Hi, >> So maybe it's the feature detection that runs when the plugin is loaded. >> But that's some simple push/pop/mov and cpuid assembler code [1]. Does >> your CPU lack support for cpuid? > > It has cpuid. Wikipedia agrees cpuid exists since 486, cpuid should be safe. > Debian 12 cpu_feature.c uses only cpuid leaf 0 and leaf 1, that should be > safe. > > Result from gdb: Illegal instruction is punpcklqdq in aesni_plugin_create. > <+52>: movd %ecs,%xmm0 > <+56>: movd %edx,%xmm0 > <+60>: punpckldq %xmm2,%xmm1 > <+64>: punpckldq %xmm3,%xmm0 > =><+68>: puncpklqdq %xmm1,%xmm0 > <+72>: movups %#mm0,(%eax) > <+75>: add $0x18,%esp > <+78>: pop %ebx > <+79>: ret
Oh, interesting. The complete `aesni` plugin is compiled with `-maes -mpclmul -mssse3`. That apparently causes the compiler to "optimize" the code there. I guess we could split the plugin into a helper library that uses those flags and compile the plugin constructor/detection code without it. I've pushed such a change to the devel branch [1]. Anyway, another question is if it actually makes sense to ship the plugin in Debian. The default crypto library is OpenSSL nowadays and since that will use AES-NI etc. anyway (likely more optimized than this plugin), there is probably not much use having `aesni` enabled as well. Regards, Tobias [1] https://github.com/strongswan/strongswan/commit/553ac5678b50729da896fa07ff8bc502a2a8f3da
