[Bug tree-optimization/97081] New: wrong code for rotate vectorization (x86 target)

[Bug tree-optimization/97081] New: wrong code for rotate vectorization (x86 target)

[igor at tachyum dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

            Bug ID: 97081
           Summary: wrong code for rotate vectorization (x86 target)
           Product: gcc
           Version: 10.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: igor at tachyum dot com
  Target Milestone: ---

Created attachment 49231
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49231&action=edit
testcase

incorrectly calculates left shift amount when breaking (a r>> sh) into (a <<
sh_l) | (a >> sh_r)

shr is calculated as (-sh & 31) while should be (-sh & 63)


        .file   "t.c"
        .text
        .p2align 4
        .globl  exec_VRORudi_i
        .type   exec_VRORudi_i, @function
exec_VRORudi_i:
.LFB0:
        .cfi_startproc
        endbr64
        andl    $63, %edx
        movdqu  (%rsi), %xmm1
        movl    %edx, %eax
        movd    %edx, %xmm0
        negl    %eax
        movdqa  %xmm1, %xmm3
        movl    %eax, %ecx
        psrlq   %xmm0, %xmm3
        andl    $31, %ecx
        movq    %rcx, %xmm2
        psllq   %xmm2, %xmm1
        por     %xmm3, %xmm1
        movups  %xmm1, (%rdi)
        movdqu  16(%rsi), %xmm1
        movdqa  %xmm1, %xmm3
        psrlq   %xmm0, %xmm1
        psllq   %xmm2, %xmm3
        por     %xmm3, %xmm1
        movups  %xmm1, 16(%rdi)
        ret
        .cfi_endproc

[Bug tree-optimization/97081] wrong code for rotate vectorization (x86 target)

[igor at tachyum dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

--- Comment #1 from Igor Shevlyakov <igor at tachyum dot com> ---
vect_recog_rotate_pattern()

seems to be the culprit

[Bug tree-optimization/97081] wrong code for rotate vectorization (x86 target)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2020-09-17
     Ever confirmed|0                           |1
             Target|                            |x86_64-*-* i?86-*-*
           Keywords|                            |needs-bisection
             Status|UNCONFIRMED                 |NEW

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Would be nice to have a executable testcase that aborts when miscompiled. 
Sofar I can confirm your observation.

[Bug tree-optimization/97081] wrong code for rotate vectorization (x86 target)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
             Status|NEW                         |ASSIGNED

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
The following aborts with vectorization enabled:

#define EXEC_ROR2(a, b, sz)     (a >> b) | (a << (64 - b))
#define TYPE unsigned long

void __attribute__((noipa))
exec_VRORudi_i(TYPE *__restrict__ pvd,
               TYPE *__restrict__ const pva, unsigned char IMM)
{
  unsigned char I2 = IMM & 63;

  for (unsigned i = 0; i < 4; i++)
    pvd[i] = EXEC_ROR2(pva[i], I2, 8);
}

int main()
{
  TYPE pvd[4], pva[4] = { 0x0102030405060708, 0x0102030405060708,
0x0102030405060708, 0x0102030405060708 };
  exec_VRORudi_i (pvd, pva, 7);
  if (pvd[0] != 0x10020406080a0c0e)
    __builtin_abort ();
  return 0;
}


creates

  patt_9 = -_25;
  patt_8 = patt_9 & 31;
  vect_patt_27.7_28 = vect__4.6_5 >> _25;
  vect_patt_22.8_29 = vect__4.6_5 << patt_8;
  vect_patt_23.9_30 = vect_patt_27.7_28 | vect_patt_22.8_29;
  _7 = _4 r>> _25;


I have a patch.

[Bug tree-optimization/97081] [8/9/10/11 Regression] wrong code for rotate vectorization (x86 target)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to fail|                            |10.1.0, 8.4.0, 9.3.0
           Priority|P3                          |P2
      Known to work|                            |7.5.0, 8.3.0, 9.2.0
   Target Milestone|---                         |8.5
            Summary|wrong code for rotate       |[8/9/10/11 Regression]
                   |vectorization (x86 target)  |wrong code for rotate
                   |                            |vectorization (x86 target)

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Interestingly GCC 7 created the correct

 patt_18 = patt_19 & 63;

but something broke it during GCC 10 development and was even backported.

