On Fri, Dec 10, 2021 at 11:38:50PM +0200, Peter Pentchev wrote: > So it turns out that the ARM AES feature uses vdupq_laneq_u32(), which > seems to be an A64-only instruction. This led to the build breaking on > the Debian armhf builder[0] (look for vdupq_laneq_u32 in the build log).
Sorry! I can't remember all the details, but it looks like my raspberry 4 (which is capable of 64-bit, but runs in 32-bit mode in raspbian; uname -m reports armv7l) has gcc-8 as the default version, and that doesn't support ARM AES (at least with the flags that we test). I see that your build log uses gcc-9; and if I test it on my system with clang-9, I get the same error message. > What do you think about the two attached patches? The first one blocks > the ARM AES feature from being enabled on non-aarch64 ARM builders, and > the second one makes sure that the tests-output/ directory is removed > after "make clean". Both patches look suitable, but I'd like to take a look at replacing vdupq_laneq_u32 in non-aarch64. Based on a quick skim, it looks like vdupq_n_u32 is supported in v7 and A32, so I imagine that it would work. (Disclaimer: it's been a few months since I did any ARM stuff.) Apologies, - Graham
