[Bug rtl-optimization/42839] [4.5 Regression] gcc.target/mips/octeon-bbit-2.c failing for -mabi=64

[rguenther at suse dot de]

------- Comment #11 from rguenther at suse dot de  2010-02-13 12:07 -------
Subject: Re:  [4.5 Regression] gcc.target/mips/octeon-bbit-2.c
 failing for -mabi=64

On Sat, 13 Feb 2010, steven at gcc dot gnu dot org wrote:

> 
> 
> ------- Comment #10 from steven at gcc dot gnu dot org  2010-02-13 11:39 -------
> In the test case of comment #2, the history of the funny PHIs is really odd. At
> -O2 we end with this in the .optimized dump:
> 
> <bb 3>:
>   # i_1 = PHI <0(2)>
>   # s_11 = PHI <0(2)>
> 
> <bb 4>:
>   # i_12 = PHI <i_1(3), i_7(4)>
>   # s_13 = PHI <s_11(3), s_6(4)>
> 
> 
> But the history of the phi of i_12 looks like this:
> 
> t.c.120t.reassoc2:  # i_12 = PHI <i_7(3), 0(2)>
> t.c.121t.vrp2:  # i_12 = PHI <i_2(3), i_7(7)>
> t.c.121t.vrp2:  # i_12 = PHI <0(3), i_7(4)>
> t.c.122t.dom2:  # i_12 = PHI <0(3), i_7(4)>
> t.c.123t.phicprop2:  # i_12 = PHI <0(2), i_7(3)>
> t.c.124t.cddce2:  # i_12 = PHI <i_1(3), i_7(6)>
> 
> Somehow we have re-introduced the special PHI "# i_1 = PHI <0(2)>" in the
> .cddce2 pass. The reason is that we created a loop pre-header.

Yep.  I wonder if cfgcloop should just not do that, and of course
I wonder why cfgcleanup doesn't get rid of single-arg PHIs (well,
probably to not break loop-closed SSA form).

Btw, the following should avoid almost all constant propagations
into PHIs (well, I guess DOM needs fixing as well)

Index: tree-ssa-propagate.c
===================================================================
--- tree-ssa-propagate.c        (revision 156666)
+++ tree-ssa-propagate.c        (working copy)
@@ -924,7 +924,9 @@ replace_phi_args_in (gimple phi, prop_va
        {
          tree val = prop_value[SSA_NAME_VERSION (arg)].value;

-         if (val && val != arg && may_propagate_copy (arg, val))
+         if (val && val != arg
+             && TREE_CODE (val) == SSA_NAME
+             && may_propagate_copy (arg, val))
            {
              if (TREE_CODE (val) != SSA_NAME)
                prop_stats.num_const_prop++;


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42839