[Bug tree-optimization/112545] New: [14 Regression] Size regression when using -Os starting with r14-569-g21e2ef2dc25

[Bug tree-optimization/112545] New: [14 Regression] Size regression when using -Os starting with r14-569-g21e2ef2dc25

[theodort at inf dot ethz.ch]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112545

            Bug ID: 112545
           Summary: [14 Regression] Size regression when using -Os
                    starting with r14-569-g21e2ef2dc25
           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/hcq8sKhvP

Given the following code:

static struct {
    int a;
    int b;
} c;
static int d, e, g;
static short f, i;
static void h(unsigned short j) {
    c.a = i;
    if (!(((j) >= 1) && ((j) <= 65535))) {
        __builtin_unreachable();
    }
}
int main1() {
    d = 6;
    for (; d != -7; d--) {
        h(d ^ d < (0 <= 6));
        if (d >= 12)
            if (e) {
                for (; f; ++f) g = c.b;
                if (g) e = 0;
            }
    }
}

gcc-trunk -Os generates significantly more complex code:

main1:
        movl    e(%rip), %ecx
        movl    $6, %eax
        xorl    %edi, %edi
        xorl    %r8d, %r8d
        movl    c+4(%rip), %r10d
        movl    g(%rip), %esi
        xorl    %r9d, %r9d
        movl    $6, d(%rip)
        movw    f(%rip), %dx
.L2:
        cmpl    $-7, %eax
        je      .L31
        xorl    %r11d, %r11d
        testl   %eax, %eax
        setle   %r11b
        cmpw    %ax, %r11w
        jne     .L3
        testb   %r8b, %r8b
        je      .L4
        movl    %eax, d(%rip)
.L4:
        testb   %dil, %dil
        je      .L5
        movl    %ecx, e(%rip)
.L5:
        xorl    %ecx, %ecx
        movw    %dx, f(%rip)
        movl    %ecx, c(%rip)
        movl    %esi, g(%rip)
.L3:
        cmpl    $11, %eax
        jle     .L6
        testl   %ecx, %ecx
        je      .L6
.L7:
        testw   %dx, %dx
        je      .L32
        incl    %edx
        movl    %r10d, %esi
        movb    $1, %r9b
        jmp     .L7
.L32:
        testl   %esi, %esi
        je      .L6
        movb    $1, %dil
        xorl    %ecx, %ecx
.L6:
        decl    %eax
        movb    $1, %r8b
        jmp     .L2
.L31:
        testb   %r8b, %r8b
        je      .L10
        movl    $-7, d(%rip)
        testb   %dil, %dil
        je      .L11
        movl    %ecx, e(%rip)
.L11:
        xorl    %eax, %eax
        movl    %eax, c(%rip)
.L12:
        testb   %r9b, %r9b
        je      .L13
        movw    %dx, f(%rip)
        movl    %esi, g(%rip)
        jmp     .L13
.L10:
        testb   %dil, %dil
        je      .L12
        movl    %ecx, e(%rip)
        jmp     .L12
.L13:
        xorl    %eax, %eax
        ret


than gcc-13.2.0 -Os:

main1:
        movl    $-7, d(%rip)
        xorl    %eax, %eax
        movl    %eax, c(%rip)
        ret



Bisects to r14-569-g21e2ef2dc25

[Bug tree-optimization/112545] [14 Regression] Size regression when using -Os starting with r14-569-g21e2ef2dc25

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112545

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

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

[Bug tree-optimization/112545] [14 Regression] Size regression when using -Os starting with r14-569-g21e2ef2dc25

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112545

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2023-11-15
             Status|UNCONFIRMED                 |NEW

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
The first difference is that the trunk deletes:
  _6 = (int) _5;

From ccp2.

And then in forwprop2 changes:
```
  _4 = _2 ^ _3;
  _5 = (short unsigned int) _4;
  _6 = (int) _5;
  c.a = 0;
  if (_5 == 0)
```

into:
```
  c.a = 0;
  if (_2 == _3)
```
Which is correct.


The problem DOM/ranger missed the above. It understood the ^ code for some
reason but not the == code.
And then does not remove the better ranges.
Someone who knows the ranger more than I do should look into it.

Anyways here is the code which shows my change just exposed the missed
optimization (13.2 also misses the issue):
```
static struct {
    int a;
    int b;
} c;
static int d, e, g;
static short f, i;

int main1() {
    d = 6;
    for (; d != -7; d--) {
        c.a = i;
        _Bool t = d <= 0;
        signed short tt = d;
        signed short ttt = t;
        if (tt == ttt)
          __builtin_unreachable();
        //h(d ^ d < (0 <= 6));
        if (d >= 12)
            if (e) {
                for (; f; ++f) g = c.b;
                if (g) e = 0;
            }
    }
}

```

[Bug tree-optimization/112545] [14 Regression] Size regression when using -Os starting with r14-569-g21e2ef2dc25

[law at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112545

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/112545] [14 Regression] Size regression when using -Os starting with r14-569-g21e2ef2dc25

[law at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112545

--- Comment #2 from Jeffrey A. Law <law at gcc dot gnu.org> ---
This looks like a missed case for singletons.

  _1 = d.9_28 <= 0;
  _2 = (signed short) _1;
  _3 = (signed short) d.9_28;
  c.a = 0;
  if (_2 == _3)
    goto <bb 4>; [0.00%]
  else
    goto <bb 5>; [100.00%]


Back substituting the condition is:

(d.9_28 <= 0) == d.9_28

d.9_28 <= 0 has a range [0,1]

The only value that can make that expression true is if d.9_28 has the value 0.

[Bug tree-optimization/112545] [14/15 Regression] Size regression when using -Os starting with r14-569-g21e2ef2dc25

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112545

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

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

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
GCC 14.1 is being released, retargeting bugs to GCC 14.2.