On Wed, Mar 28, 2012 at 3:54 PM, Jakub Jelinek <ja...@redhat.com> wrote:
> Hi!
>
> The attached testcase shows that already for 32x signed char
> vector shuffles using the index element type for computing
> BIT_FIELD_REF positions is wrong - bytes 16 and above
> in the vector are bits 128 and above, and if idx
> has signed char type, that overflows.
>
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux,
> ok for trunk/4.7?

Ok with using ...

> 2012-03-28  Jakub Jelinek  <ja...@redhat.com>
>
>        PR middle-end/52750
>        * tree-vect-generic.c (vector_element): Perform multiplication
>        for pos in bitsizetype type instead of idx type.
>
>        * gcc.c-torture/compile/pr52750.c: New test.
>
> --- gcc/tree-vect-generic.c.jj  2012-03-20 08:51:32.000000000 +0100
> +++ gcc/tree-vect-generic.c     2012-03-28 11:20:16.715685934 +0200
> @@ -567,8 +567,9 @@ vector_element (gimple_stmt_iterator *gs
>       else
>         {
>          tree size = TYPE_SIZE (vect_elt_type);
> -          tree pos = fold_build2 (MULT_EXPR, TREE_TYPE (idx), idx, size);
> -          return fold_build3 (BIT_FIELD_REF, vect_elt_type, vect, size, pos);
> +         tree pos = fold_build2 (MULT_EXPR, bitsizetype,
> +                                 fold_convert (bitsizetype, idx), size);

  bitsize_int (index)

here.

Thanks,
Richard.

> +         return fold_build3 (BIT_FIELD_REF, vect_elt_type, vect, size, pos);
>         }
>     }
>
> --- gcc/testsuite/gcc.c-torture/compile/pr52750.c.jj    2012-03-28 
> 11:51:12.110286829 +0200
> +++ gcc/testsuite/gcc.c-torture/compile/pr52750.c       2012-03-28 
> 11:49:44.000000000 +0200
> @@ -0,0 +1,11 @@
> +/* PR middle-end/52750 */
> +
> +typedef signed char V __attribute__((vector_size (32)));
> +
> +void
> +foo (V *x)
> +{
> +  V m = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
> +         16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31};
> +  *x = __builtin_shuffle (*x, m);
> +}
>
>        Jakub

Reply via email to