[lemire]

Updates the amalgamated CRoaring from 5.1.0 to 5.5.1 and the expected version in TestVersion. No API changes.

Performance

Measured on a Xeon Gold 6548N (Emerald Rapids), comparing CRoaring 5.1.0 with 5.5.1. The suite is the package's benchmarks plus 47 set-operation benchmarks (61 in all), with 10 interleaved runs each, pinned to one core. The AVX2 build compiles out the AVX-512 code with CGO_CFLAGS=-DCROARING_COMPILER_SUPPORTS_AVX512=0.

	AVX-512	AVX2 only
geometric mean	−9.4%	−9.5%
Or / Xor / AndNot, small vs. large bitmap	−68% to −73%	−70% to −73%
FastXor	−47%	−47%
Flip (sparse)	−58%	−55%
Or / Xor, two sparse bitmaps	−9% to −14%	−8% to −10%
Add many (sorted)	−14%	−10%
FastOr	−3.9%	unchanged

CRoaring 5.1.1 through 5.4.0 made FastOr up to 1.42× slower on AVX-512 hardware. That was fixed upstream in RoaringBitmap/CRoaring#952 and RoaringBitmap/CRoaring#953 (5.5.0 and 5.5.1).

A few benchmarks are 1–3% slower, such as Intersect on arrays (+2 ns) and AndNot on two sparse arrays (+1.5%). The C code on those paths is identical to 5.1.0, apart from a skew check in the array difference that makes the small-vs-large cases 3.6× faster. With functions and loops aligned to 64 bytes, all of them but one fall into noise; the remaining one is +1.5 ns on Intersect between a run and an array container. So I attribute them to code layout.

Tests

go vet and go test ./... pass on linux/amd64, with and without AVX-512, and on darwin/arm64.