[Bug tree-optimization/110919] New: [14 Regression] Dead Code Elimination Regression at -O3 since r14-1691-gbc5a2c2e793

[Bug tree-optimization/110919] New: [14 Regression] Dead Code Elimination Regression at -O3 since r14-1691-gbc5a2c2e793

[scherrer.sv at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110919

            Bug ID: 110919
           Summary: [14 Regression] Dead Code Elimination Regression at
                    -O3 since r14-1691-gbc5a2c2e793
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: scherrer.sv at gmail dot com
  Target Milestone: ---

static short a, b;
static int e;
static int *j = &e;
void bar151_(void);
void foo(void);
void bar162_(void);
static int(d)(int k, int l) { return l == k && l - 1 ? k : 0; }
static void f();
static void g(unsigned short, unsigned);
static void h() {
  g(0, 4);
  f();
  int i = -1;
  g(0, i);
}
static void f(int k) { g(k, 0); }
static void g(unsigned short k, unsigned l) {
  if (k)
    bar151_();
  if (l)
    if (k)
      bar162_();
  short c = l;
  b = c == 0 || a;
  if (d(*j, b))
    foo();
}
int main() { h(); }

gcc-87b0749cfb9 (trunk) -O3 cannot eliminate the call to foo but
gcc-releases/gcc-13.1.0 -O3 can.
-----------------------------------------------------------------------
gcc-87b0749cfb9406790b108208b466cf507ae3c431 -O3 case.c -S -o case.s
--------- OUTPUT ---------
main:
.LFB4:
        .cfi_startproc
        subq    $8, %rsp
        .cfi_def_cfa_offset 16
        movl    $4, %esi
        xorl    %edi, %edi
        call    g
        xorl    %esi, %esi
        xorl    %edi, %edi
        call    g
        movl    $-1, %esi
        xorl    %edi, %edi
        call    g
        xorl    %eax, %eax
        addq    $8, %rsp
        .cfi_def_cfa_offset 8
        ret
---------- END OUTPUT ---------

-----------------------------------------------------------------------
gcc-2b98cc24d6af0432a74f6dad1c722ce21c1f7458 -O3 case.c -S -o case.s
--------- OUTPUT ---------
main:
.LFB4:
        .cfi_startproc
        subq    $8, %rsp
        .cfi_def_cfa_offset 16
        movl    $4, %esi
        xorl    %edi, %edi
        call    g
        xorl    %esi, %esi
        xorl    %edi, %edi
        call    g
        movl    $-1, %esi
        xorl    %edi, %edi
        call    g
        xorl    %eax, %eax
        addq    $8, %rsp
        .cfi_def_cfa_offset 8
        ret
---------- END OUTPUT ---------

-----------------------------------------------------------------------
Bisects to r14-1691-gbc5a2c2e793

[Bug tree-optimization/110919] [14 Regression] Dead Code Elimination Regression at -O3 since r14-1691-gbc5a2c2e793

[scherrer.sv at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110919

--- Comment #1 from Scherrer Svenja <scherrer.sv at gmail dot com> ---
My apologies, the assembly above is incomplete. The following shows the bug

gcc-87b0749cfb9406790b108208b466cf507ae3c431 -O3 case.c -S -o case.s
--------- OUTPUT ---------
        .file   "case.c"
        .text
        .p2align 4
        .type   g, @function
g:
.LFB3:
        .cfi_startproc
        pushq   %rbx
        .cfi_def_cfa_offset 16
        .cfi_offset 3, -16
        movl    %esi, %ebx
        testw   %di, %di
        jne     .L2
        xorl    %eax, %eax
        testl   %esi, %esi
        sete    %al
.L3:
        movl    e(%rip), %edx
        cmpl    %eax, %edx
        sete    %cl
        testl   %edx, %edx
        setne   %dl
        testb   %dl, %cl
        je      .L1
        testl   %eax, %eax
        je      .L15
.L1:
        popq    %rbx
        .cfi_remember_state
        .cfi_def_cfa_offset 8
        ret
.L2:
        .cfi_restore_state
        call    bar151_
        testl   %ebx, %ebx
        je      .L1
        call    bar162_
        xorl    %eax, %eax
        jmp     .L3
.L15:
        popq    %rbx
        .cfi_def_cfa_offset 8
        jmp     foo
        .cfi_endproc
.LFE3:
        .size   g, .-g
        .section        .text.startup,"ax",@progbits
        .p2align 4
        .globl  main
        .type   main, @function
main:
.LFB4:
        .cfi_startproc
        subq    $8, %rsp
        .cfi_def_cfa_offset 16
        movl    $4, %esi
        xorl    %edi, %edi
        call    g
        xorl    %esi, %esi
        xorl    %edi, %edi
        call    g
        movl    $-1, %esi
        xorl    %edi, %edi
        call    g
        xorl    %eax, %eax
        addq    $8, %rsp
        .cfi_def_cfa_offset 8
        ret
        .cfi_endproc
.LFE4:
        .size   main, .-main
        .local  e
        .comm   e,4,4
        .ident  "GCC: (GNU) 14.0.0 20230806 (experimental)"
        .section        .note.GNU-stack,"",@progbits

