[Bug tree-optimization/111799] New: [14 Regression] Missed Dead Code Elimination since r14-2365-g2e406f0753e

[Bug tree-optimization/111799] New: [14 Regression] Missed Dead Code Elimination since r14-2365-g2e406f0753e

[theodort at inf dot ethz.ch]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111799

            Bug ID: 111799
           Summary: [14 Regression] Missed Dead Code Elimination since
                    r14-2365-g2e406f0753e
           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/bqcvMcbqn

Given the following code:

void foo(void);
static struct d {
    short e;
} g = {205}, *h = &g;
static int i;
static int *j, *a = &i;
static int **k = &j;
static int ****l;
static int *****m;
static char(n)(char b, char c) { return b + c; }
static char(o)(char b, char c) { return b * c; }
static short(p)(short f) {
    if (!(((f) >= 1) && ((f) <= 65459))) {
        __builtin_unreachable();
    }
    return 0;
}
static int *q(short);
static void s(struct d) { *k = q(i); }
static int *q(short ad) {
    int b = *a;
    ad = -21;
    for (; ad; ad = n(ad, 7)) p((ad ^ b && *a) <= *a);
    return *k;
}
int main() {
    i = 0;
    for (;; i = 1) {
        q(3);
        char r = o(126 | 1, g.e);
        p(r);
        s(*h);
        if (i) break;
        m = &l;
    }
    if (m)
        ;
    else
        foo();
    ;
}

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

main:
        movl    $2, %esi
        xorl    %ecx, %ecx
        xorl    %r9d, %r9d
        movl    $8, %edi
        movl    $0, i(%rip)
        .p2align 4,,10
        .p2align 3
.L10:
        movl    $-21, %eax
.L2:
        testl   %ecx, %ecx
        je      .L29
        leal    7(%rax), %edx
        movsbw  %dl, %r8w
        testb   %dl, %dl
        je      .L23
.L5:
        movl    $8, %eax
        cmpb    $1, %dl
        je      .L2
        movl    %r8d, %eax
        leal    7(%rax), %edx
        movsbw  %dl, %r8w
        testb   %dl, %dl
        jne     .L5
        .p2align 4,,10
        .p2align 3
.L23:
        movl    $-21, %eax
        addb    $7, %al
        je      .L30
.L9:
        movsbl  %al, %edx
        cmpl    %edx, %ecx
        cmove   %edi, %eax
        cbtw
        addb    $7, %al
        jne     .L9
.L30:
        cmpl    $1, %esi
        je      .L31
        movl    $1, %esi
        movl    $1, %ecx
        movl    $1, %r9d
        jmp     .L10
        .p2align 4,,10
        .p2align 3
.L29:
        testb   $1, %al
        je      .L4
        leal    7(%rax), %edx
        movsbw  %dl, %ax
        testb   %dl, %dl
        je      .L23
.L4:
        leal    14(%rax), %edx
        movsbw  %dl, %ax
        testb   %dl, %dl
        jne     .L4
        jmp     .L23
        .p2align 4,,10
        .p2align 3
.L31:
        testb   %r9b, %r9b
        je      .L11
        movq    $l, m(%rip)
        movl    %ecx, i(%rip)
.L25:
        xorl    %eax, %eax
        ret
.L11:
        cmpq    $0, m(%rip)
        jne     .L25
        pushq   %rax
        call    foo
        xorl    %eax, %eax
        popq    %rdx
        ret

gcc-13.2.0 -O2 eliminates the call to foo:

main:
        movq    $l, m(%rip)
        xorl    %eax, %eax
        movl    $1, i(%rip)
        ret

Bisects to r14-2365-g2e406f0753e

[Bug tree-optimization/111799] [14 Regression] Missed Dead Code Elimination since r14-2365-g2e406f0753e

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111799

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

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

[Bug tree-optimization/111799] [14 Regression] Missed Dead Code Elimination since r14-2365-g2e406f0753e

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111799

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Right now I get an ICE while doing -fdump-tree-ccp1-details even:during GIMPLE
pass: ccp
dump file: /app/output.cpp.034t.ccp1
<source>: In function 'q':
<source>:42:1: internal compiler error: Segmentation fault
   42 | }
      | ^
0x22fdf0e internal_error(char const*, ...)
        ???:0
0x14e8779 print_hex(generic_wide_int<wide_int_ref_storage<false, true> >
const&, char*)
        ???:0
0x14e8beb print_hex(generic_wide_int<wide_int_ref_storage<false, true> >
const&, _IO_FILE*)
        ???:0
g++: internal compiler error: Segmentation fault signal terminated program cc1
Please submit a full bug report, with preprocessed source (by using
-freport-bug).
See <https://gcc.gnu.org/bugs/> for instructions.
Compiler returned: 4

[Bug tree-optimization/111799] [14 Regression] Missed Dead Code Elimination since r14-2365-g2e406f0753e

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111799

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #1)
> Right now I get an ICE while doing -fdump-tree-ccp1-details even:during

Filed PR 111800 for that ...

[Bug tree-optimization/111799] [14 Regression] Missed Dead Code Elimination since r14-2365-g2e406f0753e

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111799

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Hmm, I think we reach __builtin_unreachable always and that is confusing
everything.

If we inline s early on (via using the always_inline attribute). main turns
into just __builtin_unreachable .

