[RFC, vectorizer] Allow single element vector types for vector reduction operations

[Christophe Lyon]

On 30 August 2017 at 17:11, Jon Beniston <jon@beniston.com> wrote:
> Hi Richard,
>
>> I think the issue is that with TARGET_VECTOR_MODE_SUPPORTED_P false
>> for V1SI you'll get a SImode vector type and the
>>
>>  if (VECTOR_BOOLEAN_TYPE_P (type_in)
>>      || VECTOR_MODE_P (TYPE_MODE (type_in)))
>>
>>check fails.  Try changing that to || VECTOR_TYPE_P (type_in).
>>The else path should be hopefully dead ...
>>
>>> It looks to me like the other call sites of optab_for_tree_code which
>>> are passing optab_default are mostly places where GCC is sure it is
>>> not a shift operation.
>>> Given TYPE_IN is returned from get_vectype_for_scalar_type, is it
>>> safe to simply pass optab_vector in vect_pattern_recog_1?
>>
>>I guess so.  Can you also try the above?
>
> Changing VECTOR_MODE_P check into VECTOR_TYPE_P check also works for me, I
> verified the following patch could vectorize my dot product case and there
> is no regression on gcc tests on my port.
>
> Meanwhile x86-64 bootstraps OK and no regression on gcc/g++ test.
>
>  Does this look OK?
>
>
> gcc/
> 2017-08-30  Jon Beniston  <jon@beniston.com>
>             Richard Biener  <rguenther@suse.de>
>
> diff --git a/gcc/tree-vect-patterns.c b/gcc/tree-vect-patterns.c
> index cfdb72c..5ebeac2 100644
> --- a/gcc/tree-vect-patterns.c
> +++ b/gcc/tree-vect-patterns.c
> @@ -4150,7 +4150,7 @@ vect_pattern_recog_1 (vect_recog_func *recog_func,
>    loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
>
>    if (VECTOR_BOOLEAN_TYPE_P (type_in)
> -      || VECTOR_MODE_P (TYPE_MODE (type_in)))
> +      || VECTOR_TYPE_P (type_in))
>      {
>        /* No need to check target support (already checked by the pattern
>           recognition function).  */
> diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c
> index 013fb1f..fc62efb 100644
> --- a/gcc/tree-vect-stmts.c
> +++ b/gcc/tree-vect-stmts.c
> @@ -9098,7 +9098,8 @@ get_vectype_for_scalar_type_and_size (tree
> scalar_type, unsigned size)
>    else
>      simd_mode = mode_for_vector (inner_mode, size / nbytes);
>    nunits = GET_MODE_SIZE (simd_mode) / nbytes;
> -  if (nunits <= 1)
> +  /* NOTE: nunits == 1 is allowed to support single element vector types.
> */
> +  if (nunits < 1)
>      return NULL_TREE;
>
>    vectype = build_vector_type (scalar_type, nunits);
>
>

Hi,

This patch makes this test fail:
gcc.dg/tree-ssa/ssa-dom-cse-2.c scan-tree-dump optimized "return 28;"
on arm-none-linux-gnueabi when using the default fpu, or on
arm-none-linux-gnueabihf --with-fpu=vfp (as opposed to
--with-fpu=neon*)

Christophe