Why use it

pext gathers a sparse set of bits into a dense index; pdep scatters those bits back into a chosen layout. That lets a mask describe the packing instead of a chain of shifts. mulx supplies both halves of a product for multiword arithmetic, while the shift operations give defined masked-count behavior.

Operations

import native.x86.bmi2; exports the operations below through native. The native.x86 and native hubs also export them. Here U is std::uint32_t or std::uint64_t.

Operation	Result
bzhi<Arch>(U value, unsigned index)	Clear bits at and above index & 255; retain all bits when that index reaches the width
mulx<Arch>(U a, U b, U* high)	Return the low product and write its high half
pdep<Arch>(U value, U mask)	Deposit consecutive low bits into positions selected by mask
pext<Arch>(U value, U mask)	Pack bits selected by mask into consecutive low positions
shlx<Arch>(U value, unsigned count)	Logical left shift, count modulo the width
shrx<Arch>(U value, unsigned count)	Logical right shift, count modulo the width
sarx<Arch>(value, unsigned count)	Signed right shift on int32_t or int64_t, count modulo the width
rorx<Arch, Imm8>(U value)	Right rotation by a compile-time byte, count modulo the width

rorx<Imm8>(value) uses the module default tag; rorx<Imm8, Arch>(value) also accepts an explicit tag. Imm8 is in 0–255. All functions are noexcept.

Caveats

The mulx output must point to a writable object; constant evaluation also requires that object to be modifiable in the constant expression. These wrappers specify results and the output write, with no C++ flag-preservation or exact-instruction-encoding promise.

Runtime calls need BMI2, a "bmi2" caller target and CPU admission. BMI1 and vector OS state are unnecessary. Arch is an isa<x86>; its scalar default is the provider's NATIVE_BASELINE, unaffected by importer target attributes. Feature-bearing overloads are constexpr; tags without BMI2 are consteval-only.

pdep and pext costs vary substantially by CPU. A compact expression is not a throughput guarantee; measure them in the loop that needs the packing.

See Intel's BMI2 instruction reference.