Skip to content

Execution decorators (spec stub)

Status: Phase 7d in progress (@gpu / @gpu(devices=N) MIR telemetry landed; LKIR/codegen open) Plan: .cursor/plans/li_execution_decorators_7c6e3b42.plan.md
Gaps: Provability gaps G-dec

No runtime semantics

Decorators are not Python-style runtime callables. There is:

  • No runtime decorator registry or dynamic dispatch
  • No wrapping functions at call time
  • No interpreting @foo when the program runs

@parallel, @vectorized, @gpu, @cpu, etc. are resolved entirely during lic build: parse → validate → elaborate to MIR (ParallelFor, SIMD lanes, device-placement tags). Illegal stacks, invalid devices=, or missing disjoint= are compile errors.

User decorator def macros expand at compile time to a whitelist of builtin decorators only.

Goal: zero user-visible decorator failures at runtime — if it builds, placement and parallelism shape are already proved.

Reserved stdlib names: cpu, gpu, tpu, user_defined, parallel, vectorized, async, serial, no_vectorize.

User decorator def names: strict package-prefixed snake_case; typosquat ban; expansion whitelist to builtins only.

GPU placement slice

@gpu is now visible in MIR and lic verify telemetry:

  • mir_gpu_def counts def declarations tagged with @gpu.
  • mir_gpu_multi_device_def counts def declarations tagged with @gpu(devices=N) where N > 1.
  • @gpu(devices=0) and non-integer devices values are rejected (gpu_devices) so impossible placement is not silently accepted.
  • Vendor/backend arguments such as vendor="cuda" are rejected; backend selection belongs to lig config/runtime gates, not source-level decorator strings.

This is placement metadata only. It does not yet lower kernels to LKIR, allocate device buffers, prove address-space separation, or emit CUDA/HIP/Metal/SPIR-V. Those remain G-gpu work.

See master plan Phase 7d and li-tests/decorator_exploits/ (to land with 7d-e).