@exodus/bytes 1.0.0-rc.4 → 1.0.0-rc.6

package/fallback/base64.js

@@ -1,5 +1,6 @@
 import { assertUint8 } from '../assert.js'
 import { nativeEncoder, nativeDecoder } from './_utils.js'
+import { encodeAscii } from './latin1.js'
 
 // See https://datatracker.ietf.org/doc/html/rfc4648
 
@@ -13,13 +14,10 @@ export const E_PADDING = 'Invalid base64 padding'
 export const E_LENGTH = 'Invalid base64 length'
 export const E_LAST = 'Invalid last chunk'
 
-// Alternatively, we could have mapped 0-255 bytes to charcodes and just used btoa(ascii),
-// but that approach is _slower_ than our toBase64js function, even on Hermes
-
 // We construct output by concatenating chars, this seems to be fine enough on modern JS engines
 export function toBase64(arr, isURL, padding) {
   assertUint8(arr)
-  const fullChunks = Math.floor(arr.length / 3)
+  const fullChunks = (arr.length / 3) | 0
   const fullChunksBytes = fullChunks * 3
   let o = ''
   let i = 0
@@ -51,21 +49,49 @@ export function toBase64(arr, isURL, padding) {
   // 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 * 2)
-    for (let j = 0; i < fullChunksBytes; i += 3) {
+    let j = 0
+    for (const last = arr.length - 11; i < last; i += 12, j += 8) {
+      const x0 = arr[i]
+      const x1 = arr[i + 1]
+      const x2 = arr[i + 2]
+      const x3 = arr[i + 3]
+      const x4 = arr[i + 4]
+      const x5 = arr[i + 5]
+      const x6 = arr[i + 6]
+      const x7 = arr[i + 7]
+      const x8 = arr[i + 8]
+      const x9 = arr[i + 9]
+      const x10 = arr[i + 10]
+      const x11 = arr[i + 11]
+      oa[j] = codepairs[(x0 << 4) | (x1 >> 4)]
+      oa[j + 1] = codepairs[((x1 & 0x0f) << 8) | x2]
+      oa[j + 2] = codepairs[(x3 << 4) | (x4 >> 4)]
+      oa[j + 3] = codepairs[((x4 & 0x0f) << 8) | x5]
+      oa[j + 4] = codepairs[(x6 << 4) | (x7 >> 4)]
+      oa[j + 5] = codepairs[((x7 & 0x0f) << 8) | x8]
+      oa[j + 6] = codepairs[(x9 << 4) | (x10 >> 4)]
+      oa[j + 7] = codepairs[((x10 & 0x0f) << 8) | x11]
+    }
+
+    // i < last here is equivalent to i < fullChunksBytes
+    for (const last = arr.length - 2; i < last; i += 3, j += 2) {
       const a = arr[i]
       const b = arr[i + 1]
       const c = arr[i + 2]
-      oa[j++] = codepairs[(a << 4) | (b >> 4)]
-      oa[j++] = codepairs[((b & 0x0f) << 8) | c]
+      oa[j] = codepairs[(a << 4) | (b >> 4)]
+      oa[j + 1] = codepairs[((b & 0x0f) << 8) | c]
     }
 
     o = nativeDecoder.decode(oa)
   } else {
+    // This can be optimized by ~25% with templates on Hermes, but this codepath is not called on Hermes, it uses btoa
+    // Check git history for templates version
     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]
+      o += pairs[(a << 4) | (b >> 4)]
+      o += pairs[((b & 0x0f) << 8) | c]
     }
   }
 
@@ -92,9 +118,8 @@ export function toBase64(arr, isURL, padding) {
 }
 
 // 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
+const mapSize = nativeEncoder ? 128 : 65_536 // we have to store 64 KiB map or recheck everything if we can't decode to byte array
 
