------- Comment #3 from rguenth at gcc dot gnu dot org 2009-03-06 13:36 -------
Confirmed with the reduced testcase only, at -O2 -Wall. The warning happens
after PTA triggered by SRA.
Relevant constraints:
__y_7 = &s.64+64
D.2839_21 = __y_7
__x.8_43 = D.2839_21
and code:
D.2839_21 = &__y_7->D.2545;
D.2832_22 ={v} &0B->D.2545;
if (D.2832_22 != 0B)
goto <bb 6>;
else
goto <bb 4>;
<bb 4>:
D.2888_42 = &__y_7->D.2545;
if (D.2839_21 == D.2888_42)
goto <bb 6>;
else
goto <bb 5>;
<bb 5>:
__x.8_43 = (const struct _Rb_tree_node *) D.2839_21;
D.2878_45 = __x.8_43->_M_value_field;
which shows that the warning is for dead code. Thus this is not a
wrong-code problem.
VRP could figure out the equivalencies but doesn't. The next DOM
pass figures out the second but not the first non-equivalency - that
survives until the next forwprop pass.
In the end the problem is that VRP doesn't do a good job cleaning up
after itself (there are single-argument PHIs left after it as well as
non-simplified comparisons):
<bb 3>:
Invalid sum of incoming frequencies 900, should be 10000
# __y_8 = PHI <0B(2)>
# __y_18 = PHI <__y_7(2)>
...
<bb 4>:
Invalid sum of incoming frequencies 3900, should be 102
D.2837 = 1;
D.2839_21 = &__y_18->D.2545;
D.2832_22 ={v} &0B->D.2545;
if (D.2832_22 != 0B)
goto <bb 8>;
else
goto <bb 5>;
<bb 5>:
D.2888_42 = &__y_7->D.2545;
if (D.2839_21 == D.2888_42)
goto <bb 8>;
else
goto <bb 6>;
--
rguenth at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Ever Confirmed|0 |1
Keywords|wrong-code |missed-optimization
Last reconfirmed|0000-00-00 00:00:00 |2009-03-06 13:36:52
date| |
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39390
