[Bug tree-optimization/37102] [4.3/4.4 Regression] possible integer codegen bug

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

------- Comment #6 from rguenth at gcc dot gnu dot org  2008-08-28 14:48 -------
It looks like that we correctly track liveliness by

Live on entry to BB2 :

Live on entry to BB3 : a_lsm.27_5

Live on entry to BB4 : a_lsm.27_5

Live on entry to BB5 :

Live on entry to BB6 :

for

  # BLOCK 2 freq:2000
  # PRED: ENTRY [100.0%]  (fallthru,exec)
  c.0_4 = c;
  a_lsm.27_5 = b;
  a_lsm.27_1 = a;
  if (c.0_4 != 0)
    goto <bb 3>;
  else
    goto <bb 4>;
  # SUCC: 4 [50.0%]  (false,exec) 3 [50.0%]  (true,exec)

  # BLOCK 3 freq:1000
  # PRED: 2 [50.0%]  (true,exec)
  # SUCC: 4 [100.0%]  (fallthru,exec)

  # BLOCK 4 freq:2000
  # PRED: 3 [100.0%]  (fallthru,exec) 2 [50.0%]  (false,exec)
  # a_lsm.27_29 = PHI <a_lsm.27_5(3), a_lsm.27_1(2)>
  a = a_lsm.27_5;
  e_9 = a_lsm.27_5 << 1;
  if (e_9 != 0)
    goto <bb 5>;
  else
    goto <bb 6>;
  # SUCC: 5 [54.0%]  (true,exec) 6 [46.0%]  (false,exec)


and based on this create the conflict graph.  But we still (blindly)
generate copies for the (dead) PHI node in BB4 which causes conflicts
that we didn't see.

So, why exactly do we generate copies for a PHI node which result is
not live-in in its BB?  Andrew - you are probably most familiar with
the out-of-SSA code, can you have a look here?  (the dead PHI node is
caused by the last DSE pass which removes the last use of it and there
is no other DCE pass to clean that up).

Jakubs testcase from comment #3 at -O3.

My humble "fix" for this would be to exchange 118.cddce with 120.dse which
should hide this problem.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |amacleod at redhat dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37102