On Wed, Dec 02, 2020 at 10:25:32AM +0100, Martin Liška wrote:
> It's a simple fix of a division by zero.
>
> Patch can bootstrap on x86_64-linux-gnu and survives regression tests.
>
> Ready to be installed?
> Thanks,
> Martin
>
> gcc/ChangeLog:
>
> PR c/98087
> * gimple-fold.c (clear_padding_type): Do not divide by zero.
>
> gcc/testsuite/ChangeLog:
>
> PR c/98087
> * gcc.c-torture/compile/pr98087.c: New test.
Ok, but can you extend the testcase a little bit to cover more cases?
> --- /dev/null
> +++ b/gcc/testsuite/gcc.c-torture/compile/pr98087.c
> @@ -0,0 +1,8 @@
> +/* PR c/98087 */
> +
> +struct S {};
> +void foo (int n)
> +{
> + struct S a[n][0];
> + __builtin_clear_padding (a);
> +}
Like:
struct S { char a; long long b; };
struct T { struct S c[0]; char d; };
void foo (int n)
{
struct S a[n][0];
__builtin_clear_padding (a);
__builtin_clear_padding (&a);
struct S b[7][0];
__builtin_clear_padding (&b);
struct T c;
__builtin_clear_padding (&c);
}
Without your patch we ICE on the first 3, the 4th not because the zero
length array satisfies is_empty_type. For the &a case, I even thought
I do handle that (the if (eltsz) conditional), but unlike my expectations,
int_size_in_bytes of the VLA with zero sized elements is not -1, but 0.
Jakub
