[Bug tree-optimization/106912] [13 Regression] ICE in vect_transform_loops, at tree-vectorizer.cc:1032 since r13-1575-gcf3a120084e94614

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106912

Richard Biener <rguenth at gcc dot gnu.org> changed:

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

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #6)
> Sure, the FUNCTION_TYPE can be shared, so shouldn't be overwritten in place,
> but can be copied with TREE_READONLY cleared on the copy.
> So, if we for whatever reason need to clear const flag on some functions,
> after clearing TREE_READONLY on the node->decl we'd also need to update the
> TREE_TYPE to a version without TREE_READONLY set (so that say if new calls
> are created they don't get wrong fntype) and then walk all call edges to
> that function, updating gimple_call_fntype to the updated type if it was
> equal to the old TREE_TYPE, or to a copy with TREE_READONLY cleared if it
> has TREE_READONLY set on the FUNCTION/METHOD_TYPE).

That will not catch indirect calls to the function.  So whatever we do
this part of the problem will remain (the out-of-sync profile).  Like

inline int __attribute__((const))
foo (int i)
{
  if (i == 0)
    return 2;
  return 1;
}

int (* __attribute__((const)) bar) () = foo;

int main()
{
  int r = 0;
  for (int i = 0; i < 10000; ++i)
    {
      r += foo (r);
      r += bar (r);
    }
  return r;
}

which, when compiled with -O2 -fno-early-inlining -fprofile-arcs results
in the edge counters from the inline copy via foo to be store-motion'ed
but the indirect call clobbering them each iteration.

Using -fprofile-use prevents this because we get an indirect call profiler
call which clobbers all memory, confusing LIM, doing early inlining will
early inline the function before profile instrumentation.

I cannot see how we can fix this without some global flag to ignore
const/pure-ness of functions or alternatively dropping DECL_NONALIASED.
Trying to support surgical un-setting of pure/const for single functions
isn't going to work.  Simply disabling IPA inlining for instrumented
builds might also work.