GLR parsers sometimes raise parse errors instead of performing the

default reduction.
Reported by Charles-Henry de Boysson.
* tests/calc.at (_AT_CHECK_CALC, _AT_CHECK_CALC_ERROR): Don't
check the length of the	traces when %glr.
(_AT_CHECK_CALC_ERROR): Also skip `^Stack' lines, coming from
GLR's traces.
(AT_CHECK_CALC_LALR, AT_CHECK_CALC_GLR): New.
Test GLR parsers.
* data/glr.c (YYLEFTMOST_STATE): Fix its value.
(yyltype): Remove the yy prefix from the member names.
(yytable): Complete its comment.
(yygetLRActions): Map error action number from YYTABLE from
YYTABLE_NINF to 0.
(yyisErrorAction): No longer compare YYACTION to YYPACT_NINF
(which was a bug: it should have been YYTABEL_NINF, and yet it was
not satisfying as we could compare an YYACTION computed from
YYDEFACT to YYTABLE_NINF although they are unrelated): 0 is the
only value for error actions.
(yyreportParseError): In verbose parse error messages, don't issue
`error' in the list of expected tokens.
* data/yacc.c (yyparse) <yybackup>: Rewrite the decoding of the
next action to perform to match glr.c's decoding.
(yytable): Complete its comment.

NEWS

@@ -7,6 +7,13 @@ Changes in version 1.50a:
 
 * Indonesian translation thanks to Tedi Heriyanto.
 
+* GLR parsers
+  Fix spurious parse errors.
+
+* Pure parsers
+  Some people redefine yyerror to steal yyparse' private variables.
+  Reenable this trick until an official feature replaces it.
+
 Changes in version 1.50, 2002-10-04:
 
 * GLR parsing