Skip to content

Commit f94594d

Browse files
committed
[improvement](be) Compact selected row indices with mask lookups
### What problem does this PR solve? Issue Number: None Problem Summary: Partially selected SIMD masks are expanded one bit at a time, creating a serial dependency for every selected row. Use a constexpr eight-row lookup table and vector stores for x86 masks with more than two selected rows. Keep the original ctz loop for sparse masks and other architectures. Preserve zero/full-mask shortcuts, tails, row order, and downstream materialization. The selector already has one slot per input row, so fixed-width stores fit in that allocation even when only the returned prefix is valid. On a Xeon Platinum 8457C, 4096-row batches and random 50% selection, the actual helper reduces standalone compaction from 0.693280 to 0.119248 ns/input row (5.81x, Clang 20 native). Including one int32 predicate comparison and an 8-byte payload copy improves from 1.170464 to 0.643518 ns/input row (1.82x). Tradeoff: the Clang 16 AVX2 random 1% compaction-only case changes from 0.104319 to 0.119946 ns/input row (15% slower); the Clang 20 native 1% case is unchanged. These measurements do not establish complete SQL query speedups. ### Release note None ### Check List (For Author) - Test: Manual test: native and AVX2 standalone benchmark; ASan+UBSan exhaustive mask, tail, row-order, output-bounds and payload verification; clang-format 16; repository build-hygiene and focused clang-tidy checks. Add SelectionTest cases for masks, large row bases and consecutive appends. Standard BE UT and full ASAN BE build were attempted with -j 48, but failed in the unrelated existing local ColumnVector span-iterator/PODArray copy constructor at column_vector.h:95 before the new GTest cases could execute. No SQL test run. - Behavior changed: No - Does this need documentation: No
1 parent 76011df commit f94594d

4 files changed

Lines changed: 872 additions & 5 deletions

File tree

0 commit comments

Comments
 (0)