Every C++ command and flag for Grok16 @ 16.1.1. Use g16 (or g++16 symlink) — auto-detects C++ and runs g16-cxx.
| # | Item | Short | Detail |
|---|---|---|---|
| 1 | g16 | Unified driver (C++) | Same bin/g16 as C. Auto-selects g16-cxx for .cpp/.cxx/.cc, -std=gnu++26, -x c++, -lstdc++, or argv[0] g++16 symlink. One command for the whole toolchain. |
| 2 | g++16 | Compat symlink | Symlink to g16. Preserves legacy scripts and muscle memory. Detection uses basename g++16 → C++ mode immediately. |
| 3 | g16-cxx | C++ backend | Relocated GCC g++ driver. Handles libstdc++, crtbegin/crtend, template parsing, gnu++26. Unified g16 execv() here in C++ mode. |
| 4 | g16 -std=gnu++26 | Default C++ standard | G16_CXX_STD=gnu++26. World_Redata field_g16.hh requires __cplusplus >= 202400. Pass explicitly when mixing with older TUs. |
| 5 | verify (C++ leg) | C++ smoke | verify.cpp compiled with g16 -std=gnu++26. Checks __cplusplus >= 202400. Optional CMake minimal-cmake-project links via same g16 binary. |
| 6 | PROFILE cxx | C++ compile flags | Full profile: -std=gnu++26 -O3 -march=native, entropy/Field macros, vectorize, fast-math per profile. |
| 7 | PROFILE link | C++ link flags | LTO thin/full from grok16_lto probe, OpenMP for field_compute. Normalized by grok16-profile-flags.py. |
| 8 | PROFILE source | Bench source path | Returns examples/*/bench .cpp for the profile. Used by _bench_run_one and PGO training. |
| 9 | PROFILE defs | -D macros only | Profile definition flags without optimization bundle — useful for CMake INTERFACE targets. |
| 10 | PROFILE cxx_pgo_gen | PGO generate | -fprofile-generate to data/pgo/. Run via grok16-toolchain.sh profile. |
| 11 | PROFILE cxx_pgo_use | PGO use | -fprofile-use after training. Enable G16_ENABLE_PGO=1 on field-bench for ~few-% win. |
| 12 | field_opt | Primary profile | FieldX86 + entropy + NEXUS. Bench: field-nexus-bench ~2.1ms kernel vs -O2 2.65ms (~19%). |
| 13 | ai | Matrix profile | NEXUS matrix scoring. Bench wall_ms ~4.0. -flto=thin link. |
| 14 | field_compute | CANVAS profile | OpenMP SIMD dispatch kernels. field-canvas-kernel bench. |
| 15 | vulkan_rtx | RTX SIMD profile | AVX2/FMA CPU prep. Shares field-nexus-bench source. |
| 16 | field-bench | Primary benchmark | G16_FIELD_SPEED=1 → field_opt. Writes data/bench/latest_*.json. |
| 17 | bench-all | All profiles | Runs field_opt, ai, field_compute, vulkan_rtx sequentially. |
| 18 | g16_field_mandate | CMake security | fortify, stack protector, RELRO, PIE on field targets. Required for World_Redata L2. |
| 19 | G16_FORCE_CXX=1 | Force C++ backend | Skip detection — always g16-cxx. Useful when linking C++ objects with ambiguous argv. |
| 20 | redata.cli parity | Parity gate | Python ↔ C++ WRDT bytes. Requires real g16 @ 16.1.1 dumpversion in manifest. |