@exodus/bytes 1.0.0-rc.2 → 1.0.0-rc.4

package/README.md

@@ -1,3 +1,11 @@
-# bytes
+# `@exodus/bytes`
 
-Data structures handling
+`Uint8Array` conversion to and from `base64`, `base32`, `hex` and `utf8`
+
+[Fast](./Performance.md)
+
+Performs proper input validation
+
+## License
+
+[MIT](./LICENSE)

package/array.js

@@ -2,7 +2,7 @@ import { assertTypedArray } from './assert.js'
 
 const { Buffer } = globalThis // Buffer is optional
 
-export function fromTypedArray(arr, format) {
+export function typedView(arr, format) {
   assertTypedArray(arr)
   switch (format) {
     case 'uint8':

package/assert.js

@@ -12,10 +12,18 @@ const makeMessage = (name, extra) => `Expected${name ? ` ${name} to be` : ''} an
 const TypedArray = Object.getPrototypeOf(Uint8Array)
 
 export function assertTypedArray(arr) {
-  assert(arr instanceof TypedArray, 'Expected a TypedArray instance')
+  if (arr instanceof TypedArray) return
+  throw new TypeError('Expected a TypedArray instance')
 }
 
-export function assertUint8(arr, { name, length, ...rest } = {}) {
+export function assertUint8(arr, options) {
+  if (!options) {
+    // fast path
+    if (arr instanceof Uint8Array) return
+    throw new TypeError('Expected an Uint8Array')
+  }
+
+  const { name, length, ...rest } = options
   assertEmptyRest(rest)
   if (arr instanceof Uint8Array && (length === undefined || arr.length === length)) return
   throw new TypeError(makeMessage(name, length === undefined ? '' : ` of size ${Number(length)}`))

package/base32.js

@@ -0,0 +1,33 @@
+import { assertEmptyRest } from './assert.js'
+import { typedView } from './array.js'
+import * as js from './fallback/base32.js'
+
+// See https://datatracker.ietf.org/doc/html/rfc4648
+
+// 8 chars per 5 bytes
+
+const { E_PADDING } = js
+
+export const toBase32 = (arr, { padding = false } = {}) => js.toBase32(arr, false, padding)
+export const toBase32hex = (arr, { padding = false } = {}) => js.toBase32(arr, true, padding)
+
+// By default, valid padding is accepted but not required
+export const fromBase32 = (str, { format = 'uint8', padding = 'both', ...rest } = {}) =>
+  fromBase32common(str, false, padding, format, rest)
+export const fromBase32hex = (str, { format = 'uint8', padding = 'both', ...rest } = {}) =>
+  fromBase32common(str, true, padding, format, rest)
+
+function fromBase32common(str, isBase32Hex, padding, format, rest) {
+  if (typeof str !== 'string') throw new TypeError('Input is not a string')
+  assertEmptyRest(rest)
+
+  if (padding === true) {
+    if (str.length % 8 !== 0) throw new SyntaxError(E_PADDING)
+  } else if (padding === false) {
+    if (str.endsWith('=')) throw new SyntaxError('Did not expect padding in base32 input')
+  } else if (padding !== 'both') {
+    throw new TypeError('Invalid padding option')
+  }
+
+  return typedView(js.fromBase32(str, isBase32Hex), format)
+}

package/base64.js

@@ -1,93 +1,134 @@
-import { assert, assertUint8 } from './assert.js'
-import { fromTypedArray } from './array.js'
+import { assertUint8, assertEmptyRest } from './assert.js'
+import { typedView } from './array.js'
+import * as js from './fallback/base64.js'
 
 // See https://datatracker.ietf.org/doc/html/rfc4648
 
-// base64:    A-Za-z0-9+/ and =
-// base64url: A-Za-z0-9_-
+// base64:    A-Za-z0-9+/ and = if padding not disabled
+// base64url: A-Za-z0-9_- and = if padding enabled
 
 const { Buffer, atob } = globalThis // Buffer is optional, only used when native
 const haveNativeBuffer = Buffer && !Buffer.TYPED_ARRAY_SUPPORT
 const { toBase64: web64 } = Uint8Array.prototype // Modern engines have this
 
