Skip to content

V1-ROADMAP.md

Latest commit

 

History

History
81 lines (69 loc) · 5.09 KB

V1-ROADMAP.md

File metadata and controls

81 lines (69 loc) · 5.09 KB
Raw

dashu Roadmap — v0.6 → v1.0

Last updated: 2026-08-10

Release strategy. v0.6 is the breaking-changes testbed: every planned breaking change ships there first, so downstream users can migrate and validate early, and v1.0 then freezes the API. v0.5.x (0.5.0 / 0.5.1 / 0.5.2) is released; the items below carry their current status. File:line references are anchors from the v0.5.x tree and may drift.

v0.6.0-rc.x already carries several breaking/contract changes — see the per-crate CHANGELOGs (ParseError::InvalidSyntax, Context::exp underflow → Err(Underflow), Ziv transcendentals requiring R: ErrorBounds, TryFrom<RBig> for FBig, …, plus the signed-isize RBig::pow/Relaxed::pow widening).

The v0.6 release mechanics — version sync, ## Unreleased → ## 0.6.0 changelog folding, MSRV bookkeeping, and the pre-publish check suite — are driven by the pre-publish-check process, not roadmap items.

v0.6 — non-breaking backlog (additive; land when ready)

Carried from the v0.5.x deferral list. Not breaking, so they are not v0.6 test content per se — they broaden the validation surface and bundle naturally with the breaking pull-forward above.

  • More transcendentals — rootofunity.
  • Test organization — clear src in-file vs tests/ boundary. Tests are scattered: many operations have both an in-file #[cfg(test)] mod tests (in src/<op>.rs) and a parallel tests/<op>.rs integration file, and the two frequently overlap. dashu-float is the worst case — add, mul, div, exp, log, trig, hyper, root, shift, round, convert, cmp, io, and iter each appear in both places. AGENTS.md already states the intended convention (in-file mod tests for algorithm/kernel correctness; tests/ for cross-cutting and public-API/property tests), but it is not consistently followed. Consolidate: assign each test to the side the convention dictates, deduplicate the overlaps, and tighten the AGENTS.md wording so the boundary is explicit and enforceable. As part of this, clarify the status of the tests/*_prop.rs property tests against the AGENTS.md "in-crate tests must use fixed, deterministic inputs" rule — fixed-seed / enum-driven, or moved to fuzz/.

v1.0 — API freeze

With the one breaking change on the 1.0 path (the signed-isize RBig::pow) shipped in v0.6, v1.0 is the stabilization point: the API freezes at what shipped through v0.6. No breaking changes are currently scheduled for v1.0 itself; anything that comes up is folded back into the v0.6 release cycle (or a v0.7) rather than deferred silently to 1.0.

Post-v1 — long-term goals

  • Full C <tgmath.h> type-generic math surface. The complete C standard math library for both real and complex — trig & inverse; hyperbolic & inverse; the exp/log family including exp2/exp10/expm1/log2/log10/log1p; power/root cbrt/hypot/ pow/sqrt; error & gamma erf/erfc/tgamma/lgamma; fma; rounding/remainder; fp-classification — available on both FBig and CBig (ref). The individual pieces shipped across v0.5.x/v0.6 are the first incremental steps toward this.
  • exp2 / exp10 for FBig and CBig, and complex log10. exp2/exp10 need their own reduction (x·log2(e) → power-of-two scale, mirroring exp's Ball engine); complex log10 follows CBig::log's ln|z| + i·arg(z) split divided by ln 10.
  • Complex agm — the arithmetic–geometric mean for CBig (and real FBig if wanted), via the standard AGM iteration with Ziv certification. The branch-cut behavior on ]−∞, 0] needs care; no algorithm has been sketched yet.
  • Vector ops — dot/mean helpers. (Sum for CBig is already exact-accumulating, matching FBig: Sum; Product for CBig remains a fold, matching FBig: Product.)
  • SIMD-optimized FFT multiplication. Leverage the wide crate for SIMD FFT-based multiplication. Not considered until v1.0.
  • High-bits (short-product) multiplication for dashu-float mul/sqr/cubic. These currently round the exact full product/square/cube — always correctly rounded, but O(M(n)) in the operand size regardless of the target precision, which is wasteful when operands far exceed precision (e.g. unlimited-precision significands multiplied to a low target precision). The efficient form keeps only the high limbs needed for rounding plus a sticky bit for the discarded tail — MPFR's mpfr_mulhigh_n (Mulders' short product) with a mpfr_round_p-style certify-and-fall-back-to-exact gate, so it never materializes the full product. Deferred post-v1.0: it needs a new short-product primitive in dashu-int (none exists today), and the case it optimizes is uncommon. (float/src/mul.rs:173.)
  • Constant-cache eviction/cap policy — revisit only if real workloads report memory pressure. (No cap today; a ConstCache grows until clear_cache()/drop, and callers own the lifetime explicitly via the CachedFBig/ConstCache handle.)