[Bug tree-optimization/108139] New: wrong code with "-O1 -ftree-vrp" on x86_64-linux-gnu

[Bug tree-optimization/108139] New: wrong code with "-O1 -ftree-vrp" on x86_64-linux-gnu

[zhendong.su at inf dot ethz.ch]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108139

            Bug ID: 108139
           Summary: wrong code with "-O1 -ftree-vrp" on x86_64-linux-gnu
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: zhendong.su at inf dot ethz.ch
  Target Milestone: ---

It appears to be a recent regression.

Compiler Explorer: https://godbolt.org/z/ocGqn8jTY


[554] % gcctk -v
Using built-in specs.
COLLECT_GCC=gcctk
COLLECT_LTO_WRAPPER=/local/suz-local/software/local/gcc-trunk/libexec/gcc/x86_64-pc-linux-gnu/13.0.0/lto-wrapper
Target: x86_64-pc-linux-gnu
Configured with: ../gcc-trunk/configure --disable-bootstrap
--enable-checking=yes --prefix=/local/suz-local/software/local/gcc-trunk
--enable-sanitizers --enable-languages=c,c++ --disable-werror --enable-multilib
--with-system-zlib
Thread model: posix
Supported LTO compression algorithms: zlib
gcc version 13.0.0 20221216 (experimental) [master r13-4736-gfb33e104f48] (GCC) 
[555] % 
[555] % gcctk -O1 small.c; ./a.out
[555] % 
[556] % gcctk -O1 -ftree-vrp small.c
[557] % timeout -s 9 5 ./a.out
Killed
[558] % 
[558] % cat small.c
int a, b;
int main() {
  int c;
  if (a > 1)
    a++;
  while (a)
    if (c == b)
      c = a;
  return 0;
}

[Bug tree-optimization/108139] [13 Regression] wrong code with "-O1 -ftree-vrp" on x86_64-linux-gnu

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108139

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |needs-bisection, wrong-code
            Summary|wrong code with "-O1        |[13 Regression] wrong code
                   |-ftree-vrp" on              |with "-O1 -ftree-vrp" on
                   |x86_64-linux-gnu            |x86_64-linux-gnu
            Version|unknown                     |13.0
   Target Milestone|---                         |13.0

[Bug tree-optimization/108139] [13 Regression] wrong code with "-O1 -ftree-vrp" on x86_64-linux-gnu since r13-1938-g87dd4c8c83768aaf

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108139

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[13 Regression] wrong code  |[13 Regression] wrong code
                   |with "-O1 -ftree-vrp" on    |with "-O1 -ftree-vrp" on
                   |x86_64-linux-gnu            |x86_64-linux-gnu since
                   |                            |r13-1938-g87dd4c8c83768aaf
           Keywords|needs-bisection             |
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2022-12-16
     Ever confirmed|0                           |1
                 CC|                            |amacleod at redhat dot com,
                   |                            |marxin at gcc dot gnu.org
           Priority|P3                          |P1

--- Comment #1 from Martin Liška <marxin at gcc dot gnu.org> ---
Started with r13-1938-g87dd4c8c83768aaf.

[Bug tree-optimization/108139] [13 Regression] wrong code with "-O1 -ftree-vrp" on x86_64-linux-gnu since r13-1938-g87dd4c8c83768aaf

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108139

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Folding statement: if (a.3_5 != 0)
Folding predicate a.3_5 != 0 to 1
Folded into: if (1 != 0)

I don't see how that could be possible. There is some issue coming from the
(not used on this path) uninitialized variable c too.

[Bug tree-optimization/108139] [13 Regression] wrong code with "-O1 -ftree-vrp" on x86_64-linux-gnu since r13-1938-g87dd4c8c83768aaf

[amacleod at redhat dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108139

--- Comment #3 from Andrew Macleod <amacleod at redhat dot com> ---
(In reply to Andrew Pinski from comment #2)
> Folding statement: if (a.3_5 != 0)
> Folding predicate a.3_5 != 0 to 1
> Folded into: if (1 != 0)
> 
> I don't see how that could be possible. There is some issue coming from the
> (not used on this path) uninitialized variable c too.

  <bb 5> [local count: 59055799]:
  a.3_5 = a;
  b.2_4 = b;

  <bb 6> [local count: 114863530]:
  # c_2 = PHI <c_9(D)(5), a.3_5(9)>

  <bb 7> [local count: 1073741824]:
  if (a.3_5 != 0)
    goto <bb 4>; [94.50%]
  else
    goto <bb 8>; [5.50%]

The problem is that c_9 is an undefined value in the PHI..  with only 2
arguments, we try to be clever (and this is important in other places), and
say, ok, we'll assume it has the same value as the other argument and we create
an equivalence between c_2 and the other argument, a.3_5

The only way to get to the 6->9 edge is to have processed the if statement
already, and therefore we'd know that a.3_5 is ~[0, 0], sowe determine that c_2
is also ~[0,0].

