[Bug ipa/93940] [8/9/10 Regression] crazy codegen bug on Og with builtin_constant_p, coverage, and pthread

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93940

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hubicka at gcc dot gnu.org,
                   |                            |jakub at gcc dot gnu.org

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
This happens in edge_set_predicate:
271       if (predicate && *predicate == false
272           /* When handling speculative edges, we need to do the redirection
273              just once.  Do it always on the direct edge, so we do not
274              attempt to resolve speculation while duplicating the edge.  */
275           && (!e->speculative || e->callee))
276         e = redirect_to_unreachable (e);
I bet the inliner assumes that __builtin_constant_p will evaluate to true,
  D.2273.b = 0;
  k = D.2273;
...
  _8 = k.b;
  _9 = __builtin_constant_p (_8);
but while fre3 with -O2 propagates the 0 into __builtin_constant_p and folds it
to 1, with -Og we do fewer optimizations and don't propagate the constant to
it.
With -O2 -fno-tree-fre it will fail the same way.