------- Comment #7 from jakub at gcc dot gnu dot org 2009-05-05 10:01 -------
Introduced during tuples merge apparently (haven't bisected the tuples branch
though).
To me this looks like a phiprop bug.
In *.alias (trunk, -O2) we have:
# BLOCK 3 freq:9100
# PRED: 4 [91.0%] (true,exec)
# VUSE <.MEMD.2716_19>
D.1614_8 = *cD.1606_1;
cD.1606_9 = &D.1614_8->aD.1593;
# SUCC: 4 [100.0%] (fallthru,dfs_back,exec)
# BLOCK 4 freq:10000
# PRED: 2 [100.0%] (fallthru,exec) 3 [100.0%] (fallthru,dfs_back,exec)
# cD.1606_1 = PHI <cD.1606_4(2), cD.1606_9(3)>
# VUSE <.MEMD.2716_19>
D.1614_7 = *cD.1606_1;
if (D.1614_7 != 0B)
goto <bb 3>;
else
goto <bb 5>;
# SUCC: 3 [91.0%] (true,exec) 5 [9.0%] (false,exec)
# BLOCK 5 freq:900
# PRED: 4 [9.0%] (false,exec)
# .MEMD.2716_20 = VDEF <.MEMD.2716_19>
D.1615_11 = fooD.1597 (yD.1602_10(D));
# .MEMD.2716_21 = VDEF <.MEMD.2716_20>
*cD.1606_1 = D.1615_11;
goto <bb 7>;
# SUCC: 7 [100.0%] (fallthru,exec)
# BLOCK 6 freq:9100
# PRED: 7 [91.0%] (true,exec)
# VUSE <.MEMD.2716_21>
D.1614_13 = *cD.1606_2;
cD.1606_14 = &D.1614_13->aD.1593;
# SUCC: 7 [100.0%] (fallthru,dfs_back,exec)
# BLOCK 7 freq:10000
# PRED: 5 [100.0%] (fallthru,exec) 6 [100.0%] (fallthru,dfs_back,exec)
# cD.1606_2 = PHI <cD.1606_1(5), cD.1606_14(6)>
# VUSE <.MEMD.2716_21>
D.1614_12 = *cD.1606_2;
if (D.1614_12 != 0B)
goto <bb 6>;
else
goto <bb 8>;
# SUCC: 6 [91.0%] (true,exec) 8 [9.0%] (false,exec)
Note that the D.1614_12 = *cD.1606_2 read depends on the *cD.1606_1 =
D.1615_11;
store. In phiprop we have though:
# BLOCK 3 freq:9100
# PRED: 4 [91.0%] (true,exec)
cD.1606_9 = &D.1614_8->aD.1593;
# VUSE <.MEMD.2716_19>
D.2745_25 = D.1614_8->aD.1593;
# SUCC: 4 [100.0%] (fallthru,dfs_back,exec)
# BLOCK 4 freq:10000
# PRED: 2 [100.0%] (fallthru,exec) 3 [100.0%] (fallthru,dfs_back,exec)
# cD.1606_1 = PHI <cD.1606_4(2), cD.1606_9(3)>
# D.1614_8 = PHI <D.2744_24(2), D.2745_25(3)>
D.1614_7 = D.1614_8;
if (D.1614_7 != 0B)
goto <bb 3>;
else
goto <bb 5>;
# SUCC: 3 [91.0%] (true,exec) 5 [9.0%] (false,exec)
# BLOCK 5 freq:900
# PRED: 4 [9.0%] (false,exec)
# .MEMD.2716_20 = VDEF <.MEMD.2716_19>
D.1615_11 = fooD.1597 (yD.1602_10(D));
# .MEMD.2716_21 = VDEF <.MEMD.2716_20>
*cD.1606_1 = D.1615_11;
goto <bb 7>;
# SUCC: 7 [100.0%] (fallthru,exec)
# BLOCK 6 freq:9100
# PRED: 7 [91.0%] (true,exec)
cD.1606_14 = &D.1614_13->aD.1593;
# VUSE <.MEMD.2716_21>
D.2746_26 = D.1614_13->aD.1593;
# SUCC: 7 [100.0%] (fallthru,dfs_back,exec)
# BLOCK 7 freq:10000
# PRED: 5 [100.0%] (fallthru,exec) 6 [100.0%] (fallthru,dfs_back,exec)
# cD.1606_2 = PHI <cD.1606_1(5), cD.1606_14(6)>
# D.1614_13 = PHI <D.1614_8(5), D.2746_26(6)>
D.1614_12 = D.1614_13;
if (D.1614_12 != 0B)
goto <bb 6>;
else
goto <bb 8>;
# SUCC: 6 [91.0%] (true,exec) 8 [9.0%] (false,exec)
Note that the D.1614_13 PHI now uses D.1614_8, which is correct *c content
only if the *cD.1606_1 = D.1615_11; store wasn't performed. This store
obviously can change what *c contains. Later passes then find out that
D.1614_8 must be 0 to reach BB5 and BB7, so the whole second while loop is
removed.
--
jakub at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
Summary|[4.4 Regression] breaks |[4.4/4.5 Regression] Alpine
|"reply to all" in alpine |miscompilation
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40022
