------- Comment #4 from fxcoudert at gcc dot gnu dot org 2007-08-10 15:31
-------
The problem with integer kind is from the string length:
character(2_8) :: c(2)
integer :: i
c = "aa"
print *, c .eq. "aa"
call foo ((/(c(i), i = 1,2)/))
print *, c .eq. "aa"
contains
subroutine foo (c)
character(*), dimension(:) :: c
end subroutine foo
end
If you compile it and set a breakpoint on the translation of the string
comparison (which is trans-expr.c:1172), and print out the tree that is the
length of c (lse.string_length), you get the first time:
<integer_cst 0x2b81534a6930 type <integer_type 0x2b81531d45b0 int4> constant
invariant 2>
$3 = void
which is correct, while the second time you get:
<integer_cst 0x2b81534a61b0 type <integer_type 0x2b81531d4750 int8> constant
invariant 2>
which is wrong: the charlen type is int4, not int8. Somehow, the conjunction
of function call and temporary produces this, because if you remove either the
function call or the need for a temporary, the code is then correct.
>From what I seem there is a cl->backend_decl somewhere (or a string_length)
that is generated refering to this int8 constant, which should never happen. I
haven't been able to find out where this happens, though.
PS: in the mean time, I've found what I think is an unrelated bug. This is the
only occurrence that I could find of an cl->backend_decl with wrong type, but
it doesn't fix this bug ;-)
Index: trans-expr.c
===================================================================
--- trans-expr.c (revision 127334)
+++ trans-expr.c (working copy)
@@ -1855,6 +1851,7 @@ gfc_conv_aliased_arg (gfc_se * parmse, g
gfc_array_index_type);
tmp = fold_build2 (MINUS_EXPR, gfc_array_index_type,
tmp, tmp_se.expr);
+ tmp = fold_convert (gfc_charlen_type_node, tmp);
expr->ts.cl->backend_decl = tmp;
break;
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32937