---------- END OUTPUT ---------

-----------------------------------------------------------------------
gcc-2b98cc24d6af0432a74f6dad1c722ce21c1f7458 -O3 case.c -S -o case.s
--------- OUTPUT ---------
        .file   "case.c"
        .text
        .p2align 4
        .type   g, @function
g:
.LFB3:
        .cfi_startproc
        testw   %di, %di
        jne     .L11
        ret
.L11:
        pushq   %rbx
        .cfi_def_cfa_offset 16
        .cfi_offset 3, -16
        movl    %esi, %ebx
        call    bar151_
        testl   %ebx, %ebx
        je      .L1
        popq    %rbx
        .cfi_remember_state
        .cfi_restore 3
        .cfi_def_cfa_offset 8
        jmp     bar162_
.L1:
        .cfi_restore_state
        popq    %rbx
        .cfi_def_cfa_offset 8
        ret
        .cfi_endproc
.LFE3:
        .size   g, .-g
        .section        .text.startup,"ax",@progbits
        .p2align 4
        .globl  main
        .type   main, @function
main:
.LFB4:
        .cfi_startproc
        subq    $8, %rsp
        .cfi_def_cfa_offset 16
        movl    $4, %esi
        xorl    %edi, %edi
        call    g
        xorl    %esi, %esi
        xorl    %edi, %edi
        call    g
        movl    $-1, %esi
        xorl    %edi, %edi
        call    g
        xorl    %eax, %eax
        addq    $8, %rsp
        .cfi_def_cfa_offset 8
        ret
        .cfi_endproc
.LFE4:
        .size   main, .-main
        .ident  "GCC: (GNU) 13.1.0"
        .section        .note.GNU-stack,"",@progbits

---------- END OUTPUT ---------

[Bug tree-optimization/110919] [14 Regression] Dead Code Elimination Regression at -O3 since r14-1691-gbc5a2c2e793

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110919

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

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

[Bug tree-optimization/110919] [14 Regression] Dead Code Elimination Regression at -O3 since r14-1691-gbc5a2c2e793

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110919

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2023-08-06
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed.


Note I think what EVRP produces for some thing makes it harder to optimize too.

EVRP does:

Folding statement: _23 = _4 != 1;
Folded into: _23 = (_Bool) _3;


And at the end we get:
  _19 = _6 == iftmp.1_8;
  _3 = iftmp.1_8 ^ 1;
  _21 = (_Bool) _3;
  _23 = _6 != 0;
  _10 = _19 & _23;
  _5 = _10 & _21;

Which is basically:
(a == b) & (a == 0) & (b != 0)

[Bug tree-optimization/110919] [14 Regression] Dead Code Elimination Regression at -O3 since r14-1691-gbc5a2c2e793

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110919

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
The first change we get is threadfull1 where there is an extra jump threading.
And then phiopt2 changes:
  <bb 3> [local count: 719407024]:
  if (l_15(D) != 0)
    goto <bb 6>; [50.00%]
  else
    goto <bb 7>; [50.00%]
into:
  <bb 3> [local count: 719407024]:
  _17 = l_15(D) == 0;
  _36 = (int) _17;

And then DOM2 does not do jump threading along:
  <bb 6> [local count: 628138968]:
  # iftmp.1_8 = PHI <_36(3), 0(5)>


As I mentioned in comment #2 we don't see that:
  _19 = _6 == iftmp.1_8;
  _3 = iftmp.1_8 ^ 1;
  _21 = (_Bool) _3;
  _22 = _19 & _21;
  _23 = _6 != 0;
  _5 = _22 & _23;

is always 0.

[Bug tree-optimization/110919] [14 Regression] Dead Code Elimination Regression at -O3 since r14-1691-gbc5a2c2e793

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110919

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

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

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
After reassoc1 we have:


  _19 = _6 == iftmp.1_8;
  _3 = iftmp.1_8 ^ 1;
  _21 = (_Bool) _3;
  _23 = _6 != 0;
  _10 = _19 & _23;
  _5 = _10 & _21;


Which is: (_6 != 0) & (_6 == iftmp.1_8) But since we know that iftmp.1_8 only
has 2 values, we could optimize this to: (_6 == 1) & (iftmp.1_8 == 1)


Reduced testcase:
```
_Bool f(int a, int b)
{
  b &= 1;
  _Bool t = a == b;
  int c = b^1;
  _Bool d = c;
  _Bool e = a;
  _Bool g = t & d;
  _Bool h = g & e;
  return h;
}
```
This should return false always.

