[Bug tree-optimization/103237] New: wrong code with -ftree-vectorize at -O1 on x86_64-linux-gnu

[Bug tree-optimization/103237] New: wrong code with -ftree-vectorize at -O1 on x86_64-linux-gnu

[haoxintu at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103237

            Bug ID: 103237
           Summary: wrong code with -ftree-vectorize at -O1 on
                    x86_64-linux-gnu
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: haoxintu at gmail dot com
  Target Milestone: ---

Hi all.

This seems a long latent bug existed after GCC-10.1.

$cat small.c
#include <stdint.h>
#include <stdio.h>
struct S0 {
  uint16_t a;
};
union U1 {
  struct S0 b;
};
int32_t g1[1] = {0};
uint32_t g2[][10][1] = {
    {}, {}, {}, {}, {}, {}, {{}, {}, {}, {}, {}, {}, {4294967295UL}}};
union U1 g3;
static int32_t func_1() {
  int32_t *l_1 = &g1[0];
  uint32_t l_2 ;
  for (g3.b.a = (27); (g3.b.a != 45);
       g3.b.a = g3.b.a + 9) {
    for (l_2 = 0; (l_2 <= 3); l_2 += 1) {
      uint32_t *l_3 = &g2[6][6][0];
      *l_1 = *l_3 ^= ((0 == 0) << 0);
    }
  }
  return 0;
}
int main() {
  func_1();
  for (int i = 0; i < 1; i++)
    printf("%d\n", g1[i]);
  return 0;
}


$gcc -O2 -ftree-vectorize small.c ; ./a.out
-1

$gcc -O1 -ftree-vectorize small.c ; ./a.out
0

$gcc -O1 small.c ; ./a.out
-1

$gcc -v
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/home/haoxin/haoxin-data/compilers/gcc/build/libexec/gcc/x86_64-pc-linux-gnu/12.0.0/lto-wrapper
Target: x86_64-pc-linux-gnu
Configured with:../configure
--prefix=/home/haoxin/haoxin-data/compilers/gcc/build/ --enable-bootstrap
--enable-checking=release --enable-multilib --enable-languages=c,c++,lto
--no-create --no-recursion
Thread model: posix
Supported LTO compression algorithms: zlib
gcc version 12.0.0 20211114 (experimental) (GCC) 

Can also check here: https://godbolt.org/z/ncar1aWYM

Thanks,
Haoxin

[Bug tree-optimization/103237] [10/11/12 Regression] wrong code with -ftree-vectorize at -O1 on x86_64-linux-gnu

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103237

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
            Summary|wrong code with             |[10/11/12 Regression] wrong
                   |-ftree-vectorize at -O1 on  |code with -ftree-vectorize
                   |x86_64-linux-gnu            |at -O1 on x86_64-linux-gnu
      Known to fail|                            |10.1.0, 10.3.0, 11.2.0
   Last reconfirmed|                            |2021-11-14
             Status|UNCONFIRMED                 |NEW
      Known to work|                            |9.1.0, 9.4.0
   Target Milestone|---                         |11.3

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed.

[Bug tree-optimization/103237] [10/11/12 Regression] wrong code with -ftree-vectorize at -O1 on x86_64-linux-gnu since r10-1361-g9f962469cabc7fdc

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103237

Martin Liška <marxin at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[10/11/12 Regression] wrong |[10/11/12 Regression] wrong
                   |code with -ftree-vectorize  |code with -ftree-vectorize
                   |at -O1 on x86_64-linux-gnu  |at -O1 on x86_64-linux-gnu
                   |                            |since
                   |                            |r10-1361-g9f962469cabc7fdc
                 CC|                            |marxin at gcc dot gnu.org,
                   |                            |uros at gcc dot gnu.org

--- Comment #2 from Martin Liška <marxin at gcc dot gnu.org> ---
Started with r10-1361-g9f962469cabc7fdc.

[Bug tree-optimization/103237] [10/11/12 Regression] wrong code with -ftree-vectorize at -O1 on x86_64-linux-gnu since r10-1361-g9f962469cabc7fdc

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103237

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
   Target Milestone|11.3                        |10.4
             Target|                            |x86_64-*-* i?86-*-*
           Priority|P3                          |P2

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Likely latent w/o this rev. since we optimize at GIMPLE level to

  <bb 2> [local count: 536870923]:
  g2__lsm.6_23 = MEM[(uint32_t *)&g2 + 264B];
  MEM[(uint32_t *)&g2 + 264B] = g2__lsm.6_23;
  MEM[(int32_t *)&g1] = 0;
  g3.b.a = 45;
  printf ("%d\n", 0);
  return 0;

I will eventually have a more closer look.

[Bug tree-optimization/103237] [10/11/12 Regression] wrong code with -ftree-vectorize at -O1 on x86_64-linux-gnu since r10-1361-g9f962469cabc7fdc

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103237

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Reduced testcase where we vectorize the outer loop

int g1;
unsigned int g2 = 4294967295U;
static void __attribute__((noipa)) func_1()
{
  int *l_1 = &g1;
  for (int g3a = 0; g3a != 2; g3a++)
    for (int l_2 = 0; l_2 <= 3; l_2++)
      {
        unsigned int *l_3 = &g2;
        *l_1 = *l_3 ^= 1;
      }
}
int main()
{
  func_1();
  if (g1 != -1)
    __builtin_abort ();
  return 0;
}

[Bug tree-optimization/103237] [10/11/12 Regression] wrong code with -ftree-vectorize at -O1 on x86_64-linux-gnu since r10-1361-g9f962469cabc7fdc

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103237

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |needs-bisection
      Known to work|9.1.0, 9.4.0                |
   Target Milestone|10.4                        |9.5
      Known to fail|                            |7.5.0, 9.4.1

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
And the following fails with GCC 9 and 7 as well (using SSE vectorization).

int g1;
unsigned int g2 = 4294967295U;
static void __attribute__((noipa)) func_1()
{
  int *l_1 = &g1;
  for (int g3a = 0; g3a != 4; g3a++)
    for (int l_2 = 0; l_2 <= 3; l_2++)
      {
        unsigned int *l_3 = &g2;
        *l_1 = *l_3 ^= 1;
      }
}
int main()
{
  func_1();
  if (g1 != -1)
    __builtin_abort ();
  return 0;
}

[Bug tree-optimization/103237] [10/11/12 Regression] wrong code with -ftree-vectorize at -O1 on x86_64-linux-gnu

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103237

Martin Liška <marxin at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |alan.hayward at arm dot com

--- Comment #6 from Martin Liška <marxin at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #5)
> And the following fails with GCC 9 and 7 as well (using SSE vectorization).
> 
> int g1;
> unsigned int g2 = 4294967295U;
> static void __attribute__((noipa)) func_1()
> {
>   int *l_1 = &g1;
>   for (int g3a = 0; g3a != 4; g3a++)
>     for (int l_2 = 0; l_2 <= 3; l_2++)
>       {
>         unsigned int *l_3 = &g2;
>         *l_1 = *l_3 ^= 1;
>       }
> }
> int main()
> {
>   func_1();
>   if (g1 != -1)
>     __builtin_abort ();
>   return 0;
> }

This started to fail with r7-1219-gb28ead45fe630b9e.

[Bug tree-optimization/103237] [10/11/12 Regression] wrong code with -ftree-vectorize at -O1 on x86_64-linux-gnu starting with r7-1219-gb28ead45fe630b9e

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103237

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
So what we get is a double reduction with two loop-closed PHIs on the inner
loop where one is not classified as double-reduction (because we work from the
outer loop PHI to do that only).

I think this case is simply not well supported and the easiest is to disallow
it when analyzing the inner loop (because then we run into that duplicate LC
PHI).

Testing a patch to do that.

[Bug tree-optimization/103237] [10/11/12 Regression] wrong code with -ftree-vectorize at -O1 on x86_64-linux-gnu starting with r7-1219-gb28ead45fe630b9e

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103237

--- Comment #8 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:220bd61874cf114667b44f9ded76ed0639eb278b

commit r12-5266-g220bd61874cf114667b44f9ded76ed0639eb278b
Author: Richard Biener <rguenther@suse.de>
Date:   Mon Nov 15 11:37:56 2021 +0100

    tree-optimization/103237 - avoid vectorizing unhandled double reductions

    Double reductions which have multiple LC PHIs in the inner loop
    are not handled correctly during transformation since those PHIs
    are not properly classified as reduction.  The following disables
    vectorizing them.