[Bug tree-optimization/97081] [8/9/10/11 Regression] wrong code for rotate vectorization (x86 target)

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Started with my r10-5358-gf5c03155aa77bdb065fed3c5a0031db45a850493 PR92723 fix,
which indeed has been backported to 9 and 8.

[Bug tree-optimization/97081] [8/9/10/11 Regression] wrong code for rotate vectorization (x86 target)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

--- Comment #6 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:9c9b88fdcff3520b2c4fb520c5d3b422eaa9a72f

commit r11-3282-g9c9b88fdcff3520b2c4fb520c5d3b422eaa9a72f
Author: Richard Biener <rguenther@suse.de>
Date:   Fri Sep 18 13:36:24 2020 +0200

    tree-optimization/97081 - fix wrong-code with vectorized shift

    This corrects the mask for creation of x << s | x >> (-x & mask)
    from a rotate x <<r s to use the precision of x.

    2020-09-18  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/97081
            * tree-vect-patterns.c (vect_recog_rotate_pattern): Use the
            precision of the shifted operand to determine the mask.

            * gcc.dg/vect/pr97081.c: New testcase.

[Bug tree-optimization/97081] [8/9/10 Regression] wrong code for rotate vectorization (x86 target)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[8/9/10/11 Regression]      |[8/9/10 Regression] wrong
                   |wrong code for rotate       |code for rotate
                   |vectorization (x86 target)  |vectorization (x86 target)
           Keywords|needs-bisection             |
      Known to work|                            |11.0

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed on trunk sofar.

[Bug tree-optimization/97081] [8/9/10 Regression] wrong code for rotate vectorization (x86 target)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

--- Comment #8 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jakub Jelinek <jakub@gcc.gnu.org>:

https://gcc.gnu.org/g:3d3fe967b0961cb59f5df03ae2a55d83dc4bbd34

commit r11-3283-g3d3fe967b0961cb59f5df03ae2a55d83dc4bbd34
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Fri Sep 18 15:05:53 2020 +0200

    testsuite: add another test for the rotate vectorization miscompilation

    This time with short and char where the used mask used to be larger
    than it should have been.

    2020-09-18  Jakub Jelinek  <jakub@redhat.com>

            PR tree-optimization/97081
            * gcc.dg/vect/pr97081-2.c: New test.

[Bug tree-optimization/97081] [8/9/10 Regression] wrong code for rotate vectorization (x86 target)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

--- Comment #9 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-10 branch has been updated by Richard Biener
<rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:ca84557f3024c75ede850007099ec9b2c19e9f8a

commit r10-8948-gca84557f3024c75ede850007099ec9b2c19e9f8a
Author: Richard Biener <rguenther@suse.de>
Date:   Fri Sep 18 13:36:24 2020 +0200

    tree-optimization/97081 - fix wrong-code with vectorized shift

    This corrects the mask for creation of x << s | x >> (-x & mask)
    from a rotate x <<r s to use the precision of x.

    2020-09-18  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/97081
            * tree-vect-patterns.c (vect_recog_rotate_pattern): Use the
            precision of the shifted operand to determine the mask.

            * gcc.dg/vect/pr97081.c: New testcase.

    (cherry picked from commit 9c9b88fdcff3520b2c4fb520c5d3b422eaa9a72f)

[Bug tree-optimization/97081] [8/9/10 Regression] wrong code for rotate vectorization (x86 target)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

--- Comment #10 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-10 branch has been updated by Richard Biener
<rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:a78cd759754c92cecbf235ac9b447dcdff6c6e2f

commit r10-8949-ga78cd759754c92cecbf235ac9b447dcdff6c6e2f
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Fri Sep 18 15:05:53 2020 +0200

    testsuite: add another test for the rotate vectorization miscompilation

    This time with short and char where the used mask used to be larger
    than it should have been.

