On Fri, Sep 22, 2017 at 10:15 AM, Richard Sandiford
<richard.sandif...@linaro.org> wrote:
> Richard Biener <richard.guent...@gmail.com> writes:
>> On Thu, Sep 14, 2017 at 4:05 PM, Richard Sandiford
>> <richard.sandif...@linaro.org> wrote:
>>> Richard Biener <richard.guent...@gmail.com> writes:
>>>> On Thu, Sep 14, 2017 at 1:23 PM, Richard Sandiford
>>>> <richard.sandif...@linaro.org> wrote:
>>>>> This patch adds a helper function for getting the number of
>>>>> bytes accessed by a scalar data reference, which helps when general
>>>>> modes have a variable size.
>>>>>
>>>>> Tested on aarch64-linux-gnu, x86_64-linux-gnu and powerpc64le-linux-gnu.
>>>>> OK to install?
>>>>
>>>> Can you put it into tree-data-ref.h?
>>>
>>> The idea (which I forgot to say) was to collect the uses within the
>>> vectoriser into one place so that we can assert in only one place
>>> that the reference is constant-sized.
>>>
>>> A general data_reference can be variable-sized, so the guarantees
>>> wouldn't be the same elsewhere.
>>>
>>> Would putting it in tree-vectorizer.h be OK?
>>
>> Maybe name it vect_get_scalar_dr_size then?
>>
>> Ok with that.
>
> As discussed on IRC, I tried also using SCALAR_TYPE_MODE instead of
> TYPE_MODE, to go with the new function name, but it turns out that
> the mode can be a single-element vector instead of a "true" scalar.
>
> This version of the patch sticks with vect_get_scalar_dr_size as
> the name, but adds a comment to say that "scalar" includes "scalar
> equivalent". Since SCALAR_TYPE_MODE is too strong an assertion,
> I went with tree_to_uhwi instead.
>
> As also discussed on IRC, it might be possible in current sources
> to use SCALAR_TYPE_MODE (TREE_TYPE (STMT_VINFO_VECTYPE ...)) instead.
> But SVE has extending loads and truncating stores, for which the DR size
> will be different from the vector element size, so if we switch it now,
> we'd have to switch back to the DR size again later.
>
> (Sorry again for labouring such a simple change!)
>
> Tested as before. OK to install?
Ok.
Richard.
> Thanks,
> Richard
>
>
> 2017-09-22 Richard Sandiford <richard.sandif...@linaro.org>
> Alan Hayward <alan.hayw...@arm.com>
> David Sherwood <david.sherw...@arm.com>
>
> gcc/
> * tree-vectorizer.h (vect_get_scalar_dr_size): New function.
> * tree-vect-data-refs.c (vect_update_misalignment_for_peel): Use it.
> (vect_enhance_data_refs_alignment): Likewise.
>
> Index: gcc/tree-vectorizer.h
> ===================================================================
> --- gcc/tree-vectorizer.h 2017-09-18 16:49:51.902170781 +0100
> +++ gcc/tree-vectorizer.h 2017-09-22 08:58:24.308199570 +0100
> @@ -1095,6 +1095,19 @@ vect_get_num_copies (loop_vec_info loop_
> / TYPE_VECTOR_SUBPARTS (vectype));
> }
>
> +/* Return the size of the value accessed by unvectorized data reference DR.
> + This is only valid once STMT_VINFO_VECTYPE has been calculated for the
> + associated gimple statement, since that guarantees that DR accesses
> + either a scalar or a scalar equivalent. ("Scalar equivalent" here
> + includes things like V1SI, which can be vectorized in the same way
> + as a plain SI.) */
> +
> +inline unsigned int
> +vect_get_scalar_dr_size (struct data_reference *dr)
> +{
> + return tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr))));
> +}
> +
> /* Source location */
> extern source_location vect_location;
>
> Index: gcc/tree-vect-data-refs.c
> ===================================================================
> --- gcc/tree-vect-data-refs.c 2017-09-18 16:42:00.553203578 +0100
> +++ gcc/tree-vect-data-refs.c 2017-09-22 08:58:24.308199570 +0100
> @@ -955,7 +955,6 @@ vect_compute_data_ref_alignment (struct
> return true;
> }
>
> -
> /* Function vect_update_misalignment_for_peel.
> Sets DR's misalignment
> - to 0 if it has the same alignment as DR_PEEL,
> @@ -975,8 +974,8 @@ vect_update_misalignment_for_peel (struc
> unsigned int i;
> vec<dr_p> same_aligned_drs;
> struct data_reference *current_dr;
> - int dr_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr))));
> - int dr_peel_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF
> (dr_peel))));
> + int dr_size = vect_get_scalar_dr_size (dr);
> + int dr_peel_size = vect_get_scalar_dr_size (dr_peel);
> stmt_vec_info stmt_info = vinfo_for_stmt (DR_STMT (dr));
> stmt_vec_info peel_stmt_info = vinfo_for_stmt (DR_STMT (dr_peel));
>
> @@ -1664,8 +1663,7 @@ vect_enhance_data_refs_alignment (loop_v
>
> vectype = STMT_VINFO_VECTYPE (stmt_info);
> nelements = TYPE_VECTOR_SUBPARTS (vectype);
> - mis = DR_MISALIGNMENT (dr) / GET_MODE_SIZE (TYPE_MODE (
> - TREE_TYPE (DR_REF (dr))));
> + mis = DR_MISALIGNMENT (dr) / vect_get_scalar_dr_size (dr);
> if (DR_MISALIGNMENT (dr) != 0)
> npeel_tmp = (negative ? (mis - nelements)
> : (nelements - mis)) & (nelements - 1);
> @@ -1937,8 +1935,7 @@ vect_enhance_data_refs_alignment (loop_v
> updating DR_MISALIGNMENT values. The peeling factor is the
> vectorization factor minus the misalignment as an element
> count. */
> - mis = DR_MISALIGNMENT (dr0);
> - mis /= GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr0))));
> + mis = DR_MISALIGNMENT (dr0) / vect_get_scalar_dr_size (dr0);
> npeel = ((negative ? mis - nelements : nelements - mis)
> & (nelements - 1));
> }
