[Bug tree-optimization/41101] [4.4/4.5 Regression] ICE in compute_antic, at tree-ssa-pre.c:2419

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

------- Comment #11 from rguenth at gcc dot gnu dot org  2009-09-06 12:38 -------
We again have ANTIC_IN/OUT oscillation

ANTIC_OUT[] := { x_2 (0001), y_3 (0002), {pointer_plus_expr,x_2,4} (0011) }
ANTIC_IN[] := { x_2 (0001), y_3 (0002), {pointer_plus_expr,x_2,4} (0011) }
ANTIC_OUT[] := { x_2 (0001), y_3 (0002), {pointer_plus_expr,y_3,4} (0025) }
ANTIC_IN[] := { x_2 (0001), y_3 (0002), {pointer_plus_expr,y_3,4} (0025) }

for blocks 3, 9 and 6 which form a conditionally infinite loop.

<bb 2>:

<bb 3>:
  # x_2 = PHI <x_4(D)(2), y_3(6)>
  # y_3 = PHI <y_5(D)(2), x_2(6)>
  if (N_7(D) != 0)
    goto <bb 8>;
  else
    goto <bb 9>;

<bb 9>:
  goto <bb 6>;

<bb 8>:

<bb 4>:
  # i_10 = PHI <i_20(10), 1(8)>
  D.1986_8 = i_10 * 4;
  D.1987_9 = y_3 + D.1986_8;
  D.1988_11 = x_2 + D.1986_8;
  D.1989_12 = *D.1988_11;
  D.1990_13 = x_2 + 4;
  D.1991_14 = *D.1990_13;
  D.1992_15 = D.1989_12 - D.1991_14;
  D.1993_16 = func (D.1992_15);
  *D.1987_9 = D.1993_16;
  if (D.1993_16 != 0)
    goto <bb 7>;
  else
    goto <bb 5>;

<bb 5>:
  i_20 = i_10 + 1;
  if (N_7(D) >= i_20)
    goto <bb 10>;
  else
    goto <bb 11>;

<bb 10>:
  goto <bb 4>;

<bb 11>:

<bb 6>:
  goto <bb 3>;

<bb 7>:
  return;

The maximal set has only {pointer_plus_expr,x_2,4} (0011),
{pointer_plus_expr,y_3,4} (0025) is the result of phi translating
{pointer_plus_expr,y_3,D.1986_8} (0007).

Injecting that pair into the cycle via BB3 -> BB8 yields to the oscillation
as from BB9 we get oscillating 11 which will mask out the other.

It seems to me that whenever we encounter such a cycle as 3 -> 9 -> 6 -> 3
we may not inject the maximal set there as it may not really be the
maximal set (which we can't really enlarge during phi translation, can we?
It does fix this case obviously though).


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dberlin at gcc dot gnu dot
                   |                            |org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41101

------- You are receiving this mail because: -------
You reported the bug, or are watching the reporter.