Intel's VFMADDRND231PD Instruction Appears in Documentation But Not in Hardware

Undocumented Behavior of VFMADDRND231PD

The Intel Software Development Manual (SDM) describes an instruction, VFMADDRND231PD, as a Fused Multiply-Add operation for packed double-precision floating-point values with rounding control. This instruction's unique 'rnd' suffix indicates per-instruction control over rounding mode, along with DAZ, FTZ, and exception suppression, defined by an immediate byte.

Unlike most x86 floating-point instructions that rely on global rounding modes, VFMADDRND231PD would allow specific rounding control for individual arithmetic operations. This capability could significantly accelerate algorithms dependent on precise rounding, as FMA can emulate addition, subtraction, and multiplication.

Absence in Compilers and Hardware

Despite its documentation, VFMADDRND231PD is not recognized by the GNU assembler. Furthermore, the Intel C/C++ Compiler intrinsics _mm_fmaddround_pd and _mm256_fmaddround_pd, mentioned in the SDM, are missing from both Clang and Intel's own C/C++ compiler.

Attempts to execute the instruction directly on a Haswell CPU, which supports the FMA instruction set, result in an "Undefined instruction" (#UD) error. This indicates the instruction is not implemented in the hardware.

Documentation Discrepancies

The instruction's opcode in the SDM includes a 'DDS' field, which is uniquely mentioned only for VFMADDRND231PD and has been removed from later SDM versions. This, combined with the absence of expected variants (e.g., for different data types or operand orders) common to other FMA instructions, raised initial suspicions.

A comparison with a more recent 2026 edition of the SDM confirms that VFMADDRND231PD is entirely absent, despite its presence in the 2024 version. This suggests the instruction was planned but ultimately not implemented or removed from the specification.

Hypothesis for Discrepancy

The detailed documentation, including unique tables and pseudocode, suggests VFMADDRND231PD was not a simple typo. A possible explanation is that Intel intended to introduce the vfmaddrnd* instruction set but encountered an issue prior to release, leading to its removal from the CPU and subsequent SDM versions, while earlier documentation was not fully corrected.