[Bug tree-optimization/125120] Missed auto-vectorization/memcpy replacement in __builtin_unreachable() context

[amacleod at redhat dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125120

--- Comment #3 from Andrew Macleod <amacleod at redhat dot com> ---
(In reply to Richard Biener from comment #2)
> unreachable() should be gone before loop opts, not sure why that doesn't
> happen.  LIM isn't hindered by multi-exits in principle, but we restrict
> store motion
> of conditionally executed stores - because we have to emit a load in the
> preheader.  We could possibly do flag-based SM to avoid this (but usually
> not a good idea due to cost).
> 
> Hmm, but indeed, we seem to remove __builtin_unreachable () from the 2nd
> VRP pass only.  Why's that so?  We did it very much earlier in the past.


We can only remove it if we can safely update global values to reflect the
conditional knowledge everywhere it is used.

It will always be removed by VRP2 now, but it will be removed earlier if it can
be proven safe to do so.   I think it has always been this way.

I went back to GCC11, and  :
void foo (void);

int
f (int i, int j)
{
  if (j == 1)
    {
      if (i > 1)
        __builtin_unreachable();
      if (i > 1)
        foo ();
    }
  if (i == 2)
    foo ();
  return i == 0;
}

Allows __builtin_unreachable to go past vrp2 by a few passes before it is just
dropped.