Aligning the out-of-line comparator removes the observed long-prefix regression from reusing the carried LCP in the two-target COLA builder. With both versions aligned, their observation ranges overlap in every fixture. I do not claim a demonstrated build speedup from LCP reuse.

The production change gives compare_common_bits a 64-byte entry alignment on Apple Clang/AArch64. It does not force the function out of line, change its instructions or relax any check. Other compiler/target combinations retain the existing declaration. Complete-query collateral measurements show overlapping ranges, and the existing profile tests pass under strict O3 ASan/UBSan.

Whole-build control

The four variants use the same C++ fixture:

B: 111a2d6, checked borrowed-writer append.
C: f12d708, append using the already computed exact LCP.
B+A / C+A: the corresponding headers with the alignment declaration.

Five fresh-process trials per variant, each with one warm-up and three timed rounds, produce 480 observations. The timed node has 2,048 native records and samples a main node containing 4,096 native records plus its own borrowed samples, and a 4,096-record secondary leaf. The main node's two routes share one native leaf. Keys have a fixed-width ordered integer suffix, with either zero or 4,096 common prefix bytes; bit policies add three trailing bits. Values are variable-width. The bit K=3 case uses Golomb-3 controls.

The timer includes builder construction, step and finish; serialization, result/builder destruction and fixture construction are outside it. Every process dumps its eight warm-up index files, which the runner compares byte-for-byte across all variants and trials. Timed rounds serialize afterward and check CRC and encoded size. These are exact-wire controls against the checked implementation, not a new independent full-key oracle.

The M2 Max run uses Apple Clang 21.0.0, -O3 -DNDEBUG -std=c++20, under the shared exclusive CPU lease. The runner records its rotating/reversed process order; five trials are not a fully position-balanced schedule. Ranges below are observed minima and maxima, not confidence intervals. All observations, including short fixture outliers, are retained.

Nanoseconds per virtual occurrence, median [minimum, maximum]:

Unit	K	Prefix bytes	B	C	B+A	C+A
byte	3	0	148.0 [142.2, 329.6]	142.9 [135.2, 344.0]	146.9 [138.5, 226.7]	141.6 [134.3, 283.6]
byte	15	0	159.2 [153.3, 282.0]	158.6 [147.2, 287.5]	163.5 [151.9, 267.4]	162.5 [155.2, 273.4]
bit	3	0	244.4 [231.8, 362.3]	249.1 [228.2, 386.6]	247.1 [229.9, 309.0]	238.1 [229.0, 368.6]
bit	15	0	277.0 [261.1, 368.1]	275.0 [257.7, 370.3]	273.2 [261.5, 356.5]	265.7 [251.9, 370.3]
byte	3	4096	1048.2 [997.4, 1189.9]	1348.7 [1328.4, 1463.8]	1043.8 [1017.8, 1097.3]	968.0 [927.6, 1080.6]
byte	15	4096	1191.6 [1106.6, 1220.1]	1593.7 [1512.7, 1697.0]	1164.0 [1108.4, 1297.4]	1130.3 [1076.9, 1232.7]
bit	3	4096	1124.1 [1095.1, 1195.4]	1424.9 [1398.5, 1487.2]	1131.9 [1092.0, 1175.6]	1074.3 [1027.6, 1104.8]
bit	15	4096	1288.6 [1232.2, 1418.3]	1692.6 [1641.5, 1813.3]	1275.0 [1230.0, 1361.6]	1244.3 [1198.2, 1301.6]

Without alignment, C is 26.8–33.7% slower on long prefixes and all four range pairs are disjoint. Alignment removes those regressions. C+A medians are 2.4–7.3% below B+A on long prefixes, but their ranges overlap; short-prefix comparisons overlap too. The useful demonstrated result is removal of this compiler-layout regression.

Instruction and size evidence

The benchmark include closure contains only two changed headers between B and C: the COLA append call and its borrowed-writer friend declaration. The revision also changes multiverse.h, which this fixture does not include. append_known remains out of line in both builds, and there is no new full-prefix copy.

All four emitted comparator bodies contain the same 209 instructions after expressing branch addresses relative to the function entry. The normalized bodies have the same SHA-256: c5f83f67e218e9f94b53e415234c3c8f8e8ad26e727d1ab01a3c56cf8560a346. The aligned bodies have another four unreachable padding instructions afterward.

Variant	Function entry	NEON loop start	Loop start mod 64
B	`0x100009db0`	`0x100009fc0`	0
C	`0x100009dcc`	`0x100009fdc`	28
B+A	`0x100009e40`	`0x10000a050`	16
C+A	`0x100009e40`	`0x10000a050`	16

The 40-byte NEON equality loop crosses a 64-byte boundary only in C. The identical instructions, repeated regression and alignment intervention strongly support a code-layout explanation. They do not isolate a particular cache or fetch mechanism; alignment also shifts other code, and no hardware-counter experiment was performed.

Alignment has a code-size cost in this complete executable:

Variant	`__text` bytes	`__TEXT` segment bytes	Exception table bytes	Unwind bytes
B	216,860	245,760	7,612	2,128
C	216,964	245,760	7,612	2,120
B+A	219,644	245,760	7,612	2,128
C+A	219,616	245,760	7,612	2,120

The 2.7 KiB increase includes padding beyond the comparator itself. The mapped text segment size is unchanged. Other compilers, compiler versions, inlined copies and caller mixes may lay out this code differently; this is a local measured alignment choice, not a general instruction-cache bound.

Query collateral and correctness

The unchanged whole-query harness compares f12d708 against the alignment-only commit a4c823e, with 4,096 records and 4,096 queries, five alternating trials, and byte/bit K=W=15 policies. Every process checks exact source, ordinal and value against its independent integer-key oracle before and after timing. Cross-version checksums, catalog visits, matches and encoded-array sizes agree.

Nanoseconds per complete query:

Unit	Prefix bytes	Before, median [min, max]	Aligned, median [min, max]	Median change
byte	0	1869.3 [1827.0, 3026.9]	1909.2 [1824.8, 2277.4]	+2.1%
bit	0	2920.7 [2821.3, 3200.6]	2870.1 [2801.1, 2963.3]	−1.7%
byte	4096	2271.4 [2208.1, 2341.7]	2218.2 [2195.9, 2267.0]	−2.3%
bit	4096	3785.9 [3731.6, 3862.1]	3850.7 [3758.9, 3914.8]	+1.7%

All query ranges overlap. Preparation medians increase 3.3–28.5%, but those short single preparations also have wide overlapping ranges; the raw data retains them. There is no demonstrated query or preparation speedup. The existing profile suite passes with -O3 -Wall -Wextra -Wpedantic -Werror -fsanitize=address,undefined; validation metadata records the source, header, compiler and command.

Reproduction and raw data

The runner snapshots fresh headers and records every header, source, executable and warm-up file hash. The archived candidate patch reconstructs the candidate with normalized names, so reproduction does not depend on retaining the experimental Git branch:

python3 bench/cola_layout.py --baseline 111a2d6 \
  --candidate-patch bench/results/cola_layout_candidate.patch \
  --build-dir build-world-layout --trials 5 --rounds 3

Run under the host's exclusive CPU/build-directory lease. The recorded runner also uses current names; its metadata retains the original measured hash. The reusable runner accepts the archived header patch and records normalized source hashes separately.

Final control: 480 CSV rows, metadata and commands, recorded runner, instruction excerpts and section sizes.
Query collateral: 40 CSV rows, metadata and commands.
Earlier measurements, retained separately rather than pooled into these tables: original 240 rows, original metadata, initial four-way 288 rows, initial metadata.