[Bug tree-optimization/111799] [14 Regression] Missed Dead Code Elimination since r14-2365-g2e406f0753e

[theodort at inf dot ethz.ch]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111799

--- Comment #4 from Theodoros Theodoridis <theodort at inf dot ethz.ch> ---
Oops, there was a bug in my reduction, here's the fixed code:

https://godbolt.org/z/shxffzs8E

void foo(void);
typedef unsigned short uint16_t;
static int b;
static int ****c;
static int *f = &b;
static int *****ad;
static char(a)(char g, char h) { return g + h; }
static char(d)(char g, char h) { return g * h; }
static void(e)(uint16_t g) {
    if (!(((g) >= 1) && ((g) <= 65459))) {
        __builtin_unreachable();
    }
}
int main() {
    b = 0;
    for (;; b = 1) {
        char i = d(126 | 1, 205);
        e(i);
        short j;
        int k = *f;
        j = -21;
        for (; j; j = a(j, 7)) e((j ^ k && *f) <= *f);
        if (b) break;
        ad = &c;
    }
    if (ad)
        ;
    else
        foo();
    ;
}



trunk -O2:

main:
        movl    $0, b(%rip)
        movl    $2, %esi
        xorl    %ecx, %ecx
        xorl    %r8d, %r8d
.L7:
        movl    $-21, %eax
.L2:
        testl   %ecx, %ecx
        je      .L26
        leal    7(%rax), %edx
        movsbw  %dl, %di
        testb   %dl, %dl
        je      .L6
.L5:
        movl    $8, %eax
        cmpb    $1, %dl
        je      .L2
        movl    %edi, %eax
        leal    7(%rax), %edx
        movsbw  %dl, %di
        testb   %dl, %dl
        jne     .L5
.L6:
        cmpl    $1, %esi
        je      .L27
.L12:
        movl    $1, %esi
        movl    $1, %ecx
        movl    $1, %r8d
        jmp     .L7
.L26:
        testb   $1, %al
        je      .L4
        leal    7(%rax), %edx
        movsbw  %dl, %ax
        testb   %dl, %dl
        je      .L6
.L4:
        leal    14(%rax), %edx
        movsbw  %dl, %ax
        testb   %dl, %dl
        jne     .L4
        cmpl    $1, %esi
        jne     .L12
.L27:
        testb   %r8b, %r8b
        je      .L28
        movq    $c, ad(%rip)
        movl    %ecx, b(%rip)
.L23:
        xorl    %eax, %eax
        ret
.L28:
        cmpq    $0, ad(%rip)
        jne     .L23
        pushq   %rax
        call    foo
        xorl    %eax, %eax
        popq    %rdx
        ret


13.2 -O2:

main:
        movq    $c, ad(%rip)
        xorl    %eax, %eax
        movl    $1, b(%rip)
        ret


the bisection remains the same: r14-2365-g2e406f0753e

[Bug middle-end/111799] [14 Regression] Missed Dead Code Elimination since r14-2365-g2e406f0753e

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111799

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
          Component|tree-optimization           |middle-end
     Ever confirmed|0                           |1
           Keywords|                            |TREE
   Last reconfirmed|                            |2023-10-18

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Theodoros Theodoridis from comment #4)
> Oops, there was a bug in my reduction, here's the fixed code:
> 
> https://godbolt.org/z/shxffzs8E
> 
> void foo(void);
> typedef unsigned short uint16_t;
> static int b;
> static int ****c;
> static int *f = &b;
> static int *****ad;
> static char(a)(char g, char h) { return g + h; }
> static char(d)(char g, char h) { return g * h; }
> static void(e)(uint16_t g) {
>     if (!(((g) >= 1) && ((g) <= 65459))) {
>         __builtin_unreachable();
>     }
> }
> int main() {
>     b = 0;
>     for (;; b = 1) {
>         char i = d(126 | 1, 205);
>         e(i);
>         short j;
>         int k = *f;
>         j = -21;
>         for (; j; j = a(j, 7)) e((j ^ k && *f) <= *f);
>         if (b) break;
>         ad = &c;
>     }
>     if (ad)
>         ;
>     else
>         foo();
>     ;
> }


Confirmed with this testcase but what is interesting is that the optimization
of calling foo away does NOT happen at the gimple level but at the RTL level
and just happens on x86_64 but NOT on aarch64.

on aarch64 for GCC 13, we even have:
        mov     w0, 1
        str     wzr, [x1, #:lo12:.LANCHOR0]
        cbz     w0, .L13

Which is obvious should have been removed ...

Anyways this is a missed jump threading that should have happened at the gimple
level. I have not looked into why it is not done.

[Bug middle-end/111799] [14 Regression] Missed Dead Code Elimination since r14-2365-g2e406f0753e

[law at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111799

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

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

[Bug middle-end/111799] [14 Regression] Missed Dead Code Elimination since r14-2365-g2e406f0753e

[law at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111799

--- Comment #6 from Jeffrey A. Law <law at gcc dot gnu.org> ---
I'd hazard a guess we need to first eliminate the ad = &c assignment.  That
should in turn allow us to realize ad is unchanging.

[Bug middle-end/111799] [14/15 Regression] Missed Dead Code Elimination since r14-2365-g2e406f0753e

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111799

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

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

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
GCC 14.1 is being released, retargeting bugs to GCC 14.2.