On Mon, Oct 21, 2019 at 3:13 AM Daniel Stenberg <dan...@haxx.se> wrote: > > On Sat, 19 Oct 2019, Jeffrey Walton wrote: > > >> The sad results is however that on this machine test 323 seems to work just > >> fine even if I try to repeat it mant times! :-O > >> > >> The particular machine I used features a "AArch64 Processor rev 1 > >> (aarch64)" in a board called "APM X-Gene Mustang". > > > > The other ones to try is GCC117 or GCC118. They are roughly the same Aarch64 > > machines, but with CRC+Crypto. The Mustang boards lack the cpu features. > > I've now also ran test 323 on gcc117 and the test works fine for me there too. > > I noticed however that if I try to use valgrind for the tests on this machine, > not a single TLS-related test seems to work. Calling RAND_status() in the > OpenSSL 1.1.0l version triggers an 'Illegal instruction' error which sounds > pretty bad! =(
That is probably either (1) Valgrind missing a VEX decode for an instruction, or (2) OpenSSL performing a feature probe. I used to encounter (1) all the time especially when compiling with -march=native. For (2), feature probes are business as usual. See the SIGILL topic at https://www.openssl.org/docs/faq.html#PROG . For (1) you stop using Valgrind and file a bug report. (Or you build Valgrind from sources hoping the VEX has been updated). For (2) you ignore it. > On gcc115 I could run 323 even with valgrind, but it sports OpenSSL 1.0.1f > which I have no idea if it makes a difference or not. > > So, to sum it up: I still have no idea why test 323 fails on travis arm64! Yeah, I would not worry too much about it until you determine the cause. It could be something wonky like a bad shell interaction or unexpected iptables result. Jeff ------------------------------------------------------------------- Unsubscribe: https://cool.haxx.se/list/listinfo/curl-library Etiquette: https://curl.haxx.se/mail/etiquette.html