dart2js: Imprecise type inference and type propagation on || condition

[rakudrama]

The code below, compiled with

sdk/bin/dart2js --package-root=out/ReleaseX64/packages ~/play1/bug6j1a.dart \
 --trust-type-annotations --dump-info

contains an interceptor call for externalUrl.startsWith('#'), i.e.

J.startsWith$1$s(externalUrl, "#")

The dumpinfo output shows AA.x is [null|exact=JSString], so global type inference does not determine externalUrl cannot be null.
Further, SSA type propagation does not clean up the type.
So both 'passes' fail to be precise enough.
Changing the first 'if' to two 'if' statements generates the expected code for the return expression:

  if (externalUrl == null) return '';
  if (externalUrl.isEmpty) return '';

---

out.js:

encodeSecureHref: [function(externalUrl) {
  if (externalUrl == null || externalUrl.length === 0)
    return "";
  T.AA(externalUrl);
  return J.startsWith$1$s(externalUrl, "#") ||
         C.JSString_methods.startsWith$1(externalUrl.toLowerCase(), "mailto:")
      ? externalUrl
      : "http://app/" +
          P._Uri__uriEncode(C.List_nxB, externalUrl, C.Utf8Codec_false, true);
}, "call$1", "bug6j1a__encodeSecureHref$closure", 2, 0, 0]

---

bug6j1a.dart:

import 'package:expect/expect.dart';

const String _SECURE_LINK_PREFIX = 'http://app/';

String encodeSecureHref(String externalUrl) {
  if (externalUrl == null || externalUrl.isEmpty) {
    return '';
  }
  AA(externalUrl);
  return (externalUrl.startsWith('#') ||
      externalUrl.toLowerCase().startsWith('mailto:')) ?
          externalUrl :
          _SECURE_LINK_PREFIX + Uri.encodeQueryComponent(externalUrl);
}

@NoInline()
AA(x) => x;

var f = (x) => x * 2;

main() {
  print(encodeSecureHref('#foo'));
  print(encodeSecureHref('foo'));
  print(encodeSecureHref(null));
  print(encodeSecureHref(''));
  print(f(100));
  f = encodeSecureHref;
  print(encodeSecureHref);
}

[rakudrama]

#28330 is one way of solving this for types_propagation.
Something similar should be incorporated into global type analysis, which should be easier since the graph has no structural constraints (i.e. need not be reducible).

[rakudrama]

#9505e95 pushes refinements past the join point in cases like the above example