


Very inconsistent names plo64 instead pLo64, p_hi instead pHi64

MaxGraey · 2026-06-17T23:48:08Z

+  while (v >= 100) {
+    const q = v / 100;
+    p -= 4;
+    store<u32>(p, load<u32>(DIGITS + (<usize>(v - q * 100) << alignof<u32>())));
+    v = q;
+  }


this loop can be unrolled

MaxGraey · 2026-06-17T23:50:35Z

+@inline function decimalLen15(v: u64): i32 {
+  if (v < 100000000) {
+    if (v < 10000) {
+      if (v < 100) return v < 10 ? 1 : 2;
+      return v < 1000 ? 3 : 4;
+    }
+    if (v < 1000000) return v < 100000 ? 5 : 6;
+    return v < 10000000 ? 7 : 8;
+  }
+  if (v < 1000000000000) {
+    if (v < 10000000000) return v < 1000000000 ? 9 : 10;
+    return v < 100000000000 ? 11 : 12;
+  }
+  if (v < 100000000000000) return v < 10000000000000 ? 13 : 14;
+  return 15;
+}


See similar apreach which slightly more optimal for perf for branch leafs:

https://github.com/AssemblyScript/assemblyscript/blob/main/std/assembly/util/number.ts#L133

MaxGraey · 2026-06-17T23:52:47Z

+@inline function writeUInt16(buf: usize, value: u64): usize {
+  const len = decimalLen16(value);
+  let p = buf + (<usize>len << 1);
+  let v = value;
+  while (v >= 100) {
+    const q = v / 100;
+    p -= 4;
+    store<u32>(p, load<u32>(DIGITS + (<usize>(v - q * 100) << alignof<u32>())));
+    v = q;
+  }
+  if (v >= 10) {
+    store<u32>(buf, load<u32>(DIGITS + (<usize>v << alignof<u32>())));
+  } else {
+    store<u16>(buf, 0x30 + <u32>v);
+  }
+  return buf + (<usize>len << 1);
+}


Also did you compared this approach with original?
https://github.com/AssemblyScript/assemblyscript/blob/main/std/assembly/util/number.ts#L221

I think original whould be faster

MaxGraey · 2026-06-17T23:55:37Z

+      if (binSig != 0) return writeNaN(buf);
+      return writeInfinity(buf, neg);
+    }
+    if (binSig == 0) { store<u16>(buf, 0x30); return buf + 2; } // +/-0 -> "0"


Suggested change

if (binSig == 0) { store<u16>(buf, 0x30); return buf + 2; } // +/-0 -> "0"

// +/-0 -> "0"

if (binSig == 0) {

store<u16>(buf, 0x30);

return buf + 2;

}

Btw I guess it should be "0.0"?

MaxGraey · 2026-06-17T23:59:00Z

+@inline export function toDigits32(value: u64): void {
+  if (ASC_FEATURE_SIMD) return toDigits32Simd(value);
+  toBcd8(value);
+  gDigHi = gBcd + ZEROS;
+  gDigNum = gBcdLen;
+}


Suggested change

@inline export function toDigits32(value: u64): void {

if (ASC_FEATURE_SIMD) return toDigits32Simd(value);

toBcd8(value);

gDigHi = gBcd + ZEROS;

gDigNum = gBcdLen;

}

@inline export function toDigits32(value: u64): void {

if (ASC_FEATURE_SIMD) {

toDigits32Simd(value);

} else {

toBcd8(value);

gDigHi = gBcd + ZEROS;

gDigNum = gBcdLen;

}

}

MaxGraey · 2026-06-17T23:59:48Z

+    if (hasExtraDigit) store<u16>(buf + 16, <u32>lastDigitChar);
+    const sig = 8 + i32(hasExtraDigit);
+    const endByte = buf + ((decExp + 1) << 1);
+    for (let z = buf + (sig << 1); z < endByte; z += 16) putBlock8(z, ZEROS);


Suggested change

for (let z = buf + (sig << 1); z < endByte; z += 16) putBlock8(z, ZEROS);

for (let z = buf + (sig << 1); z < endByte; z += 16) {

putBlock8(z, ZEROS);

}

MaxGraey · 2026-06-18T00:00:41Z

+  }
+
+  const n = numDigits + i32(hasExtraDigit);
+  const endPos = decExp >= 0 ? (n > decExp + 1 ? n + 1 : decExp + 1) : n;


nit: double ternary

MaxGraey · 2026-06-18T00:03:37Z

I think I’ve roughly point style direction. There are still quite a lot of tweaks & nits to make. Ping me when you’re done, and I’ll give it another pass.

…ript into jairus/switch-to-xjb

JairusSW added 3 commits June 8, 2026 21:45

chore: add xjb-as to util

5705f06

Signed-off-by: Jairus Tanaka <me@jairus.dev>

chore: add liceses to NOTICE

c778be9

Signed-off-by: Jairus Tanaka <me@jairus.dev>

chore: update changed tests

1fdc61d

Signed-off-by: Jairus Tanaka <me@jairus.dev>

JairusSW marked this pull request as draft June 9, 2026 06:04

JairusSW added 5 commits June 16, 2026 16:00

chore: simplify and use smaller tables

09c3ef8

chore: use table from number instead

7cd14a4

chore: update notice

e0646ff

chore: clean up

650a7c3

chore: clean up

759c01f

JairusSW marked this pull request as ready for review June 17, 2026 18:13