[not-matthias]

Summary

• Price amd64 REP MOVS/STOS through 48 measured curves covering widths, directions, and alignment classes instead of per-element instruction costs.
• Carry the full hot-copy cost in existing Ct/Cl columns. Suppress only REP data L1-hit references; preserve instruction counters, cache simulation, and every data miss.
• Enable with cache and cycle estimation by default. --rep-bulk-cost=no restores legacy accounting. No new output events or downstream format changes.

Verification

• Built on Linux amd64 with GCC 13 and vendored Capstone.
• All 31 Callgrind tests passed across the suite run and a targeted rerun after building its missing demangling fixture. Both new default/legacy regressions pass.
• A separate 256-case smoke passed natively and in default/legacy modes: all MOVS/STOS widths, both directions, four alignments, and zero/nonzero counts. Regression fixtures also verify backward overlap and self-copy results.
• Generated table reproduces byte for byte from the measured input; 48 curves validate.
• Standalone REP/AVX2 copy smoke verifies destination bytes and unchanged AVX counters, instruction counters, and cache-miss counts.

Modeled cycles per copy, including the same loop overhead:

Copy size	REP legacy	REP bulk	AVX2, unchanged
1 KiB	13,343.80	63.64	489.03
10 KiB	133,151.80	209.50	4,608.04
64 KiB	923,749.80	83,246.62	101,071.16

Modeling limits requiring review

This is deliberately a draft, not a claim of calibrated end-to-end speedups. The native 10 KiB comparison was 34.97 ns REP versus 41.46 ns AVX2, whereas the new score favors REP by about 22x. At 1 KiB native AVX2 is faster, but the model favors REP. Existing scalar/vector pricing is unchanged.

The table was measured on Zen 5; the existing static instruction LUT is Zen 4. Calibration targets disjoint L1-resident buffers, up to 16 KiB MOVS / 32 KiB STOS. Larger sizes extrapolate the final slope and retain existing miss penalties. Overlap performance is outside the calibration domain, although copy semantics are preserved. Held-out L1 samples at least 1 KiB have 0.203% median absolute relative error but 65.1% p95 and 154.8% maximum error; smaller sizes also have substantial relative-error outliers. Mixed REP/non-REP Ct/Cl aggregation retains the existing geometric-mean cross terms.

Dr/Dw become modeled miss-only references for supported REP instructions, not raw access counts. Results across this model boundary should not be interpreted as application performance changes.

[codspeed]

Merging this PR will improve performance by 10.56%

⚡ 1 improved benchmark
✅ 83 untouched benchmarks
⏩ 60 skipped benchmarks¹

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	test_valgrind[valgrind.codspeed, stress-ng --cpu 1 --cpu-ops 10, full-with-inline]	7.5 s	6.8 s	+10.56%

Tip

Curious why performance improved? Comment @codspeedbot explain why performance improved on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing rep-bulk-cost-model (a315af8) with master (d034e80)}

Footnotes

1. 60 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