When we visit the if statement, we know a.3_5 is varying, but we also look at
what equivalences a.3_5 has, and voila c_2 is equivlant o a.3_5, so we refine
the value to match c_2 and then decide we can fold away the if as always true.

we don't seem to run into this at hihger optimization levels because bb6 and
bb7 are merged, and with the PHI in the same block,we don't trigger the
outgoing edge equivalence evaluations :-P  so we pick up the original VARYING
result for a.3_5 and don't adjust it with its "equivalence c_2". 

hum.

[Bug tree-optimization/108139] [13 Regression] wrong code with "-O1 -ftree-vrp" on x86_64-linux-gnu since r13-1938-g87dd4c8c83768aaf

[amacleod at redhat dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108139

--- Comment #4 from Andrew Macleod <amacleod at redhat dot com> ---
9->6 edge, not 6->9...

[Bug tree-optimization/108139] [13 Regression] wrong code with "-O1 -ftree-vrp" on x86_64-linux-gnu since r13-1938-g87dd4c8c83768aaf

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108139

--- Comment #5 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Andrew Macleod <amacleod@gcc.gnu.org>:

https://gcc.gnu.org/g:0bdd2261c254f08b0f4437c156b79711d68c6e7f

commit r13-4806-g0bdd2261c254f08b0f4437c156b79711d68c6e7f
Author: Andrew MacLeod <amacleod@redhat.com>
Date:   Fri Dec 16 16:53:31 2022 -0500

    Don't use PHI equivalences in range-on-entry.

    If there is only one argument to a PHI which is defined, an equivalency is
    created between the def and the argument.  It is safe to consider the def
    equal to the argument, but it is dangerous to assume the argument is also
    equivalent to the def as there may be branches which change the range on
the
    path to the PHI on that argument

    This patch avoid using that relation in range-on-entry calculations.

            PR tree-optimization/108139
            gcc/
            * gimple-range-cache.cc (ranger_cache::fill_block_cache): Do not
            use equivalences originating from PHIS.

            gcc/testsuite/
            * gcc.dg/pr108139.c: New.

[Bug tree-optimization/108139] [13 Regression] wrong code with "-O1 -ftree-vrp" on x86_64-linux-gnu since r13-1938-g87dd4c8c83768aaf

[amacleod at redhat dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108139

Andrew Macleod <amacleod at redhat dot com> changed:

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

--- Comment #6 from Andrew Macleod <amacleod at redhat dot com> ---
fixed.

[Bug tree-optimization/108139] [13 Regression] wrong code with "-O1 -ftree-vrp" on x86_64-linux-gnu since r13-1938-g87dd4c8c83768aaf

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108139

--- Comment #7 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Andrew Macleod <amacleod@gcc.gnu.org>:

https://gcc.gnu.org/g:24af552876eff707f75d30d3f0f0e7a5d62dd857

commit r13-7150-g24af552876eff707f75d30d3f0f0e7a5d62dd857
Author: Andrew MacLeod <amacleod@redhat.com>
Date:   Tue Apr 11 17:29:03 2023 -0400

    Don't use ANY PHI equivalences in range-on-entry.

    PR 108139 dissallows PHI equivalencies in the on-entry calculator, but
    it was only checking if the equivlaence was a PHI.  In this case, NAME
    itself is a PHI with an equivlaence caused by an undefined value, so we
    also need to check that case.  Unfortunately this un-fixes 101912.

            PR tree-optimization/109462
            gcc/
            * gimple-range-cache.cc (ranger_cache::fill_block_cache): Don't
            check for equivalences if NAME is a phi node.

            gcc/testsuite/
            * gcc.dg/uninit-pr101912.c: XFAIL the warning.

[Bug tree-optimization/108139] [13 Regression] wrong code with "-O1 -ftree-vrp" on x86_64-linux-gnu since r13-1938-g87dd4c8c83768aaf

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108139

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

https://gcc.gnu.org/g:9c2a5db997446a9438a3e01f5229dec3f78b09e7

commit r13-7170-g9c2a5db997446a9438a3e01f5229dec3f78b09e7
Author: Andrew MacLeod <amacleod@redhat.com>
Date:   Wed Apr 12 13:10:55 2023 -0400

    Ensure PHI equivalencies do not dominate the argument edge.

    When we create an equivalency between a PHI definition and an argument,
    ensure the definition does not dominate the incoming argument edge.

            PR tree-optimization/108139
            PR tree-optimization/109462
            * gimple-range-cache.cc (ranger_cache::fill_block_cache): Remove
            equivalency check for PHI nodes.
            * gimple-range-fold.cc (fold_using_range::range_of_phi): Ensure def
            does not dominate single-arg equivalency edges.