[PATCH][ARM] Enable auto-vectorization for copysignf

[Jiong Wang]

On 11/09/14 14:55, Jiong Wang wrote:
> On 11/09/14 14:43, Christophe Lyon wrote:
>> Hi Jiong,
>>
>> On 9 September 2014 12:59, Ramana Radhakrishnan
>> <ramana.gcc@googlemail.com> wrote:
>>> On Mon, Aug 18, 2014 at 11:31 AM, Jiong Wang <jiong.wang@arm.com> wrote:
>>>> this patch enable auto-vectorization for copysignf by using vector
>>>> bit selection instruction on arm32 when neon available.
>>>>
>> I've noticed that your new testcase fails (the scan-tree-dump-times
>> line), in the following cases:
>> * forcing -march=armv5t in RUNTESTFLAGS (targets
>> arm-none-linux-gnueabi and arm-none-linux-gnueabihf)
>> * target armeb-none-linux-gnueabihf
>>
>> You can have a look at:
>> http://cbuild.validation.linaro.org/build/cross-validation/gcc/trunk/215067/report-build-info.html
>>
>> If you go 1 level up at
>> http://cbuild.validation.linaro.org/build/cross-validation/gcc/trunk/215067/
>> you'll be able to browse into the per-target subdirs and get the .sum
>> files if you need them.
> Christophe,
>
>     the auto-test system is great!
>
>     the testcase only pass when both hardware & abi options meet requirement.
>
>     I tried to skip those environment where neon is not available by "dg-require-effective-target arm_neon_hw"
>
>     there maybe something not covered. I'll have a look.
>
>     thanks.

    the scan of

         "/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */

    is a little bit fragile, it needs -mfpu, -mfloat-abi both meet requirement, and thus cause
    trouble if the test environment has complicated options combinations like the Linaro test farm.

    currently, I haven't found any good way in Dejagnu to accurately detect what's options used.

    I was trying specify options like -march=armv7 -mfloat-abi=hard, and do compile test only,
    but even this, these options may be override by options specified explicitly by user, and thus
    pass those dejagnu check-*-target while fail on the later actual compile.

    the one way I can think of which is 100% safe on all test environment is:

      * keep the test as a "run" test only, as the correctness is important.
      * remove the scan of "vectorized 1 loops".

-- Jiong

>
> -- Jiong
>
>> Christophe.
>>
>>
>>>> for a simple testcase:
>>>>
>>>>     for (i = 0; i < N; i++)
>>>>       r[i] = __builtin_copysignf (a[i], b[i]);
>>>>
>>>>
>>>> assuming vector factor be 4, the generated instruction sequences is:
>>>>
>>>>           vmov.i32        q10, #2147483648  @ v4si
>>>> .L2:
>>>>           vld1.64 {d18-d19}, [ip:64]
>>>>           add     r3, r3, #16
>>>>           add     ip, ip, #16
>>>>           vldr    d16, [r3, #-16]
>>>>           vldr    d17, [r3, #-8]
>>>>           vbif    q8, q9, q10
>>>
>>> Ok.
>>>
>>> Ramana
>>>
>>>
>>>> thanks.
>>>>
>>>> gcc/
>>>>     * config/arm/arm.c (NEON_COPYSIGNF): New enum.
>>>>     (arm_init_neon_builtins): Support NEON_COPYSIGNF.
>>>>     (arm_builtin_vectorized_function): Likewise.
>>>>     * config/arm/arm_neon_builtins.def: New macro for copysignf.
>>>>     * config/arm/neon.md (neon_copysignf<mode>): New pattern for vector
>>>> copysignf.
>>>>
>>>> gcc/testsuite/
>>>>     * gcc.target/arm/vect-copysignf.c: New testcase.
>
>