So something like:
(simplify
 (bit_and:c (eq:c @0 @1) (ne:c @0 INTEGER_CST@2))
 (if (has_two_value_range (@1))
  (with {
     tree other_value = ...;
   }
   (bit_and (eq @0 { other_value; }) (eq @1 { other_value; }))
  )
 )
)

[Bug tree-optimization/110919] [14 Regression] Dead Code Elimination Regression at -O3 since r14-1691-gbc5a2c2e793

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110919

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Or take:
```
_Bool g(int a, int b)
{
  b &= 1;
  _Bool t = a == b;
  _Bool e = a != 0;
  _Bool g = t & e;
  return g;
}
int g0(int a, int b)
{
  b &= 1;
  _Bool t = 1 == b;
  _Bool e = a == 1;
  _Bool g = t & e;
  return g;
}
```

These two should produce the same code generation.

[Bug tree-optimization/110919] [14 Regression] Dead Code Elimination Regression at -O3 since r14-1691-gbc5a2c2e793

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110919

--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
For -O1 we have:
  _24 = c_16 == 0;
  _6 = (int) _24;
  _2 = (short int) _24;
  _3 = (int) _24;
  _5 = e;
  _20 = _3 == _5;
  _21 = c_16 != 0;
  _22 = _20 & _21;

((c_16 == 0) == _5) & (c_16 != 0)
((!b) ? 0 : 1) == _5 & (b)
(c_16 != 0) & (0 == _5)

This fixes -O1 (which didn't work before):

(simplify
 (bit_and:c zero_one_valued_p@0 (eq:c @1 (convert @2)))
 (with { bool wascmp; }
  (if (bitwise_inverted_equal_p (@0, @2, wascmp))
   (bit_and @0 (eq @1 { build_zero_cst (TREE_TYPE (@1)); })))))

But -O2 is still broken.

[Bug tree-optimization/110919] [14 Regression] Dead Code Elimination Regression at -O3 since r14-1691-gbc5a2c2e793

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110919

--- Comment #7 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Generalized for -O1 case:
// (a & (b & ~a)) -> 0
// (a & (b | ~a)) -> a & b
// (a & (b ^ ~a)) -> a & b
// (a & (b cmp !a)) -> a & (b cmp 0)
(for op (tcc_comparison bit_and bit_xor bit_and)
 (simplify
  (bit_and:c zero_one_valued_p@0 (op:c @1 (convert? @2)))
  (with { bool wascmp; }
   (if (bitwise_inverted_equal_p (@0, @2, wascmp))
    (bit_and @0 (op @1 { build_zero_cst (TREE_TYPE (@1)); })))))


Note I filed PR 111282 for `(a & (b ^ ~a))` also.

[Bug tree-optimization/110919] [14 Regression] Dead Code Elimination Regression at -O3 since r14-1691-gbc5a2c2e793

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110919

--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #7)
> Generalized for -O1 case:
> // (a & (b & ~a)) -> 0
> // (a & (b | ~a)) -> a & b
> // (a & (b ^ ~a)) -> a & b
> // (a & (b cmp !a)) -> a & (b cmp 0)
> (for op (tcc_comparison bit_and bit_xor bit_and)
>  (simplify
>   (bit_and:c zero_one_valued_p@0 (op:c @1 (convert? @2)))
>   (with { bool wascmp; }
>    (if (bitwise_inverted_equal_p (@0, @2, wascmp))
>     (bit_and @0 (op @1 { build_zero_cst (TREE_TYPE (@1)); })))))
> 
> 
> Note I filed PR 111282 for `(a & (b ^ ~a))` also.

The corrected version is:
// (a & (b & ~a)) -> 0
// (a & (b | ~a)) -> a & b
// (a & (b ^ ~a)) -> a & b
// (a & (b cmp !a)) -> a & (b cmp 0)
(for op (tcc_comparison bit_and bit_xor bit_and)
 (simplify
  (bit_and:c zero_one_valued_p@0 (op:c @1 (convert? zero_one_valued_p@2)))
  (with { bool wascmp; }
   (if (bitwise_inverted_equal_p (@0, @2, wascmp))
    (bit_and @0 (op @1 { build_zero_cst (TREE_TYPE (@1)); }))))))

[Bug tree-optimization/110919] [14 Regression] Dead Code Elimination Regression at -O3 since r14-1691-gbc5a2c2e793

[law at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110919

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/110919] [14/15 Regression] Dead Code Elimination Regression at -O3 since r14-1691-gbc5a2c2e793

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110919

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

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

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
GCC 14.1 is being released, retargeting bugs to GCC 14.2.