[Bug tree-optimization/108574] New: wrong code at -Os and above on x86_64-linux-gnu

[Bug tree-optimization/108574] New: wrong code at -Os and above on x86_64-linux-gnu

[zhendong.su at inf dot ethz.ch]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108574

            Bug ID: 108574
           Summary: wrong code at -Os and above 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 from 12.2.

Compiler Explorer: https://godbolt.org/z/4T3xMvsab


[578] % 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/13.0.1/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
--disable-multilib
Thread model: posix
Supported LTO compression algorithms: zlib
gcc version 13.0.1 20230127 (experimental) [master r12-4647-g3f861a5c8fd] (GCC) 
[579] % 
[579] % gcctk -O3 small.c
[580] % ./a.out
Floating point exception
[581] % 
[581] % cat small.c
int a = 1, b, c = 2, d;
int main() {
  if (b)
    goto L2;
 L1:
  {
    int e = c;
    a = 1 % a;
    while (e && 1 <= d)
      ;
    d >= b;
  L2:
    if (1 >= e)
      goto L1;
  }
  return 0;
}

[Bug tree-optimization/108574] wrong code at -Os and above on x86_64-linux-gnu

[zhendong.su at inf dot ethz.ch]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108574

--- Comment #1 from Zhendong Su <zhendong.su at inf dot ethz.ch> ---
Here is another reproducer likely for the same bug.

Compiler Explorer: https://godbolt.org/z/Efvczc6f8


[556] % gcctk -O3 small.c
[557] % ./a.out
Floating point exception
[558] % 
[558] % cat small.c
int a, b, c, d, e, f, g = -1, h;
void l() {
  if (!e)
    goto i;
  for (; g; g++) {
    b = ~d;
    int j = 0, k = 1;
    if (k && (b || f))
      j = b;
  i:
    a = ~j;
  }
}
int main() {
  h = 3;
  for (; h; h--) {
    e = 1;
    int m = ~a, n = 1 % m;
    c = n;
    l();
  }
  return 0;
}

[Bug tree-optimization/108574] wrong code at -Os and above on x86_64-linux-gnu

[zhendong.su at inf dot ethz.ch]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108574

--- Comment #2 from Zhendong Su <zhendong.su at inf dot ethz.ch> ---
Another likely related test case.

Compiler Explorer: https://godbolt.org/z/G514hnMeY


[537] % gcctk -O3 small.c
[538] % timeout -s 9 5 ./a.out
Killed
[539] % gcctk -O1 small.c
[540] % ./a.out
[541] % 
[541] % cat small.c
int a = 3557301289, d;
char b, f;
unsigned short c = 241;
short e, g;
static void h() {
  if (!a)
    goto i;
  b = a;
  for (; a < 2; a = b) {
    unsigned short j;
    if (c || !g) {
      j = c;
    i:
      e = j;
    }
    f = j;
    d = ~(f & ~2880764155);
    while (d > -2316069)
      ;
  }
}
int main() {
  h();
  return 0;
}

[Bug tree-optimization/108574] [13 Regression] wrong code at -Os and above on x86_64-linux-gnu

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108574

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
            Summary|wrong code at -Os and above |[13 Regression] wrong code
                   |on x86_64-linux-gnu         |at -Os and above on
                   |                            |x86_64-linux-gnu
   Target Milestone|---                         |13.0
           Keywords|                            |wrong-code
   Last reconfirmed|                            |2023-01-27
             Status|UNCONFIRMED                 |NEW

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Zhendong Su from comment #0)
> It appears to be a recent regression from 12.2.

In the testcase from comment #0, pre is messing it up because of an
uninitialized variable on one path (which is not taken).

Note somehow the statement:
d >= b;
changes the IR slightly ...

testcase in comment #1 also depends on an uninitialized variable on a path not
taken too (in that case j).

testcase in comment #2 also looks similar and depends on an unintialized
variable on a path not taken too.

[Bug tree-optimization/108574] [13 Regression] wrong code at -Os and above on x86_64-linux-gnu

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108574

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
           Priority|P3                          |P1
             Status|NEW                         |ASSIGNED
            Version|unknown                     |13.0

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Likely caused by r13-5348-gc29d85359add80

We do

marking outgoing edge 12 -> 6 executable
Looking for changed values of backedge 12->6 destination PHIs
Predication says 0 and e_10 are equal on edge 11 -> 6
Setting value number of e_5 to 0

I will have a more detailed look.

[Bug tree-optimization/108574] [13 Regression] wrong code at -Os and above on x86_64-linux-gnu

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108574

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
*** Bug 108584 has been marked as a duplicate of this bug. ***

[Bug tree-optimization/108574] [13 Regression] wrong code at -Os and above on x86_64-linux-gnu

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108574

--- 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:7ac3e69e311351b70407d7f87a0169c4d463e57b

commit r13-5492-g7ac3e69e311351b70407d7f87a0169c4d463e57b
Author: Richard Biener <rguenther@suse.de>
Date:   Mon Jan 30 09:25:23 2023 +0100

    tree-optimization/108574 - wrong-code with PRE PHI node processing

    The PR108523 was too optimistic in replacing the same value with
    an equivalence from a possibly not taken edge.  The following
    rectifies this and instead refrains from using the equivalence in
    the problematic cases.

            PR tree-optimization/108574
            * tree-ssa-sccvn.cc (visit_phi): Instead of swapping
            sameval and def, ignore the equivalence if there's the
            danger of oscillating between two values.

            * gcc.dg/torture/pr108574-1.c: New testcase.
            * gcc.dg/torture/pr108574-2.c: Likewise.
            * gcc.dg/torture/pr108574-3.c: Likewise.

[Bug tree-optimization/108574] [13 Regression] wrong code at -Os and above on x86_64-linux-gnu

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108574

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

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

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