ld script fill pattern expression

It turns out we do need to backtrack when parsing after all.  The
fill_opt component in the section rule swiches to EXPRESSION and back
to SCRIPT, and to find the end of an expression it is necessary to
look ahead one token.

	* ldgram.y (section): Throw away lookahead NAME token.
	(overlay_section): Likewise.
	* testsuite/ld-elf/overlay.t: Add fill pattern on overlays.
	Test fill pattern before stupidly named normal sections too,
	and before /DISCARD/.

1 files changed, 22 insertions, 0 deletions

diff --git a/ld/ldgram.y b/ld/ldgram.y
index 1f6c44a..a0c881e 100644
--- a/ld/ldgram.y
+++ b/ld/ldgram.y
@@ -1063,6 +1063,18 @@ section:	NAME
 			{ ldlex_popstate (); }
 		memspec_opt memspec_at_opt phdr_opt fill_opt
 			{
+			  /* fill_opt may have switched the lexer into
+			     expression state, and back again, but in
+			     order to find the end of the fill
+			     expression the parser must look ahead one
+			     token.  If it is a NAME, throw it away as
+			     it will have been lexed in the wrong
+			     state.  */
+			  if (yychar == NAME)
+			    {
+			      yyclearin;
+			      ldlex_backup ();
+			    }
 			  lang_leave_output_section_statement ($17, $14,
 							       $16, $15);
 			}
@@ -1079,6 +1091,11 @@ section:	NAME
 		'}'
 		memspec_opt memspec_at_opt phdr_opt fill_opt
 			{
+			  if (yychar == NAME)
+			    {
+			      yyclearin;
+			      ldlex_backup ();
+			    }
 			  lang_leave_overlay ($5, (int) $4,
 					      $15, $12, $14, $13);
 			}
@@ -1179,6 +1196,11 @@ overlay_section:
 			{ ldlex_popstate (); }
 		phdr_opt fill_opt
 			{
+			  if (yychar == NAME)
+			    {
+			      yyclearin;
+			      ldlex_backup ();
+			    }
 			  lang_leave_overlay_section ($9, $8);
 			}
 		opt_comma