This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Fix PR32628, bogus overflow flag on sizetype consts
On Wed, 16 Jan 2008, Richard Kenner wrote:
> > It's the subtle overflow checking code in the E_Array_Subtype case of
> > gnat_to_gnu_entity, more specifically:
> >
> > /* See if the base array type is already flat. If it is, we
> > are probably compiling an ACVC test, but it will cause the
> > code below to malfunction if we don't handle it specially. */
> > if (TREE_CODE (gnu_base_min) == INTEGER_CST
> > && TREE_CODE (gnu_base_max) == INTEGER_CST
> > && !TREE_OVERFLOW (gnu_base_min)
> > && !TREE_OVERFLOW (gnu_base_max)
> > && tree_int_cst_lt (gnu_base_max, gnu_base_min))
> > gnu_high = size_zero_node, gnu_min = size_one_node;
> >
> > which now triggers on the attached testcase, whereas it used not to.
>
> Ah, right.
>
> I guess we do have to have TREE_OVERFLOW on sizetypes for that purpose.
>
> I must say, then, that I missed the whole point of this change. If
> it's just for constants, why have special handling for TREE_OVEFLOW?
> If we KNOW the constant overflows, why not say so? What happens at run time
> has nothing to do with this.
The whole point of this excercise was to not set TREE_OVERFLOW for
the offset constant in the POINTER_PLUS_EXPR that is built for
int f(char *device)
{
return device == ((char *)0 + ~0UL);
}
because we warn then from the C frontend:
#2 0x000000000044fd5f in parser_build_binary_op (code=PLUS_EXPR, arg1=
{value = 0x2ad17856d5a0, original_code = ERROR_MARK}, arg2=
{value = 0x2ad177bd1b40, original_code = ERROR_MARK})
at /space/rguenther/src/svn/trunk/gcc/c-typeck.c:2785
2782 if (TREE_OVERFLOW_P (result.value)
2783 && !TREE_OVERFLOW_P (arg1.value)
2784 && !TREE_OVERFLOW_P (arg2.value))
2785 overflow_warning (result.value);
and of course (sizetype)~0UL did _not_ have overflow. The value just
got sign-extended but fit_double_type detects overflow if _any_ change
is made.
So, conversion to sizetype bogusly sets TREE_OVERFLOW. I can even more
constrain the fix to only ignore overflow if also the signedness
of both types match.
Richard.