[Bug rtl-optimization/103006] [9/10/11/12 Regression] wrong code at -O1 or -O2 on x86_64-linux-gnu by r7-7101

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103006

--- Comment #14 from Richard Biener <rguenth at gcc dot gnu.org> ---
There's an interesting case,

  a = BIRTH
loop:
  b = DEATH
  a = DEATH
  b = BIRTH
  goto loop;

where we end up having both a and b in the live-in set at the loop label
but a is removed before we see the BIRTH of b which is where we add
conflicts based on the current set of active vars.

In the case I'm running into this I have tail-recursion do

  a = BIRTH
  b = BIRTH
...
  a = DEATH
  b = DEATH

into

loop:
  a = BIRTH
  b = BIRTH
  goto loop;
  a = DEATH
  b = DEATH

leading to a similar issue.  The issue above can for example arise from
loop rotation.

In all cases live from backedges confuse the "optimization" done to only
record conflicts when we add a var to the live set (and it is not already set).

The previous code had

              /* If this is the first real instruction in this BB we need
                 to add conflicts for everything live at this point now.
                 Unlike classical liveness for named objects we can't
                 rely on seeing a def/use of the names we're interested in.
                 There might merely be indirect loads/stores.  We'd not add any
                 conflicts for such partitions.  */

and the easiest is to do the same here (we don't see the backedge "use"),
but we could possibly improve by noting which vars are only live from
a backedge here.