I recently learned about the wonders of non deterministic results of floats between compilers and platforms. I'd be targeting x64 (SSE4.2. No 32-bit build. Windows) and ARM NEON (Android) with precise floats enabled, because I need to be able to exactly reproduce previous results with the same inputs. Can I be sure that because of me avoiding x86, the x87 registers won't get used and SSE will be used instead? Meaning that the floats are deterministic and won't be using the varying size intermediates? And are the trig/pow/exp/etc. functions still non deterministic as well? Of course rcp and rsqrt are off the table in this strict float mode. I would ofc have to ensure it's IEEE754 compliant using _mm_setcsr for sse.
Thanks for the help.