More optimizations, remove history preservation

Sources/_StringProcessing/ByteCodeGen.swift

@@ -665,15 +665,19 @@ fileprivate extension Compiler.ByteCodeGen {
       }
       // If there's a capture transform, apply it now.
       if let transform = transform {
-        let fn = builder.makeTransformFunction { input, storedCapture in
+        let fn = builder.makeTransformFunction { input, cap in
           // If it's a substring capture with no custom value, apply the
           // transform directly to the substring to avoid existential traffic.
-          if let cap = storedCapture.latest, cap.value == nil {
-            return try transform(input[cap.range])
+          //
+          // FIXME: separate out this code path. This is fragile,
+          // slow, and these are clearly different constructs
+          if let range = cap.range, cap.value == nil {
+            return try transform(input[range])
           }
+
           let value = constructExistentialOutputComponent(
              from: input,
-             component: storedCapture.latest,
+             component: cap.deconstructed,
              optionalCount: 0)
           return try transform(value)
         }

Sources/_StringProcessing/Engine/MECapture.swift

@@ -32,59 +32,48 @@
 
 extension Processor {
   struct _StoredCapture {
-    // Set whenever we push the very first capture, allows us
-    // to theoretically re-compute anything we want to later.
-    fileprivate var startState: SavePoint? = nil
-
-    // Save the entire history as we go, so that backtracking
-    // can just lop-off aborted runs.
-    //
-    // Backtracking entries can specify a per-capture stack
-    // index so that we can abort anything that came after.
-    //
-    // By remembering the entire history, we waste space, but
-    // we get flexibility for now.
-    //
-    fileprivate var history: Array<(range: Range<Position>, value: Any?)> = []
+    var range: Range<Position>? = nil
+
+    var value: Any? = nil
 
     // An in-progress capture start
     fileprivate var currentCaptureBegin: Position? = nil
 
     fileprivate func _invariantCheck() {
-      if startState == nil {
-        assert(history.isEmpty)
-        assert(currentCaptureBegin == nil)
-      } else if currentCaptureBegin == nil {
-        assert(!history.isEmpty)
+      if range == nil {
+        assert(value == nil)
       }
     }
 
     // MARK: - IPI
 
-    var isEmpty: Bool { history.isEmpty }
-
-    var latest: (range: Range<Position>, value: Any?)? { history.last }
+    var deconstructed: (range: Range<Position>, value: Any?)? {
+      guard let r = range else { return nil }
+      return (r, value)
+    }
 
     /// Start a new capture. If the previously started one was un-ended,
-    /// will clear it and restart. If this is the first start, will save `initial`.
+    /// will clear it and restart.
     mutating func startCapture(
-      _ idx: Position, initial: SavePoint
+      _ idx: Position
     ) {
       _invariantCheck()
       defer { _invariantCheck() }
 
-      if self.startState == nil {
-        self.startState = initial
-      }
       currentCaptureBegin = idx
     }
 
     mutating func endCapture(_ idx: Position) {
       _invariantCheck()
-      assert(currentCaptureBegin != nil)
       defer { _invariantCheck() }
 
-      history.append((currentCaptureBegin! ..< idx, value: nil))
+      guard let low = currentCaptureBegin else {
+        fatalError("Invariant violated: ending unstarted capture")
+      }
+
+      range = low..<idx
+      value = nil // TODO: cleaner IPI around this...
+      currentCaptureBegin = nil
     }
 
     mutating func registerValue(
@@ -93,28 +82,15 @@ extension Processor {
     ) {
       _invariantCheck()
       defer { _invariantCheck() }
-      if let sp = overwriteInitial {
-        self.startState = sp
-      }
-      history[history.endIndex - 1].value = value
-    }
-
-    mutating func fail(truncatingAt stackIdx: Int) {
-      _invariantCheck()
-      assert(stackIdx <= history.endIndex)
-      defer { _invariantCheck() }
 
-      history.removeSubrange(stackIdx...)
-      if history.isEmpty {
-        startState = nil
-      }
+      self.value = value
     }
   }
 }
 
 extension Processor._StoredCapture: CustomStringConvertible {
   var description: String {
-    return String(describing: history)
+    return String(describing: self)
   }
 }
 
@@ -124,10 +100,10 @@ struct MECaptureList {
 
   func latestUntyped(from input: String) -> Array<Substring?> {
     values.map {
-      guard let last = $0.latest else {
+      guard let range = $0.range else {
         return nil
       }
-      return input[last.0]
+      return input[range]
     }
   }
 }

