perf(compute): vectorize mixed 64-bit filters

[fallintoplace]

What changed

• Add AVX2 and ARM64 NEON paths for null-free 64-bit fixed-width filters.
• Compact two groups of four uint64 values from each filter byte with table-driven permutations and exact-size stores.
• Keep the existing path for nulls, unsupported CPUs, short or unaligned inputs, dense masks, and noasm builds.
• Add exhaustive mask tests, guard tests, and int64/float64 benchmarks.

Benchmark

Apple M1 Pro, macOS arm64, Go 1.26.3. Median of five runs. The baseline uses GODEBUG=cpu.all=off.

Case	Baseline	NEON	Speedup
int64, 1M, alternating	3.97 ms	0.45 ms	8.7x
int64, 1M, random50	2.86 ms	1.61 ms	1.8x
float64, 1M, alternating	4.07 ms	0.47 ms	8.6x
float64, 1M, random50	2.92 ms	1.65 ms	1.8x

The mixed-mask cases also go from 33 to 32 allocations per operation. The output allocation is still the main B/op cost.

Tests

• Full go test suite with Parquet test data
• go test ./arrow/compute/...
• go test -tags noasm ./arrow/compute/...
• go test -race ./arrow/compute/internal/kernels ./arrow/compute
• amd64 cross-build of the kernel package

[zeroshade]

No SIMD data-path correctness issue found. Focused AVX2 kernel tests and compute filter tests passed, including the noasm fallback. ARM64 was inspected but not executed. One optional build-constraint correction below; this is not a runtime blocker.

[zeroshade]

Non-blocking: the space in this legacy +build constraint means OR, so noasm-only and appengine-only builds still include this assembly source. go list -tags noasm -f '{{.SFiles}}' ./arrow/compute/internal/kernels confirms it. The Go caller is excluded, so it remains unreachable. Please match the uint32 sibling with //go:build go1.18 && amd64 && !noasm && !appengine and preserve the constraint during regeneration.

[zeroshade]

Thanks — the build constraint is fixed, and the Makefile change goes further than I asked.

filter_uint64_avx2_amd64.s now carries //go:build go1.18 && amd64 && !noasm && !appengine, matching the uint32 sibling exactly. Verified by build rather than by eye — go list -tags noasm and -tags appengine both now exclude the file, on amd64 and arm64 alike.

The Makefile rewrite is a nice addition: regeneration will no longer clobber the constraint. The uint32 sibling still has no such guard and would be clobbered; worth a follow-up, not this PR.

Correctness re-verified. preallocateData sizes the output buffer to exactly outputLength * 8 and GetSpanValues applies out.Offset, so len(outData) is precisely the popcount; the exact-size stores stay in bounds, and the 0xdeadbeef sentinel in the test asserts that directly. Input side, len(valuesData) == values.Len == length and the explicit filterByteOffset+numBytes > len(filterData) check covers the filter. The uint64 wrapper is logic-identical to the already-merged uint32 one.

NEON read, not executed. I traced the hand-written kernel against the Go table generator: positions 0-1 land in V4 and 2-3 in V5, the 0x80 fill yields zeroed lanes that are never stored, popcount lives at tables[512+mask], stores are exactly 8/16/24/32 bytes, the pointer advances 64 bytes per filter byte, and g/FP/LR are untouched so NOFRAME|NOSPLIT is valid. I have no ARM64 hardware or qemu here, so the exhaustive mask test on NEON is still unexecuted on my side — arm64 cross-build and vet are clean, but that is not the same thing.

Tests. All 256 mask bytes, the aligned-offset path, and seven guard rejections. I confirmed the AVX2 tests actually run on this host rather than skipping. go test, -tags noasm, and -race all pass across ./arrow/compute/....

Perf on AVX2, since the PR reports only NEON. 1M elements, -benchtime=200x -count=5, medians:

Case	main	PR	Speedup
alternating	10.13 ms	1.57 ms	6.5x
random50	6.77 ms	1.32 ms	5.1x
clustered50	1.03 ms	0.92 ms	1.1x

No regressions. A first pass at -benchtime=50x looked like clustered50 regressed ~14%, but that was noise. Worth noting separately: main's timings swing a lot (alternating ranged 4.7-18.4 ms across runs) while this stays tight at 1.43-1.76 ms — the SIMD path is branch-predictor-independent, which is a real benefit beyond the median.

Approving on the code. One process note: CI has never run here — head 6089dc0a has zero check-runs and combined status pending. I would not merge until the matrix is green, particularly since the ARM64 path is the one neither of us has executed.

Non-blocking: the perl substitution matches only legacy //+build. If c2goasm is ever updated to emit //go:build directly, the rewrite silently no-ops and the constraint reverts to whatever the generator produced. A check that the line was actually replaced would make it durable.