TSC Clocks + Benchmark Methodology Rework

[MoonFlowww]

Summary

Adds three RDTSC-family clock backends behind compile-time macros, with a runtime calibration cascade and chrono fallback. Reworks the benchmark to remove self-measurement bias and -O3 collapsing of the no-track baseline. A few smaller cleanups in the stats path.

---

Major changes

1. TSC clock backends (ctrack.hpp)

New Clock_RDTSC / Clock_RDTSCP / Clock_RDTSCP_LFENCE structs selected via CTRACK_CLOCK_RDTSC* macros. Selected backend aliased to ActiveClock; chrono is the default when no macro is set. Hard #error on non-x86_64 if a TSC macro is defined.

All std::chrono::high_resolution_clock::time_point and duration_cast<nanoseconds> calls in Event, Simple_Event, EventGroup, store, ctrack_result*, and EventHandler are replaced with ActiveClock::time_point and ActiveClock::duration_ns(). The interface contract: NOW(), duration_ns(s,e), to_string(tp) is uniform across all four clocks.

Calibration cascade in calibrate_tsc():

Source	Path	Coverage
C1	CPUID 0x15	Intel Skylake+ (exact TSC Hz)
C2	CPUID 0x16	Intel Haswell+ (base MHz, assumed = TSC)
C3	/sys/.../cpu0/cpufreq/base_frequency	Linux + intel_pstate
C4	HKLM\...\CentralProcessor\0\~MHz	Windows registry
Fallback	3× 1ms __rdtsc vs steady_clock, median	AMD / VMs

Calibration runs once via a function-local static const bool _ = (calibrate_tsc(), true); inside EventHandler's ctor. Anchors tsc_anchor_cycles + tsc_anchor_system for to_string conversion.

If every source returns 0 the library calls std::abort() with a message pointing at the macro to remove. No silent fallback to chrono — wrong-by-frequency-ratio numbers are worse than an abort.

2. Benchmark methodology (ctrack_benchmark.cpp)

Three fixes that together remove the bias the old harness had:

BENCHMARK_NOINLINE on busy_wait_ns and every *_no_track helper. The tracked variants are naturally barriered: each EventHandler ctor/dtor mutates thread-local event state, so the compiler cannot reorder or fuse adjacent calls. The *_no_track variants have no such barrier — at -O3 the entire call tree gets inlined into the worker loop as a flat sequence of busy_wait_ns calls, which the scheduler can then reorder asymmetrically vs the tracked path. NOINLINE on the no-track side restores call-site symmetry with the tracked side, so the delta reflects CTRACK overhead and not asymmetric optimization.

raw_clock_ns() — CLOCK_MONOTONIC_RAW on POSIX, QueryPerformanceCounter on Windows. Replaces std::chrono::high_resolution_clock as the outer timer in measure_overhead(). Measuring a solution with itself adds bias; the outer timer needs a path independent of whatever ctrack uses internally.

measure_overhead() restructure:

• Warmup pass before the trials so caches and branch predictors are hot. Its result is not used in the median.
• 5 trials with alternating order (no_track first vs track first per parity) so any first-trial effect cancels across pairs.
• ctrack::result_as_string() moved outside the timed window: pre-clear of accumulated state happens before t0, post-clear after t1. The old harness called it inside the timed window, charging stats-flush cost to overhead.
• Median across trials rejects scheduler outliers.
• Negative raw_diff clamped to 0 with a verbose-mode note (noise floor).
  Outer clock measures the track − no_track delta.

3. Bench results

Variant	Accuracy err	Overhead
chrono	12.84%	17.79%
RDTSC	5.85%	8.55%
RDTSCP	1.57%	15.36%
RDTSCP + LFENCE	0.31%	19.73%

---

Minor changes

• load_child_events_simple: parent_event lookup hoisted out of the inner child loop (was redundantly fetched per child).
• EventHandler ctor signature dropped the defaulted start_time parameter; start_time is now captured after register_event() and after the write_events_locked spin, so the spin no longer counts toward the measured interval.
• EventHandler dtor: removed manual capacity()-size() < 1 check before emplace_back (let vector handle it). Sub-events still use an explicit reserve(max(4, cap*4)) growth pattern.
• BeautifulTable::table_time unit string "mcs" → "us". parse_function_timing in the benchmark updated to match. us is the standard ASCII fallback for µs in low-latency perf tooling; mcs is non-standard.
• table_timepoint rewritten to dispatch through ActiveClock::to_string.
• result_print / result_as_string print TSC frequency: X GHz when a TSC backend is active.