https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121192
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |needs-bisection
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
So vector lowering does the straight-forward component-based operation. At
-O0 we do not have the { -1, -1 } condition constant propagated and thus
it isn't folded. We turn
VEC_COND_EXPR <_1, { 0, 0 }, { 4294967295, 4294967295 }>
into
_9 = BIT_FIELD_REF <_1, 32, 0>;
_10 = _9 != 0;
_11 = _10 ? 0 : 4294967295;
and the COND_EXPR is what gets simplified wrongly.
unsigned int __GIMPLE (ssa,startwith("forwprop"))
foo (int a)
{
_Bool c;
unsigned int _1;
__BB(2):
c_3 = a_2(D) != 0;
_1 = c_3 ? 0u : 4294967295u;
return _1;
}
turns into (with -O -fgimple), triggering the same patterns:
<bb 2> :
_5 = a_2(D) == 0;
_6 = (unsigned int) _5;
_1 = -_6;
return _1;
Hmm, but that's all correct. Also the sparc veclower case - so it must
go wrong on the RTL side, possibly because of dead SSA defs?
