[Weijun-H]

Which issue does this PR close?

Closes #27280.

Rationale for this change

CASE WHEN f(x) = ... THEN ... ELSE f(x) END repeats f(x) for ELSE rows. This opt-in API reuses its first result at the original evaluation position, avoiding extra work for expensive operands. Direct physical-plan builders can use it without logical CSE, while preserving evaluation position.

What changes are included in this PR?

• Add RequiredOperandReusePolicy and CaseExpr::with_required_operand_reuse. Callers identify an unconditional, exactly-once operand and guarantee that column replacement, projection and a private column preserve expression semantics; unknown leaves decline reuse.
• Evaluate the actual first-WHEN instance once. Expand scalar results after argument rewrites without reevaluation. Rebuilds retain permission only if the original call is preserved and WHEN/ELSE arguments still match.
• Restrict reuse to CASE without a base expression. Permission participates in Eq/Hash; try_to_proto returns Ok(None) for a CASE with permission.
• Add a reproducible complete-CASE Criterion benchmark, including first-use preparation and repeated batches.

What is the testing strategy for this PR?

Validated tree 42c33d272af2: 77 CASE tests (10 reuse), whole-workspace clippy, formatting, and extended workspace tests pass (13,632 standard tests, 8 ignored; 547/547 SQLLogicTest files). Both new regressions fail on the prior implementation and pass with the instance-selection and scalar-rewrite fixes.

The table compares ON/OFF within the same corrected version, not this version against an unpatched baseline. Physical-expression benchmark: Rust 1.99.1, Apple arm64, release-nonlto; 8,192 rows. Each OFF/ON mode has three rounds of 30 samples. Times below are medians of three per-round arithmetic means of time/iterations; percentages use those OFF/ON values. First includes CASE construction, authorization and preparation, with inputs/operand nodes prebuilt. Repeated reuses the prepared plan. Output and evaluated-row checks run outside timing. Two input batches alternate; at ELSE 50% each sends a different contiguous half to ELSE. Each iteration evaluates one 8,192-row batch; times are per batch.

Operand / total input columns (time unit)	ELSE	First OFF → ON (time saved)	Repeated OFF → ON (time saved)
add / 1 (µs)	0%	2.312 → 3.735 (-61.5%)	1.615 → 1.851 (-14.6%)
add / 1 (µs)	50%	29.692 → 30.206 (-1.7%)	28.730 → 28.011 (+2.5%)
add / 1 (µs)	100%	2.887 → 3.705 (-28.3%)	2.164 → 1.844 (+14.8%)
add / 100 (µs)	0%	2.306 → 4.780 (-107.3%)	1.602 → 2.877 (-79.6%)
add / 100 (µs)	50%	29.560 → 31.546 (-6.7%)	28.603 → 29.316 (-2.5%†)
add / 100 (µs)	100%	2.916 → 4.772 (-63.6%)	2.149 → 2.835 (-31.9%)
MD5 of 4096 bytes / 1 (ms)	0%	39.576 → 39.506 (+0.2%†)	38.492 → 38.576 (-0.2%†)
MD5 of 4096 bytes / 1 (ms)	50%	58.864 → 38.571 (+34.5%)	58.516 → 38.777 (+33.7%)
MD5 of 4096 bytes / 1 (ms)	100%	77.472 → 38.666 (+50.1%)	76.809 → 38.586 (+49.8%)

Positive percentages mean less time. † Paired rounds disagree in direction; no directional conclusion. The 100-column input adds 99 unused Int32 columns sharing buffers; only widths 1 and 100 were tested. These single-machine microbenchmarks establish neither statistical significance nor end-to-end query speedups.

Are there any user-facing changes?

New opt-in physical-expression API. Conventional SQL planning does not enable it automatically. Callers must verify semantic safety and whether avoided computation outweighs added costs; cheap operands can become slower.

[Copilot]

🟡 Changes recommended

Active findings identify incorrect evaluation of structurally equal but distinct volatile instances and a rewrite path that can turn valid CASE evaluation into an array-precondition error.

2 open issues

• Traverse the first WHEN instance, not the equal ELSE instance · New
• Revalidate array-valued operands after argument rewrites · New

What changed in this PR

Adds an opt-in API for physical-plan producers to reuse a CASE’s first required WHEN operand as its ELSE value, avoiding duplicate evaluation without relying on logical optimization.

Changes:

• Introduces a caller-supplied reuse policy and prepared CASE evaluation.
• Tracks permission through expression rewrites, equality, hashing, and serialization eligibility.
• Adds tests for reuse, branch selection, projection, rewrites, and error ordering.

File	Description
datafusion/​physical-expr/​src/​expressions/​mod.rs	Re-exports RequiredOperandReusePolicy.
datafusion/​physical-expr/​src/​expressions/​case.rs	Implements operand reuse, permission handling, and regression tests.

_{Review effort: Balanced}