@thomasfosterau/effect-parse 0.0.0-stage → 0.1.0

package/src/internal/numbers.ts

@@ -0,0 +1,291 @@
+/**
+ * Number conversions shared by the `Ascii` and `Text` parsers and the
+ * compilers.
+ */
+import type { Input } from "../Input.js"
+import { Fail } from "../Input.js"
+import { latin1 } from "./text.js"
+
+const toText = (s: Input): string =>
+  typeof s === "string" ? s : s instanceof Uint8Array ? latin1.text(s, 0, s.length) : s.join("")
+
+/** The value of the text matched by `Ascii.float` and `Text.float`. */
+export const toFloat = (s: Input): number => {
+  if (typeof s === "string") {
+    // Plain decimal notation (as opposed to `inf` or `nan`) ends in a digit or "."
+    const last = s.charCodeAt(s.length - 1)
+    if ((last >= 48 && last <= 57) || last === 46) return Number(s)
+  }
+  const t = toText(s).toLowerCase()
+  const negative = t.startsWith("-")
+  const body = t.startsWith("-") || t.startsWith("+") ? t.slice(1) : t
+  if (body === "nan") return NaN
+  if (body.startsWith("inf")) return negative ? -Infinity : Infinity
+  return Number(t)
+}
+
+/**
+ * The parsers that `Ascii.float` and `Text.float` map with `toFloat`: the
+ * compilers replace them by {@link scanFloatString} and
+ * {@link scanFloatBytes} when the text is in decimal notation.
+ */
+export const floatTexts = new WeakSet<object>()
+
+/** Adds a parser to {@link floatTexts}. */
+export const floatText = <P extends object>(p: P): P => {
+  floatTexts.add(p)
+  return p
+}
+
+/** The powers of ten that doubles represent exactly. */
+const exact = Array.from({ length: 23 }, (_, i) => Number(`1e${i}`))
+
+/**
+ * Where the latest scan ended, when it succeeded, and the furthest offset it
+ * read (when it succeeded or failed): streaming builds re-run the parser
+ * when that is near the end of a partial input.
+ */
+export const floatScan = { end: 0, read: 0 }
+
+const isDigit = (c: number): boolean => c >= 48 && c <= 57
+
+// Exact products of doubles (Dekker), without allocating: `a * b` is
+// `productHi + productLo` exactly, barring overflow and underflow
+const Split = 134217729 // 2^27 + 1
+let productHi = 0
+let productLo = 0
+const twoProduct = (a: number, b: number): void => {
+  const p = a * b
+  let t = Split * a
+  const ah = t - (t - a)
+  const al = a - ah
+  t = Split * b
+  const bh = t - (t - b)
+  const bl = b - bh
+  productHi = p
+  productLo = ((ah * bh - p) + ah * bl + al * bh) + al * bl
+}
+
+/** 2^-96: a bound on the relative error of {@link decimal}'s 19-digit path. */
+const Error19 = 1 / 2 ** 96
+
+/**
+ * The double nearest to `±(high * 10^low digits + low) * 10^e10`, where
+ * `high` holds the first 15 significant digits, and `low` the next `lowDigits`
+ * (at most 4). Returns `undefined` when this can't tell it apart from the
+ * neighbouring doubles, for the caller to use `Number`.
+ *
+ * Up to 15 digits, the digits and a power of ten up to 10^22 are exact
+ * doubles, so one multiplication or division rounds correctly (Clinger's
+ * fast path). Up to 19 digits, the digits are an exact sum of two doubles,
+ * and multiplying or dividing that by an exact power of ten gives the value
+ * to within a relative 2^-96 (double-double arithmetic). It is then rounded
+ * once, and kept only if the bounds of that error round to the same double:
+ * this is never the case for an exact tie, so ties go to `Number` too.
+ */
+const decimal = (negative: boolean, high: number, low: number, lowDigits: number, e10: number): number | undefined => {
+  let v: number
+  if (lowDigits === 0) {
+    if (e10 > 22 || e10 < -22) return undefined
+    v = e10 === 0 ? high : e10 > 0 ? high * exact[e10]! : high / exact[-e10]!
+  } else {
+    if (e10 > 22 || e10 < -22) return undefined
+    // The digits: high * 10^lowDigits + low = wh + wl, exactly
+    twoProduct(high, exact[lowDigits]!)
+    const wh = productHi + low
+    const wl = productLo + (low - (wh - productHi))
+    let xh: number
+    let xl: number
+    if (e10 >= 0) {
+      const p = exact[e10]!
