Relanding: avoiding creating strings for canonicalization

pkg/front_end/lib/src/fasta/scanner/canonicalizer.dart

Index: pkg/front_end/lib/src/fasta/scanner/canonicalizer.dart
diff --git a/pkg/front_end/lib/src/fasta/scanner/canonicalizer.dart b/pkg/front_end/lib/src/fasta/scanner/canonicalizer.dart
new file mode 100644
index 0000000000000000000000000000000000000000..154f8fad94e794ba9f640ded95b174c137cba8c6
--- /dev/null
+++ b/pkg/front_end/lib/src/fasta/scanner/canonicalizer.dart
@@ -0,0 +1,120 @@
+// Copyright (c) 2017, the Dart project authors.  Please see the AUTHORS file

[ahe, 2017/02/20 09:18:30]

Consider renaming this file to string_canonicalizer.dart.

+// for details. All rights reserved. Use of this source code is governed by a
+// BSD-style license that can be found in the LICENSE file.
+
+library fasta.scanner.canonicalizer;
+
+import 'dart:convert';
+
+class Node {
+  var /* String | List<int> */ data; 
+  int start;
+  int end;
+  String payload;
+  Node next;
+  Node(this.data, this.start, this.end, this.payload, this.next);
+}
+
+/**

[ahe, 2017/02/20 09:18:30]

FWIW, I think we're shifting towards preferring three slashes over /**.

+ * A hash table for triples:
+ *  (list of bytes, start, end) --> canonicalized string
+ * Using triples avoids allocating string slices before checking if they
+ * are canonical.
+ * 
+ * Gives about 3% speedup on dart2js.
+ */
+class StringCanonicalizer {
+  /// Mask away top bits to keep hash calculation within 32-bit SMI range.
+  static const int MASK = 16 * 1024 * 1024 - 1;
+
+  static const int INITIAL_SIZE = 8 * 1024;
+
+  /// Linear size of a hash table.
+  int _size = INITIAL_SIZE;
+  /// Items in a hash table.
+  int _count = 0;
+  /// The table itself.
+  List<Node> _nodes = new List<Node>(INITIAL_SIZE);
+
+  static String decode(List<int> data, int start, int end, bool asciiOnly) {
+    var s;
+    if (asciiOnly) {
+      s = new String.fromCharCodes(data, start, end);
+    } else {
+      s = new Utf8Decoder(allowMalformed: true).convert(data, start, end);
+    }
+    return s;
+  }
+
+  static int hashBytes(List<int> data, int start, int end) {
+    int h = 5381;
+    for (int i = start; i < end; i++) {
+      h = ((h << 5) + h + data[i]) & MASK;
+    }
+    return h;
+  }
+
+  static int hashString(String data, int start, int end) {
+    int h = 5381;
+    for (int i = start; i < end; i++) {
+      h = ((h << 5) + h + data.codeUnitAt(i)) & MASK;
+    }
+    return h;
+  }
+
+  rehash() {
+    var newSize = _size * 2;
+    var newNodes = new List<Node>(newSize);
+    for (int i = 0; i < _size; i++) {
+      Node t = _nodes[i];
+      while (t != null) {
+        Node n = t.next;
+        int newIndex = t.data is String ?
+                      hashString(t.data, t.start, t.end) & (newSize - 1)
+                      : hashBytes(t.data, t.start, t.end) & (newSize - 1);
+        Node s = newNodes[newIndex];
+        t.next = s;
+        newNodes[newIndex] = t;
+        t = n;
+      }
+    }
+    _size = newSize;
+    _nodes = newNodes;
+  }
+
+  String canonicalize(data, int start, int end, bool asciiOnly) {
+    if (_count > _size) rehash();
+    int index = data is String
+        ? hashString(data, start, end)
+        : hashBytes(data, start, end);
+    index = index & (_size - 1);
+    Node s = _nodes[index];
+    Node t = s;
+    int len = end - start;
+    while (t != null) {
+      if (t.end - t.start == len) {
+        int i = start, j = t.start;
+        while (i < end && data[i] == t.data[j]) {
+          i++;
+          j++;
+        }
+        if (i == end) {
+          return t.payload;
+        }
+      }
+      t = t.next;
+    }
+    String payload;
+    if (data is String)
+      payload = data.substring(start, end);
+    else
+      payload = decode(data, start, end, asciiOnly);
+    _nodes[index] = new Node(data, start, end, payload, s);
+    _count++;
+    return payload;
+  }
+
+  clear() {
+    _nodes = new List<Node>(_size);
+  }
+}