https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79938
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Severity|normal |enhancement
Last reconfirmed| |2021-08-02
Status|UNCONFIRMED |NEW
Component|target |tree-optimization
Ever confirmed|0 |1
--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
I think we could be better than what we producing on the trunk:
_1 = BIT_FIELD_REF <a_49(D), 8, 0>;
_3 = BIT_FIELD_REF <a_49(D), 8, 8>;
_6 = BIT_FIELD_REF <a_49(D), 8, 16>;
_8 = BIT_FIELD_REF <a_49(D), 8, 24>;
_13 = BIT_FIELD_REF <a_49(D), 8, 32>;
_15 = BIT_FIELD_REF <a_49(D), 8, 40>;
_18 = BIT_FIELD_REF <a_49(D), 8, 48>;
_20 = BIT_FIELD_REF <a_49(D), 8, 56>;
_25 = BIT_FIELD_REF <a_49(D), 8, 64>;
_27 = BIT_FIELD_REF <a_49(D), 8, 72>;
_30 = BIT_FIELD_REF <a_49(D), 8, 80>;
_32 = BIT_FIELD_REF <a_49(D), 8, 88>;
_37 = BIT_FIELD_REF <a_49(D), 8, 96>;
_63 = {_1, _13, _25, _37};
vect__2.10_22 = (vector(4) int) _63;
_39 = BIT_FIELD_REF <a_49(D), 8, 104>;
_29 = {_3, _15, _27, _39};
vect__4.11_60 = (vector(4) int) _29;
_69 = vect__2.10_22 + vect__4.11_60;
_42 = BIT_FIELD_REF <a_49(D), 8, 112>;
_10 = {_6, _18, _30, _42};
vect__7.9_17 = (vector(4) int) _10;
_44 = BIT_FIELD_REF <a_49(D), 8, 120>;
_5 = {_8, _20, _32, _44};
vect__9.8_66 = (vector(4) int) _5;
_70 = vect__7.9_17 + vect__9.8_66;
vect__11.14_57 = _69 + _70;
_55 = VIEW_CONVERT_EXPR<__m128i>(vect__11.14_57);
------ CUT ----
We could produce a shuffle from a_49(D) and then do extracts to get _63, _29,
_10, and _5.
clang does:
pshufb .LCPI1_0(%rip), %xmm2 # xmm2 =
xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero,xmm2[3],zero,zero,zero,xmm2[2],zero,zero,zero
pshufd $238, %xmm2, %xmm0
