[Bug tree-optimization/25243] Jump threading opportunity missed in tree-ssa but caught in jump1

[steven at gcc dot gnu dot org]

--- Comment #1 from steven at gcc dot gnu dot org  2005-12-03 14:22 ---
I should have said, this is at -O1 -fthread-jumps.  I guess VRP catches this
at -O2 and better.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25243

[Bug tree-optimization/25243] Jump threading opportunity missed in tree-ssa but caught in jump1

[steven at gcc dot gnu dot org]

--- Comment #2 from steven at gcc dot gnu dot org  2005-12-03 14:37 ---
Actually VRP doesn't catch it.

Do:
-if (e[i] == 16)
+if (e[i] == 16)
so that store-CCP doesn't load e[0] anymore to find that it is 16.  With that,
the .vrp dump at -O2 looks like this:

baz (r)
{
  long unsigned int i;
  int D.1638;
  long unsigned int i.0;

bb 0:
  goto bb 3 (L2);

L0:;
  i_4 = i_1;
  D.1638_6 = e[i_1];
  if (D.1638_6 == 17) goto L3; else goto L1;

L1:;
  i_7 = i_1 + 1;

  # i_1 = PHI 0(0), i_7(2);
L2:;
  if (i_1 = 3) goto L0; else goto L3;

L3:;
  if (i_1 == 4) goto L4; else goto L5;

L4:;
  foo (r_3);

L5:;
  return;

}

The tree loop optimizers turn (i_1 = 3) into (i_1 != 4), but there are no
passes after VRP that use the ASSERT_EXPRs to propagate i_1 == 4 down along
the false-edge going to the block starting with label L3.

So even at -O2 we don't catch this jump threading opportunity at the tree
level.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25243

[Bug tree-optimization/25243] Jump threading opportunity missed in tree-ssa but caught in jump1

[steven at gcc dot gnu dot org]

--- Comment #3 from steven at gcc dot gnu dot org  2005-12-03 15:46 ---
With a minor hack, we optimize the test case in dom3:

Index: tree-ssa-dom.c
===
--- tree-ssa-dom.c  (revision 107822)
+++ tree-ssa-dom.c  (working copy)
@@ -2776,7 +2776,9 @@ cprop_operand (tree stmt, use_operand_p
   /* If the operand has a known constant value or it is known to be a
  copy of some other variable, use the value or copy stored in
  CONST_AND_COPIES.  */
-  val = SSA_NAME_VALUE (op);
+  val = op;
+  while (TREE_CODE (val) == SSA_NAME  SSA_NAME_VALUE (val))
+val = SSA_NAME_VALUE (val);
   if (val  val != op  TREE_CODE (val) != VALUE_HANDLE)
 {
   tree op_type, val_type;


Jeff, thoughts?


-- 

steven at gcc dot gnu dot org changed:

   What|Removed |Added

 Status|UNCONFIRMED |NEW
 Ever Confirmed|0   |1
   Last reconfirmed|-00-00 00:00:00 |2005-12-03 15:46:50
   date||


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25243