Add arczi84's NFS3 floating-point kernels

[codewiz]

Adds two floating-point benchmarks from arczi84's 68k port of Need for Speed III, contributed under 0BSD:

• nfs3-project: projectObjectVertex(), a vertex through a 3x3 matrix, a translation and the perspective divide, every result checked by testing its float bits.
• nfs3-clip: PolygonClipWork::run(), a triangle clipped against the near plane and the four screen edges, with float compares, divides and interpolation.

The kernels are C++, so only gcc builds them. Built with -fno-exceptions -fno-rtti -fno-threadsafe-statics, they need nothing from libstdc++, and the existing C link takes them. vbcc and SAS/C skip both rows. kernel.cpp is arczi84's file with his main() replaced by two entry points. Those set the FPCR to single precision, as the game does, and restore it afterwards. source.sha256 records the hash of the file as received.

Results, 25 MHz 68040 under volamos:

• 1064 and 951 ms at -O2; gcc 16.2-rc14 beats 6.5.0b by 8.90% and 5.40%.
• -Os makes both kernels 2.5 times slower: it keeps the GuardedFloatMath operations out of line, so every float operation becomes a call.
• The check values agree across every build, soft float and gcc 6.5.0b included. Five rounds of arczi84's sizes give 449f5330 and 4352ff86, the sums his README lists for x86-64 rather than his PiStorm ones. Every result is rounded to IEEE single, as SSE does, so the PiStorm (Emu68) may not apply the FPCR's precision control like a real FPU.

Also in this PR:

• RESULTS.md gets the new rows, and its gcc columns are rerun on the current tree. The compiler comparison's geomean leaves the C++ rows out (new --geomean-skip in the results tools), so it still covers the same 16 benchmarks for every compiler.
• The harness stops on Ctrl-C between benchmarks. libnix's stdio used to do this; newlib's does not.
• The name column is widened to 16 characters, and the header is now uppercase and aligned.

Generated with Claude Code

---

Note

Low Risk
Benchmark and build/doc additions; no changes to existing benchmark logic beyond optional registration and reporting layout.

Overview
Adds two gcc-only C++ benchmarks from arczi84’s NFS3 68k port: nfs3-project (vertex projection with guarded float math) and nfs3-clip (triangle clipping). Vendored code lives under third_party/nfs3/ with 0BSD license; the build wires in NFS3_SRCS, a %.cpp compile rule, and BENCH_HAVE_NFS3 so vbcc/SAS/C builds skip these rows.

The harness widens benchmark names to 16 columns, checks Ctrl-C between benchmarks (for newlib), and registers the new benches when NFS3 is enabled. Results tooling gains --geomean-skip so compiler-comparison geomeans exclude benchmarks not every toolchain builds; README, RESULTS.md, charts, and Aminet readme are updated with the new rows and refreshed timings.

^{Reviewed by Cursor Bugbot for commit 11a7984. Bugbot is set up for automated code reviews on this repo. Configure here.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-12T16:24:29.548574Z	d80e266	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[Copilot]

🟡 Changes recommended

Skipped benchmarks with checksum mismatches incorrectly suppress the chart’s otherwise valid geomean.

1 open finding

• Skipped checksum mismatches incorrectly mark chart incomplete · New

What changed in this PR

Adds two gcc-only NFS3 floating-point benchmarks and integrates them into builds, reporting, and documentation.

Changes:

• Adds projection and clipping kernels with benchmark harnesses.
• Extends build and results tooling for C++ and selective geomeans.
• Refreshes benchmark results and adds Ctrl-C handling.

File	Description
tools/​results-table.py	Adds geomean exclusions.
tools/​results-chart.py	Charts excluded benchmarks.
third_party/​nfs3/​source.sha256	Records upstream source hash.
third_party/​nfs3/​README.md	Documents kernel provenance.
third_party/​nfs3/​nfs3.h	Declares C-compatible entry points.
third_party/​nfs3/​LICENSE	Adds the 0BSD license.
third_party/​nfs3/​kernel.cpp	Implements both floating-point kernels.
third_party/​LICENSE.md	Registers the NFS3 license.
src/​main.c	Registers benchmarks and handles Ctrl-C.
src/​bench.h	Exposes benchmark descriptors.
results/​opt-levels-size-040.svg	Updates code-size results.
results/​opt-levels-040.svg	Updates optimization charts.
results/​compilers-040.svg	Updates compiler comparison chart.
RESULTS.md	Documents refreshed measurements.
README.md	Lists the new benchmarks.
Makefile	Adds C++ object compilation.
compiler.mk	Configures gcc-only C++ builds.
benches/​nfs3.c	Adapts kernels to the harness.
aminet/​benchwork.readme.in	Updates package documentation.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.