https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101912
Aldy Hernandez <aldyh at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |aldyh at gcc dot gnu.org,
| |rguenth at gcc dot gnu.org
--- Comment #2 from Aldy Hernandez <aldyh at gcc dot gnu.org> ---
ISTM that the warning is correct, given the IL the uninit pass gets.
On the 7->3 entry to the loop prevcorr_23 can be undefined, and it is read in
BB4 with no guard whatsoever:
<bb 7> [local count: 85931698]:
<bb 3> [local count: 1034442873]:
# prevcorr_23 = PHI <corr_14(8), prevcorr_12(D)(7)>
# leapcnt_24 = PHI <leapcnt_15(8), 0(7)>
corr_14 = getint ();
if (corr_14 <= 0)
goto <bb 10>; [3.66%]
else
goto <bb 4>; [96.34%]
<bb 10> [local count: 37860610]:
goto <bb 6>; [100.00%]
<bb 4> [local count: 996582262]:
_1 = corr_14 == 1;
_2 = leapcnt_24 != 0;
_7 = (unsigned int) prevcorr_23;
_5 = _7 + 4294967295;
_4 = _5 > 1;
_22 = _1 & _2;
_18 = _4 & _22;
if (_18 != 0)
goto <bb 11>; [1.21%]
else
goto <bb 5>; [98.79%]
However, in the original source we shouldn't read from prevcorr if leapcnt ==
0, which is the case on entry to the loop:
&& !(leapcnt == 0
|| (prevcorr < corr
? corr == prevcorr + 1
: (corr == prevcorr
|| corr == prevcorr - 1)))))
It looks like we lost the leapcnt gate protecting the prevcorr_23 read.
Unless I'm missing some subtle language thinggie, we loose the gate in the
ifcombine pass. Before it, the read from prevcorr was protected by:
<bb 4> [local count: 996582262]:
_1 = corr_14 == 1;
_2 = leapcnt_5 != 0;
_3 = _1 & _2;
if (_3 != 0)
goto <bb 5>; [50.00%]
else
goto <bb 7>; [50.00%]
and after ifcombine we have:
<bb 4> [local count: 996582262]:
_1 = corr_14 == 1;
_2 = leapcnt_5 != 0;
_3 = _1 & _2;
_19 = prevcorr_4 != 1;
_16 = _3 & _19;
_21 = prevcorr_4 != 2;
_18 = _16 & _21;
if (_18 != 0)
goto <bb 7>; [1.21%]
else
goto <bb 5>; [98.79%]
Unfortunately, adding -fdisable-tree-ifcombine doesn't make the warning go
away, but at least the IL by the uninit pass seems correct.
