Re: [PATCH][i386] PR82941/PR82942 - Adding vzeroupper generation for SKX

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Tue, Nov 14, 2017 at 4:17 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
> On Tue, Nov 14, 2017 at 3:18 AM, Peryt, Sebastian
> <sebastian.peryt@intel.com> wrote:
>> I have updated tests and changelog according to Jakub's suggestions.
>> Please find attached v2 of my patch.
>>
>>
>> 14.11.2017  Sebastian Peryt  <sebastian.peryt@intel.com>
>>
>> gcc/
>>
>>         PR target/82941
>>         PR target/82942
>>         * config/i386/i386.c (pass_insert_vzeroupper): Modify gate condition
>>         to return true on Xeon and not on Xeon Phi.
>>         (ix86_check_avx256_register): Changed to ...
>>         (ix86_check_avx_upper_register): ... this. Add extra check for
>>         VALID_AVX512F_REG_OR_XI_MODE.
>>         (ix86_avx_u128_mode_needed): Changed
>>         ix86_check_avx256_register to ix86_check_avx_upper_register.
>>         (ix86_check_avx256_stores): Changed to ...
>>         (ix86_check_avx_upper_stores): ... this. Changed
>>         ix86_check_avx256_register to ix86_check_avx_upper_register.
>>         (ix86_avx_u128_mode_after): Changed
>>         avx_reg256_found to avx_upper_reg_found. Changed
>>         ix86_check_avx256_stores to ix86_check_avx_upper_stores.
>>         (ix86_avx_u128_mode_entry): Changed
>>         ix86_check_avx256_register to ix86_check_avx_upper_register.
>>         (ix86_avx_u128_mode_exit): Ditto.
>>         * config/i386/i386.h: (host_detect_local_cpu): New define.
>
> @@ -2497,7 +2497,7 @@ public:
>    /* opt_pass methods: */
>    virtual bool gate (function *)
>      {
> -      return TARGET_AVX && !TARGET_AVX512F
> +      return TARGET_AVX && !TARGET_AVX512PF && !TARGET_AVX512ER
>                                                 ^^^^^^^^^^^^^^^^^^^^^
>  Please remove  this.
>
> From glibc commit:
>
> commit 4cb334c4d6249686653137ec273d081371b3672d
> Author: H.J. Lu <hjl.tools@gmail.com>
> Date:   Tue Apr 18 14:01:45 2017 -0700
>
>     x86: Use AVX2 memcpy/memset on Skylake server [BZ #21396]
>
>     On Skylake server, AVX512 load/store instructions in memcpy/memset may
>     lead to lower CPU turbo frequency in certain situations.  Use of AVX2
>     in memcpy/memset has been observed to have improved overall performance
>     in many workloads due to the higher frequency.
>
>     Since AVX512ER is unique to Xeon Phi, this patch sets Prefer_No_AVX512
>     if AVX512ER isn't available so that AVX2 versions of memcpy/memset are
>     used on Skylake server.
>
> Only AVX512ER is really unique to Xeon Phi.
>
>        && TARGET_VZEROUPPER && flag_expensive_optimizations
>        && !optimize_size;
>      }

Just a thought, should we have a separate patch to add -mprefer-vzeroupper
to cover all bases in the future, like

  /* opt_pass methods: */
  virtual bool gate (function *)
    {
      return TARGET_AVX && (!TARGET_AVX512ER || TARGET_PREFER_VZEROUPPER)
             && TARGET_VZEROUPPER && flag_expensive_optimizations
             && !optimize_size;
    }

>> 14.11.2017  Sebastian Peryt  <sebastian.peryt@intel.com>
>>
>> gcc/testsuite/
>>
>>         PR target/82941
>>         PR target/82942
>>         * gcc.target/i386/pr82941-1.c: New test.
>>         * gcc.target/i386/pr82941-2.c: New test.
>>         * gcc.target/i386/pr82942-1.c: New test.
>>         * gcc.target/i386/pr82942-2.c: New test.
>>
>>
>> Thanks,
>> Sebastian
>>
>>> -----Original Message-----
>>> From: Jakub Jelinek [mailto:jakub@redhat.com]
>>> Sent: Tuesday, November 14, 2017 10:51 AM
>>> To: Peryt, Sebastian <sebastian.peryt@intel.com>
>>> Cc: gcc-patches@gcc.gnu.org; Uros Bizjak <ubizjak@gmail.com>; Kirill Yukhin
>>> <kirill.yukhin@gmail.com>; Lu, Hongjiu <hongjiu.lu@intel.com>
>>> Subject: Re: [PATCH][i386] PR82941/PR82942 - Adding vzeroupper generation
>>> for SKX
>>>
>>> On Tue, Nov 14, 2017 at 09:45:12AM +0000, Peryt, Sebastian wrote:
>>> > Hi,
>>> >
>>> > This patch fixes PR82941 and PR82942 by adding vzeroupper generation on
>>> SKX.
>>> > Bootstrapped and tested.
>>> >
>>> > 14.11.2017  Sebastian Peryt  <sebastian.peryt@intel.com>
>>> >
>>> > gcc/
>>>
>>> In that case the ChangeLog entry should list the PRs, i.e.
>>>       PR target/82941
>>>       PR target/82942
>>> >     * config/i386/i386.c (pass_insert_vzeroupper): Modify gate condition
>>> >     to return true on Xeon and not on Xeon Phi.
>>> >     (ix86_check_avx256_register): Changed to ...
>>> >     (ix86_check_avx_upper_register): ... this.
>>> >     (ix86_check_avx_upper_register): Add extra check for
>>> >     VALID_AVX512F_REG_OR_XI_MODE.
>>>
>>> The way this is usually written is instead:
>>>       (ix86_check_avx256_register): Changed to ...
>>>       (ix86_check_avx_upper_register): ... this.  Add extra check for
>>>       VALID_AVX512F_REG_OR_XI_MODE.
>>> i.e. don't duplicate the function name, just continue mentioning further changes.
>>>
>>> >     (ix86_avx_u128_mode_needed): Changed
>>> >     ix86_check_avx256_register to ix86_check_avx_upper_register.
>>> >     (ix86_check_avx256_stores): Changed to ...
>>> >     (ix86_check_avx_upper_stores): ... this.
>>> >     (ix86_check_avx_upper_stores): Changed
>>> >     ix86_check_avx256_register to ix86_check_avx_upper_register.
>>>
>>> Likewise.
>>>
>>> > gcc/testsuite/
>>> >     * gcc.target/i386/pr82941.c: New test.
>>> >     * gcc.target/i386/pr82942.c: New test.
>>>
>>> Shouldn't there be also a test that if using -march=knl and another one if using -
>>> mavx512f -mavx512er that we don't emit any vzeroupper?
>>>
>>>       Jakub
>
>
>
> --
> H.J.



-- 
H.J.