+      twoProduct(wh, p)
+      xh = productHi
+      xl = productLo + wl * p
+    } else {
+      const p = exact[-e10]!
+      xh = wh / p
+      twoProduct(xh, p)
+      xl = ((wh - productHi) - productLo + wl) / p
+    }
+    v = xh + xl
+    const error = (xh < 0 ? -xh : xh) * Error19
+    if (xh + (xl + error) !== v || xh + (xl - error) !== v) return undefined
+  }
+  return negative ? -v : v
+}
+
+/**
+ * Matches and converts what `Ascii.float` matches at `i`, when it is in
+ * decimal notation: an optional sign, then digits or "." (the other forms,
+ * `nan` and `inf`, start with a letter). Returns the value, setting
+ * `floatScan.end`; `Fail` if the float parser fails there (always with
+ * `Backtrack`); or `undefined` if the text does not start like a decimal
+ * number, for the caller to run the parser.
+ *
+ * Numbers with up to 19 significant digits and a decimal exponent up to 22
+ * are converted by {@link decimal}, the rest by `Number`, so the value is
+ * always what `Number` gives for the text.
+ */
+export const scanFloatString = (s: string, i: number, end: number): number | Fail | undefined => {
+  const start = i
+  let c = i < end ? s.charCodeAt(i) : -1
+  if (c === 45 || c === 43) {
+    c = i + 1 < end ? s.charCodeAt(i + 1) : -1
+    if (!isDigit(c) && c !== 46) return undefined
+    i++
+  } else if (!isDigit(c) && c !== 46) {
+    return undefined
+  }
+  // The first 15 significant digits, the next 4, how many there are in all,
+  // and the decimal exponent of the last one kept
+  let high = 0
+  let low = 0
+  let digits = 0
+  let e10 = 0
+  if (c !== 46) {
+    do {
+      if (digits < 15) {
+        if (high !== 0 || c !== 48) {
+          high = high * 10 + (c - 48)
+          digits++
+        }
+      } else {
+        if (digits < 19) low = low * 10 + (c - 48)
+        else e10++
+        digits++
+      }
+      c = ++i < end ? s.charCodeAt(i) : -1
+    } while (isDigit(c))
+    if (c === 46) c = ++i < end ? s.charCodeAt(i) : -1
+  } else {
+    // "." must be followed by a digit
+    c = ++i < end ? s.charCodeAt(i) : -1
+    if (!isDigit(c)) {
+      floatScan.read = i
+      return Fail
+    }
+  }
+  while (isDigit(c)) {
+    if (digits < 15) {
+      if (high !== 0 || c !== 48) {
+        high = high * 10 + (c - 48)
+        digits++
+      }
+      e10--
+    } else {
+      if (digits < 19) {
+        low = low * 10 + (c - 48)
+        e10--
+      }
+      digits++
+    }
+    c = ++i < end ? s.charCodeAt(i) : -1
+  }
+  let read = i
+  if (c === 101 || c === 69) {
+    // The exponent is optional: without digits, it is not consumed
+    let j = i + 1
+    c = j < end ? s.charCodeAt(j) : -1
+    const negative = c === 45
+    if (negative || c === 43) c = ++j < end ? s.charCodeAt(j) : -1
+    if (isDigit(c)) {
+      let e = 0
+      do {
+        // Past this, the number can only overflow or underflow
+        if (e < 1e6) e = e * 10 + (c - 48)
+        c = ++j < end ? s.charCodeAt(j) : -1
+      } while (isDigit(c))
+      e10 += negative ? -e : e
+      i = j
+    }
+    read = j
+  }
+  floatScan.end = i
+  floatScan.read = read
+  const negative = s.charCodeAt(start) === 45
+  if (high === 0) return negative ? -0 : 0
+  const v = digits > 19 ? undefined : decimal(negative, high, low, digits > 15 ? digits - 15 : 0, e10)
+  return v === undefined ? Number(s.slice(start, i)) : v
+}
+
+/** {@link scanFloatString} on bytes, or the units of an ASCII compatible encoding. */
+export const scanFloatBytes = (b: Uint8Array, i: number, end: number): number | Fail | undefined => {
+  const start = i
+  let c = i < end ? b[i]! : -1
+  if (c === 45 || c === 43) {
+    c = i + 1 < end ? b[i + 1]! : -1
+    if (!isDigit(c) && c !== 46) return undefined
+    i++
+  } else if (!isDigit(c) && c !== 46) {
+    return undefined
+  }
+  let high = 0
+  let low = 0
+  let digits = 0
+  let e10 = 0
+  if (c !== 46) {
+    do {
+      if (digits < 15) {
+        if (high !== 0 || c !== 48) {
+          high = high * 10 + (c - 48)
+          digits++
+        }
+      } else {