-// Last chunk is rechecked at API
 export function fromBase64(str, isURL) {
   let inputLength = str.length
   while (str[inputLength - 1] === '=') inputLength--
@@ -121,26 +146,31 @@ export function fromBase64(str, isURL) {
   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) {
-      const a = (m[codes[i++]] << 18) | (m[codes[i++]] << 12) | (m[codes[i++]] << 6) | m[codes[i++]]
+    const codes = encodeAscii(str, E_CHAR)
+    for (; i < mainLength; i += 4) {
+      const c0 = codes[i]
+      const c1 = codes[i + 1]
+      const c2 = codes[i + 2]
+      const c3 = codes[i + 3]
+      const a = (m[c0] << 18) | (m[c1] << 12) | (m[c2] << 6) | m[c3]
       if (a < 0) throw new SyntaxError(E_CHAR)
-      arr[at++] = a >> 16
-      arr[at++] = (a >> 8) & 0xff
-      arr[at++] = a & 0xff
+      arr[at] = a >> 16
+      arr[at + 1] = (a >> 8) & 0xff
+      arr[at + 2] = a & 0xff
+      at += 3
     }
   } else {
-    while (i < mainLength) {
-      const a =
-        (m[str.charCodeAt(i++)] << 18) |
-        (m[str.charCodeAt(i++)] << 12) |
-        (m[str.charCodeAt(i++)] << 6) |
-        m[str.charCodeAt(i++)]
+    for (; i < mainLength; i += 4) {
+      const c0 = str.charCodeAt(i)
+      const c1 = str.charCodeAt(i + 1)
+      const c2 = str.charCodeAt(i + 2)
+      const c3 = str.charCodeAt(i + 3)
+      const a = (m[c0] << 18) | (m[c1] << 12) | (m[c2] << 6) | m[c3]
       if (a < 0) throw new SyntaxError(E_CHAR)
-      arr[at++] = a >> 16
-      arr[at++] = (a >> 8) & 0xff
-      arr[at++] = a & 0xff
+      arr[at] = a >> 16
+      arr[at + 1] = (a >> 8) & 0xff
+      arr[at + 2] = a & 0xff
+      at += 3
     }
   }
 

package/fallback/hex.js

@@ -1,37 +1,105 @@
 import { assertUint8 } from '../assert.js'
-import { nativeEncoder } from './_utils.js'
+import { nativeDecoder, nativeEncoder } from './_utils.js'
+import { encodeAscii } from './latin1.js'
 
-let hexArray
+let hexArray // array of 256 bytes converted to two-char hex strings
+let hexCodes // hexArray converted to u16 code pairs
 let dehexArray
+const _00 = 0x30_30 // '00' string in hex, the only allowed char pair to generate 0 byte
+const _ff = 0x66_66 // 'ff' string in hex, max allowed char pair (larger than 'FF' string)
+const allowed = '0123456789ABCDEFabcdef'
 
 export const E_HEX = 'Input is not a hex string'
 
-function toHexPart(arr, start, end) {
+function toHexPartAddition(a, start, end) {
   let o = ''
   let i = start
-  const last3 = end - 3
-  // Unrolled loop is faster
-  while (i < last3) {
-    const a = arr[i++]
-    const b = arr[i++]
-    const c = arr[i++]
-    const d = arr[i++]
-    o += hexArray[a]
-    o += hexArray[b]
-    o += hexArray[c]
-    o += hexArray[d]
+  const h = hexArray
+  for (const last3 = end - 3; i < last3; i += 4) {
+    const x0 = a[i]
+    const x1 = a[i + 1]
+    const x2 = a[i + 2]
+    const x3 = a[i + 3]
+    o += h[x0]
+    o += h[x1]
+    o += h[x2]
+    o += h[x3]
   }
 
-  while (i < end) o += hexArray[arr[i++]]
+  while (i < end) o += h[a[i++]]
   return o
 }
 
