Summary

Use _m (merge) instead of _x (don't-care) for SVE multiply-accumulate operations (svmla, svmls) on accumulator vectors in spatial.h and dot.h.

The _x variant leaves inactive lanes undefined when the predicate is false (i.e., for elements past the end of the input array). However, the final horizontal reduction svaddv uses svptrue (all lanes active), which sums all lanes — including those with undefined values from _x. This causes MemorySanitizer to report use-of-uninitialized-value when the input vector length is not a multiple of the SVE register width.

The _m (merge) variant preserves the first operand (the accumulator) for inactive lanes, keeping them at their initialized zero value. This is correct because:

• Accumulators start at 0
• Active lanes accumulate normally
• Inactive lanes retain 0 (or their previous accumulated value)
• The final svaddv(svptrue, ...) correctly sums all lanes

Affected functions

spatial.h: simsimd_l2sq_f32_sve, simsimd_cos_f32_sve, simsimd_l2sq_f64_sve, simsimd_cos_f64_sve, simsimd_l2sq_f16_sve, simsimd_cos_f16_sve, simsimd_l2sq_bf16_sve

dot.h: simsimd_dot_f32_sve, simsimd_dot_f32c_sve, simsimd_vdot_f32c_sve, simsimd_dot_f64_sve, simsimd_dot_f64c_sve, simsimd_vdot_f64c_sve, simsimd_dot_f16_sve, simsimd_dot_f16c_sve, simsimd_vdot_f16c_sve

Reproduction

Detected as MemorySanitizer: use-of-uninitialized-value in simsimd_cos_f32_sve during ClickHouse CI stress tests on ARM with MSan.