-export function toBase64(x) {
+const { E_CHAR, E_PADDING, E_LENGTH, E_LAST } = js
+
+const shouldUseAtob = atob && Boolean(globalThis.HermesInternal) // faster only on Hermes (and a little in old Chrome), js path beats it on normal engines
+
+// For native Buffer codepaths only
+const isBuffer = (x) => x.constructor === Buffer && Buffer.isBuffer(x)
+const toBuffer = (x) => (isBuffer(x) ? x : Buffer.from(x.buffer, x.byteOffset, x.byteLength))
+
+export function toBase64(x, { padding = true } = {}) {
   assertUint8(x)
-  if (web64 && x.toBase64 === web64) return x.toBase64() // Modern
-  if (!haveNativeBuffer) return toBase64js(x, BASE64, true) // Fallback
-  if (x.constructor === Buffer && Buffer.isBuffer(x)) return x.toString('base64') // Older Node.js
-  return Buffer.from(x.buffer, x.byteOffset, x.byteLength).toString('base64') // Older Node.js
+  if (web64 && x.toBase64 === web64) {
+    return padding ? x.toBase64() : x.toBase64({ omitPadding: !padding }) // Modern, optionless is slightly faster
+  }
+
+  if (!haveNativeBuffer) return js.toBase64(x, false, padding) // Fallback
+  const res = toBuffer(x).toString('base64') // Older Node.js
+  if (padding) return res
+  const at = res.indexOf('=', res.length - 3)
+  return at === -1 ? res : res.slice(0, at)
 }
 
-// NOTE: base64url omits padding
-export function toBase64url(x) {
+// NOTE: base64url omits padding by default
+export function toBase64url(x, { padding = false } = {}) {
   assertUint8(x)
-  if (web64 && x.toBase64 === web64) return x.toBase64({ alphabet: 'base64url', omitPadding: true }) // Modern
-  if (!haveNativeBuffer) return toBase64js(x, BASE64URL, false) // Fallback
+  if (web64 && x.toBase64 === web64) {
+    return x.toBase64({ alphabet: 'base64url', omitPadding: !padding }) // Modern
+  }
+
+  if (!haveNativeBuffer) return js.toBase64(x, true, padding) // Fallback
   if (x.constructor === Buffer && Buffer.isBuffer(x)) return x.toString('base64url') // Older Node.js
-  return Buffer.from(x.buffer, x.byteOffset, x.byteLength).toString('base64url') // Older Node.js
+  const res = toBuffer(x).toString('base64url') // Older Node.js
+  return padding && res.length % 4 !== 0 ? res + '='.repeat(4 - (res.length % 4)) : res
 }
 
 // Unlike Buffer.from(), throws on invalid input (non-base64 symbols and incomplete chunks)
 // Unlike Buffer.from() and Uint8Array.fromBase64(), does not allow spaces
 // NOTE: Always operates in strict mode for last chunk
 
-// Accepts both padded and non-padded variants, only strict base64
-export function fromBase64(str, format = 'uint8') {
-  if (typeof str !== 'string') throw new TypeError('Input is not a string')
+// By default accepts both padded and non-padded variants, only strict base64
+export function fromBase64(str, options = {}) {
+  if (typeof options === 'string') options = { format: options } // Compat due to usage, TODO: remove
+  const { format = 'uint8', padding = 'both', ...rest } = options
+  return fromBase64common(str, false, padding, format, rest)
+}
 
-  // These checks should be needed only for Buffer path, not Uint8Array.fromBase64 path, but JSC lacks proper checks
-  assert(str.length % 4 !== 1, 'Invalid base64 length') // JSC misses this in fromBase64
-  if (str.endsWith('=')) {
-    assert(str.length % 4 === 0, 'Invalid padded length') // JSC misses this too
-    assert(str[str.length - 3] !== '=', 'Excessive padding') // no more than two = at the end
-  }
+// By default accepts only non-padded strict base64url
+export function fromBase64url(str, { format = 'uint8', padding = false, ...rest } = {}) {
+  return fromBase64common(str, true, padding, format, rest)
+}
 