Sources/_StringProcessing/Engine/Processor.swift

@@ -103,6 +103,12 @@ extension Processor {
     input[bounds]
   }
 
+  // Advance in our input, without any checks or failure signalling
+  mutating func _uncheckedForcedConsumeOne() {
+    assert(currentPosition != end)
+    input.formIndex(after: &currentPosition)
+  }
+
   // Advance in our input
   //
   // Returns whether the advance succeeded. On failure, our
@@ -145,30 +151,26 @@ extension Processor {
     return slice
   }
 
-  mutating func match(_ e: Element) {
+  // Match against the current input element. Returns whether
+  // it succeeded vs signaling an error.
+  mutating func match(_ e: Element) -> Bool {
     guard let cur = load(), cur == e else {
       signalFailure()
-      return
-    }
-    if consume(1) {
-      controller.step()
+      return false
     }
+    _uncheckedForcedConsumeOne()
+    return true
   }
+
+  // Match against the current input prefix. Returns whether
+  // it succeeded vs signaling an error.
   mutating func matchSeq<C: Collection>(
     _ seq: C
-  ) where C.Element == Input.Element {
-    let count = seq.count
-
-    guard let inputSlice = load(count: count),
-          seq.elementsEqual(inputSlice)
-    else {
-      signalFailure()
-      return
-    }
-    guard consume(.init(count)) else {
-      fatalError("unreachable")
+  ) -> Bool where C.Element == Input.Element {
+    for e in seq {
+      guard match(e) else { return false }
     }
-    controller.step()
+    return true
   }
 
   mutating func signalFailure() {
@@ -356,18 +358,24 @@ extension Processor {
 
     case .match:
       let reg = payload.element
-      match(registers[reg])
+      if match(registers[reg]) {
+        controller.step()
+      }
 
     case .matchSequence:
       let reg = payload.sequence
       let seq = registers[reg]
-      matchSeq(seq)
+      if matchSeq(seq) {
+        controller.step()
+      }
 
     case .matchSlice:
       let (lower, upper) = payload.pairedPosPos
       let range = registers[lower]..<registers[upper]
       let slice = input[range]
-      matchSeq(slice)
+      if matchSeq(slice) {
+        controller.step()
+      }
 
     case .consumeBy:
       let reg = payload.consumer
@@ -439,19 +447,19 @@ extension Processor {
       //   Should we assert it's not finished yet?
       //   What's the behavior there?
       let cap = storedCaptures[capNum]
-      guard let range = cap.latest?.range else {
+      guard let range = cap.range else {
         signalFailure()
         return
       }
-      matchSeq(input[range])
+      if matchSeq(input[range]) {
+        controller.step()
+      }
 
     case .beginCapture:
       let capNum = Int(
         asserting: payload.capture.rawValue)
 
-       let sp = makeSavePoint(self.currentPC)
-       storedCaptures[capNum].startCapture(
-         currentPosition, initial: sp)
+       storedCaptures[capNum].startCapture(currentPosition)
        controller.step()
 
      case .endCapture:

Sources/_StringProcessing/Engine/Structuralize.swift

@@ -12,7 +12,7 @@ extension CaptureList {
     for (i, (cap, meStored)) in zip(captures, list.values).enumerated() {
       let element = AnyRegexOutput.ElementRepresentation(
         optionalDepth: cap.optionalDepth,
-        content: meStored.latest,
+        content: meStored.deconstructed,
         name: cap.name,
         referenceID: list.referencedCaptureOffsets.first { $1 == i }?.key
       )

Sources/_StringProcessing/Regex/Match.swift

@@ -140,14 +140,19 @@ extension Regex {
     // FIXME: Something more efficient, likely an engine interface, and we
     // should scrap the RegexConsumer crap and call this
 
+    let executor = Executor(program: regex.program.loweredProgram)
+    let graphemeSemantic = regex.initialOptions.semanticLevel == .graphemeCluster
+
     var low = inputRange.lowerBound
     let high = inputRange.upperBound
     while true {
-      if let m = try _match(input, in: low..<high, mode: .partialFromFront) {
+      if let m: Regex<Output>.Match = try executor.match(
+        input, in: low..<high, .partialFromFront
+      ) {
         return m
       }
       if low >= high { return nil }
-      if regex.initialOptions.semanticLevel == .graphemeCluster {
+      if graphemeSemantic {
         input.formIndex(after: &low)
       } else {
         input.unicodeScalars.formIndex(after: &low)