Emit profiles (default / harden / lean)

in compile and in build accept an emit profile that reshapes Core IR optimization and (for harden) native codegen fingerprints.

in compile --path examples/compile/antidecomp_sample.in \
  --target native --target-triple x86_64-unknown-none \
  --linkage static-lib --entry main \
  --out /tmp/sample.o --profile harden

in build --path examples/compile/antidecomp_sample.in --out /tmp/sample --lean
# shorthands:
in compile ... --harden
in compile ... --lean

Profiles

Default details (fast, non-classic)

Default is the runtime profile. It does not try to look like gcc/clang:

• Frame pointer via lea, not mov rbp, rsp / mov rsp, rbp
• No per-call rep stosq of the scratch frame (locals are written before use)
• More inlining than a textbook two-stmt helper pass

Harden details

Harden uses the same Core IR → native lowering → JIT/artifact pipeline as other profiles. It preserves typed JIT results and the requested target and linkage. Native compilation defers program execution until runtime.

IR passes add MBA expressions, opaque predicates, _pc dispatch, and internal _H names. Native emit adds profile-specific instruction shapes. These are obfuscation measures, with no claim of cryptographic protection.

Lean details

Lean raises the inliner threshold (2 → 12 stmts) and recursion depth (10 → 24), runs two inline waves, then the usual fold/DCE/dead-fn cleanup. Prefer this when you want smaller call graphs inside a module without harden noise.

Dual-emit

From one CLI invocation, emit two artifacts:

1. runtime — default or lean only (fast; no harden transforms — no CFG _pc dispatch, no alien MBA/junk harden shapes)
2. distribution / anti-decomp sample — always EmitProfile::Harden, written to --harden-out (or derived from --out by inserting -harden before the extension)

The hot-path production/runtime emit must not enable harden IR passes. Profiles stay orthogonal: lean ≠ harden. Dual-emit is rejected when the only requested profile is harden (--harden / --profile harden).

--harden-out together with --out implies dual-emit even without --dual-emit. --out is required when dual-emit / --harden-out is set (in compile already requires --out; in build requires it in this mode).

in compile --path examples/compile/antidecomp_sample.in \
  --target native --target-triple x86_64-unknown-none \
  --linkage static-lib --entry main \
  --out /tmp/sample.o --harden-out /tmp/sample-harden.o
# or:
in compile ... --out /tmp/sample.o --dual-emit
in build --path ... --out /tmp/sample --dual-emit --lean   # runtime=lean, harden beside it

Derived harden paths: foo.o → foo-harden.o, foo → foo-harden.

Honest limits

• Harden is limited to the same supported native subset as other profiles. Lite CFG dispatch skips try/match/throw.
• No cryptographic string encryption; string obscuring is best-effort.
• Does not defeat a determined reverse engineer with dynamic tracing.
• Debug stripping: Core IR has no debug_value; SIL helpers already strip them when SIL is materialised. Harden does not add DWARF.
• Correctness first: transforms must preserve observable semantics for the owned subset.
• Break is not used for dispatch exit because native lower currently treats Break as a no-op; the dispatcher exits via _pc bounds instead.

Ghidra / objdump smoke

See scripts/ghidra-antidecomp-smoke.sh. Prefer dual-emit (--out + --harden-out, or --dual-emit) as the one-shot way to produce both metrics artifacts. Separate --profile default / --profile harden compiles still work. When Ghidra + Java are absent the script still writes objdump/nm metrics under docs/benchmarks/, exiting 0 with a skip note for CI.