[Bug tree-optimization/92645] Hand written vector code is 450 times slower when compiled with GCC compared to Clang

[Bug tree-optimization/92645] Hand written vector code is 450 times slower when compiled with GCC compared to Clang

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92645

--- Comment #23 from Richard Biener <rguenth at gcc dot gnu.org> ---
So the issue is we're both not doing enough and too much, the half way early
optimizations do confuse us later.  Another such opportunity would maybe
be:

   short unsigned int _950;
  _950 = BIT_FIELD_REF <_58, 16, 240>;
  _253 = (unsigned char) _950;

where this is the only use of _950.  It might be tempting to "optimize"
this into

  _253 = BIT_FIELD_REF <_58, 8, 240>;

forwprop does similar transforms for loads of complex and vector (though
the above is not a load but the transform would extend to loads as well).

[Bug tree-optimization/92645] Hand written vector code is 450 times slower when compiled with GCC compared to Clang

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92645

--- Comment #24 from Richard Biener <rguenth at gcc dot gnu.org> ---
Created attachment 49958
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49958&action=edit
unincluded GCC source

The GCC source no longer compiles due to missing changes in the x86 intrinsic
includes in the preprocessed source:

...
/aux/hubicka/trunk-install2/lib/gcc/x86_64-pc-linux-gnu/10.0.0/include/avx512vlbwintrin.h:
In function 'void _mm_mask_cvtsepi16_storeu_epi8(void*, __mmask8, __m128i)':
/aux/hubicka/trunk-install2/lib/gcc/x86_64-pc-linux-gnu/10.0.0/include/avx512vlbwintrin.h:258:38:
error: cannot convert '__v8qi*' to 'long long unsigned int*'
<built-in>: note:   initializing argument 1 of 'void
__builtin_ia32_pmovswb128mem_mask(long long unsigned int*, __vector(8) short
int, unsigned char)'
In file included from
/aux/hubicka/trunk-install2/lib/gcc/x86_64-pc-linux-gnu/10.0.0/include/immintrin.h:69,
                 from
/aux/hubicka/firefox-2019-2/gfx/skia/skia/src/opts/SkOpts_
...

attached unincluded source that can be compiled with trunk and GCC 10
when using -march=haswell

[Bug tree-optimization/92645] Hand written vector code is 450 times slower when compiled with GCC compared to Clang

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92645

--- Comment #25 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:3ddc18251a821fe69d6229abbf83d77284d2340a

commit r11-6644-g3ddc18251a821fe69d6229abbf83d77284d2340a
Author: Richard Biener <rguenther@suse.de>
Date:   Wed Jan 13 12:40:01 2021 +0100

    tree-optimization/92645 - improve SLP with existing vectors

    This improves SLP discovery in the face of existing vectors allowing
    punning of the vector shape (or even punning from an integer type).
    For punning from integer types this does not yet handle lane zero
    extraction being represented as conversion rather than BIT_FIELD_REF.

    2021-01-13  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/92645
            * tree-vect-slp.c (vect_build_slp_tree_1): Relax supported
            BIT_FIELD_REF argument.
            (vect_build_slp_tree_2): Record the desired vector type
            on the external vector def.
            (vectorizable_slp_permutation): Handle required punning
            of existing vector defs.

            * gcc.target/i386/pr92645-6.c: New testcase.

[Bug tree-optimization/92645] Hand written vector code is 450 times slower when compiled with GCC compared to Clang

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92645

--- Comment #26 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:285fa338b06b804e72997c4d876ecf08a9c083af

commit r11-6649-g285fa338b06b804e72997c4d876ecf08a9c083af
Author: Richard Biener <rguenther@suse.de>
Date:   Wed Jan 13 13:48:31 2021 +0100

    tree-optimization/92645 - avoid harmful early BIT_FIELD_REF
canonicalization

    This avoids canonicalizing BIT_FIELD_REF <T1> (a, <sz>, 0) to
    (T1)a on integer typed a.  This confuses the vectorizer SLP matching.

    With this delayed to after vector lowering the testcase in PR92645
    from Skia is now finally optimized to reasonable assembly.

    2021-01-13  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/92645
            * match.pd (BIT_FIELD_REF to conversion): Delay canonicalization
            until after vector lowering.

            * gcc.target/i386/pr92645-7.c: New testcase.
            * gcc.dg/tree-ssa/ssa-fre-54.c: Adjust.
            * gcc.dg/pr69047.c: Likewise.

[Bug tree-optimization/92645] Hand written vector code is 450 times slower when compiled with GCC compared to Clang

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92645

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|---                         |FIXED

--- Comment #27 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed.

[Bug tree-optimization/92645] Hand written vector code is 450 times slower when compiled with GCC compared to Clang

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92645

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |11.0