https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103144
--- Comment #3 from rguenther at suse dot de <rguenther at suse dot de> --- On Wed, 1 Dec 2021, crazylht at gmail dot com wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103144 > > --- Comment #2 from Hongtao.liu <crazylht at gmail dot com> --- > Another issue is for SLP, when trip count is small and loop is completely > unrolled. SLP failed to generate vlshr_optab. > > #include<stdint.h> > void > foo (uint64_t* __restrict pdst, uint64_t* psrc, uint64_t shift) > { > for (int64_t i = 0; i != 4; i++) > { > pdst[i] = psrc[i] + shift; > shift >>= 1; > } > } > > .175t.slp1 dump: > > _8 = *psrc_12(D); > _28 = _8 + shift_10(D); > shift_30 = shift_10(D) >> 1; > _39 = psrc_12(D) + 8; > _40 = *_39; > _41 = pdst_13(D) + 8; > _42 = shift_30 + _40; > shift_44 = shift_30 >> 1; > _53 = psrc_12(D) + 16; > _54 = *_53; > _55 = pdst_13(D) + 16; > _56 = shift_44 + _54; > shift_58 = shift_44 >> 1; > _3 = psrc_12(D) + 24; > _4 = *_3; > _5 = pdst_13(D) + 24; > _6 = _4 + shift_58; > _15 = {_28, _42, _56, _6}; > vectp.8_18 = pdst_13(D); > MEM <vector(4) long unsigned int> [(uint64_t *)vectp.8_18] = _15; The issue is that the shifts are dependent on each other: shift_30 = shift_10(D) >> 1; shift_44 = shift_30 >> 1; shift_58 = shift_44 >> 1; and one addend is shift_10(D) which isn't shifted (SLP build does not support inserting no-op operations into some lanes like +- 0 or */ 1 or in this case >> 0). So first of all we either need to simplify the above to shift_30 = shift_10(D) >> 1; shift_44 = shift_10(D) >> 2; shift_58 = shift_10(D) >> 3; then the shift_10(D) >> 0 issue remains. With void foo (uint64_t* __restrict pdst, uint64_t* psrc, uint64_t shift) { for (int64_t i = 0; i != 4; i++) { shift >>= 1; pdst[i] = psrc[i] + shift; } } you see SLP discovery works but we vectorize it as shift_23 = shift_10(D) >> 1; shift_37 = shift_23 >> 1; shift_51 = shift_37 >> 1; _17 = {shift_10(D), shift_23, shift_37, shift_51}; vect_shift_23.10_8 = _17 >> 1; which eats from the profitability. We also fail to improve such code (if it were written by a user) later realizing we could merge the scalar shifts with the vector shift.
