http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60363
--- Comment #6 from bin.cheng <amker.cheng at gmail dot com> ---
After investigation, I think the root cause is:
For given mergephi2 dump (before vrp), there are latch(bb13) and header(bb14)
of a loop:
<bb 13>:
# changed_18 = PHI <changed_26(12), changed_1(7), changed_1(8), changed_1(9),
changed_1(11)>
<bb 14>:
# changed_1 = PHI <changed_18(13), 0(2)>
# kill_elt_4 = PHI <kill_elt_3(13), kill_elt_7(D)(2)>
if (a_elt_9(D) != 0B)
goto <bb 3>;
else
goto <bb 15>;
The latch is removed by the bogus patch, and the code turns into:
<bb 14>:
# changed_1 = PHI <changed_1(7), 0(2), changed_26(12), changed_1(11),
changed_1(9), changed_1(8)>
# kill_elt_4 = PHI <kill_elt_3(7), kill_elt_7(D)(2), kill_elt_3(12),
kill_elt_3(11), kill_elt_3(9), kill_elt_3(8)>
if (a_elt_9(D) != 0B)
goto <bb 3>;
else
goto <bb 15>;
Since VRP requires LOOP_HAVE_SIMPLE_LATCH, it will be initialized into:
<bb 14>:
# changed_19 = PHI <changed_1(7), changed_1(22), changed_26(12),
changed_1(11), changed_1(9)>
# kill_elt_18 = PHI <kill_elt_3(7), kill_elt_33(22), kill_elt_32(12),
kill_elt_32(11), kill_elt_32(9)>
<bb 17>:
# changed_1 = PHI <changed_19(14), 0(2)>
# kill_elt_4 = PHI <kill_elt_18(14), kill_elt_7(D)(2)>
if (a_elt_9(D) != 0B)
goto <bb 24>;
else
goto <bb 15>;
After all works in vrp (including jump threading), the code is transformed into
below form:
<bb 18>:
# kill_elt_2 = PHI <kill_elt_10(5), kill_elt_14(4)>
if (kill_elt_2 != 0B)
goto <bb 6>;
else
goto <bb 21>;
<bb 21>:
goto <bb 27>;
...
<bb 14>:
# changed_19 = PHI <changed_1(7), changed_1(22), changed_26(12),
changed_1(11), changed_1(9), changed_1(27), changed_1(29)>
# kill_elt_18 = PHI <kill_elt_3(7), 0B(26), kill_elt_3(12), kill_elt_3(11),
kill_elt_3(9), kill_elt_3(27), kill_elt_3(29)>
<bb 17>:
# changed_1 = PHI <changed_19(14), 0(2)>
# kill_elt_4 = PHI <kill_elt_18(14), kill_elt_7(D)(2)>
if (a_elt_9(D) != 0B)
goto <bb 24>;
else
goto <bb 15>;
...
<bb 26>:
goto <bb 14>;
<bb 27>:
# kill_elt_41 = PHI <0B(21)>
if (b_elt_11(D) != 0B)
goto <bb 26>;
else
goto <bb 14>;
When trying to remove bb26, it finds out bb27 is pred of both bb26 and bb14,
then checks whether the phi args are consistent for both edges (<26, 14> and
<27, 14>). Apparently "0B(26)" != "kill_elt_3(27)", resulting in keeping bb27.
Actually, kill_elt_3(27) equals to "0", because of check condition in bb18.
So this might be a missed opportunity in vrp?
