Re: [PATCH, PR52252] Alternative way of vectorization for load groups of size 2 and 3.

[Richard Biener <richard dot guenther at gmail dot com>]

On Thu, Jun 12, 2014 at 6:04 AM, Evgeny Stupachenko <evstupac@gmail.com> wrote:
> Testing finished. No new regressions.
> Is the following patch ok?

+  if (targetm.sched.reassociation_width (VEC_PERM_EXPR, mode) > 1 ||
+      !vect_shift_permute_load_chain (dr_chain, size, stmt, gsi,
&result_chain))

||s and &&s go to the next line.

I miss testcases that make sure the vectorizer/backend code-paths are
both exercised.  Put them in gcc.target/i386 and provide an appropriate
-march.

The vectorizer changes are ok with the above fixed, I defer to backend
maintainers for the i386 changes.

Richard.

> 2014-06-11  Evgeny Stupachenko  <evstupac@gmail.com>
>
>         * config/i386/i386.c (ix86_reassociation_width): Add alternative for
>         vector case.
>         * config/i386/i386.h (TARGET_VECTOR_PARALLEL_EXECUTION): New.
>         * config/i386/x86-tune.def (X86_TUNE_VECTOR_PARALLEL_EXECUTION): New.
>         * tree-vect-data-refs.c (vect_shift_permute_load_chain): New.
>         Introduces alternative way of loads group permutaions.
>         (vect_transform_grouped_load): Try alternative way of permutations.
>
> Thanks,
> Evgeny
>
> On Tue, Jun 10, 2014 at 4:43 PM, Evgeny Stupachenko <evstupac@gmail.com> wrote:
>> ix86_reassociation_width checks INTEGRAL_MODE_P and FLOAT_MODE_P which
>> include vector mode.
>> I'll try to separate this into scalar and vector part, but it will
>> require more testing (under the testing now).
>> What about the rest of the patch?
>>
>> Thanks,
>> Evgeny
>>
>> On Thu, Jun 5, 2014 at 3:54 PM, Ramana Radhakrishnan
>> <ramana.radhakrishnan@arm.com> wrote:
>>> On 06/05/14 12:43, Evgeny Stupachenko wrote:
>>>>
>>>> New hook is related to vector instructions only. Vector instructions
>>>> could be sequential in pipeline, but scalar - parallel. For x86
>>>> architectures TARGET_SCHED_REASSOC_WIDTH does not give required
>>>> differentiation.
>>>> General hooks could be potentially reused in other algorithms/by other
>>>> architectures.
>>>
>>>
>>> It already takes a "mode" argument. Couldn't you use a vector mode to work
>>> this out ?
>>>
>>> If it is not enough then please be more specific about the documentation of
>>> this hook about where it is useful so that it's easy for people reading the
>>> documentation to understand at a glance what purpose it serves.
>>>
>>>
>>> Ramana
>>>
>>>
>>>>
>>>> Thanks,
>>>> Evgeny
>>>>
>>>> On Thu, Jun 5, 2014 at 2:04 PM, Ramana Radhakrishnan
>>>> <ramana.gcc@googlemail.com> wrote:
>>>>>
>>>>> On Wed, May 28, 2014 at 2:09 PM, Evgeny Stupachenko <evstupac@gmail.com>
>>>>> wrote:
>>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> The patch introduces alternative way of permutations for load groups
>>>>>> of size 2 and 3 which should be faster on architectures with low
>>>>>> parallelism.
>>>>>> The patch gives 2 times gain on Silvermont to the test from PR52252
>>>>>> (in addition to already committed 3 times gain).
>>>>>>
>>>>>> Patch passes bootstrap on x86. Make check is in progress.
>>>>>
>>>>>
>>>>> Why do we need a new hook ? Can't you derive this information from
>>>>> something which is equally badly named TARGET_SCHED_REASSOC_WIDTH
>>>>> though used in the reassociation logic but also serves a similar
>>>>> purpose ?
>>>>>
>>>>> Also the documentation of this hook is incomplete at best and wrong at
>>>>> worst as this is not applied everywhere in the vectorizer but just for
>>>>> this special case for load store permuting. Implying this is useful
>>>>> everywhere in the vectorizer does not appear to be correct.
>>>>>
>>>>> regards
>>>>> Ramana
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>>
>>>>>> ChangeLog:
>>>>>>
>>>>>> 2014-05-28  Evgeny Stupachenko  <evstupac@gmail.com>
>>>>>>
>>>>>>          * config/i386/i386.c (ix86_have_vector_parallel_execution):
>>>>>> New.
>>>>>>          (TARGET_VECTORIZE_HAVE_VECTOR_PARALLEL_EXECUTION): New.
>>>>>>          * config/i386/i386.h (TARGET_VECTOR_PARALLEL_EXECUTION): New.
>>>>>>          * config/i386/x86-tune.def
>>>>>> (X86_TUNE_VECTOR_PARALLEL_EXECUTION): New.
>>>>>>          * target.def (have_vector_parallel_execution): New.
>>>>>>          * doc/tm.texi.in (have_vector_parallel_execution)): New.
>>>>>>          * doc/tm.texi: Regenerate.
>>>>>>          * targhooks.c (default_have_vector_parallel_execution): New.
>>>>>>          * tree-vect-data-refs.c (vect_shift_permute_load_chain): New.
>>>>>>          Introduces alternative way of loads group permutaions.
>>>>>>          (vect_transform_grouped_load): Try alternative way of
>>>>>> permutaions.
>>>>>>
>>>>>> Evgeny
>>>>
>>>>
>>>