Native execution surface (syntax + ergonomics)¶

Date: 2026-05-25
Depends on: Execution resource model, Execution decorators
Plan: WP8 in compiler_lint_and_parallelism plan

Design goal¶

Concurrency, parallelism, threads, cores, and SIMD feel simple, flexible, powerful, and native — one mental model: source → li.toml → lic build → binary.

Principles¶

User says	Li surface	Proof
use 8 cores	`parallel for` + `--cores=8` or `[execution] cores = 8`	disjoint required (G-par)
vectorize inner	`@vectorized(lanes=8)` on inner `for`	no extra OS threads
CI fast check	`lic check --workspace --jobs=8`	not in `.li` source

Layer 1 — Keywords¶

parallel for i in 0..<N is primary (MIR ParallelFor → li_parallel_for_*). Requires disjoint_* or inherited disjoint= — else compile error.

Layer 2 — Decorators (stackable, top-to-bottom)¶

@parallel(disjoint=...), @vectorized(lanes=N) on def / for / parallel for. SIMD = one thread, N lanes. Parallel = cores × threads_per_core. Compile farm = --jobs only — never in source.

Layer 3 — Config and CLI¶

[execution]
cores = 8
threads_per_core = 1
memory_mb = 4096

lic build app.li -o app --cores=8 --threads-per-core=1
lic check --workspace --jobs=8 --max-memory=4096

Deprecated: --threads alias; LI_* env (one-release warn).

Phase 2 — `team(cores=4) { ... }`¶

Spec only until parser lands.

Learned from¶

System	Li adaptation
Nim `parallel for`	keyword + proved disjoint
Rust rayon	`@parallel` inherits; static chunks v1
numba	prange vs SIMD = separate knobs
Chapel	`team()` phase 2

Handbook: parallelism.md.