    2020-09-18  Jakub Jelinek  <jakub@redhat.com>

            PR tree-optimization/97081
            * gcc.dg/vect/pr97081-2.c: New test.

    (cherry picked from commit 3d3fe967b0961cb59f5df03ae2a55d83dc4bbd34)

[Bug tree-optimization/97081] [8/9 Regression] wrong code for rotate vectorization (x86 target)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

--- Comment #11 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-9 branch has been updated by Richard Biener
<rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:86b25a1a5e1956c30fe7eaee80ebf719b759d631

commit r9-9089-g86b25a1a5e1956c30fe7eaee80ebf719b759d631
Author: Richard Biener <rguenther@suse.de>
Date:   Fri Sep 18 13:36:24 2020 +0200

    tree-optimization/97081 - fix wrong-code with vectorized shift

    This corrects the mask for creation of x << s | x >> (-x & mask)
    from a rotate x <<r s to use the precision of x.

    2020-09-18  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/97081
            * tree-vect-patterns.c (vect_recog_rotate_pattern): Use the
            precision of the shifted operand to determine the mask.

            * gcc.dg/vect/pr97081.c: New testcase.

    (cherry picked from commit 9c9b88fdcff3520b2c4fb520c5d3b422eaa9a72f)

[Bug tree-optimization/97081] [8/9 Regression] wrong code for rotate vectorization (x86 target)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

--- Comment #12 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-9 branch has been updated by Richard Biener
<rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:f75352bdb139c40abb400746c03d9242a0c30bd6

commit r9-9090-gf75352bdb139c40abb400746c03d9242a0c30bd6
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Fri Sep 18 15:05:53 2020 +0200

    testsuite: add another test for the rotate vectorization miscompilation

    This time with short and char where the used mask used to be larger
    than it should have been.

    2020-09-18  Jakub Jelinek  <jakub@redhat.com>

            PR tree-optimization/97081
            * gcc.dg/vect/pr97081-2.c: New test.

    (cherry picked from commit 3d3fe967b0961cb59f5df03ae2a55d83dc4bbd34)

[Bug tree-optimization/97081] [8 Regression] wrong code for rotate vectorization (x86 target)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

--- Comment #13 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-8 branch has been updated by Richard Biener
<rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:16e2f38167eb90a4dc977b806dcc0dc9011cc456

commit r8-10801-g16e2f38167eb90a4dc977b806dcc0dc9011cc456
Author: Richard Biener <rguenther@suse.de>
Date:   Fri Sep 18 13:36:24 2020 +0200

    tree-optimization/97081 - fix wrong-code with vectorized shift

    This corrects the mask for creation of x << s | x >> (-x & mask)
    from a rotate x <<r s to use the precision of x.

    2020-09-18  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/97081
            * tree-vect-patterns.c (vect_recog_rotate_pattern): Use the
            precision of the shifted operand to determine the mask.

            * gcc.dg/vect/pr97081.c: New testcase.

    (cherry picked from commit 86b25a1a5e1956c30fe7eaee80ebf719b759d631)

[Bug tree-optimization/97081] [8 Regression] wrong code for rotate vectorization (x86 target)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

--- Comment #14 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-8 branch has been updated by Richard Biener
<rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:e407585a3ae48a25f031450565cf2b657d431cee

commit r8-10802-ge407585a3ae48a25f031450565cf2b657d431cee
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Fri Sep 18 15:05:53 2020 +0200

    testsuite: add another test for the rotate vectorization miscompilation

    This time with short and char where the used mask used to be larger
    than it should have been.

    2020-09-18  Jakub Jelinek  <jakub@redhat.com>

            PR tree-optimization/97081
            * gcc.dg/vect/pr97081-2.c: New test.

    (cherry picked from commit f75352bdb139c40abb400746c03d9242a0c30bd6)

[Bug tree-optimization/97081] [8 Regression] wrong code for rotate vectorization (x86 target)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97081

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to work|                            |8.4.1
             Status|ASSIGNED                    |RESOLVED
         Resolution|---                         |FIXED

--- Comment #15 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed.