[Bug c++/108008] New: Compiler mis-optimization with posix_memalign

[Bug c++/108008] New: Compiler mis-optimization with posix_memalign

[hahnjo at hahnjo dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

            Bug ID: 108008
           Summary: Compiler mis-optimization with posix_memalign
           Product: gcc
           Version: 12.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: hahnjo at hahnjo dot de
  Target Milestone: ---

Created attachment 54033
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=54033&action=edit
init.cpp

Consider the attached example, reduced from some real code, and also available
on Compiler Explorer: https://godbolt.org/z/x4Gsjo7rT

The expectation is that sameZ[i] = true if zPlanes[i] == zPlanes[i + 1] which
should be the case for i = 1. However, when compiling the code with GCC 11.2.0
or 12.2.0 and -std=c++17 -O3, the output is sameZ = 0 for all three iterations
of the second loop. The problem disappears when changing anything in the
slightest manner: compiling with -O2, using -std=gnu++17, using a different
allocation function, commenting out one of the other two statements in the
first loop, using the unconditional assignment to sameZ[i], or removing the
exit(1).

Could this be a bug in the optimizer?

[Bug c++/108008] Compiler mis-optimization with posix_memalign

[amonakov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

Alexander Monakov <amonakov at gcc dot gnu.org> changed:

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

--- Comment #1 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
-fno-split-loops "cures" it (of course it might just be an enabling transform
for an incorrect optimization later on)

Bisecting trunk for which commit fixes/hides it may be useful.

[Bug tree-optimization/108008] Compiler mis-optimization with posix_memalign

[dhsbz at chaw dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

dhsbz at chaw dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dhsbz at chaw dot com

--- Comment #2 from dhsbz at chaw dot com ---
I tried the example with gcc master, and it seems to behave as expected (i.e.
sameZ=1 for i=1). Tracking back it seems the commit which restored this was:

commit ce92603fbe3b4870e0a38efee1ee766d62942065 (HEAD)
Author: Richard Biener
Date:   Mon Jul 18 12:06:00 2022 +0200

    Improve common reduction vs builtin code generation in loop distribution
[...]

(using gcc as of the prior commit shows the behavior as in the bug
description). I can't judge if the above commit had fixed an underlying problem
that the example was triggering, or if it changed the behavior in an indirect
way.

[Bug tree-optimization/108008] Compiler mis-optimization with posix_memalign

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to dhsbz from comment #2)
> I tried the example with gcc master, and it seems to behave as expected
> (i.e. sameZ=1 for i=1). Tracking back it seems the commit which restored
> this was:

r13-1728-gce92603fbe3b48

[Bug tree-optimization/108008] Compiler mis-optimization with posix_memalign

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Created attachment 54041
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=54041&action=edit
Slightly more reduced testcase

Slightly more reduced and self contained testcase.
I was able to remove the exit from Init as it just caused the function not to
be inlined.
There is more to be done to reduce it further though.

[Bug tree-optimization/108008] Compiler mis-optimization with posix_memalign

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #54041|0                           |1
        is obsolete|                            |

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Created attachment 54042
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=54042&action=edit
Removed headers and removed difference from -std=c++17/-std=gnu++17

-std=c++17 was causing posix_memalign not to be known as a allocator and such.
Hide it using the gnu::noipa attribute instead.

[Bug tree-optimization/108008] Compiler mis-optimization with posix_memalign

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #54042|0                           |1
        is obsolete|                            |

--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Created attachment 54043
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=54043&action=edit
Reduced further but needs -O3 -fno-tree-vectorize to hit the bug

[Bug tree-optimization/108008] [12 Regression] Compiler mis-optimization with posix_memalign

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2022-12-07
            Summary|Compiler mis-optimization   |[12 Regression] Compiler
                   |with posix_memalign         |mis-optimization with
                   |                            |posix_memalign
   Target Milestone|---                         |12.3
     Ever confirmed|0                           |1
      Known to work|                            |8.1.0

--- Comment #7 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to dhsbz from comment #2)
> (using gcc as of the prior commit shows the behavior as in the bug
> description). I can't judge if the above commit had fixed an underlying
> problem that the example was triggering, or if it changed the behavior in an
> indirect way.

I think it just changes the IR just ever so slightly which makes the problem go
latent. As shown sometimes adding and removing vectorization.

