[Bug tree-optimization/110963] New: [14 Regression] Dead Code Elimination Regression since r14-2946-g46c8c225455

[Bug tree-optimization/110963] New: [14 Regression] Dead Code Elimination Regression since r14-2946-g46c8c225455

[theodort at inf dot ethz.ch]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110963

            Bug ID: 110963
           Summary: [14 Regression] Dead Code Elimination Regression since
                    r14-2946-g46c8c225455
           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/devdMf1qd

Given the following code:

void foo(void);
static int c = 76, f, g;
static int *h, *j, *k = &g;
static int **i = &h;
static short a;
static char(l)(char b) {
    if (!(((b) >= 77) && ((b) <= 77))) {
        __builtin_unreachable();
    }
    return 0;
}
static short(m)(short d, short e) { return d + e; }
static short n(char) {
    j = *i;
    if (j == 0)
        ;
    else
        *i = 0;
    *k = 0;
    return 0;
}
static char o() {
    l(0);
    return 0;
}
static char p(int ad) {
    a = m(!0, ad);
    l(a);
    if (f) {
        *i &&n(o());
        *i = 0;
    } else
        n(0);
    if (h == &f || h == 0)
        ;
    else
        foo();
    return 0;
}
int main() {
    p(c);
    c = 8;
}

gcc-trunk -O3 does not eliminate the call to foo:

main:
        movl    f(%rip), %edi
        testl   %edi, %edi
        je      .L11
.L7:
        movl    $8, c(%rip)
        xorl    %eax, %eax
        ret
.L11:
        cmpq    $0, h(%rip)
        je      .L4
        xorl    %esi, %esi
        movq    %rsi, h(%rip)
.L4:
        movq    h(%rip), %rax
        xorl    %ecx, %ecx
        movl    %ecx, g(%rip)
        testq   %rax, %rax
        je      .L7
        cmpq    $f, %rax
        je      .L7
        pushq   %rax
        call    foo
        xorl    %eax, %eax
        movl    $8, c(%rip)
        popq    %rdx
        ret

gcc-13.2.0 -O3 eliminates the call to foo:

main:
        movl    f(%rip), %ecx
        movq    h(%rip), %rax
        testl   %ecx, %ecx
        jne     .L3
        testq   %rax, %rax
        je      .L4
        xorl    %edx, %edx
        movq    %rdx, h(%rip)
.L4:
        xorl    %eax, %eax
        movl    %eax, g(%rip)
.L3:
        movl    $8, c(%rip)
        xorl    %eax, %eax
        ret

Bisects to r14-2946-g46c8c225455

[Bug tree-optimization/110963] [14 Regression] Dead Code Elimination Regression since r14-2946-g46c8c225455

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110963

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |14.0
                 CC|                            |pinskia at gcc dot gnu.org
           Keywords|                            |missed-optimization

[Bug tree-optimization/110963] [14 Regression] Dead Code Elimination Regression since r14-2946-g46c8c225455

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110963

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2023-08-09
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed.
PRE seems to have missed an obvious redundant load here though:
The IR after PRE:
```
  pretmp_12 = h;
  if (f.2_9 != 0)
    goto <bb 3>; [50.00%]
  else
    goto <bb 6>; [50.00%]
...
  <bb 6> [local count: 536870912]:
  if (pretmp_12 == 0B)
    goto <bb 8>; [30.00%]
  else
    goto <bb 7>; [70.00%]

  <bb 7> [local count: 375809640]:
  h = 0B;

  <bb 8> [local count: 536870912]:
  g = 0;
  h.6_15 = h;
```

The IR before PRE:
```
  <bb 2> [local count: 1073741824]:
  # VUSE <.MEM_2(D)>
  f.2_9 = f;
  if (f.2_9 != 0)
    goto <bb 3>; [50.00%]
  else
    goto <bb 5>; [50.00%]

  <bb 3> [local count: 536870912]:
  # VUSE <.MEM_2(D)>
  _11 = h;
  if (_11 != 0B)
    goto <bb 4>; [0.00%]
  else
    goto <bb 9>; [100.00%]
...
  <bb 5> [local count: 536870912]:
  # VUSE <.MEM_2(D)>
  _13 = h;
  if (_13 == 0B)
    goto <bb 7>; [30.00%]
  else
    goto <bb 6>; [70.00%]

  <bb 6> [local count: 375809640]:
  # .MEM_20 = VDEF <.MEM_2(D)>
  h = 0B;

  <bb 7> [local count: 536870912]:
  # .MEM_16 = PHI <.MEM_2(D)(5), .MEM_20(6)>
  # .MEM_21 = VDEF <.MEM_16>
  g = 0;
  # VUSE <.MEM_21>
  h.6_15 = h;
```

[Bug tree-optimization/110963] [14 Regression] Dead Code Elimination Regression since r14-2946-g46c8c225455

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110963

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
I will have a look.

[Bug tree-optimization/110963] [14 Regression] Dead Code Elimination Regression since r14-2946-g46c8c225455

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110963

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
So we're now doing an additional sinking:

Sinking # VUSE <.MEM_2(D)>
pretmp_12 = h;
 from bb 2 to bb 4
Removing basic block 9

that's the only IL difference on GIMPLE, without this we optimize away foo
during RTL optimization during RTL PRE.

