Test fails with gcc 10.2.0 about highwayhash

[thrasibule]

I've made some progress on this.

• test passes with clang. The fact that it wasn't trying avx2 was because I was running it in virtuabox, and even though virtualbox lets avx2 flag thought the host, it doesn't let fma and bmi throughwhich are also required.
• test passes with gcc if I use the -O2 flag instead of -O3. So it must be some over agressive optimization somewhere
• unfortunately, I've tried to run with -O2 -ftree-loop-vectorize -fversion-loops-for-strides -fipa-cp-clone -fgcse-after-reload -floop-interchange -floop-unroll-and-jam -fpeel-loops -fpredictive-commoning -fsplit-loops -fsplit-paths -ftree-loop-distribution -ftree-partial-pre -ftree-slp-vectorize -funswitch-loops -fvect-cost-model -fvect-cost-model=dynamic which to my understanding should be quivalent to -O3 but the test still passes, so there is something else that I'm missing.

[jan-wassenberg]

Thanks for letting us know and sharing your results. It does sound like a compiler bug, which is unfortunately not uncommon in my experience.

avx2 was because I was running it in virtuabox, and even though virtualbox lets avx2 flag thought the host, it doesn't let fma and bmi throughwhich are also required.

Yes, this is unfortunate. We also ran into that with JPEG XL. There the FMA is helpful but for HighwayHash we could remove those extra flag requirements.

Interesting that -O3 is not the same as its constituent flags (at least as defined at https://gcc.gnu.org/onlinedocs/gcc/Optimize-Options.html - I verified it's the same list). Would you like to report this as a potential a bug to GCC?

[thrasibule]

I've checked that it also works with -O3 -fno-strict-aliasing, but I don't get any aliasing warnings, so it must be something quite subtle. Does it ring a bell? It also works with gcc 9.3. I'm a bit wary to report it to gcc without narrowing down the issue further.