# Keep the existing CMOC Release defaults November 12, 2026. Retain generated C projects' Release `-O2` default and Debug `-O0 -fno-omit-frame-pointer`. Keep `-fomit-frame-pointer` available as an explicit project compiler option. No VCC Plus product version, candidate or project Makefile was changed by this qualification. ## Measured outcome All **864 executions passed**, comparing complete serialized result state with independent Python expectations. Repeated cycle counts agreed exactly. arm64 and x86_64 compiler slices emitted identical guest binaries for equivalent builds. | Comparison | Result | | --- | --- | | Release versus Debug, MC6809 | 14.17% fewer cycles, equal-workload geometric mean | | Omission versus existing Release, MC6809 | 0.69% fewer cycles | | Omission versus existing Release, HD6309 emulation | 0.69% fewer cycles | | Omission versus existing Release, HD6309 native | 0.84% fewer cycles | | Omission versus existing Release, initialized size | Approximately 3% fewer bytes, summed across eight independent programs | | Largest MC6809 omission speed gain | Input handling, 2.88% fewer cycles | | Per-program initialized size savings | 17–32 bytes | [Per-workload measurements](MEASUREMENTS.md) and [complete structured receipts](results.json) retain every seed, repeat, mode, output and binary hash. The geometric mean gives each workload equal weight. It is not a model of a CoCo Man frame. Linked size includes startup, runtime helpers, setup and result collection. The RAM buffer and collection helpers are test harness overhead, so the percentages are not a prediction of whole-game size savings. The existing Release optimization provides a useful improvement over Debug. Frame-pointer omission produces small additional improvements here, but the compiler labels that option experimental. These measurements support leaving it available for deliberate per-project use; they do not justify changing the application's default for everyone. The benchmark has no native UI changes. ## Exact compiler and patches The compiler was taken from the retained, signed/notarized VCC Plus **3.6.2 build 55** app. Its bytes match `/Applications/VCC Plus.app` as inspected for this run. - CMOC 0.1.100, LWTools 4.25, mcpp 2.7.2. - Universal `libexec/cmoc` SHA-256: `c46271d42037420d3f9d294c670f2d6130c469cf6fd2c41ec2338d990e313512`. - Shipped manifest markers `vccAddZeroFlagsFix`, `vccIndexedByteSubtractionFix`, `vccIndexedArgumentLoadFix` and `vccDebugMetadata` are all 1. - No patches were applied during benchmarking and no toolchain was rebuilt: those fixes are already in the supplied release compiler. - The mutable local staging compiler had different bytes and was not used. - Each compiler architecture was selected explicitly on the unchanged native compiler, with the shipped wrapper's equivalent data-alias environment. A stock-CMOC timing comparison is deliberately outside this decision. Previous VCC Plus regressions established that the optimizer patches correct wrong results, including indexed argument loads failing at `-O2` on both guest CPUs. The input benchmark also exercises adjacent array/calibration arguments. Removing correctness fixes to obtain a faster number would not establish a usable default. This suite can be rerun against a separately identified stock runtime if the question later becomes the patches' performance cost. ## Execution and confidence - Eight authored workloads; seeds 1, 165 and 65521; two repeats per cell. - Three profiles, two compiler host slices and three guest CPU modes. - 6809 machine code in every profile; the harness explicitly enables HD6309 native mode for that separate execution dimension. - Correctness covers the serialized mutable workload state, not just a digest. - Start/stop breakpoints exclude setup and collection, include the call/return, and mask guest IRQ/FIRQ. No baseline subtraction or host elapsed timing. - Independent three-NOP calibration passed: 6/6/3 cycles respectively. - No live native app or personal media was operated. The benchmark loads its own ROM-backed headless machine in a separate process and attaches no disks. - All four evidence-accounting unit checks and all 67 shared Release core checks passed. No tests were skipped in these runs. The linked Release core is identified by its recorded library SHA-256. The developer directory was selected for the build command without changing the Mac's global Xcode selection. The x86_64 compiler ran under Rosetta on Apple Silicon. There was no physical Intel, macOS 14, physical CoCo, audible output or UI accessibility qualification. The graphics, audio, input and disk workloads are CPU-side synthetic buffers. They do not measure GIME presentation, real DAC timing, analog controls, disk latency or device-library compatibility. Complex games, interrupts and inline assembly remain separate requirements before adopting omission globally. Debug frame preservation continues to support VCC Plus's verified local variables and call-stack reconstruction; this run makes no broader optimized-debugging claim. ## Public evidence The publication view of the [numerical record](https://vccplus.brianholman.com/publications/vcc-plus-architecture/evidence/research/results/cmoc-benchmark-recorded-json/) preserves the measured outputs, cycles and limits. This companion projection does not repeat the experiment.