Why use it

A 52-bit limb leaves twelve spare bits in a qword for accumulation. IFMA multiplies two such limbs and adds either half of their 104-bit product into 64-bit lanes. This is useful for batched multi-precision multiplication and modular arithmetic, where the algorithm controls carry propagation.

Operations

import native.x86.ifma; supplies operations on simd<std::uint64_t,N,Arch> with N = 2, 4, 8. Link native::native; native.x86 and native also export them.

Operation	Per-lane result
madd52lo<Arch>(accumulator, a, b)	Add product bits 0–51 to the accumulator
madd52hi<Arch>(accumulator, a, b)	Add product bits 52–103 to the accumulator

The unsigned product uses only the low 52 bits of each multiplicand. Every accumulator bit participates, and addition wraps modulo 2⁶⁴.

Both names have mask_NAME<Arch>(accumulator, mask, a, b) to retain inactive accumulator lanes and maskz_NAME<Arch>(mask, accumulator, a, b) to clear them. Masks are predicate<N,Arch>.

Caveats

There is no carry propagation between lanes, saturation or modular reduction. The caller must account for accumulated carries before a 64-bit lane wraps. The top twelve bits of multiplicands are ignored even when they contain a carry.

Form	Runtime features	Caller target
Unmasked 128/256-bit VEX	AVX-IFMA, AVX	"avxifma"
128/256-bit EVEX, all mask modes	AVX512IFMA, AVX512F, AVX512VL	"avx512ifma,avx512vl"
512-bit EVEX, all mask modes	AVX512IFMA, AVX512F	"avx512ifma"

Short unmasked forms prefer AVX-IFMA when the tag supports it. Masked runtime forms always need EVEX features. BW and DQ are unnecessary. AVX-IFMA and AVX512IFMA are independent features, absent from the general AVX2 and AVX-512 profiles. Clang's AVX-IFMA target also enables AVX2; use target_features to include the compiler prerequisites. Admit the complete target and OS vector state before entering a matching caller.

Feature-bearing overloads are constexpr with native runtime paths. Weaker tags have consteval overloads if the chosen SSE2, AVX or AVX512F register storage exists, with prerequisites. Constant evaluation gives exact integer results and preserves the tag.

See Intel's Software Developer's Manual and Clang's AVX-IFMA, AVX512IFMA and VL IFMA headers.