https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111003
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |RESOLVED
Resolution|--- |FIXED
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
Of course that's not really a "fix", but it's restoring of not doing the extra
hoisting. So it does fix the regression but not maybe the underlying issue.
We're not anticipating more imprecise range-info on DEFs when hoisting here.
The "real" fix would be to improve conditional hoisting - we have
<bb 12> [local count: 29170419070]:
# prephitmp_61 = PHI <pretmp_60(11), prephitmp_59(10)>
_13 = (unsigned int) _3;
if (_3 > 6)
goto <bb 73>; [50.00%]
else
goto <bb 13>; [50.00%]
<bb 73> [local count: 14585209535]:
goto <bb 14>; [100.00%]
<bb 13> [local count: 14585209535]:
iftmp.9_18 = _13 << 6;
_63 = _13 ^ iftmp.9_18;
_65 = _63 != 10;
<bb 14> [local count: 35145083376]:
# prephitmp_66 = PHI <1(73), _65(13), 1(74)>
# prephitmp_67 = PHI <prephitmp_61(73), prephitmp_61(13), prephitmp_59(74)>
and the diamond with if (_3 > 6) including the precceding def of _13
(which wasn't sunk) is invariant. There's an existing bug about LIM
missing a way to hoist a sub-CFG (it can only hoist PHIs as COND_EXPRs).
It might not exactly fit "LIM" but could be integrated with this pass.
I'd say we close this bug, I'll add the testcase.