tl;dr: The Intel SDM documents vfmaddrnd231pd, an FMA instruction that would allow per-instruction rounding control without needing AVX-512, potentially speeding up rounding-dependent algorithms. However, the instruction doesn't actually work: assemblers don't recognize it, compiler intrinsics are missing, and running the hand-assembled opcode on FMA-capable CPUs raises #UD. The author speculates Intel planned the instruction, pulled it late, but left stale documentation behind—it's now absent from the 2026 SDM.