[Bug tree-optimization/110963] [14 Regression] Dead Code Elimination Regression since r14-2946-g46c8c225455

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110963

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #1)
> Confirmed.
> PRE seems to have missed an obvious redundant load here though:

Skipping partial redundancy for expression {mem_ref<0B>,addr_expr<&h>}@.MEM_21
(0008), no redundancy on to be optimized for speed edge

renaming 'main' into 'bar' fixes this:

Found partial redundancy for expression {mem_ref<0B>,addr_expr<&h>}@.MEM_22
(0007)
Created phi prephitmp_7 = PHI <_10(11), 0B(6)>
 in block 7 (0007)
...


then the call to 'foo' is eliminated already in FRE5 (before the sinking pass
affected by r14-2946-g46c8c225455).

IMHO this isn't a bug worth pursuing.

Alternatively you can put __attribute__((hot)) on 'main' which we otherwise
consider cold.  Maybe that's a good idea for future DCE regression searches?

Maybe we can consider the special case worthwhile also when optimizing for
size.

[Bug tree-optimization/110963] [14 Regression] Dead Code Elimination Regression since r14-2946-g46c8c225455

[theodort at inf dot ethz.ch]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110963

--- Comment #5 from Theodoros Theodoridis <theodort at inf dot ethz.ch> ---
(In reply to Richard Biener from comment #4)
> Alternatively you can put __attribute__((hot)) on 'main' which we otherwise
> consider cold.  Maybe that's a good idea for future DCE regression searches?


Thanks for the feedback. IIRC, I have reported at least one more case where
renaming 'main' would make the issue go away. So if adding __attribute__((hot))
makes sense, then I'll include it when searching.

[Bug tree-optimization/110963] [14 Regression] Dead Code Elimination Regression since r14-2946-g46c8c225455

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110963

--- Comment #6 from rguenther at suse dot de <rguenther at suse dot de> ---
On Thu, 10 Aug 2023, theodort at inf dot ethz.ch wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110963
> 
> --- Comment #5 from Theodoros Theodoridis <theodort at inf dot ethz.ch> ---
> (In reply to Richard Biener from comment #4)
> > Alternatively you can put __attribute__((hot)) on 'main' which we otherwise
> > consider cold.  Maybe that's a good idea for future DCE regression searches?
> 
> 
> Thanks for the feedback. IIRC, I have reported at least one more case where
> renaming 'main' would make the issue go away. So if adding __attribute__((hot))
> makes sense, then I'll include it when searching.

I think it makes sense, it properly distinguishes -Os from -O2 where
otherwise those would behave too similar.

[Bug tree-optimization/110963] [14 Regression] Dead Code Elimination Regression since r14-2946-g46c8c225455

[hubicka at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110963

--- Comment #7 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
We don't consider main cold, but executed once: code out of loops is optimized
for size, but anything in loops is optimized according to -O setting.  I did
not really think of users overwriting it by hot attribute, but looking into
implementation it will work :)

[Bug tree-optimization/110963] [14 Regression] Dead Code Elimination Regression since r14-2946-g46c8c225455

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110963

--- 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:4d6132e59327e809a4d4e39fb9465dbd43775b7c

commit r14-3217-g4d6132e59327e809a4d4e39fb9465dbd43775b7c
Author: Richard Biener <rguenther@suse.de>
Date:   Thu Aug 10 13:55:36 2023 +0200

    tree-optimization/110963 - more PRE when optimizing for size

    The following adjusts the heuristic when we perform PHI insertion
    during GIMPLE PRE from requiring at least one edge that is supposed
    to be optimized for speed to also doing insertion when the expression
    is available on all edges (but possibly with different value) and
    we'd at most have one copy from a constant.  The first ensures
    we optimize two computations on all paths to one plus a possible
    copy due to the PHI, the second makes sure we do not need to insert
    many possibly large copies from constants, disregarding the
    cummulative size cost of the register copies when they are not
    coalesced.

    The case in the testcase is

      <bb 5>
      _14 = h;
      if (_14 == 0B)
        goto <bb 7>;
      else
        goto <bb 6>;

      <bb 6>
      h = 0B;

      <bb 7>
      h.6_12 = h;

    and we want to optimize that to

      <bb 7>
      # h.6_12 = PHI <_14(5), 0B(6)>

    If we want to consider the cost of the register copies I think the
    only simplistic enough way would be to restrict the special-case to
    two incoming edges - we'd assume one register copy is coalesced
    leaving one copy from a register or from a constant.

    As with every optimization the downstream effects are probably
    bigger than what we can locally estimate.

            PR tree-optimization/110963
            * tree-ssa-pre.cc (do_pre_regular_insertion): Also insert
            a PHI node when the expression is available on all edges
            and we insert at most one copy from a constant.

            * gcc.dg/tree-ssa/ssa-pre-34.c: New testcase.

[Bug tree-optimization/110963] [14 Regression] Dead Code Elimination Regression since r14-2946-g46c8c225455

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110963

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

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

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed nevertheless.