vyre-std — the vyre standard library

Status: shipping, workspace member, published alongside vyre and vyre-conform. Layer: L2 — compositional helpers built from vyre core primitives. Scope: GPU-native pattern matching (DFA assembly pipeline) and arithmetic helpers.

What ships today

Pattern compilation — full GPU DFA assembly pipeline

regex_to_nfa → Nfa → nfa_to_dfa → Dfa → dfa_minimize → Dfa → dfa_pack → PackedDfa

One-line API — patterns in, GPU-ready bytes out:

use vyre_std::pattern::{dfa_assemble, AssembleOptions, Pattern};

let packed = dfa_assemble(
    &[
        Pattern::Literal(b"GET /api"),
        Pattern::Literal(b"POST /api"),
        Pattern::Regex("PUT /api/v[0-9]+"),
    ],
    AssembleOptions::default(),
)?;
// `packed.bytes` is ready for a backend-specific dispatcher to upload.

Individual stages are also exposed for callers who need finer control:

• [pattern::regex_to_nfa] — Thompson construction. Supports literal, concat, |, *, +, ?, [abc] (with ranges and negation), escape.
• [pattern::nfa_to_dfa] — subset construction with dead-state pruning.
• [pattern::dfa_minimize] — Hopcroft's partition refinement. Idempotent.
• [pattern::dfa_pack] — transition-table compression. Two formats:
  • Dense — fastest scan, 256 × 4 bytes per state.
  • EquivClass — byte-equivalence compression for narrow alphabets.
• [pattern::dfa_assemble] — composite entry: patterns → PackedDfa.
• [pattern::cache] — content-addressed compilation cache so the second compile of a pattern set is a file read.

Arithmetic helpers — ~80 compositional Cat A ops

Every helper is a pure composition of core primitives; the conform gate proves byte-identical agreement with the CPU reference:

• saturating_{add,sub,mul}_{u8,u16,u32,u64,i8,i16,i32,i64} — 24 ops.
• wrapping_{add,sub,mul}_{u8,u16,u32,u64,i8,i16,i32,i64} — 24 ops.
• min_*, max_*, clamp_*, midpoint_*, abs_diff_* — unsigned variants, 16+ ops.
• div_ceil_*, div_floor_*, div_round_* — unsigned variants, 6 ops.
• lerp_f32, lerp_f64 — linear interpolation.

Aho-Corasick DFA build

Literal-pattern-set construction (trie + failure links + output chain) lives at pattern::aho_corasick_build. The CPU reference and WGSL kernel agree byte-for-byte; the conform gate maintains five GOLDEN samples and twenty KAT vectors as regression anchors.

Compositional discipline

Every std:: op is Category A. That means:

• It lowers to the same IR nodes a hand-optimizer would write.
• It introduces no new hardware intrinsic — those live in vyre core.
• The conform gate verifies the lowered output is byte-identical to the hand-written composition on every witness.

Ops that require a hardware intrinsic belong in core (vyre), not here.

Design gates for new std:: ops

A PR adding a new std:: op must:

1. Reference a coordination note or issue describing the composition.
2. Declare category = "A" and a non-empty composition_of list naming the core primitives it builds from.
3. Pass conform-certify at L1 minimum.
4. Add a KAT vector set and an adversarial witness set.

Related

• [vyre] — the core IR, primitives, and lowering contract.
• [vyre-conform] — the immune system that verifies this crate.
• [vyre-spec] — frozen data contracts shared across all three crates.
• [LEGENDARY_PLAN.md] — the build-out roadmap this crate follows.