------- Comment #6 from sgk at troutmask dot apl dot washington dot edu
2006-08-03 05:09 -------
Subject: Re: [4.2 Regression]: Optional argument failed
On Thu, Aug 03, 2006 at 04:49:33AM -0000, kargl at gcc dot gnu dot org wrote:
>
> (gdb) print *isym
> $28 = {name = 0x8869b61 "__convert_i4_r4",
> lib_name = 0x8869b61 "__convert_i4_r4", formal = 0x0, ts = {type = BT_REAL,
> kind = 4, derived = 0x0, cl = 0x0}, elemental = 1, pure = 0, generic = 0,
> specific = 0, actual_ok = 0, standard = 1, noreturn = 0, simplify = {
>
> Note, the __convert function is marked as elemental. I don't know why
> this is so.
>
OK, I understand the problem. Even though the __convert functions
are not elemental fortran procedures, gfortran marks these as
elemental to use the machinary to deal with both scalar and array
arguments. The upshot, AFAICT, is that we need to special case
the __convert functions in resolve.c.
The following patch allows a mangled version of HJ's testcase to
compile.
Index: resolve.c
===================================================================
--- resolve.c (revision 115827)
+++ resolve.c (working copy)
@@ -1012,7 +1012,8 @@
&& arg->expr->symtree->n.sym->attr.optional
&& formal_optional
&& arg->expr->rank
- && (set_by_optional || arg->expr->rank != rank))
+ && (set_by_optional || arg->expr->rank != rank)
+ && !(isym && isym->generic_id == GFC_ISYM_CONVERSION))
{
gfc_error ("'%s' at %L is an array and OPTIONAL; it cannot "
"therefore be an actual argument of an ELEMENTAL "
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28548
