[Bug middle-end/16458] PowerPC - redundant compare

[bergner at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16458

--- Comment #5 from Peter Bergner <bergner at gcc dot gnu.org> 2011-12-14 22:42:57 UTC ---
Created attachment 26091
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26091
Another partial fix...

Answering Nathan's comment from his patch:

+    /* These are unsigned values, perhaps there will be a later
+       ordering compare that can be shared with this one.
+       Unfortunately we cannot detect the signedness of the operands
+       for non-subregs.  */

Actually, there is a way to tell whether some of our non-subreg operands are
unsigned or not.  For register operands of user defined variables:

  TYPE_UNSIGNED (TREE_TYPE (REG_EXPR (rtx)))

will be true if the operand is unsigned.  I've attached a patch that implements
that and it does fix the test case.  It also fixes a problem I encountered with
switch statements and -fno-jump-tables:

  enum CASE_VALUES {
    CASE0 = 0,
    CASE1,
    CASE2,
    CASE3,
    CASE4
  };

  int
  test_switch (enum CASE_VALUES index)
  {
    switch (index)
      {
        case CASE0: return case0 ();
        case CASE1: return case1 ();
        case CASE2: return case2 ();
        case CASE3: return case3 ();
        case CASE4: return case4 ();
        default:    return cased ();
      }
  }

Currently, GCC generates (-fno-jump-tables) the following, because it notices
that the enum contains no negative values, so it marks index as unsigned and we
end up with unsigned compares (modulo the equality compares) so we cannot CSE
them:

    test_switch:
        cmpwi 7,3,2
        beq 7,.L5
        cmplwi 7,3,2
        ble 7,.L10
        cmpwi 7,3,3
        beq 7,.L6
        ...

With the attached patch, we generate all unsigned compares and CSE is able to
remove the redundant compares:

    test_switch:
        cmplwi 7,3,2
        beq 7,.L5
        ble 7,.L10
        cmplwi 7,3,3
        beq 7,.L6
        ...

The patch still does not handle:

  if (*a == *b)
    ...
  else if (*a > *b)
    ...

or:

  switch (*index)
  ...

but we couldn't handle those before anyway.