[Bug tree-optimization/110538] New: [14 Regression] Dead Code Elimination Regression since r14-368-ge1366a7e4ce

[Bug tree-optimization/110538] New: [14 Regression] Dead Code Elimination Regression since r14-368-ge1366a7e4ce

[theodort at inf dot ethz.ch]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110538

            Bug ID: 110538
           Summary: [14 Regression] Dead Code Elimination Regression since
                     r14-368-ge1366a7e4ce
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: theodort at inf dot ethz.ch
  Target Milestone: ---

https://godbolt.org/z/53Ys78do1

Given the following code:

void foo(void);
static int b, d, f;
static int *c, *e = &d;
static unsigned g;
void a();
int main() {
    g = -19;
    for (; g; ++g) {
        int h = g;
        int *i = &b;
        int **j = &i;
        if (d) {
            int **k = &i;
            j = &c;
            *k = &f;
        } else
            *e = 0;
        if (!(((h) >= -19) && ((h) <= -1))) {
            __builtin_unreachable();
        }
        if (i)
            ;
        else
            a();
        if (j == &i || j == &c)
            ;
        else
            foo();
        ;
    }
}

gcc-trunk -O2 does not eliminate the call to foo:

main:
        subq    $24, %rsp
        movl    $-19, g(%rip)
        .p2align 4,,10
        .p2align 3
.L4:
        movl    d(%rip), %edx
        movl    $c, %eax
        testl   %edx, %edx
        jne     .L2
        leaq    8(%rsp), %rax
.L2:
        cmpq    $c, %rax
        je      .L5
        leaq    8(%rsp), %rdx
        cmpq    %rdx, %rax
        je      .L5
        call    foo
.L5:
        addl    $1, g(%rip)
        jne     .L4
        xorl    %eax, %eax
        addq    $24, %rsp
        ret

gcc-13.1.0 -O2 eliminates the call to foo:

main:
        movl    $-19, g(%rip)
        .p2align 4,,10
        .p2align 3
.L2:
        movl    g(%rip), %eax
        addl    $1, %eax
        movl    %eax, g(%rip)
        jne     .L2
        ret

Bisects to r14-368-ge1366a7e4ce

[Bug tree-optimization/110538] [14 Regression] Dead Code Elimination Regression since r14-368-ge1366a7e4ce

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110538

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |14.0
           Keywords|                            |missed-optimization

[Bug tree-optimization/110538] [14 Regression] Dead Code Elimination Regression since r14-368-ge1366a7e4ce

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110538

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2023-07-07
             Status|UNCONFIRMED                 |NEW

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

In .optimized we get:
  # j_24 = PHI <&i(7), &c(3)>
  _2 = j_24 == &c;
  _22 = &i == j_24;
  _23 = _2 | _22;

Obvious _23 is always 1.

[Bug tree-optimization/110538] [14 Regression] Dead Code Elimination Regression since r14-368-ge1366a7e4ce

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110538

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
So dom3 was able to optimize that via a jump threading before in GCC 13 but no
longer on the trunk (I don't understand why though).

Anyways the only pass which is able to optimize:
```
int f123(int a, int c, int i)
{
  int *d;
  if (a) d = &i; else d = &c;
  int e = d == &i;
  int f = d == &c;
  return e | f;
}
```
to `1` is PRE ...

But note
```
int g123();
int h123();
int f123_1(int a, int c, int i)
{
  int *d;
  if (a) d = &i; else d = &c;
  int e = d == &i;
  int f = d == &c;
  if (e | f)
    return h123();
  return g123();
}
```
Still can be optimized by dom2 on the trunk (via jump threading).

So how did dom3 miss the original testcase ....

[Bug tree-optimization/110538] [14 Regression] Dead Code Elimination Regression since r14-368-ge1366a7e4ce

[law at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110538

Jeffrey A. Law <law at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |law at gcc dot gnu.org
           Priority|P3                          |P2

[Bug tree-optimization/110538] [14 Regression] Dead Code Elimination Regression since r14-368-ge1366a7e4ce

[law at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110538

--- Comment #3 from Jeffrey A. Law <law at gcc dot gnu.org> ---
Thread references removed edge:   Cancelling jump thread: (9, 10) incoming
edge;  (10, 8) joiner (8, 6) normal;

Thread references removed edge:   Cancelling jump thread: (2, 3) incoming edge;
 (3, 9) joiner (9, 10) nocopy (10, 8) normal;

Thread references removed edge:   Cancelling jump thread: (2, 3) incoming edge;
 (3, 10) joiner (10, 8) normal;

Thread references removed edge:   Cancelling jump thread: (3, 10) incoming
edge;  (10, 8) normal;

[ ... ]

Merging blocks 10 and 8
Removing basic block 11
;; basic block 11, loop depth 1
;;  pred:         
goto <bb 3>; [100.00%]
;;  succ:       3

So one of the edges in the jump threading path gets removed.  As a result the
threader throws the path away.  At least that's my best guess.

We could possibly defer block merging until after threading.  No idea how hard
that might be.  The sequencing is a bit painful, but IIRC the key is that we
must remove unreachables in the CFG before threading (due to how unreachables
interact with the dominator tree updates).

[Bug tree-optimization/110538] [14/15 Regression] Dead Code Elimination Regression since r14-368-ge1366a7e4ce

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110538

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|14.0                        |14.2

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
GCC 14.1 is being released, retargeting bugs to GCC 14.2.