018. Lower direct predicates for native Float columns

• Status: Implemented
• Date: 2026-10-14

Problem and mechanism

Eligible predicates on directly bound native Float columns previously passed through LATERAL projections with OFFSET 0 barriers. The direct path compares the column with same-Float literals and supports scalar truth and null tests. It also lowers in() when the receiver and literal needles are Float. Mixed Int/Float comparisons retain the exact general path; computed expressions, guards, JSON, and structured representations are not eligible.

The direct comparisons preserve Float64 behavior for schema-valid finite Float values, including signed-zero equality, subnormal values, adjacent representable values, and maximum finite values. Nullable input handling remains explicit, while no-value and null behavior match the general lowering.

Applicability

Reproduction

The baseline is commit d448008. The implementation is included in commit 21bd6ef. The comparison is between the previous projection lowering and the direct Float predicates. Representative filters are factor<1.5, factor=-0.0, !factor<1.5, and age>18 factor<1.5. The comparisons keep the same bound parameters; lengths count generated WHERE characters and barriers count OFFSET 0 projections. The candidate lowering is implemented in backends/sql/scalar_where.go.

Unit and targeted local/PostgreSQL parity tests passed for positive and negative zero, the minimum subnormal, adjacent Float values, maximum finite Float, and large-Int mixed-comparison regressions. Full go test ./..., go vet ./..., and go vet -tags=integration ./backends/sql passed.

Plain EXPLAIN with default planner settings on an empty temporary indexed Float table showed a Seq Scan with a SubPlan/Nested Loop for the baseline and a Bitmap Index Scan for the direct factor<1.5 predicate. Its index condition included factor IS NOT NULL AND factor < 1.5::double precision. This observation is limited to that fixture. No EXPLAIN ANALYZE or execution timing was collected.

Results

These are SQL-shape observations, not runtime measurements. The direct form removes projection barriers in the listed cases, but this does not establish faster execution. The plan observation above applies only to the tested factor<1.5 predicate and empty temporary fixture.

Correctness and tradeoffs

The direct path is limited to a native Float binding and statically known Float operands. In filter context, a nullable receiver is false; in() treats null needles as non-matches, and every non-null needle must be Float for the direct path. Mixed Int/Float operations, computed operands, guards, JSON, and structured fields retain the existing general lowering so it can preserve exact cross-type semantics and representation recovery. Parity coverage also includes the mixed Int/Float fallback near large integer boundaries.

Decision and limits

Use the direct path only for the eligible native Float shapes above. The SQL length and barrier counts do not predict PostgreSQL plan choice or execution time. The single fixture plan does not establish index use on other schemas or data, and controlled measurements would be needed before making speed claims.