linq_fold: peel key blocks (group_by + distinct_by) — 16-28 ns/op wins

[borisbat]

Summary

Bench-driven follow-up after PR #2822 closed the Wave 5 terminator splice ladder. A sweep over benchmarks/sql flagged group_by lanes as consistently 1.2–1.6× slower on the decs splice (m4) than the array splice (m3f). Root cause: plan_group_by_core and plan_distinct's distinct_by arm emitted invoke($keyBlock, sourceVar) per element. Switching to peel_lambda_rename_var / peel_lambda_replace_var inlines _.field as a direct field read, with the existing fallback to invoke(...) for non-peelable blocks.

The win is twofold:

• Per-element invoke dispatch eliminated (~3-5 ns/elem in INTERP on both lanes).
• On the decs side, the whole-decs_tup lambda argument no longer pins the pruner — build_decs_inner_for_pruned now drops the components only the keyBlock referenced through the tuple shape (e.g. tuple<id:int;name:string;price:int;brand:int;year:int;dealer_id:int> shrinks to tuple<brand:int>), removing per-iter string copies and matching get_ro slots.

Bench (INTERP, 100K rows, ns/op, before → after)

bench	m3f before	m3f after	m4 before	m4 after
groupby_count	36.3	19.5 (-46%)	49.9	22.5 (-55%)
groupby_sum	36.0	18.6	50.1	24.4
groupby_min	42.9	25.5	57.1	29.3
groupby_max	42.8	24.9	56.5	29.6
groupby_average	52.1	30.4	62.7	34.7
groupby_having_count	36.4	19.3	50.0	22.2
groupby_having_hidden_sum	39.9	24.2	57.3	30.1
groupby_multi_reducer	53.3	32.3	63.6	36.6
groupby_select_sum	58.6	37.2	61.4	42.4
groupby_first	35.3	18.5	47.3	29.9
groupby_where_count	23.7	14.7	37.0	19.8
groupby_where_sum	23.3	14.1	36.4	19.7
distinct_by_count (NEW)	33.7	15.3 (-55%)	44.3	19.7 (-56%)

distinct_by_count is the new bench that covers the path the existing distinct_count doesn't touch (it preselects to a scalar first). All other non-groupby/distinct_by lanes unchanged within bench noise.

Bundled

The branch carries two adjacent improvements that surfaced while shipping this:

1. dastest bench format with 1-decimal ns/op (commit 1). The native column truncated to int64 ns/op; go_export and the JSON message line already used :.1f. Switched native to the same shape, keeping column alignment + zero-allocs green/yellow coloring.

2. dastest lint enablement (commit 3). dastest/dastest_clargs.das was missing options gen2, which made it parse fine via direct daslang invocation (defaults to v2) but fail under the lint pipeline's compile_file path (defaults to v1) — @clarg_doc = ... is v2-only syntax. That parse error rippled to every lint of dastest/suite.das and blocked pre-push for any dastest touch. Added options gen2, then addressed the ~30 pre-existing PERF/STYLE/LINT warnings it surfaced in dastest/suite.das + 2 STYLE005 in dastest/log.das (the latter via suite's require chain). No behavior change.

Splice-shape test update

test_wave4_pruned_groupby_count_splice_shape previously documented "pruning DISABLED — bucket needs full tuple" as a known limitation; that limitation is now fixed for the count-by-group shape, so the gate asserts 1 get_ro (brand) + 0 for price/year. The parity test (test_wave4_pruned_groupby_count_parity) is unchanged and still green.

Test plan

• mcp__daslang__lint clean on daslib/linq_fold.das, benchmarks/sql/distinct_by_count.das, tests/linq/test_linq_from_decs.das, and all touched dastest files
• tests/linq 1382/1382 INTERP + 1382/1382 AOT
• tests/decs 245/245 INTERP
• tests/dasSQLITE 782/782 INTERP
• test_aot full lane 2409/2409
• Bench sweep over benchmarks/sql (199 benchmarks) — diff shows gains on 12 groupby + new distinct_by; all others unchanged
• dastest --bench --bench_format native formatter visual check (3 sample benches)
• CI 9-lane matrix
• Copilot review

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Improves linq_fold’s group-by and distinct-by planning by peeling simple key lambdas so _.field becomes a direct field read, eliminating per-element invoke(...) overhead and enabling better decs tuple-slot pruning.

Changes:

• Update plan_group_by_core and the distinct_by arms of plan_distinct / plan_decs_distinct to use peel_lambda_rename_var / peel_lambda_replace_var instead of per-element invoke(...).
• Update the Wave4 pruning splice-shape test to assert the newly-enabled pruning behavior for count-by-group.
• Add a new SQL benchmark (distinct_by_count) and tweak dastest benchmark native formatting to print 1-decimal ns/op; enable gen2 parsing for dastest_clargs.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tests/linq/test_linq_from_decs.das	Updates splice-shape assertion to reflect newly-working pruning for group_by count-by-group.
dastest/suite.das	Refactors for lint/style, adds reserve for message cloning, and prints native ns/op with 1 decimal.
dastest/log.das	Minor early-return style cleanups.
dastest/dastest_clargs.das	Enables options gen2 so gen2-only clargs annotations parse under lint/compile_file.
daslib/linq_fold.das	Uses peeled key expressions for group_by and distinct_by to avoid per-element invoke(...).
benchmarks/sql/distinct_by_count.das	Adds a benchmark that exercises the `distinct_by

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.