Skip to content

Phase 7: Native HPC (SIMD + parallel for)

Depends on: Phases 3–5b (MIR/LLVM, benchmarks harness)
Blocks: Pure-Li Tier 2 perf tables, OpenMP scaling columns

Proof gaps (Doc-c): G-par · G-dec · G-math · still open

Goal: Built-in simd[T, N] and proved parallel for without user-level parallel/math libraries. Toolchain links LLVM + libomp only.

7a — SIMD vertical slice

Task Exit
simd[T,N] in typechecker (TyKind::Simd) li-tests/simd/ pass
MIR: splat, binop, horizontal sum LLVM <N x double>
simd_dot benchmark pure Li li_pure=True in harness

7b — parallel for + OpenMP

Task Exit
Stmt::ParallelFor AST + parser Parses exploit fixtures
Outlined par body + li_omp_parallel_for in runtime/li_rt.c -fopenmp link
Replace policy.cpp string hacks with structured overlap check (keep fixtures) race_shared_memory green
lic build --threads=N / LI_OMP_THREADS CSV threads column

7c — Benchmark truthfulness

Task Exit
md_lennard_jones pure Li driver No LI_EXTRA_C for li label
Tier 2 verify checksum bench.py smoke

7d — Execution decorators (decorator-first HPC)

Depends on: 2g (def), 7a (SIMD), 7b (parallel for + structured disjoint)
Plan: .cursor/plans/li_execution_decorators_7c6e3b42.plan.md
Spec (to land): docs/superpowers/specs/2026-05-16-li-execution-decorators.md

Goal: Primary surface for parallelism, vectorization, and device placement is stackable @ decorators on def and on for/while — elaborating to the same proved cores as keywords (parallel for, simd, future gpu proc).

Sub Task Exit
7d-a Lexer @, decorator lists on def/for/while, AST attrs Parse tests — done
7d-e (partial) Policy: reserved_name, typosquat, parallel_requires_disjoint decorator_exploits/ CI
7d-b Elaboration → ParallelFor / simd / host placement MIR tags li-tests/decorators/ positive — partial: @vectorized(lanes=4), @no_vectorize
7d-c Structured disjoint=; scoped @vectorized on for (ArraySimdScope) vectorized_for_scope_ok.li — partial: disjoint= still G-par open
7d-d std/execution/decorators.li + docs/language/decorators.md Handbook + gallery
7d-e decorator def with strict naming (package prefix, typosquat ban), expansion whitelist li-tests/decorator_exploits/ all fail except control; CI on every PR

Policy (binding):

  • Stdlib names (parallel, vectorized, async, cpu, gpu, tpu, user_defined, …) are reserved — no Python-style shadowing via import or user decorator def.
  • User decorators use their own multi-segment names (li_math_tiled_parallel in package li-math).
  • @user_defined(...) is the stdlib hook for custom chips, not a user-chosen decorator name.

CI (parallel track)

  • Windows: build + run_all.sh --ci + security
  • Fuzz: daily fuzz.yml, merge_fuzz_corpus.sh, corpus artifact + optional bot PR
  • scripts/ci.sh: log race_shared_memory and decorator_exploits explicitly
  • TSan nightly (post-7b): optional memory.yml job

Exit gate (phase complete)

7a–7c (Phase 7 core):

  • ./li-tests/run_all.sh simd race_shared_memory
  • bench.py --tier 0 in CI; tier 1/2 perf runs advisory via bench.py
  • Fuzz workflow present (.github/workflows/fuzz.yml); scripts/export-fuzz-status.sh

7d (decorators — can ship after 7b; recommended before calling HPC “done” for users):

  • ./li-tests/run_all.sh decorators decorator_exploits
  • Tier 2 MD example uses @cpu @parallel @vectorized on def (elaborates to same MIR as keywords)
  • Fuzz corpus includes @ decorator stacks and reserved-name parse seeds (compiler/fuzz/corpus/seed_decorators)

7e (mathematical surface — user writes formulas, not simd(...)):

Plan: 2026-05-16-li-math-linalg-surface.md

Sub Task Exit
7e-a Lower *, +, dot, sum to 7a SIMD MIR partial: simd_dot pure-Li a @ b (#148)
7e-b Lower A @ B for Tier 1 matmul benches partial: matmul_naive / matmul_blocked pure-Li; ≤1.2× C++ advisory
7e-c docs/language/linear-algebra.md, docs/guide/math-hpc-examples.md done on main
7e-d/e ArrayDotF64 / ArrayBinOpF64 gather SIMD partial on main (#148)
2f / P-linalg Contract corpus for dot/sum/matmul entry partial: #151 closed + loop open — G-math, G-lean
  • ./li-tests/run_all.sh math_linalg
  • Tier 1 Li sources: math notation only (a @ b, C = A @ B — no user __li_simd_*)
  • Tier 1 perf: Li within 1.2× C++ on same machine (investigate reds on dashboard)