https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116979
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jakub at gcc dot gnu.org
--- Comment #10 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
This changed in GCC 12 with r12-4240, with additional -ftree-vectorize flag
with r12-1912 aka PR95045 fix when vec_addsubv2sf3 has been introduced.
Before those, mul just wasn't vectorized at all, so widening_mul saw
_12 = __r$_M_value$real_8 * _10;
_13 = __r$_M_value$imag_9 * _11;
_14 = __r$_M_value$real_8 * _11;
_15 = __r$_M_value$imag_9 * _10;
_16 = _12 - _13;
_17 = _14 + _15;
and turned that into
_13 = __r$_M_value$imag_9 * _11;
_15 = __r$_M_value$imag_9 * _10;
_5 = -_13;
_16 = .FMS (__r$_M_value$real_8, _10, _13);
_17 = .FMA (__r$_M_value$real_8, _11, _15);
So, 2 scalar multiplications, 2 scalar fmas.
Now, vectorization turns that into
vect__12.26_31 = vect___r__M_value_real_8.22_24 * vect__10.25_28;
vect__10.34_40 = VEC_PERM_EXPR <vect__10.25_28, vect__10.25_28, { 1, 0 }>;
vect__13.35_43 = vect___r__M_value_real_8.30_35 * vect__10.34_40;
vect__4.36_44 = .VEC_ADDSUB (vect__12.26_31, vect__13.35_43);
So 2 vector multiplications, one permutation and one addsub.
Now, since r12-2051 we have the .VEC_FMADDSUB and .VEC_FMSUBADD ifns.
Shall the widening_mul pass turn from that
vect__12.26_31 = vect___r__M_value_real_8.22_24 * vect__10.25_28;
...
vect__4.36_44 = .VEC_ADDSUB (vect__12.26_31, vect__13.35_43);
into
vect__4.36_44 = .VEC_FMADDSUB (vect___r__M_value_real_8.22_24,
vect__10.25_28, vect__13.35_43);
?
