Re: PATCH: PR target/46519: Missing vzeroupper

[Richard Guenther <richard.guenther@gmail.com>]

On Sat, Nov 20, 2010 at 12:31 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
> On Fri, Nov 19, 2010 at 2:48 PM, Richard Guenther
> <richard.guenther@gmail.com> wrote:
>> On Fri, Nov 19, 2010 at 10:30 PM, H.J. Lu <hjl.tools@gmail.com> wrote:
>>> On Thu, Nov 18, 2010 at 1:11 AM, Uros Bizjak <ubizjak@gmail.com> wrote:
>>>> On Thu, Nov 18, 2010 at 12:36 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
>>>>
>>>>> Here is the patch for
>>>>>
>>>>> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46519
>>>>>
>>>>> We have 2 blocks pointing to each others. This patch first scans
>>>>> all blocks without moving vzeroupper so that we can have accurate
>>>>> information about upper 128bits at block entry.
>>>>
>>>> This introduces another insn scanning pass, almost the same as
>>>> existing vzeroupper pass (modulo CALL_INSN/JUMP_INSN handling).
>>>>
>>>> So, if I understand correctly:
>>>> - The patch removes the detection if the function ever touches AVX registers.
>>>> - Due to this, all call_insn RTXes have to be decorated with
>>>> CALL_NEEDS_VZEROUPPER.
>>>> - A new pre-pass is required that scans all functions in order to
>>>> detect functions with live AVX registers at exit, and at the same time
>>>> marks the functions that *do not* use AVX registers.
>>>> - Existing pass then re-scans everything to again detect functions
>>>> with live AVX registers at exit and handles vzeroupper emission.
>>>>
>>>> I don't think this approach is acceptable. Maybe a LCM infrastructure
>>>> can be used to handle this case?
>>>>
>>>
>>> Here is the rewrite of the vzeroupper optimization pass.
>>> To avoid circular dependency, it has 2 passes.  It
>>> delays the circular dependency to the second pass
>>> and avoid rescan as much as possible.
>>>
>>> I compared the bootstrap times with/wthout this patch
>>> on 64bit Sandy Bridge with multilib and --with-fpmath=avx.
>>> I enabled c,c++,fortran,java,lto,objc
>>>
>>> Without patch:
>>>
>>> 12378.70user 573.02system 41:54.21elapsed 515%CPU
>>>
>>> With patch
>>>
>>> 12580.56user 578.07system 42:25.41elapsed 516%CPU
>>>
>>> The overhead is about 1.6%.
>>
>> That's a quite big overhead for something that doesn't use FP
>> math (and thus no AVX).
>
> AVX256 vector insns are independent of FP math.  They can be
> generated by vectorizer as well as loop unroll.  We can limit
> it to -O2 or -O3 if overhead is a big concern.

Limiting it to -fexpensive-optimizations would be a good start.  Btw,
how is code-size affected?  Does it make sense to disable it when
optimizing a function for size?  As it affects performance of callees
whether the caller is optimized for size or speed probably isn't the
best thing to check.

Richard.

> H.J.
> ---
>> Richard.
>>
>>>
>>> --
>>> H.J.
>>> ---
>>> gcc/
>>>
>>> 2010-11-19  H.J. Lu  <hongjiu.lu@intel.com>
>>>
>>>        PR target/46519
>>>        * config/i386/i386.c (upper_128bits_state): New.
>>>        (block_info_def): Remove upper_128bits_set and done.  Add state,
>>>        referenced, count, processed and rescanned.
>>>        (check_avx256_stores): Updated.
>>>        (move_or_delete_vzeroupper_2): Updated. Handle deleted BB_END.
>>>        Call note_stores only if needed.  Set referenced and count.
>>>        (move_or_delete_vzeroupper_1): Updated.  Set rescan_vzeroupper_p.
>>>        (rescan_move_or_delete_vzeroupper): New.
>>>        (move_or_delete_vzeroupper):  Process and rescan all all basic
>>>        blocks instead of predecessor blocks of all exit points.
>>>        (use_avx256_p): Removed.
>>>        (init_cumulative_args): Don't set use_avx256_p.
>>>        (ix86_function_arg): Likewise.
>>>        (ix86_expand_move): Likewise.
>>>        (ix86_expand_vector_move_misalign): Likewise.
>>>        (ix86_local_alignment): Likewise.
>>>        (ix86_minimum_alignment): Likewise.
>>>        (ix86_expand_epilogue): Don't check use_avx256_p when generating
>>>        vzeroupper.
>>>        (ix86_expand_call): Likewise.
>>>
>>>        * config/i386/i386.h (machine_function): Remove use_vzeroupper_p
>>>        and use_avx256_p.  Add rescan_vzeroupper_p.
>>>
>>> gcc/testsuite/
>>>
>>> 2010-11-17  H.J. Lu  <hongjiu.lu@intel.com>
>>>
>>>        PR target/46519
>>>        * gcc.target/i386/avx-vzeroupper-10.c: Expect no avx_vzeroupper.
>>>        * gcc.target/i386/avx-vzeroupper-11.c: Likewise.
>>>
>>>        * gcc.target/i386/avx-vzeroupper-20.c: New.
>>>        * gcc.target/i386/avx-vzeroupper-21.c: Likewise.
>>>        * gcc.target/i386/avx-vzeroupper-22.c: Likewise.
>>>        * gcc.target/i386/avx-vzeroupper-23.c: Likewise.
>>>        * gcc.target/i386/avx-vzeroupper-24.c: Likewise.
>>>
>>
>
>
>
> --
> H.J.
>