VM: Fix incorrect canonicalization rule for `UnboxedIntConverterInstr`.

tests/language/vm/regress_29137_vm_test.dart

+// Copyright (c) 2017, the Dart project authors.  Please see the AUTHORS file
+// 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.
+//
+// Check that optimizer correctly handles (x << y) & MASK_32 pattern on 32-bit
+// platforms: given the pattern
+//
+//     v1 <- UnboxedIntConverter([tr] mint->uint32, v0)
+//     v2 <- UnboxedIntConverter(uint32->mint, v1)
+//
+// optimizer must *not* replace v2 with v0 because the first convertion is
+// truncating and is erasing the high part of the mint value.
+//
+// VMOptions=--optimization-counter-threshold=5 --no-background-compilation
+
+import "package:expect/expect.dart";
+
+const _MASK_32 = 0xffffffff;
+int _rotl32(int val, int shift) {
+  var mod_shift = shift & 31;

[kustermann, 2017/04/03 15:31:48]

final :)

+  return ((val << mod_shift) & _MASK_32) |
+      ((val & _MASK_32) >> (32 - mod_shift));
+}
+
+rot8(v) => _rotl32(v, 8);
+
+main() {
+  // Note: value is selected in such a way that (value << 8) is not a smi - this
+  // triggers emittion of BinaryMintOp instructions for shifts.
+  const value = 0xF0F00000;
+  const rotated = 0xF00000F0;
+  Expect.equals(rotated, rot8(value));
+  for (var i = 0; i < 10; i++) {
+    Expect.equals(rotated, rot8(value));
+  }
+}