https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106346
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> --- (In reply to Tamar Christina from comment #4) > I believe the problem is actually g:27842e2a1eb26a7eae80b8efd98fb8c8bd74a68e > > We added an optab for the widening left shift pattern there however the > operation requires a uniform shift constant to work. See > https://godbolt.org/z/4hqKc69Ke > > The existing pattern that deals with this is vect_recog_widen_shift_pattern > which is a scalar pattern. during build_slp it validates that constants are > the same and when they're not it aborts SLP. This is why we lose > vectorization. Eventually we hit V4HI for which we have no widening shift > optab for and it vectorizes using that low VF. > > This example shows a number of things wrong: > > 1. The generic costing seems off, this sequence shouldn't have been > generated, as a vector sequence it's more inefficient than the scalar > sequence. Using -mcpu=neover-n1 or any other costing structure correctly > only gives scalar. > > 2. vect_recog_widen_shift_pattern is implemented in the wrong place. It > predates the existence of the SLP pattern matcher. Because of the uniform > requirements it's better to use the SLP pattern matcher where we have access > to all the constants to decide whether the pattern is a match or not. That > way we don't abort SLP. Are you ok with this as a fix Richi? patterns are difficult beasts - I think vect_recog_widen_shift_pattern is at the correct place but instead what is lacking is SLP discovery support for scrapping it - that is, ideally the vectorizer would take patterns as a hint and ignore them when they are not helpful. Now - in theory, for SLP vectorization, all patterns could be handled as SLP patterns and scalar patterns disabled. But that isn't easy to do either. I fear to fight this regression the easiest route is to pretend the ISA can do widen shift by vector and fixup in the expander ... > 3. The epilogue costing seems off.. > > This example https://godbolt.org/z/YoPcWv6Td ends up generating an > exceptionally high epilogue cost and so thinks vectorization at the higher > VF is not profitable. > > *src1_18(D) 1 times vec_to_scalar costs 2 in epilogue > MEM[(uint16_t *)src1_18(D) + 2B] 1 times vec_to_scalar costs 2 in epilogue > MEM[(uint16_t *)src1_18(D) + 4B] 1 times vec_to_scalar costs 2 in epilogue > MEM[(uint16_t *)src1_18(D) + 6B] 1 times vec_to_scalar costs 2 in epilogue > MEM[(uint16_t *)src1_18(D) + 8B] 1 times vec_to_scalar costs 2 in epilogue > MEM[(uint16_t *)src1_18(D) + 10B] 1 times vec_to_scalar costs 2 in epilogue > MEM[(uint16_t *)src1_18(D) + 12B] 1 times vec_to_scalar costs 2 in epilogue > MEM[(uint16_t *)src1_18(D) + 14B] 1 times vec_to_scalar costs 2 in epilogue > /app/example.c:16:12: note: Cost model analysis for part in loop 0: > Vector cost: 23 > Scalar cost: 17 I don't see any epilogue cost - the example doesn't have a loop. With BB vect you could see no epilogue costs? > For some reason it thinks it needs a scalar epilogue? using > -fno-vect-cost-model gives the desired codegen.
