Does anyone know if using higher optimizations like -O2 and -O3 and –msse (or-msse2) produces optimized code using SSE instructions to speed the optimizations? If that were the case, building CPU Detect versions of GnuBG with optimizations could generate SSE instructions elsewhere in the code – and that would likely crash the GnuBG (On systems without SSE)
Because of this, I think CPU Detect versions of Gnubg should be compiled with SSE2(2) compile options, and the Makefiles modified to build the neural net files separately (with either –msse or –msse2).
Or, we could simplify things – maybe we could consider removing the autodetect feature from the build process completely. Is it very useful? And do the gains outweigh the possible risks?