[PATCH] Avoid sharing ts.cl if ts.cl->length == NULL

[Jakub Jelinek]

Hi!

The attached testcase is miscompiled by gfortran 4.1.x and 4.2.x.
While it works with 4.3.x, I still see incorrect ts.cl on EXPR_ARRAY etc.

The problem is that ts.cl struct is shared among all variables
declared on the same line, in this case
character(*), intent(in) :: bar1, bar2
While trans-decl.c (create_function_arglist) unshares them for the dummy
arguments, at that point it is too late.  Between variable_decl and
create_function_arglist bar2's ts has been copied over to many other
expressions (e.g. EXPR_ARRAY) and while bar2's ts is changed afterwards,
all the expressions still share ts.cl with bar1.  And
create_function_arglist sets f->sym->ts.cl->backend_decl to _bar1
argument and so anything that uses EXPR_ARRAY's etc. ts.cl->backend_decl
will use wrong length (_bar1 instead of _bar2).

Tested on 4.2/x86_64-linux and 4.3/x86_64-linux (the patch below is 4.2
version, there are whitespace differences in the trunk patch).
Ok for 4.1/4.2/trunk?

2007-08-17  Jakub Jelinek  <jakub@redhat.com>

	* decl.c (variable_decl): Don't share charlen structs if
	length == NULL.
	* trans-decl.c (create_function_arglist): Assert
	f->sym->ts.cl->backend_decl is NULL instead of unsharing
	charlen struct here.

	* gfortran.dg/assumed_charlen_sharing.f90: New test.

--- gcc/fortran/decl.c.jj	2007-03-20 00:22:06.000000000 +0100
+++ gcc/fortran/decl.c	2007-08-17 15:51:25.000000000 +0200
@@ -1139,10 +1139,11 @@ variable_decl (int elem)
 	  break;
 
 	/* Non-constant lengths need to be copied after the first
-	   element.  */
+	   element.  Also copy assumed lengths.  */
 	case MATCH_NO:
-	  if (elem > 1 && current_ts.cl->length
-		&& current_ts.cl->length->expr_type != EXPR_CONSTANT)
+	  if (elem > 1
+	      && (current_ts.cl->length == NULL
+		  || current_ts.cl->length->expr_type != EXPR_CONSTANT))
 	    {
 	      cl = gfc_get_charlen ();
 	      cl->next = gfc_current_ns->cl_list;
--- gcc/fortran/trans-decl.c.jj	2007-03-02 09:27:51.000000000 +0100
+++ gcc/fortran/trans-decl.c	2007-08-17 16:16:47.000000000 +0200
@@ -1428,25 +1428,8 @@ create_function_arglist (gfc_symbol * sy
 	  if (!f->sym->ts.cl->length)
 	    {
 	      TREE_USED (length) = 1;
-	      if (!f->sym->ts.cl->backend_decl)
-		f->sym->ts.cl->backend_decl = length;
-	      else
-		{
-		  /* there is already another variable using this
-		     gfc_charlen node, build a new one for this variable
-		     and chain it into the list of gfc_charlens.
-		     This happens for e.g. in the case
-		     CHARACTER(*)::c1,c2
-		     since CHARACTER declarations on the same line share
-		     the same gfc_charlen node.  */
-		  gfc_charlen *cl;
-	      
-		  cl = gfc_get_charlen ();
-		  cl->backend_decl = length;
-		  cl->next = f->sym->ts.cl->next;
-		  f->sym->ts.cl->next = cl;
-		  f->sym->ts.cl = cl;
-		}
+	      gcc_assert (!f->sym->ts.cl->backend_decl);
+	      f->sym->ts.cl->backend_decl = length;
 	    }
 
 	  hidden_typelist = TREE_CHAIN (hidden_typelist);
--- gcc/testsuite/gfortran.dg/assumed_charlen_sharing.f90.jj	2007-08-17 16:20:56.000000000 +0200
+++ gcc/testsuite/gfortran.dg/assumed_charlen_sharing.f90	2007-08-17 16:23:24.000000000 +0200
@@ -0,0 +1,29 @@
+! This testcase was miscompiled, because ts.cl
+! in function bar was initially shared between both
+! dummy arguments.  Although it was later unshared,
+! all expressions which copied ts.cl from bar2
+! before that used incorrectly bar1's length
+! instead of bar2.
+! { dg-do run }
+
+subroutine foo (foo1, foo2)
+  implicit none
+  integer, intent(in) :: foo2
+  character(*), intent(in) :: foo1(foo2)
+end subroutine foo
+
+subroutine bar (bar1, bar2)
+  implicit none
+  character(*), intent(in) :: bar1, bar2
+
+  call foo ((/ bar2 /), 1)
+end subroutine bar
+
+program test
+  character(80) :: str1
+  character(5) :: str2
+
+  str1 = 'String'
+  str2 = 'Strng'
+  call bar (str2, str1)
+end program test

	Jakub