Why use it

When bitsets occupy several words, a packed population count handles each word without unpacking the vector into scalar registers. XOR followed by counting measures Hamming distance; AND followed by counting measures intersections. The caller can then reduce the lane counts or keep one result per word.

Operations

import native.x86.vpopcntdq; exports the operations below through native::native. native.x86 and native re-export the module.

Operation	Input and result	Lane counts	Count range
vpopcntd<Arch>(value)	simd<std::uint32_t,N,Arch>	4, 8, 16	0–32
vpopcntq<Arch>(value)	simd<std::uint64_t,N,Arch>	2, 4, 8	0–64

Each name also has mask_NAME<Arch>(source, mask, value) to keep inactive source lanes and maskz_NAME<Arch>(mask, value) to clear them. Masks are predicate<N,Arch>; bit i selects lane i. Results retain their element width and exact architecture tag. All calls are noexcept.

Caveats

These are per-lane counts, without an implicit horizontal sum. For byte and word counts use BITALG; for one scalar word use POPCNT. Their feature bits are independent.

Runtime calls need AVX512F and AVX512VPOPCNTDQ, plus AVX512VL for 128/256 bits. Use target_features<native::x86>("avx512vpopcntdq"), adding avx512vl for shorter forms. Admit the caller's whole compiler target and OS AVX-512 state before entry, including for VL forms. The general AVX-512 profile does not imply VPOPCNTDQ.

Feature-bearing overloads are constexpr with native runtime paths. Weaker tags accept consteval calls if the vector storage exists: SSE2 for 128 bits, AVX for 256 bits or AVX512F for 512 bits, with prerequisites. They supply no runtime software fallback.

See Clang's 512-bit and VL intrinsic interfaces.