Mark URLs with empty schemes as invalid.

components/omnibox/browser/autocomplete_input.cc

Index: components/omnibox/browser/autocomplete_input.cc
diff --git a/components/omnibox/browser/autocomplete_input.cc b/components/omnibox/browser/autocomplete_input.cc
index 7207fc8b5e68ba44e4e354bbc79b4f3317e4d5d1..73e8844c351e8d2faf08c0335cf367561f913183 100644
--- a/components/omnibox/browser/autocomplete_input.cc
+++ b/components/omnibox/browser/autocomplete_input.cc
@@ -155,6 +155,12 @@ metrics::OmniboxInputType::Type AutocompleteInput::Parse(
   if (first_non_white == base::string16::npos)
     return metrics::OmniboxInputType::INVALID;  // All whitespace.
 
+  // Treat input that begins with a colon as a query. Otherwise the URL
+  // formatter will attempt to fix it by prepending a scheme, which the user
+  // probably did not want if they explicitly typed a colon at the beginning.
+  if (text[first_non_white] == ':')
+    return metrics::OmniboxInputType::QUERY;

[Peter Kasting, 2016/10/05 00:56:30]

I'm fine with doing what you did to the URLFixerUpper and I agree that changing
GetValidScheme() would be wrong.

I'm still a little uncomfortable adding this up here though.  I think it's
probably OK if the fixer upper fixes ":w" into "http:w".  The result of that
will go in |canonicalized_url|, and we don't automatically navigate just because
the canonicalized URL contains something valid (because there are lots of other
cases where the input can be canonicalized, but was intended as a search -- we
distinguish those below, which is why I'm hoping to catch this case below as
well).

What I was trying to ask before was, if you don't touch this file, and you apply
the rest of this CL, then where is the specific exit point below) from this
function for ":w"?  I'd assume we'd drop all the way to the bottom of the
function and return UNKNOWN, which seems like a perfectly fine behavior to have.
 (Though it might require changing some tests.)

This doesn't duplicate the previous behavior precisely, but that's OK -- the old
conditional was there to prevent a crash trying to read from a nonexistent host,
not because we specifically wanted to ensure inputs starting with a colon were
forced to be queries.

[brettw, 2016/10/05 05:10:12]

It does return UNKNOWN at the bottom, and I was trying to keep the current
behavior.

GURL will return an empty string if the host doesn't exist, so you don't need to
check beforehand. I removed the empty host check below.

+
   // Ask our parsing back-end to help us understand what the user typed.  We
   // use the URLFixerUpper here because we want to be smart about what we
   // consider a scheme.  For example, we shouldn't consider www.google.com:80
@@ -254,13 +260,6 @@ metrics::OmniboxInputType::Type AutocompleteInput::Parse(
   // between an HTTP URL and a query, or the scheme is HTTP or HTTPS, in which
   // case we should reject invalid formulations.
 
-  // If we have an empty host it can't be a valid HTTP[S] URL.  (This should
-  // only trigger for input that begins with a colon, which GURL will parse as a
-  // valid, non-standard URL; for standard URLs, an empty host would have
-  // resulted in an invalid |canonicalized_url| above.)
-  if (!canonicalized_url->has_host())
-    return metrics::OmniboxInputType::QUERY;
-
   // Determine the host family.  We get this information by (re-)canonicalizing
   // the already-canonicalized host rather than using the user's original input,
   // in case fixup affected the result here (e.g. an input that looks like an