LLVM code generation

The LLVM backend consumes the lowered lowering::BoundProgram; parsing, name lookup, generic specialization, semantic checks, borrow checking, and allocation analysis have already completed. The native compiler configures the host target triple and data layout. LlvmCodeGenerator then declares functions before emitting their bodies and verifies every generated module with LLVM. Before object emission, the native compiler runs LLVM’s target-aware default O2 pipeline, including scalar simplification, loop optimization, and vectorization. Textual --emit-llvm output remains the direct verified lowering and is useful for inspecting frontend code generation.

Implemented scope

The backend emits the language features exercised by the native compiler and end-to-end test suites:

• primitive integers, floating-point values, Booleans, characters, none, and null raw pointers;
• functions, recursion, direct calls, extern declarations, locals, assignments, numeric conversions, and short-circuit Boolean operations;
• conditional and loop control flow, also, break, proceed, positions, and checked jumps;
• strings and string views, including conversions, content comparison, checked view operations, output, and deterministic cleanup;
• raw pointers, checked references, fixed arrays, checked slices, tuples, error values, enums and exhaustive matches, and structures with lifecycle hooks;
• direct deferred calls and cleanup on the supported scope-exit edges;
• system allocation, runtime-sized arrays, typed destruction, and free;
• concrete generic specializations produced before semantic analysis; and
• target-width isize and usize, C ABI integer extension attributes for supported extern calls, object emission, and host-native executable linking.

Aggregate calling conventions are internal to Flecha unless a feature guide explicitly documents C compatibility. The backend diagnoses any unsupported lowered node or conversion instead of emitting unverified IR.

Commands and smoke test

Emit textual LLVM IR with the compiler driver:

make build
build/flecha --emit-llvm examples/llvm_poc.fl -o build/llvm_poc.ll

Run the IR smoke target with:

make llvm-poc

The target emits the example module, executes it with lli, and verifies exit code 42. Native object and executable generation are documented in native code generation.

Current limits

The installed compiler targets its host machine and currently fixes native object optimization at O2. Explicit cross-compilation triples and sysroots, optimization-level selection, source-level debug information, a stable general-purpose Flecha ABI, and a C source backend are not part of the 0.1 implementation. Language feature guides document additional type- and operation-specific limits before code generation.