-  return fromTypedArray(fromBase64common(str, false), format)
+// By default accepts both padded and non-padded variants, base64 or base64url
+export function fromBase64any(str, { format = 'uint8', padding = 'both', ...rest } = {}) {
+  const isBase64url = !str.includes('+') && !str.includes('/') // likely to fail fast, as most input is non-url, also double scan is faster than regex
+  return fromBase64common(str, isBase64url, padding, format, rest)
 }
 
-// Accepts both only non-padded strict base64url
-export function fromBase64url(str, format = 'uint8') {
+function fromBase64common(str, isBase64url, padding, format, rest) {
   if (typeof str !== 'string') throw new TypeError('Input is not a string')
+  assertEmptyRest(rest)
+  const auto = padding === 'both' ? str.endsWith('=') : undefined
+  // Older JSC supporting Uint8Array.fromBase64 lacks proper checks
+  if (padding === true || auto === true) {
+    if (str.length % 4 !== 0) throw new SyntaxError(E_PADDING) // JSC misses this
+    if (str[str.length - 3] === '=') throw new SyntaxError(E_PADDING) // no more than two = at the end
+  } else if (padding === false || auto === false) {
+    if (str.length % 4 === 1) throw new SyntaxError(E_LENGTH) // JSC misses this in fromBase64
+    if (padding === false && str.endsWith('=')) {
+      throw new SyntaxError('Did not expect padding in base64 input') // inclusion is checked separately
+    }
+  } else {
+    throw new TypeError('Invalid padding option')
+  }
 
-  // These checks should be needed only for Buffer path, not Uint8Array.fromBase64 path, but JSC lacks proper checks
-  assert(str.length % 4 !== 1, 'Invalid base64 length') // JSC misses this in fromBase64
-  assert(!str.endsWith('='), 'Did not expect padding in base64url input') // inclusion is checked separately
-
-  return fromTypedArray(fromBase64common(str, true), format)
+  return typedView(fromBase64impl(str, isBase64url), format)
 }
 
-let fromBase64common
+// ASCII whitespace is U+0009 TAB, U+000A LF, U+000C FF, U+000D CR, or U+0020 SPACE
+const ASCII_WHITESPACE = /[\t\n\f\r ]/ // non-u for JSC perf
+
+let fromBase64impl
 if (Uint8Array.fromBase64) {
   // NOTICE: this is actually slower than our JS impl in older JavaScriptCore and (slightly) in SpiderMonkey, but faster on V8 and new JavaScriptCore
-  fromBase64common = (str, isBase64url) => {
+  fromBase64impl = (str, isBase64url) => {
     const alphabet = isBase64url ? 'base64url' : 'base64'
-    assert(!/\s/u.test(str), `Invalid character in ${alphabet} input`) // all other chars are checked natively
+    if (ASCII_WHITESPACE.test(str)) throw new SyntaxError(E_CHAR) // all other chars are checked natively
     const padded = str.length % 4 > 0 ? `${str}${'='.repeat(4 - (str.length % 4))}` : str
     return Uint8Array.fromBase64(padded, { alphabet, lastChunkHandling: 'strict' })
   }
 } else {
-  fromBase64common = (str, isBase64url) => {
-    if (isBase64url) {
-      assert(!/[^0-9a-z_-]/iu.test(str), 'Invalid character in base64url input')
-    } else {
-      assert(!/[^0-9a-z=+/]/iu.test(str), 'Invalid character in base64 input')
-    }
-
+  fromBase64impl = (str, isBase64url) => {
     let arr
-    if (!haveNativeBuffer && atob) {
+    if (haveNativeBuffer) {
+      const invalidRegex = isBase64url ? /[^0-9a-z=_-]/iu : /[^0-9a-z=+/]/iu
+      if (invalidRegex.test(str)) throw new SyntaxError(E_CHAR)
+      const at = str.indexOf('=')
+      if (at >= 0 && /[^=]/iu.test(str.slice(at))) throw new SyntaxError(E_PADDING)
+      arr = Buffer.from(str, 'base64')
+    } else if (shouldUseAtob) {
       // atob is faster than manual parsing on Hermes
-      const raw = atob(isBase64url ? str.replaceAll('-', '+').replaceAll('_', '/') : str)
+      if (isBase64url) {
+        if (/[\t\n\f\r +/]/.test(str)) throw new SyntaxError(E_CHAR) // atob verifies other invalid input
+        str = str.replaceAll('-', '+').replaceAll('_', '/')
+      } else {
+        if (ASCII_WHITESPACE.test(str)) throw new SyntaxError(E_CHAR) // all other chars are checked natively
+      }
+
+      let raw
+      try {
+        raw = atob(str)
+      } catch {
+        throw new SyntaxError(E_CHAR) // convert atob errors
+      }
+
       const length = raw.length
       arr = new Uint8Array(length)
       for (let i = 0; i < length; i++) arr[i] = raw.charCodeAt(i)
     } else {
-      // base64url is already checked to have no padding via a regex above
-      if (!isBase64url) {
-        const at = str.indexOf('=')
-        if (at >= 0) assert(!/[^=]/iu.test(str.slice(at)), 'Invalid padding')
-      }
-
-      arr = haveNativeBuffer ? Buffer.from(str, 'base64') : fromBase64js(str)
+      return js.fromBase64(str, isBase64url) // early return to skip last chunk verification, it's already validated in js
     }
 
     if (arr.length % 3 !== 0) {
@@ -95,111 +136,9 @@ if (Uint8Array.fromBase64) {
       const expected = toBase64(arr.subarray(-(arr.length % 3)))
       const end = str.length % 4 === 0 ? str.slice(-4) : str.slice(-(str.length % 4)).padEnd(4, '=')
       const actual = isBase64url ? end.replaceAll('-', '+').replaceAll('_', '/') : end
-      if (expected !== actual) throw new Error('Invalid last chunk')
+      if (expected !== actual) throw new SyntaxError(E_LAST)
     }
 
     return arr
   }
 }
-
-const BASE64 = [...'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/']
-const BASE64URL = [...'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789-_']
-
-const BASE64_PAIRS = []
-const BASE64URL_PAIRS = []
-
-// We construct output by concatenating chars, this seems to be fine enough on modern JS engines
-function toBase64js(arr, alphabet, padding) {
-  assertUint8(arr)
-  const fullChunks = Math.floor(arr.length / 3)
-  const fullChunksBytes = fullChunks * 3
-  let o = ''
-  let i = 0
-
-  const pairs = alphabet === BASE64URL ? BASE64URL_PAIRS : BASE64_PAIRS
-  if (pairs.length === 0) {
-    for (let i = 0; i < 64; i++) {
-      for (let j = 0; j < 64; j++) pairs.push(`${alphabet[i]}${alphabet[j]}`)
-    }
-  }
-
-  // Fast path for complete blocks
-  // This whole loop can be commented out, the algorithm won't change, it's just an optimization of the next loop
-  for (; i < fullChunksBytes; i += 3) {
-    const a = arr[i]
-    const b = arr[i + 1]
-    const c = arr[i + 2]
-    o += pairs[(a << 4) | (b >> 4)] + pairs[((b & 0x0f) << 8) | c]
-  }
-
-  // If we have something left, process it with a full algo
-  let carry = 0
-  let shift = 2 // First byte needs to be shifted by 2 to get 6 bits
-  const length = arr.length
-  for (; i < length; i++) {
-    const x = arr[i]
-    o += alphabet[carry | (x >> shift)] // shift >= 2, so this fits
-    if (shift === 6) {
-      shift = 0
-      o += alphabet[x & 0x3f]
-    }
-
-    carry = (x << (6 - shift)) & 0x3f
-    shift += 2 // Each byte prints 6 bits and leaves 2 bits
-  }
-
-  if (shift !== 2) o += alphabet[carry] // shift 2 means we have no carry left
-  if (padding) o += ['', '==', '='][length - fullChunksBytes]
-
-  return o
-}
-
-// Assumes no chars after =, checked
-let fromBase64jsMap
-
-function fromBase64js(str) {
-  const map = fromBase64jsMap || new Array(256)
-  if (!fromBase64jsMap) {
-    fromBase64jsMap = map
-    BASE64.forEach((c, i) => (map[c.charCodeAt(0)] = i))
-    map['-'.charCodeAt(0)] = map['+'.charCodeAt(0)] // for base64url
-    map['_'.charCodeAt(0)] = map['/'.charCodeAt(0)] // for base64url
-  }
-
-  let inputLength = str.length
-  while (str[inputLength - 1] === '=') inputLength--
-
-  const arr = new Uint8Array(Math.floor((inputLength * 3) / 4))
-  const tailLength = inputLength % 4
-  const mainLength = inputLength - tailLength // multiples of 4
-
-  let at = 0
-  let i = 0
-  let tmp
-
-  while (i < mainLength) {
-    tmp =
-      (map[str.charCodeAt(i)] << 18) |
-      (map[str.charCodeAt(i + 1)] << 12) |
-      (map[str.charCodeAt(i + 2)] << 6) |
-      map[str.charCodeAt(i + 3)]
-    arr[at++] = tmp >> 16
-    arr[at++] = (tmp >> 8) & 0xff
-    arr[at++] = tmp & 0xff
-    i += 4
-  }
-
-  if (tailLength === 3) {
-    tmp =
-      (map[str.charCodeAt(i)] << 10) |
-      (map[str.charCodeAt(i + 1)] << 4) |
-      (map[str.charCodeAt(i + 2)] >> 2)
-    arr[at++] = (tmp >> 8) & 0xff
-    arr[at++] = tmp & 0xff
-  } else if (tailLength === 2) {
-    tmp = (map[str.charCodeAt(i)] << 2) | (map[str.charCodeAt(i + 1)] >> 4)
-    arr[at++] = tmp & 0xff
-  }
-
-  return arr
-}

package/fallback/_utils.js

@@ -0,0 +1,6 @@
+const { Buffer, TextEncoder, TextDecoder } = globalThis
+const haveNativeBuffer = Buffer && !Buffer.TYPED_ARRAY_SUPPORT
+const isNative = (x) => x && (haveNativeBuffer || `${x}`.includes('[native code]')) // we consider Node.js TextDecoder/TextEncoder native
+const nativeEncoder = isNative(TextEncoder) ? new TextEncoder() : null
+const nativeDecoder = isNative(TextDecoder) ? new TextDecoder('utf8', { ignoreBOM: true }) : null
+export { nativeEncoder, nativeDecoder }

package/fallback/base32.js

@@ -0,0 +1,198 @@
+import { assertUint8 } from '../assert.js'
+import { nativeEncoder, nativeDecoder } from './_utils.js'
+
+// See https://datatracker.ietf.org/doc/html/rfc4648
+
+const BASE32 = [...'ABCDEFGHIJKLMNOPQRSTUVWXYZ234567'] // RFC 4648, #6
+const BASE32HEX = [...'0123456789ABCDEFGHIJKLMNOPQRSTUV'] // RFC 4648, #7
+const BASE32_HELPERS = {}
+const BASE32HEX_HELPERS = {}
+
+export const E_CHAR = 'Invalid character in base32 input'
+export const E_PADDING = 'Invalid base32 padding'
+export const E_LENGTH = 'Invalid base32 length'
+export const E_LAST = 'Invalid last chunk'
+
+// We construct output by concatenating chars, this seems to be fine enough on modern JS engines
+export function toBase32(arr, isBase32Hex, padding) {
+  assertUint8(arr)
+  const fullChunks = Math.floor(arr.length / 5)
+  const fullChunksBytes = fullChunks * 5
+  let o = ''
+  let i = 0
+
+  const alphabet = isBase32Hex ? BASE32HEX : BASE32
+  const helpers = isBase32Hex ? BASE32HEX_HELPERS : BASE32_HELPERS
+  if (!helpers.pairs) {
+    helpers.pairs = []
+    if (nativeDecoder) {
+      // Lazy to save memory in case if this is not needed
+      helpers.codepairs = new Uint16Array(32 * 32)
+      const u16 = helpers.codepairs
+      const u8 = new Uint8Array(u16.buffer, u16.byteOffset, u16.byteLength) // write as 1-byte to ignore BE/LE difference
+      for (let i = 0; i < 32; i++) {
+        const ic = alphabet[i].charCodeAt(0)
+        for (let j = 0; j < 32; j++) u8[(i << 6) | (j << 1)] = u8[(j << 6) | ((i << 1) + 1)] = ic
+      }
+    } else {
+      const p = helpers.pairs
+      for (let i = 0; i < 32; i++) {
+        for (let j = 0; j < 32; j++) p.push(`${alphabet[i]}${alphabet[j]}`)
+      }
+    }
+  }
+
+  const { pairs, codepairs } = helpers
+
+  // Fast path for complete blocks
+  // This whole loop can be commented out, the algorithm won't change, it's just an optimization of the next loop
+  if (nativeDecoder) {
+    const oa = new Uint16Array(fullChunks * 4)
+    for (let j = 0; i < fullChunksBytes; i += 5) {
+      const a = arr[i]
+      const b = arr[i + 1]
+      const c = arr[i + 2]
+      const d = arr[i + 3]
+      const e = arr[i + 4]
+      oa[j++] = codepairs[(a << 2) | (b >> 6)] // 8 + 8 - 5 - 5 = 6 left
+      oa[j++] = codepairs[((b & 0x3f) << 4) | (c >> 4)] // 6 + 8 - 5 - 5 = 4 left
+      oa[j++] = codepairs[((c & 0xf) << 6) | (d >> 2)] // 4 + 8 - 5 - 5 = 2 left
+      oa[j++] = codepairs[((d & 0x3) << 8) | e] // 2 + 8 - 5 - 5 = 0 left
+    }
+
+    o = nativeDecoder.decode(oa)
+  } else {
+    for (; i < fullChunksBytes; i += 5) {
+      const a = arr[i]
+      const b = arr[i + 1]
+      const c = arr[i + 2]
+      const d = arr[i + 3]
+      const e = arr[i + 4]
+      o += pairs[(a << 2) | (b >> 6)] // 8 + 8 - 5 - 5 = 6 left
+      o += pairs[((b & 0x3f) << 4) | (c >> 4)] // 6 + 8 - 5 - 5 = 4 left
+      o += pairs[((c & 0xf) << 6) | (d >> 2)] // 4 + 8 - 5 - 5 = 2 left
+      o += pairs[((d & 0x3) << 8) | e] // 2 + 8 - 5 - 5 = 0 left
+    }
+  }
+
+  // If we have something left, process it with a full algo
+  let carry = 0
+  let shift = 3 // First byte needs to be shifted by 3 to get 5 bits
+  for (; i < arr.length; i++) {
+    const x = arr[i]
+    o += alphabet[carry | (x >> shift)] // shift >= 3, so this fits
+    if (shift >= 5) {
+      shift -= 5
+      o += alphabet[(x >> shift) & 0x1f]
+    }
+
+    carry = (x << (5 - shift)) & 0x1f
+    shift += 3 // Each byte prints 5 bits and leaves 3 bits
+  }
+
+  if (shift !== 3) o += alphabet[carry] // shift 3 means we have no carry left
+  if (padding) o += ['', '======', '====', '===', '='][arr.length - fullChunksBytes]
+
+  return o
+}
+
+// TODO: can this be optimized? This only affects non-Hermes barebone engines though
+const mapSize = nativeEncoder ? 256 : 65_536 // we have to store 64 KiB map or recheck everything if we can't decode to byte array
+
+export function fromBase32(str, isBase32Hex) {
+  let inputLength = str.length
+  while (str[inputLength - 1] === '=') inputLength--
+  const paddingLength = str.length - inputLength
+  const tailLength = inputLength % 8
+  const mainLength = inputLength - tailLength // multiples of 8
+  if (![0, 2, 4, 5, 7].includes(tailLength)) throw new SyntaxError(E_LENGTH) // fast verification
+  if (paddingLength > 7 || (paddingLength !== 0 && str.length % 8 !== 0)) {
+    throw new SyntaxError(E_PADDING)
+  }
+
+  const alphabet = isBase32Hex ? BASE32HEX : BASE32
+  const helpers = isBase32Hex ? BASE32HEX_HELPERS : BASE32_HELPERS
+
+  if (!helpers.fromMap) {
+    helpers.fromMap = new Int8Array(mapSize).fill(-1) // no regex input validation here, so we map all other bytes to -1 and recheck sign
+    alphabet.forEach((c, i) => {
+      helpers.fromMap[c.charCodeAt(0)] = helpers.fromMap[c.toLowerCase().charCodeAt(0)] = i
+    })
+  }
+
+  const m = helpers.fromMap
+
+  const arr = new Uint8Array(Math.floor((inputLength * 5) / 8))
+  let at = 0
+  let i = 0
+
+  if (nativeEncoder) {
+    const codes = nativeEncoder.encode(str)
+    if (codes.length !== str.length) throw new SyntaxError(E_CHAR) // non-ascii
+    while (i < mainLength) {
+      // each 5 bits, grouped 5 * 4 = 20
+      const a = (m[codes[i++]] << 15) | (m[codes[i++]] << 10) | (m[codes[i++]] << 5) | m[codes[i++]]
+      const b = (m[codes[i++]] << 15) | (m[codes[i++]] << 10) | (m[codes[i++]] << 5) | m[codes[i++]]
+      if (a < 0 || b < 0) throw new SyntaxError(E_CHAR)
+      arr[at++] = a >> 12
+      arr[at++] = (a >> 4) & 0xff
+      arr[at++] = ((a << 4) & 0xff) | (b >> 16)
+      arr[at++] = (b >> 8) & 0xff
+      arr[at++] = b & 0xff
+    }
+  } else {
+    while (i < mainLength) {
+      // each 5 bits, grouped 5 * 4 = 20
+      const a =
+        (m[str.charCodeAt(i++)] << 15) |
+        (m[str.charCodeAt(i++)] << 10) |
+        (m[str.charCodeAt(i++)] << 5) |
+        m[str.charCodeAt(i++)]
+      const b =
+        (m[str.charCodeAt(i++)] << 15) |
+        (m[str.charCodeAt(i++)] << 10) |
+        (m[str.charCodeAt(i++)] << 5) |
+        m[str.charCodeAt(i++)]
+      if (a < 0 || b < 0) throw new SyntaxError(E_CHAR)
+      arr[at++] = a >> 12
+      arr[at++] = (a >> 4) & 0xff
+      arr[at++] = ((a << 4) & 0xff) | (b >> 16)
+      arr[at++] = (b >> 8) & 0xff
+      arr[at++] = b & 0xff
+    }
+  }
+
+  // Last block, valid tailLength: 0 2 4 5 7, checked already
+  // We check last chunk to be strict
+  if (tailLength < 2) return arr
+  const ab = (m[str.charCodeAt(i++)] << 5) | m[str.charCodeAt(i++)]
+  if (ab < 0) throw new SyntaxError(E_CHAR)
+  arr[at++] = ab >> 2
+  if (tailLength < 4) {
+    if (ab & 0x3) throw new SyntaxError(E_LAST)
+    return arr
+  }
+
+  const cd = (m[str.charCodeAt(i++)] << 5) | m[str.charCodeAt(i++)]
+  if (cd < 0) throw new SyntaxError(E_CHAR)
+  arr[at++] = ((ab << 6) & 0xff) | (cd >> 4)
+  if (tailLength < 5) {
+    if (cd & 0xf) throw new SyntaxError(E_LAST)
+    return arr
+  }
+
+  const e = m[str.charCodeAt(i++)]
+  if (e < 0) throw new SyntaxError(E_CHAR)
+  arr[at++] = ((cd << 4) & 0xff) | (e >> 1) // 4 + 4
+  if (tailLength < 7) {
+    if (e & 0x1) throw new SyntaxError(E_LAST)
+    return arr
+  }
+
+  const fg = (m[str.charCodeAt(i++)] << 5) | m[str.charCodeAt(i++)]
+  if (fg < 0) throw new SyntaxError(E_CHAR)
+  arr[at++] = ((e << 7) & 0xff) | (fg >> 3) // 1 + 5 + 2
+  // Can't be 8, so no h
+  if (fg & 0x7) throw new SyntaxError(E_LAST)
+  return arr
+}