[Bug tree-optimization/108684] [13 Regression] ICE: verify_ssa failed

["amacleod at redhat dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108684

--- Comment #3 from Andrew Macleod <amacleod at redhat dot com> ---
OK, its been a while.   Why is there a VUSE on the return?  this is the IL
right from into-ssa:

int f (int a)
{
  int _4;

  <bb 2> :
  # .MEM_2 = VDEF <.MEM_1(D)>
  __asm__(" " : "=X" a_3 :  : "memory");
  if (a_3 != 0)
    goto <bb 3>; [INV]
  else
    goto <bb 4>; [INV]

  <bb 3> :
  _4 = 0;
  // predicted unlikely by early return (on trees) predictor.
  # VUSE <.MEM_2>
  return _4;

  <bb 4> :
  __builtin_unreachable ();

}

The problem is that in VRP, we change a_3's global range to be ~[0,0], rewrite
the condition to remove the block with the builtin_unreachable... which there
leaves the IL something like:

 <bb 2> [local count: 1073741824]:
  # .MEM_2 = VDEF <.MEM_1(D)>
  __asm__(" " : "=X" a_3 :  : "memory");

  <bb 3> [local count: 1073741824]:
  # VUSE <.MEM_2>
  return 0;

and since there are no more uses of a_3, the asm is considered dead and removed
by simple_dce_from_worklist.

BUt it does not adjust the vdefs at all.. so we when its deleted, we are still
eft with:
  <bb 3> [local count: 1073741824]:
  # VUSE <.MEM_2>
  return 0;

but now there is no longer a definition of MEM_2  and verify_ssa fails.

I can fix it by not trying to dce any defs that have VDEFS on them... but that
doesnt seem  like the correct long term solution... Why is there a VUSE on that
return anyway?  to keep it from being hoisted above the asm?

Whats the correct thing to do here?   Im surprised that somewhere in the stmt
deletion process, it doesnt rewrite all uses of MEM_2 to MEM_1 before deliting
it.   Or is that a manual step?