    2021-11-15  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/103237
            * tree-vect-loop.c (vect_is_simple_reduction): Fail for
            double reductions with multiple inner loop LC PHI nodes.

            * gcc.dg/torture/pr103237.c: New testcase.

[Bug tree-optimization/103237] [9/10/11 Regression] wrong code with -ftree-vectorize at -O1 on x86_64-linux-gnu starting with r7-1219-gb28ead45fe630b9e

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103237

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

https://gcc.gnu.org/g:fb1bc2c0865f2b15a219e11270b5c5fca565e169

commit r11-9261-gfb1bc2c0865f2b15a219e11270b5c5fca565e169
Author: Richard Biener <rguenther@suse.de>
Date:   Mon Nov 15 11:37:56 2021 +0100

    tree-optimization/103237 - avoid vectorizing unhandled double reductions

    Double reductions which have multiple LC PHIs in the inner loop
    are not handled correctly during transformation since those PHIs
    are not properly classified as reduction.  The following disables
    vectorizing them.

    2021-11-15  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/103237
            * tree-vect-loop.c (vect_is_simple_reduction): Fail for
            double reductions with multiple inner loop LC PHI nodes.

            * gcc.dg/torture/pr103237.c: New testcase.

    (cherry picked from commit 220bd61874cf114667b44f9ded76ed0639eb278b)

[Bug tree-optimization/103237] [9/10 Regression] wrong code with -ftree-vectorize at -O1 on x86_64-linux-gnu starting with r7-1219-gb28ead45fe630b9e

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103237

--- 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:c4015dab1a4697177f5801c8149f7ce15eeeb9e1

commit r10-10465-gc4015dab1a4697177f5801c8149f7ce15eeeb9e1
Author: Richard Biener <rguenther@suse.de>
Date:   Mon Nov 15 11:37:56 2021 +0100

    tree-optimization/103237 - avoid vectorizing unhandled double reductions

    Double reductions which have multiple LC PHIs in the inner loop
    are not handled correctly during transformation since those PHIs
    are not properly classified as reduction.  The following disables
    vectorizing them.

    2021-11-15  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/103237
            * tree-vect-loop.c (vect_is_simple_reduction): Fail for
            double reductions with multiple inner loop LC PHI nodes.

            * gcc.dg/torture/pr103237.c: New testcase.

    (cherry picked from commit 220bd61874cf114667b44f9ded76ed0639eb278b)

[Bug tree-optimization/103237] [9 Regression] wrong code with -ftree-vectorize at -O1 on x86_64-linux-gnu starting with r7-1219-gb28ead45fe630b9e

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103237

--- 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:68c80396906ab10051a69a2181e114c61dd4ee8a

commit r9-9954-g68c80396906ab10051a69a2181e114c61dd4ee8a
Author: Richard Biener <rguenther@suse.de>
Date:   Mon Nov 15 11:37:56 2021 +0100

    tree-optimization/103237 - avoid vectorizing unhandled double reductions

    Double reductions which have multiple LC PHIs in the inner loop
    are not handled correctly during transformation since those PHIs
    are not properly classified as reduction.  The following disables
    vectorizing them.

    2021-11-15  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/103237
            * tree-vect-loop.c (vect_is_simple_reduction): Fail for
            double reductions with multiple inner loop LC PHI nodes.

            * gcc.dg/torture/pr103237.c: New testcase.

    (cherry picked from commit 220bd61874cf114667b44f9ded76ed0639eb278b)

[Bug tree-optimization/103237] [9 Regression] wrong code with -ftree-vectorize at -O1 on x86_64-linux-gnu starting with r7-1219-gb28ead45fe630b9e

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103237

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|---                         |FIXED
      Known to work|                            |10.3.1

--- Comment #12 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed.