Why use it

A byte permutation can rearrange packed records or implement a small table lookup without being confined to 128-bit blocks. VBMI adds whole-vector byte selection from one or two tables. Multishift extracts several eight-bit windows from each qword, useful when values arrive in a bit-packed representation.

Operations

import native.x86.vbmi; exports the following through native::native. native.x86 and native also export them. Byte vectors have type simd<std::uint8_t,N,Arch> with N = 16, 32, 64.

Operation	Operands	Result byte i
vpermb	indices, value	value[indices[i] % N]
vpermt2b	a, indices, b	Selection from the concatenation of a and b
vpermi2b	indices, a, b	Same two-table selection, with indices as tied destination
vpmultishiftqb	control, value	Eight wrapping bits from qword value[i / 8]

Single-table indices use their low log2(N) bits. Two-table indices use log2(2*N) bits: indices below N choose a, the rest choose b. Selection spans the complete vector.

Multishift control is a byte vector, while value is simd<std::uint64_t,N/8,Arch>. Byte i selects eight bits starting at control[i] & 63 of qword value[i / 8]. Positions wrap from 63 to zero within that qword; the first selected bit becomes result bit zero.

All names take an explicit <Arch> and provide merging and zeroing forms with predicate<N,Arch> masks:

mask_vpermb(source, mask, indices, value) and mask_vpmultishiftqb(source, mask, control, value) keep inactive source bytes.
mask_vpermt2b(a, mask, indices, b) keeps inactive bytes from a.
mask_vpermi2b(indices, mask, a, b) keeps inactive index bytes.
maskz_NAME(mask, ...) clears inactive bytes and otherwise uses the ordinary operand order.

Caveats

Writemasks control output bytes. A masked-off table byte can still supply data to an active output byte. The two merging permutations differ in which operand survives; the compiler may choose either encoding for unmasked or zeroing calls.

Multishift never crosses a qword boundary and ignores control bits 6 and 7. It is a wrapping bit window, not a whole-vector variable shift.

Runtime wrappers need AVX512F, AVX512BW and AVX512VBMI, plus AVX512VL for short forms. Use target_features<native::x86>("avx512vbmi"), adding avx512vl for 128/256 bits. BW reflects Clang's target prerequisites. DQ, BITALG and VBMI2 are independent. Admit the full matching caller target and OS AVX-512 state before entry.

Feature-bearing overloads are constexpr with native runtime paths; weaker tags are consteval-only with complete SSE2, AVX or AVX512F storage for the chosen width. Results preserve the exact architecture tag.

See Intel's Software Developer's Manual and Clang's VBMI definitions.