[SQL] Derive pushdown predicates from ANSI integral arithmetic - #58705

What changes were proposed in this pull request?

Add a Catalyst optimizer rule that derives a data-source-pushable predicate from comparisons over ANSI integral addition and subtraction by literals.

For example, i - 10 > 100 gains the conjunct i > 110 OR i < INT_MIN + 10. The second range retains inputs for which the original arithmetic must raise an overflow error. The original comparison remains as a residual predicate and continues to define exact evaluation semantics.

The rule supports byte, short, integer, and long arithmetic with literal constants and ordered comparisons or equality. It does not change legacy arithmetic, floating-point arithmetic, multiplication, or arithmetic with a nonliteral operand.

Why are the changes needed?

Arithmetic around a column currently prevents Parquet and other data sources from using their min/max statistics for predicate pruning. The derived predicate exposes safe ranges over the underlying column while preserving ANSI overflow behavior.

Does this PR introduce any user-facing change?

Yes. In ANSI mode, eligible arithmetic filters can now be pushed into data sources for pruning. Query results and overflow errors are unchanged because the original predicate is retained.

How was this patch tested?

Added Catalyst optimizer coverage for ordered comparisons, equality, reversed operands, unsupported expressions, integral range boundaries, and both overflow directions. Added Parquet V2 tests that verify the derived predicate is pushed into the scan and that an overflow input still raises ARITHMETIC_OVERFLOW.

Ran:

./build/sbt "catalyst/testOnly org.apache.spark.sql.catalyst.optimizer.BinaryComparisonSimplificationSuite"
./build/sbt 'sql/testOnly org.apache.spark.sql.execution.datasources.parquet.ParquetV2FilterSuite -- -z "push down derived predicate for ANSI integral arithmetic"'
./build/sbt 'sql/testOnly org.apache.spark.sql.execution.datasources.parquet.ParquetV2FilterSuite -- -z "derived predicate preserves ANSI integral overflow"'
./dev/lint-scala

Was this patch authored or co-authored using generative AI tooling?

Generated-by: OpenAI Codex 0.147.0