* data/glr.c (yyresolveLocations): Remove bogus YYASSERT that the

state before an empty RHS is always resolved here.  Only the location
of that state is guaranteed to be resolved, and that's enough.  This
fixes the remaining bug reported by Derek M. Jones in
<http://lists.gnu.org/archive/html/bug-bison/2006-05/msg00027.html>.
* tests/glr-regression.at (Uninitialized location when reporting
ambiguity): Test the above case.
Also, the embedded comments in this test case claim it checks the case
of an empty RHS that has inherited the initial location.  However, the
corresponding LHS was already resolved, so yyresolveLocations didn't
actually have reason to modify it.  Fix this by forcing
nondeterministic operation at the beginning of the parse.

data/glr.c

@@ -201,7 +201,7 @@ typedef struct YYLTYPE
 ]])
 
 b4_defines_if([#include @output_header_name@],
-              [b4_shared_declarations])[
+	      [b4_shared_declarations])[
 
 /* Enabling traces.  */
 #ifndef YYDEBUG
@@ -1836,8 +1836,14 @@ yyresolveLocations (yyGLRState* yys1, int yyn1,
 	    }
 	  else
 	    {
+	      /* Both yyresolveAction and yyresolveLocations traverse the GSS
+		 in reverse rightmost order.  It is only necessary to invoke
+		 yyresolveLocations on a subforest for which yyresolveAction
+		 would have been invoked next had an ambiguity not been
+		 detected.  Thus the location of the previous state (but not
+		 necessarily the previous state itself) is guaranteed to be
+		 resolved already.  */
 	      yyGLRState *yyprevious = yyoption->yystate;
-	      YYASSERT (yyprevious->yyresolved);
 	      yyrhsloc[0].yystate.yyloc = yyprevious->yyloc;
 	    }
 	  yychar_current = yychar;