|
Everett
|
I compared ordinary cursor advancement with a compile-time choice between ordinary advancement and advancement that also computes the exact adjacent-key comparison. In this run the change is broadly flat in complete sampling and pipeline timings. I do not infer an end-to-end speedup from removing the runtime comparison-pointer branch.
The baseline is main d1cc57368df10ffcade606b14ab40d8e6428e97d, which already contains the accepted sample-encoder comparison reuse and validated-subview change. The candidate is 35be933dffba607e3a482a50f07a3f8c983e2ce7: the complete header trees differ only by four substitutions in profile_cursor:
advance() calls advance_impl<false>(nullptr).advance_comparison() calls advance_impl<true>(&result).advance_impl has the compile-time bool Compare parameter.if constexpr (Compare).The candidate has the same Git tree as 7918b1c2b768d76e21efe445c8fd790c6af2db25 (tree 7a09160749190760163bec5bc9e708421b6ebd00). These names identify the same content, including the common encoder change. The private call sites supply a valid output pointer exactly when that comparison block exists; the previous validation, reconstruction, errors and endpoint handling remain in both modes.
I use the unchanged sampling harness and runner, with the fixtures and independent oracles described in the sampling report. There are 4,096 native keys, byte/bit profiles, interleaved and duplicate-borrow fixtures, and 0/64/4,096 shared prefix bytes. Five alternating process trials with three rounds each produce 720 rows. Every occurrence, counter and exact pipeline-encoding check passed. Separate scoped O3 ASan/UBSan sampling and native-merge tests also passed for this header change.
These are median ns per augmented source occurrence on the same native arm64 M2 Max host, AppleClang 21 -O3 -DNDEBUG, under the exclusive CPU/build-directory lease. Sampling changes range from −3.0% to +2.0%. Short-prefix pipeline changes range from −4.1% to +4.3%; long-prefix changes range from −0.7% to +0.7%. Every baseline/candidate observation range overlaps. I find no material regression in this matrix and no defensible speedup claim.
| Prefix bytes | Fixture | Profile | Sample before | Sample bool | Pipeline before | Pipeline bool | Pipeline change |
|---|---|---|---|---|---|---|---|
| 0 | interleaved | byte | 35.48 | 35.49 | 60.79 | 61.28 | +0.8% |
| 0 | interleaved | bit | 66.87 | 67.65 | 101.16 | 105.51 | +4.3% |
| 0 | duplicates | byte | 35.22 | 35.92 | 57.16 | 58.69 | +2.7% |
| 0 | duplicates | bit | 60.99 | 60.87 | 93.52 | 93.41 | -0.1% |
| 64 | interleaved | byte | 37.60 | 37.33 | 65.99 | 64.56 | -2.2% |
| 64 | interleaved | bit | 71.60 | 69.44 | 110.83 | 106.33 | -4.1% |
| 64 | duplicates | byte | 37.28 | 37.80 | 61.72 | 61.81 | +0.1% |
| 64 | duplicates | bit | 64.21 | 63.36 | 101.38 | 98.48 | -2.9% |
| 4,096 | interleaved | byte | 198.33 | 198.29 | 309.19 | 308.43 | -0.2% |
| 4,096 | interleaved | bit | 231.96 | 229.07 | 350.62 | 353.03 | +0.7% |
| 4,096 | duplicates | byte | 197.20 | 197.01 | 298.12 | 299.56 | +0.5% |
| 4,096 | duplicates | bit | 227.09 | 226.76 | 333.58 | 331.34 | -0.7% |
The complete executable's __text section is 168,764 bytes before and 168,128 after (636 bytes smaller, 0.38%). Its __TEXT segment remains 196,608 bytes; exception tables and unwind information are unchanged. These are this harness's code-size results; another program can instantiate a different mix of the two advance modes. The raw size output and executable hashes make the measurement explicit. I did not separately measure compilation time.
A separate small object-file probe explicitly instantiates byte/bit native cursor advance and advance_comparison. Its ordinary function bodies shrink from 548/504 bytes to 208/208 bytes, with stack frames shrinking from 224 to 144 bytes. The comparison-mode bodies are 544/500 bytes. Instantiating both modes increases this probe's total __text from 9,972 to 10,452 bytes (+480) and compact unwind from 3,648 to 3,904 bytes (+256); exception tables and strings are unchanged. That object-file result and the complete harness's smaller code section describe different consumers. The probe source, commands, symbols and hashes and its baseline assembly and candidate assembly preserve the evidence. No timing is inferred from those function or frame sizes.
The baseline above already contains the accepted bit_view::subview change. Its separate earlier probe compares f072b1c with 37b8b28: after checking the requested range, a subview reuses its validated backing span instead of repeating the original view constructor's capacity checks. In that small arm64 probe the successful path goes from three conditional branches to one and from a 16-byte stack frame to no frame. This is assembly evidence, with no attributed timing measurement. The probe metadata and exact source, before assembly and after assembly keep it separate from the bool-specialization benchmark.
The CSV preserves all 720 observations; the metadata records full revisions, every header hash, the shared harness/runner hashes, compiler commands, host and trial order. Only CSV line endings are normalized to LF. Run inside the host's configured CPU/build-directory resource lease:
The sample timer includes the five-scalar result consumer; the pipeline timer includes construction and finalization. Full correctness checks and destruction remain outside the timers. These warm in-memory observations do not establish cold mmap, storage-I/O, other-architecture or workload-independent performance.
1.9.8