On Tue, 26 Feb 2019 at 21:44, Nick Desaulniers <ndesaulni...@google.com> wrote: > > On Mon, Feb 25, 2019 at 11:19 PM Ard Biesheuvel > <ard.biesheu...@linaro.org> wrote: > > > > On Tue, 26 Feb 2019 at 05:03, <ndesaulni...@google.com> wrote: > > > > > > Clang warns: vector initializers are not compatible with NEON intrinsics > > > in big endian mode [-Wnonportable-vector-initialization] > > > > > > While this is usually the case, it's not an issue for this case since > > > we're initializing the uint8x16_t (16x uint8_t's) with the same value. > > > > > > Instead, use vdupq_n_u8 which both compilers lower into a single movi > > > instruction: https://godbolt.org/z/vBrgzt > > > > > > This avoids the static storage for a constant value. > > > > > > Link: https://github.com/ClangBuiltLinux/linux/issues/214 > > > Suggested-by: Nathan Chancellor <natechancel...@gmail.com> > > > Signed-off-by: Nick Desaulniers <ndesaulni...@google.com> > > > > Much better, thanks, > > > > Did you double check that the intrinsic exists on 32-bit ARM as well? > > I assume it does, but please make sure if you haven't yet. > > Thanks for the review!
My pleasure. > Looking through Clang's generated arm_neon.h, vdupq_n_u8 seems to have > 2 definitions predicated on __LITTLE_ENDIAN__ (not __arch64__ or > __ARM_ARCH >= 8 like some of the other types and functions). > > So NEON got some additions in v8? Basically, yes. One example is right there in the NEON recovery code, guarded by #ifdef CONFIG_ARM. New intrinsics were also introduced for the crypto instructions, although I think those were also added to the 32-bit version of arm_neon.h > Is there a doc that lists them? Not that I know of. > http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dui0491g/BABDBBJB.html > is where I found vdupq_n_u8, but it doesn't seem to mention > compatibility (so I assume it's been around since the introduction of > NEON?). Yes, it appears that the 32-bit arm_neon.h header on my box defines this intrinsic, so this should be fine.