[Bug tree-optimization/90385] [9/10 Regression] ICE: tree check: expected ssa_name, have real_cst in transform_to_exit_first_loop_alt, at tree-parloops.c:1772

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90385

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |jakub at gcc dot gnu.org

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 46322
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=46322&action=edit
gcc10-pr90385.patch

Untested fix.  In PR90211 I thought but without serious analysis that we handle
those invariant exit PHI args fine, but as this testcase shows, that would
require more work.
So this modification of the PR90211 change just punts on those, if somebody has
cycles and important use case (not -fno-* testcases), the support can always be
added.