[Bug tree-optimization/108008] [12 Regression] Compiler mis-optimization with posix_memalign

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #7)
> (In reply to dhsbz from comment #2)
> > (using gcc as of the prior commit shows the behavior as in the bug
> > description). I can't judge if the above commit had fixed an underlying
> > problem that the example was triggering, or if it changed the behavior in an
> > indirect way.
> 
> I think it just changes the IR just ever so slightly which makes the problem
> go latent. As shown sometimes adding and removing vectorization.

Even my example does not show the issue with GCC 11 but does show it with GCC
12.

[Bug tree-optimization/108008] [12 Regression] wrong code with -O3 and posix_memalign

[amonakov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

--- Comment #9 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
I think this is tree-ldist placing memset(sameZ, 0, zPlaneCount) after the
loop, overwriting conditional 'sameZ[i] = true' assignments that happen in the
loop.

For the smaller testcase from comment #6, -O2 -ftree-loop-distribution is
enough, namely:

works:

gcc-12 -O2 -ftree-loop-distribution -fno-tree-vectorize
-fno-tree-loop-distribute-patterns

breaks:

gcc-12 -O2 -ftree-loop-distribution -fno-tree-vectorize

[Bug tree-optimization/108008] [12 Regression] wrong code with -O3 and posix_memalign

[amonakov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

--- Comment #10 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
Looks similar to PR 107323, but needs explicit -ftree-loop-distribution to
trigger.

[Bug tree-optimization/108008] [12 Regression] wrong code with -O3 and posix_memalign

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

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

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

--- Comment #11 from Martin Liška <marxin at gcc dot gnu.org> ---
> 
> gcc-12 -O2 -ftree-loop-distribution -fno-tree-vectorize

This got fixed with r13-1732-g87f46a16ec05beb5 which is a costing revision.

[Bug tree-optimization/108008] [12 Regression] wrong code with -O3 and posix_memalign since r12-4526-gd8edfadfc7a9795b

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |aldyh at gcc dot gnu.org
           Keywords|needs-bisection             |
            Summary|[12 Regression] wrong code  |[12 Regression] wrong code
                   |with -O3 and posix_memalign |with -O3 and posix_memalign
                   |                            |since
                   |                            |r12-4526-gd8edfadfc7a9795b

--- Comment #12 from Martin Liška <marxin at gcc dot gnu.org> ---
And it started with r12-4526-gd8edfadfc7a9795b.

[Bug tree-optimization/108008] [12 Regression] wrong code with -O3 and posix_memalign since r12-4526-gd8edfadfc7a9795b

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P2
                 CC|                            |rguenth at gcc dot gnu.org
           See Also|                            |https://gcc.gnu.org/bugzill
                   |                            |a/show_bug.cgi?id=107323
      Known to work|                            |13.0
           Keywords|                            |needs-bisection

--- Comment #13 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Alexander Monakov from comment #9)
> I think this is tree-ldist placing memset(sameZ, 0, zPlaneCount) after the
> loop, overwriting conditional 'sameZ[i] = true' assignments that happen in
> the loop.
> 
> For the smaller testcase from comment #6, -O2 -ftree-loop-distribution is
> enough, namely:
> 
> works:
> 
> gcc-12 -O2 -ftree-loop-distribution -fno-tree-vectorize
> -fno-tree-loop-distribute-patterns
> 
> breaks:
> 
> gcc-12 -O2 -ftree-loop-distribution -fno-tree-vectorize

Note I can not reproduce on the branch after r12-8855-g19a9b5e587e87f52,
so maybe that's truly the duplicate.

Bisecting on the branch might reveal that.

[Bug tree-optimization/108008] [12 Regression] wrong code with -O3 and posix_memalign since r12-4526-gd8edfadfc7a9795b

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|needs-bisection             |

--- Comment #14 from Martin Liška <marxin at gcc dot gnu.org> ---
> Note I can not reproduce on the branch after r12-8855-g19a9b5e587e87f52,
> so maybe that's truly the duplicate.

gcc-12 branch is fixed since the mentioned revision.

[Bug tree-optimization/108008] [12 Regression] wrong code with -O3 and posix_memalign since r12-4526-gd8edfadfc7a9795b

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108008

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |DUPLICATE

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

*** This bug has been marked as a duplicate of bug 107323 ***