+// Optimiziation for Hermes which is the main user of fallback
+function toHexPartTemplates(a, start, end) {
+  let o = ''
+  let i = start
+  const h = hexArray
+  for (const last15 = end - 15; i < last15; i += 16) {
+    const x0 = a[i]
+    const x1 = a[i + 1]
+    const x2 = a[i + 2]
+    const x3 = a[i + 3]
+    const x4 = a[i + 4]
+    const x5 = a[i + 5]
+    const x6 = a[i + 6]
+    const x7 = a[i + 7]
+    const x8 = a[i + 8]
+    const x9 = a[i + 9]
+    const x10 = a[i + 10]
+    const x11 = a[i + 11]
+    const x12 = a[i + 12]
+    const x13 = a[i + 13]
+    const x14 = a[i + 14]
+    const x15 = a[i + 15]
+    o += `${h[x0]}${h[x1]}${h[x2]}${h[x3]}${h[x4]}${h[x5]}${h[x6]}${h[x7]}${h[x8]}${h[x9]}${h[x10]}${h[x11]}${h[x12]}${h[x13]}${h[x14]}${h[x15]}`
+  }
+
+  while (i < end) o += h[a[i++]]
+  return o
+}
+
+// Using templates is significantly faster in Hermes and JSC
+// It's harder to detect JSC and not important anyway as it has native impl, so we detect only Hermes
+const toHexPart = globalThis.HermesInternal ? toHexPartTemplates : toHexPartAddition
+
 export function toHex(arr) {
   assertUint8(arr)
 
   if (!hexArray) hexArray = Array.from({ length: 256 }, (_, i) => i.toString(16).padStart(2, '0'))
   const length = arr.length // this helps Hermes
 
+  // Only old browsers use this, barebone engines don't have TextDecoder
+  // But Hermes can use this when it (hopefully) implements TextDecoder
+  if (nativeDecoder) {
+    if (!hexCodes) {
+      hexCodes = new Uint16Array(256)
+      const u8 = new Uint8Array(hexCodes.buffer, hexCodes.byteOffset, hexCodes.byteLength)
+      for (let i = 0; i < 256; i++) {
+        const pair = hexArray[i]
+        u8[2 * i] = pair.charCodeAt(0)
+        u8[2 * i + 1] = pair.charCodeAt(1)
+      }
+    }
+
+    const oa = new Uint16Array(length)
+    let i = 0
+    for (const last3 = arr.length - 3; ; i += 4) {
+      if (i >= last3) break // loop is fast enough for moving this here to be useful on JSC
+      const x0 = arr[i]
+      const x1 = arr[i + 1]
+      const x2 = arr[i + 2]
+      const x3 = arr[i + 3]
+      oa[i] = hexCodes[x0]
+      oa[i + 1] = hexCodes[x1]
+      oa[i + 2] = hexCodes[x2]
+      oa[i + 3] = hexCodes[x3]
+    }
+
+    for (; i < length; i++) oa[i] = hexCodes[arr[i]]
+    return nativeDecoder.decode(oa)
+  }
+
   if (length > 30_000) {
     // Limit concatenation to avoid excessive GC
     // Thresholds checked on Hermes
@@ -51,54 +119,74 @@ export function toHex(arr) {
   return toHexPart(arr, 0, length)
 }
 
-// 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 fromHex(str) {
   if (typeof str !== 'string') throw new TypeError('Input is not a string')
   if (str.length % 2 !== 0) throw new SyntaxError(E_HEX)
 
-  // We don't use native Buffer impl, as rechecking input make it slower than pure js
-  // This path is used only on older engines though
-
-  if (!dehexArray) {
-    dehexArray = new Int8Array(mapSize).fill(-1) // no regex input validation here, so we map all other bytes to -1 and recheck sign
-    for (let i = 0; i < 16; i++) {
-      const s = i.toString(16)
-      dehexArray[s.charCodeAt(0)] = dehexArray[s.toUpperCase().charCodeAt(0)] = i
-    }
-  }
-
   const length = str.length / 2 // this helps Hermes in loops
   const arr = new Uint8Array(length)
-  let j = 0
+
+  // Native encoder path is beneficial even for small arrays in Hermes
   if (nativeEncoder) {
-    // Native encoder path is beneficial even for small arrays in Hermes
-    const codes = nativeEncoder.encode(str)
-    if (codes.length !== str.length) throw new SyntaxError(E_HEX) // non-ascii
-    const last3 = length - 3 // Unroll nativeEncoder path as this is what modern Hermes takes and a small perf improvement is nice there
+    if (!dehexArray) {
+      dehexArray = new Uint8Array(_ff + 1) // 26 KiB cache, >2x perf improvement on Hermes
+      const u8 = new Uint8Array(2)
+      const u16 = new Uint16Array(u8.buffer, u8.byteOffset, 1) // for endianess-agnostic transform
+      const map = [...allowed].map((c) => [c.charCodeAt(0), parseInt(c, 16)])
+      for (const [ch, vh] of map) {
+        u8[0] = ch // first we read high hex char
+        for (const [cl, vl] of map) {
+          u8[1] = cl // then we read low hex char
+          dehexArray[u16[0]] = (vh << 4) | vl
+        }
+      }
+    }
+
+    const codes = encodeAscii(str, E_HEX)
+    const codes16 = new Uint16Array(codes.buffer, codes.byteOffset, codes.byteLength / 2)
     let i = 0
-    while (i < last3) {
-      const a = (dehexArray[codes[j++]] << 4) | dehexArray[codes[j++]]
-      const b = (dehexArray[codes[j++]] << 4) | dehexArray[codes[j++]]
-      const c = (dehexArray[codes[j++]] << 4) | dehexArray[codes[j++]]
-      const d = (dehexArray[codes[j++]] << 4) | dehexArray[codes[j++]]
-      if (a < 0 || b < 0 || c < 0 || d < 0) throw new SyntaxError(E_HEX)
-      arr[i++] = a
-      arr[i++] = b
-      arr[i++] = c
-      arr[i++] = d
+    for (const last3 = length - 3; i < last3; i += 4) {
+      const ai = codes16[i]
+      const bi = codes16[i + 1]
+      const ci = codes16[i + 2]
+      const di = codes16[i + 3]
+      const a = dehexArray[ai]
+      const b = dehexArray[bi]
+      const c = dehexArray[ci]
+      const d = dehexArray[di]
+      if ((!a && ai !== _00) || (!b && bi !== _00) || (!c && ci !== _00) || (!d && di !== _00)) {
+        throw new SyntaxError(E_HEX)
+      }
+
+      arr[i] = a
+      arr[i + 1] = b
+      arr[i + 2] = c
+      arr[i + 3] = d
     }
 
     while (i < length) {
-      const res = (dehexArray[codes[j++]] << 4) | dehexArray[codes[j++]]
-      if (res < 0) throw new SyntaxError(E_HEX)
-      arr[i++] = res
+      const ai = codes16[i]
+      const a = dehexArray[ai]
+      if (!a && ai !== _00) throw new SyntaxError(E_HEX)
+      arr[i++] = a
     }
   } else {
+    if (!dehexArray) {
+      // no regex input validation here, so we map all other bytes to -1 and recheck sign
+      // non-ASCII chars throw already though, so we should process only 0-127
+      dehexArray = new Int8Array(128).fill(-1)
+      for (let i = 0; i < 16; i++) {
+        const s = i.toString(16)
+        dehexArray[s.charCodeAt(0)] = dehexArray[s.toUpperCase().charCodeAt(0)] = i
+      }
+    }
+
+    let j = 0
     for (let i = 0; i < length; i++) {
-      const res = (dehexArray[str.charCodeAt(j++)] << 4) | dehexArray[str.charCodeAt(j++)]
-      if (res < 0) throw new SyntaxError(E_HEX)
+      const a = str.charCodeAt(j++)
+      const b = str.charCodeAt(j++)
+      const res = (dehexArray[a] << 4) | dehexArray[b]
+      if (res < 0 || (0x7f | a | b) !== 0x7f) throw new SyntaxError(E_HEX) // 0-127
       arr[i] = res
     }
   }

