https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114324
--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> ---
With -Ofast there's still grouped stores we fail two-lane SLP discovery for,
possibly due to re-association. In particular for example
_54 = REALPART_EXPR <(*a_19(D))[_27]>;
_55 = IMAGPART_EXPR <(*a_19(D))[_27]>;
_56 = REALPART_EXPR <(*a_19(D))[_30]>;
_57 = IMAGPART_EXPR <(*a_19(D))[_30]>;
_60 = _54 - _56;
_8 = _57 - _55;
cannot be discovered, we fail to consider swapping operands of a minus by
negating it in the parent (of its parent)
b2$real_63 = _8 * ci$imag_10;
b2$imag_64 = ci$imag_10 * _60;
_71 = a2$real_52 - b2$real_63;
_72 = a2$imag_53 - b2$imag_64;
the parent (of its parent) would then turn into a plusminus operation.
The less efficient way to swap operands in the {_60, _8} compute would
be to insert a conditional negate - either via negate + merge or via
multiplication by { 1, -1 }, for FP a lane-speific negate can be
implemented as XOR. This might be more efficient for targets without
a native plusminus op.
Even simply not giving up when running into two different DR groups
would be possible - we can insert an interleaving operation.
That said, I think we have accumulated some duplicates around SLP discovery
issues for _Complex ops after re-association.
