[Bug middle-end/108992] Regression: Branch direction canonicalization leads to pointless tail duplication / CSE/sinking by inverting branch

[pinskia at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108992

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |INVALID
             Status|UNCONFIRMED                 |RESOLVED

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
I suspect this is not a bug, GCC tries to optimze the fast path into straight
line code without any waste of space.
In the first case GCC predicts that the cond is going to be true 66% of the
time because there is comparison against 0 prediction going in the heurstics.

For the first testcase if you do:
void use(int *);
void use2(int *);

void foo(bool cond, int * p)
{
    if (__builtin_expect(cond, 1)) {
        use(p);
    }
    use2(p);
}

Then you get the result you want.
Adding the builtin_expect for the second case you get the same too.
Basically GCC is pushing what it thinks as cold code away from the original
path.