Treat negative utf8-units the same way as other broken encodings.

sdk/lib/convert/utf.dart

Index: sdk/lib/convert/utf.dart
diff --git a/sdk/lib/convert/utf.dart b/sdk/lib/convert/utf.dart
index 80f7a7837f15ece8586f67385f46b0d3b5f3fd03..5a6cfe6ade59c9c2ffc7248eec783c9694f04492 100644
--- a/sdk/lib/convert/utf.dart
+++ b/sdk/lib/convert/utf.dart
@@ -455,7 +455,14 @@ class _Utf8Decoder {
 
       while (i < endIndex) {
         int unit = codeUnits[i++];
-        if (unit <= _ONE_BYTE_LIMIT) {
+        if (unit < 0) {

[Lasse Reichstein Nielsen, 2013/08/22 06:10:42]

Would it be cheaper/faster to do:
  if ((unit & ~0x7f) == 0) {
    // one-byte
  } else if ((unit & ~0xdf) == 0) {
    // two-byte
  } else if ((unit & ~0xef) == 0) {
    // three-byte
  } else if ((unit & ~0xf7) == 0) {
    // four-byte
  } else {
    // malformed
  }

This puts the most common case first, instead of wasting time checking for
malformedness first. Each check should be a single bit-and + branch-if-not-zero
in optimized smi code, so it's hard to make cheaper. Negative numbers are tested
the same way as any other numbers - they have bits set outside the first octet.

[floitsch, 2013/08/22 15:40:25]

Rapidly tried it, but somehow introduced a bug. Now added a TODO.

+          // TODO(floitsch): should this be unit <= 0 ?

[Lasse Reichstein Nielsen, 2013/08/22 06:10:42]

No, zero is a valid UTF-8 code unit encoding the U+0000 code point.

[floitsch, 2013/08/22 15:40:25]

Done.

+          if (!_allowMalformed) {
+            throw new FormatException(
+                "Negative UTF-8 code unit: -0x${(-unit).toRadixString(16)}");
+          }
+          _stringSink.writeCharCode(_REPLACEMENT_CHARACTER);
+        } else if (unit <= _ONE_BYTE_LIMIT) {
           _isFirstCharacter = false;
           _stringSink.writeCharCode(unit);
         } else {