+        if (digits < 19) low = low * 10 + (c - 48)
+        else e10++
+        digits++
+      }
+      c = ++i < end ? b[i]! : -1
+    } while (isDigit(c))
+    if (c === 46) c = ++i < end ? b[i]! : -1
+  } else {
+    c = ++i < end ? b[i]! : -1
+    if (!isDigit(c)) {
+      floatScan.read = i
+      return Fail
+    }
+  }
+  while (isDigit(c)) {
+    if (digits < 15) {
+      if (high !== 0 || c !== 48) {
+        high = high * 10 + (c - 48)
+        digits++
+      }
+      e10--
+    } else {
+      if (digits < 19) {
+        low = low * 10 + (c - 48)
+        e10--
+      }
+      digits++
+    }
+    c = ++i < end ? b[i]! : -1
+  }
+  let read = i
+  if (c === 101 || c === 69) {
+    let j = i + 1
+    c = j < end ? b[j]! : -1
+    const negative = c === 45
+    if (negative || c === 43) c = ++j < end ? b[j]! : -1
+    if (isDigit(c)) {
+      let e = 0
+      do {
+        if (e < 1e6) e = e * 10 + (c - 48)
+        c = ++j < end ? b[j]! : -1
+      } while (isDigit(c))
+      e10 += negative ? -e : e
+      i = j
+    }
+    read = j
+  }
+  floatScan.end = i
+  floatScan.read = read
+  const negative = b[start] === 45
+  if (high === 0) return negative ? -0 : 0
+  const v = digits > 19 ? undefined : decimal(negative, high, low, digits > 15 ? digits - 15 : 0, e10)
+  return v === undefined ? Number(asciiText(b, start, i)) : v
+}
+
+/** The text of ASCII bytes. */
+const asciiText = (b: Uint8Array, start: number, end: number): string =>
+  String.fromCharCode.apply(null, b.subarray(start, end) as unknown as Array<number>)

package/src/internal/resume.ts

@@ -0,0 +1,55 @@
+/**
+ * Checkpoints for resumable (streaming) parses.
+ *
+ * In partial mode a parser only succeeds on what it has seen, so a value
+ * parsed from a prefix of the input stays the same when more input
+ * arrives. A repetition that runs out of input mid-way can therefore record
+ * how far it got (its accumulator and position) before the failure
+ * propagates. When the parse is attempted again with more input, the same
+ * repetition, entered at the same offset, takes its checkpoint and carries
+ * on from there instead of re-parsing every element.
+ *
+ * Checkpoints are keyed by a per-parser key and the offset the repetition
+ * started at. Taking a checkpoint removes it, so its accumulator has a
+ * single owner; the repetition saves a new one if it runs out again.
+ *
+ * A resumed repetition that then completes leaves a {@link Done}
+ * checkpoint with its result, since the input may still run out after it
+ * (in trailing whitespace, say) and the next attempt would otherwise parse
+ * it all again. `Done` checkpoints stay in place when taken.
+ */
+import type { State } from "../Input.js"
+
+/** The checkpoints of a resumable parse. */
+export type Checkpoints = Map<object, Map<number, unknown>>
+
+/** Takes (and removes) the checkpoint of `key` entered at `offset`, if any. */
+export const take = (state: State<any>, key: object, offset: number): unknown => {
+  const byKey = state.resume!.get(key)
+  if (byKey === undefined) return undefined
+  const checkpoint = byKey.get(offset)
+  if (checkpoint !== undefined && !(checkpoint instanceof Done)) byKey.delete(offset)
+  return checkpoint
+}
+
+/** The result of a repetition that completed, and where it ended. */
+export class Done {
+  constructor(readonly value: unknown, readonly pos: number) {}
+}
+
+/**
+ * Returns `value`, the result of a repetition entered at `entry`; when the
+ * repetition was resumed from a checkpoint, records it as {@link Done}.
+ */
+export const done = <A>(state: State<any>, key: object, entry: number, resumed: boolean, value: A): A => {
+  if (resumed && state.resume !== undefined) save(state, key, entry, new Done(value, state.pos))
+  return value
+}
+
+/** Records the checkpoint of `key` entered at `offset`. */
+export const save = (state: State<any>, key: object, offset: number, checkpoint: unknown): void => {
+  const resume = state.resume!
+  let byKey = resume.get(key)
+  if (byKey === undefined) resume.set(key, byKey = new Map())
+  byKey.set(offset, checkpoint)
+}