Use AVX512_VPOPCNTDQ for the population-count kernels

[lemire]

Description

Adds AVX-512 counterparts to the five AVX2 population-count kernels, selected when the CPU has AVX512_VPOPCNTDQ.

Type of Change

• Performance improvement

Changes Made

AVX2 has no vector population count, so popcnt_avx2_amd64.s builds one out of a VPSHUFB nibble lookup plus VPSADBW, handling 256 bits per iteration. AVX512_VPOPCNTDQ provides VPOPCNTQ, which counts all eight 64-bit lanes of a ZMM in a single instruction, so the new kernels handle 2048 bits per iteration: four ZMM loads, four VPOPCNTQ, four adds, with four accumulators to keep the adds off one dependency chain.

The five dispatchers in popcnt_avx2_amd64.go gain one branch each; no call sites change. CPUs without VPOPCNTQ keep the existing AVX2 path, and non-amd64 and appengine builds are untouched.

Detection is cpu.X86.HasAVX512VPOPCNTDQ from golang.org/x/sys/cpu, which already verifies that the OS saves the opmask and ZMM state and honors GODEBUG=cpu.avx512vpopcntdq=off to disable the path at run time. This makes golang.org/x/sys a direct dependency; it was already in the module graph.

Performance Impact

One 8 KB bitmap container (1024 words) on a Xeon Gold 6548N (Emerald Rapids), pinned with taskset, Go 1.26.3:

kernel	pure Go	AVX2	AVX-512	vs AVX2
popcntSlice	303.2 ns	179.9 ns	44.7 ns	4.02x
popcntAndSlice	449.0 ns	225.9 ns	68.5 ns	3.30x
popcntOrSlice	-	226.9 ns	68.5 ns	3.31x

These back GetCardinality, computeCardinality, AndCardinality, OrCardinality, XorCardinality, Intersects, and the cardinality half of every bitmap-container boolean operation.

Testing

go test ./..., roaring64, the smat corpus and smat-hit checks, gofmt, go vet, and cross-builds for 386/arm/arm64 and -tags appengine, all on VPOPCNTQ hardware so the new path actually executes.

TestPopcntAVX512MatchesGo compares all five kernels against the portable implementations at lengths 0,1,2,3,4,5,7,8,15,16,17,31,32,33,63,64,65,1023,1024,1025 - covering the empty case, the sub-block cases, exact multiples of the 32-word main loop, and the scalar tail. TestPopcntDispatchMatchesGo checks the dispatchers themselves on any CPU.

Breaking Changes

None.

Additional Notes

Part of a series applying AVX-512 to roaring, following CRoaring's kernels. It overlaps with #557 and the follow-ups only in go.mod/go.sum (all add the same golang.org/x/sys requirement); the code is disjoint. Worth deciding as a group whether these should share one AVX-512 feature gate rather than one flag per feature - CRoaring gates on VBMI2 + VPOPCNTDQ + BITALG together.

https://claude.ai/code/session_0123ePBefqPjrCvhxdwWFakj