[Regression] LICM not duplicating instruction exposes code size regressions at -Oz

[dcci]

Reduced testcase:

int f(int* a, int b, int c, int size) {
  union {
    int value;
    struct {
      int flag : 1;
      int data : 31;
    } parts;
  } u;

  for (int i = 0; i < size; i++) {
    if (a[i] > 0 && b < 10) {
      if (c > 5 && a[i] % 3 == 0) {
        c += a[i];
        a[i] = 0;
        u.value = c;
        if (u.parts.flag) {
          b += 2;
        }
        if (c > 10) {
          u.parts.flag = -1;
        }
        if (b > 5 && (c % 4 == 0 || c % 7 == 0)) {
          a[i] = c - b;
          if ((c - b) % 2 == 0 && (c - b) > 5) {
            a[i + 1] = (c - b) * 3;
          }
          if (a[i] > 10) {
            a[i] = a[i] / 2;
          }
        }
        // Additional union code usage
        u.parts.data = a[i] * 2;
        a[i] = u.value;
        if (u.parts.flag) {
          a[i] += 1;
        }
      }
      return 0;
    }
  }
  return (c % 2 == 0) ? 0 : 1;
}

Size after/before:

[davidino@alcor]$ size /tmp/with 
   text    data     bss     dec     hex filename
    287       0       0     287     11f /tmp/with

[davidino@alcor]$ size /tmp/without 
   text    data     bss     dec     hex filename
    264       0       0     264     108 /tmp/without

This bisects to:

commit 0659000ff79decc1173aac140d4b0325fe696c57 (HEAD)
Author: Nikita Popov <npopov@redhat.com>
Date:   Wed Apr 26 10:00:42 2023 +0200

    [LICM] Don't duplicate instructions just because they're free

Unclear to me if this is the real regression or it's exposing something else, but I wanted to report anyway.

@nikic