https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117574
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Priority|P3 |P2
--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
for (g = 40; g != 124; g += 50)
.optimized has
d.3_5 = d; // 0
_6 = (int) d.3_5;
_9 = _6 + 40;
_10 = (long int) _9; // 40
_43 = (unsigned int) d.3_5;
_23 = _43 + 41;
_29 = (int) _23;
_50 = _43 + 76;
_39 = (int) _50;
_18 = _29 >= _39; // 41 >= 76
_37 = _18 ? 1 : 368934881474191034; // 368934881474191034
_42 = (unsigned long) d.3_5;
_41 = _37 * 50; // 84
_17 = _42 + 40;
_38 = _17 + _41; // 124
_46 = (long int) _38;
<bb 4> [local count: 955630224]:
# g_33 = PHI <g_21(4), _10(3)>
# b_lsm.12_28 = PHI <_20(4), b_lsm.12_30(3)>
_20 = b_lsm.12_28 + g_33;
g_21 = g_33 + 50;
if (g_21 != _46)
goto <bb 4>; [89.00%]
it is IVCANON replacing the inner loop IV and thus in the end this looks like
a niter issue. Before IVCANON we have
d.3_5 = d;
_6 = (int) d.3_5;
_7 = _6 + 76;
_8 = (long int) _7;
_9 = _6 + 40;
_10 = (long int) _9;
<bb 6> [local count: 955630224]:
# g_33 = PHI <g_21(15), _10(13)>
# b_lsm.12_28 = PHI <_20(15), b_lsm.12_30(13)>
_20 = b_lsm.12_28 + g_33;
g_21 = g_33 + 50;
if (_8 >= g_21) // 76 >= { 40, +, 50 }
goto <bb 15>; [89.00%]
else
goto <bb 17>; [11.00%]
<bb 15> [local count: 850510900]:
goto <bb 6>; [100.00%]
so we wrongly replaced a >= IV check with a != one, the generated code suggests
a flipped condition somewhere:
_18 = _29 >= _39; // 41 >= 76
_37 = _18 ? 1 : 368934881474191034; // 368934881474191034
with _18 = _29 <= _39 it would have been correct.
