Let's set the record straight here. Largely the x87 has not been changed in a long time and functions more as a slow legacy unit for long double computations and special cases. Even the trig functions have accuracy issues and dont meet the 0.5ulp of IEEE not to mention are horrendously slow. Its faster to use the vectorized units with scalars and tolerate 1ulp accuracy for better performance and SIMD possibility. So largely as the FPU is slow including the interface to load and store data into FPU registers, it simply became more convenient to let libraries switch to the new SSE and AVX.

Now also be careful of bad and misleading information in a post here. FMA for integers would change the precision. As there are sign, carry and overflow flags whose values would absolutely be different in a fused vs unfused scenario and emulating it would require extra instructions. One reason to be warry of advice from those not well versed in hardware.