Type specialization

Speed mode (-O2) may clone a function into representation variants guarded by runtime tag checks. -O0 and --no-type-specialization produce generic code only.

Policy

• Profiles come from inferred call-site facts. Today only exact, target-representable integer singletons prove a small-integer argument. Specs, broad integer(), unions and unknowns stay generic.
• A variant is justified only if it removes an implemented check: an exact low-tag comparison on an argument load in the entry block, before side effects.
• Limits: 3 variants per function, 32 per module, 128 per target. Estimated and then actual growth (clone + guards + dispatch + fallback) must stay within 2x the generic instruction count per function and module.
• Generic bodies are always kept. Over-budget drafts are discarded silently; the program never fails because of specialization.
• Selection is deterministic (module, symbol, profile order).

Lowering

LLVM cloning/simplification replaces proven checks inside the clone. A bounded dispatcher tests every constrained tag and forwards context and arguments to a variant or the generic body; the public symbol and ABI do not change. No unchecked unboxing, spec assumptions or fast-math flags are introduced.

Current effect

The current source subset has no removable representation checks, so real Erlang source receives zero variants even at O2; specializing an identity would only grow code. Focused LLVM fixtures inject repeated tag checks to exercise hits, fallbacks and growth rollback with native execution.

codegen_measurements writes measurements/<config>/measurements.json in the codegen test build directory (compile time, IR/object bytes, variants, native time). Times are descriptive, never thresholds. Windows x64 / LLVM 23.1.2 sample (2026-09-29): source workload 232,876 IR / 40,498 object bytes at O0 and 204,408 / 35,886 at both O2 policies, zero variants; the synthetic fixture added 30 instructions to a 55-instruction function for one accepted variant.