[Bug tree-optimization/111167] New: swapping around duplicated conditionals can produce better code

[Bug tree-optimization/111167] New: swapping around duplicated conditionals can produce better code

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111167

            Bug ID: 111167
           Summary: swapping around duplicated conditionals can produce
                    better code
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: pinskia at gcc dot gnu.org
  Target Milestone: ---

Take reduced from GCC's sources for lookup_attribute:
```
void f(int);
inline void canonicalize_attr_name(const char *&s, int l) {
  if (l > 4 && s[0] == '_' && s[l - 1] == '_' ) 
    s += 2;
}
void lookup_attribute(const char *attr_ns, int list, int t) {
  if (attr_ns && attr_ns[0] != '_') {
    canonicalize_attr_name(attr_ns, 5);
  }
  if (list == 0)
  {
    short attr_ns_len = attr_ns ? t : 0;
    f(attr_ns_len);
  }
}

```

In optimized we have:
```
  if (attr_ns_25(D) != 0B)
    goto <bb 4>; [70.00%]
  else
    goto <bb 3>; [30.00%]

  <bb 3> [local count: 322122543]:
  if (list_10(D) == 0)
    goto <bb 6>; [50.00%]
  else
    goto <bb 7>; [50.00%]

  <bb 4> [local count: 751619279]:
  if (list_10(D) == 0)
    goto <bb 5>; [50.00%]
  else
    goto <bb 7>; [50.00%]

```

But we should really just have:
if (list_10(D) == 0) goto bb7;
if (attr_ns_25(D) != 0B) goto bb5; else bb6;

[Bug tree-optimization/111167] swapping around duplicated conditionals can produce better code

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111167

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement

[Bug tree-optimization/111167] swapping around duplicated conditionals can produce better code

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111167

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
In this case the problem seems to be jump threading related.

If we did:
```
void f(int);
void lookup_attribute1(const char *attr_ns, int list, int t) {
  short t1;
  if (attr_ns)  { if (list) {t1 = 0; goto L1;} return; }
  if (list){t1 = t;goto L1;} else return;
L1:
  f(t1);
}
```

I Notice clang does not do this either even though the original testcase seems
ok.

[Bug tree-optimization/111167] swapping around duplicated conditionals can produce better code

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111167

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
>Take reduced from GCC's sources for lookup_attribute:

Note it shows up even not in a reduced version of lookup_attribute too.