package/fallback/latin1.js

@@ -0,0 +1,89 @@
+import { nativeEncoder } from './_utils.js'
+
+// See http://stackoverflow.com/a/22747272/680742, which says that lowest limit is in Chrome, with 0xffff args
+// On Hermes, actual max is 0x20_000 minus current stack depth, 1/16 of that should be safe
+const maxFunctionArgs = 0x20_00
+
+export function asciiPrefix(arr) {
+  let p = 0 // verified ascii bytes
+  const length = arr.length
+  // Threshold tested on Hermes (worse on <=48, better on >=52)
+  // Also on v8 arrs of size <=64 might be on heap and using Uint32Array on them is unoptimal
+  if (length > 64) {
+    // Speedup with u32
+    const u32start = (4 - (arr.byteOffset & 3)) % 4 // offset start by this many bytes for alignment
+    for (; p < u32start; p++) if (arr[p] >= 0x80) return p
+    const u32length = ((arr.byteLength - u32start) / 4) | 0
+    const u32 = new Uint32Array(arr.buffer, arr.byteOffset + u32start, u32length)
+    let i = 0
+    for (const last3 = u32length - 3; ; p += 16, i += 4) {
+      if (i >= last3) break // loop is fast enough for moving this here to be _very_ useful, likely due to array access checks
+      const a = u32[i]
+      const b = u32[i + 1]
+      const c = u32[i + 2]
+      const d = u32[i + 3]
+      if (a & 0x80_80_80_80 || b & 0x80_80_80_80 || c & 0x80_80_80_80 || d & 0x80_80_80_80) break
+    }
+
+    for (; i < u32length; p += 4, i++) if (u32[i] & 0x80_80_80_80) break
+  }
+
+  for (; p < length; p++) if (arr[p] >= 0x80) return p
+  return length
+}
+
+export function decodeLatin1(arr, start = 0, stop = arr.length) {
+  start |= 0
+  stop |= 0
+  const total = stop - start
+  if (total === 0) return ''
+  if (total > maxFunctionArgs) {
+    let prefix = ''
+    for (let i = start; i < stop; ) {
+      const i1 = Math.min(stop, i + maxFunctionArgs)
+      prefix += String.fromCharCode.apply(String, arr.subarray(i, i1))
+      i = i1
+    }
+
+    return prefix
+  }
+
+  const sliced = start === 0 && stop === arr.length ? arr : arr.subarray(start, stop)
+  return String.fromCharCode.apply(String, sliced)
+}
+
+export const encodeLatin1 = globalThis.HermesInternal
+  ? (str) => {
+      const length = str.length
+      const arr = new Uint8Array(length)
+      if (length > 64) {
+        const at = str.charCodeAt.bind(str) // faster on strings from ~64 chars on Hermes, but can be 10x slower on e.g. JSC
+        for (let i = 0; i < length; i++) arr[i] = at(i)
+      } else {
+        for (let i = 0; i < length; i++) arr[i] = str.charCodeAt(i)
+      }
+
+      return arr
+    }
+  : (str) => {
+      const length = str.length
+      const arr = new Uint8Array(length)
+      // Can be optimized with unrolling, but this is not used on non-Hermes atm
+      for (let i = 0; i < length; i++) arr[i] = str.charCodeAt(i)
+      return arr
+    }
+
+// Expects nativeEncoder to be present
+export const encodeAscii = globalThis.HermesInternal
+  ? (str, ERR) => {
+      // Much faster in Hermes
+      const codes = new Uint8Array(str.length + 4) // overshoot by a full utf8 char
+      const info = nativeEncoder.encodeInto(str, codes)
+      if (info.read !== str.length || info.written !== str.length) throw new SyntaxError(ERR) // non-ascii
+      return codes.subarray(0, str.length)
+    }
+  : (str, ERR) => {
+      const codes = nativeEncoder.encode(str)
+      if (codes.length !== str.length) throw new SyntaxError(ERR) // non-ascii
+      return codes
+    }