Phase 7d-c — @vectorized on for enables scoped array SIMD¶
Summary¶
@vectorized(lanes=4) on a for loop emits ArraySimdScope MIR so array @ / element-wise ops inside the body use f64×4 codegen even when the enclosing def has @no_vectorize.
Agent continuation¶
- Read
ArraySimdScopeinmir.hpp,lower.cppForlowering,emit.cppscope stack. - Run
./li-tests/run_all.sh(172 pass). - Then benchmarks ingest;
**on arrays (2i). - Blocked on auto-detecting reduction loops without explicit
@vectorized.
Changed¶
compiler/mir/,compiler/codegen/emit.cppli-tests/decorators/vectorized_for_scope_ok.li
Not changed¶
@parallelonfor; matmul auto-SIMD inner loops.
Breaking¶
N/A.
Security¶
N/A.
Performance¶
Inner-loop array ops regain SIMD under @no_vectorize + @vectorized for.
Downstream¶
- AXPY-style loops can annotate
@vectorizedon the serialfor.