On Wed, Jul 26, 2023 at 4:52 AM Kewen.Lin <li...@linux.ibm.com> wrote:
>
> Hi,
>
> PR110776 exposes one issue that we could query unaligned
> load for vector type but actually no unaligned vector load
> is supported there. The reason is that the costed load is
> with single-lane vector type and its memory access type is
> VMAT_ELEMENTWISE, we actually take it as scalar load and
> set its alignment_support_scheme as dr_unaligned_supported.
>
> To avoid the ICE as exposed, following Rich's suggestion,
> this patch is to make VMAT_ELEMENTWISE be costed as scalar
> load.
>
> Bootstrapped and regress-tested on x86_64-redhat-linux,
> powerpc64-linux-gnu P8/P9 and powerpc64le-linux-gnu P9/P10.
>
> Is it ok for trunk?
OK.
> BR,
> Kewen
> -----
>
> Co-authored-by: Richard Biener <rguent...@suse.de>
>
> PR tree-optimization/110776
>
> gcc/ChangeLog:
>
> * tree-vect-stmts.cc (vectorizable_load): Always cost VMAT_ELEMENTWISE
> as scalar load.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/powerpc/pr110776.c: New test.
> ---
> gcc/testsuite/gcc.target/powerpc/pr110776.c | 22 +++++++++++++++++++++
> gcc/tree-vect-stmts.cc | 5 ++++-
> 2 files changed, 26 insertions(+), 1 deletion(-)
> create mode 100644 gcc/testsuite/gcc.target/powerpc/pr110776.c
>
> diff --git a/gcc/testsuite/gcc.target/powerpc/pr110776.c
> b/gcc/testsuite/gcc.target/powerpc/pr110776.c
> new file mode 100644
> index 00000000000..749159fd675
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/powerpc/pr110776.c
> @@ -0,0 +1,22 @@
> +/* { dg-require-effective-target powerpc_altivec_ok } */
> +/* { dg-options "-O2 -mdejagnu-cpu=power6 -maltivec" } */
> +
> +/* Verify there is no ICE. */
> +
> +int a;
> +long *b;
> +int
> +c ()
> +{
> + long e;
> + int d = 0;
> + for (long f; f; f++)
> + {
> + e = b[f * a];
> + if (e)
> + d = 1;
> + }
> + if (d)
> + for (;;)
> + ;
> +}
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index ed28fbdced3..09705200594 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -9840,7 +9840,10 @@ vectorizable_load (vec_info *vinfo,
> {
> if (costing_p)
> {
> - if (VECTOR_TYPE_P (ltype))
> + /* For VMAT_ELEMENTWISE, just cost it as scalar_load to
> + avoid ICE, see PR110776. */
> + if (VECTOR_TYPE_P (ltype)
> + && memory_access_type != VMAT_ELEMENTWISE)
> vect_get_load_cost (vinfo, stmt_info, 1,
> alignment_support_scheme,
> misalignment,
> false, &inside_cost, nullptr,
> cost_vec,
> --
> 2.39.1
