------- Comment #1 from fxcoudert at gcc dot gnu dot org 2007-02-01 13:51
-------
With a little variation, you get a better error message, although "augument"
part could be improved :)
$ cat u.f90
integer :: i
i = -1
print *, repeat ("", i)
end
$ gfortran u.f90 && ./a.out
Fortran runtime error: Augument NCOPIES is negative.
The following patch removes the check from the library, and makes the front-end
generate it instead: removes the weird error message, adds error location, more
easily optimized away.
More important: examination of the generated code revealed that REPEAT
evaluates it ncopies arguments multiple times (up to 3 times), which is
wrong-code:
integer :: i
i = -1
print *, repeat ("a", f())
contains
integer function f()
integer :: x = 5
save x
x = x - 1
f = x
end function f
end
The above code prints out "aa" instead of "aaaa"; it segfaults if "x = 5" is
changed into "x = 3", for example. The patch below does also fix this
wrong-code issue. (We're not alone on this one, I found the same bug in Sun and
Portland compilers)
Index: gcc/fortran/trans-intrinsic.c
===================================================================
--- gcc/fortran/trans-intrinsic.c (revision 121280)
+++ gcc/fortran/trans-intrinsic.c (working copy)
@@ -3357,18 +3357,32 @@
tree ncopies;
tree var;
tree type;
+ tree cond;
args = gfc_conv_intrinsic_function_args (se, expr);
len = TREE_VALUE (args);
tmp = gfc_advance_chain (args, 2);
ncopies = TREE_VALUE (tmp);
+
+ /* Check that ncopies is not negative. */
+ ncopies = gfc_evaluate_now (ncopies, &se->pre);
+ cond = fold_build2 (LT_EXPR, boolean_type_node, ncopies,
+ build_int_cst (TREE_TYPE (ncopies), 0));
+ gfc_trans_runtime_check (cond,
+ "Argument NCOPIES of REPEAT intrinsic is negative",
+ &se->pre, &expr->where);
+
+ /* Compute the destination length. */
len = fold_build2 (MULT_EXPR, gfc_int4_type_node, len, ncopies);
type = gfc_get_character_type (expr->ts.kind, expr->ts.cl);
var = gfc_conv_string_tmp (se, build_pointer_type (type), len);
+ /* Create the argument list and generate the function call. */
arglist = NULL_TREE;
arglist = gfc_chainon_list (arglist, var);
- arglist = chainon (arglist, args);
+ arglist = gfc_chainon_list (arglist, TREE_VALUE(args));
+ arglist = gfc_chainon_list (arglist, TREE_VALUE(TREE_CHAIN(args)));
+ arglist = gfc_chainon_list (arglist, ncopies);
tmp = build_function_call_expr (gfor_fndecl_string_repeat, arglist);
gfc_add_expr_to_block (&se->pre, tmp);
Index: libgfortran/intrinsics/string_intrinsics.c
===================================================================
--- libgfortran/intrinsics/string_intrinsics.c (revision 121280)
+++ libgfortran/intrinsics/string_intrinsics.c (working copy)
@@ -362,14 +362,8 @@
{
int i;
- /* See if ncopies is valid. */
- if (ncopies < 0)
- {
- /* The error is already reported. */
- runtime_error ("Augument NCOPIES is negative.");
- }
-
- /* Copy characters. */
+ /* We don't need to check that ncopies is non-negative here, because
+ the front-end already generates code for that check. */
for (i = 0; i < ncopies; i++)
{
memmove (dest + (i * slen), src, slen);
--
fxcoudert at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |fxcoudert at gcc dot gnu dot
| |org
Status|UNCONFIRMED |NEW
Ever Confirmed|0 |1
Keywords| |wrong-code
Known to fail| |4.3.0 4.2.0 4.1.2
Last reconfirmed|0000-00-00 00:00:00 |2007-02-01 13:51:13
date| |
Target Milestone|--- |4.2.0
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30611
