PERF: vectorize is_range_indexer and is_sequence_range scans

[jbrockmendel]

Split out from GH-65298 (the has_sentinel half stays there).

4x-unrolls the inner scan loops of is_range_indexer and is_sequence_range in lib.pyx, combining lanes with bitwise | (not or) so the compiler can emit vectorized comparisons, with a scalar tail loop for the remainder. This mirrors the existing has_nans/all_nans unrolling (GH-65192).

has_infs was also unrolled in GH-65298 but is dropped here — benchmarking showed it flat (clang already vectorizes the float-compare loop, and it's memory-bound at scale), so the extra code wasn't earning its keep.

Benchmarks

Both the old (scalar) and new (unrolled) variants compiled into one extension with pandas' build flags (-O3 -std=c17 -DNDEBUG, baseline NEON on Apple Silicon), timed in-process, best-of-9. Common full-scan path (array is a range / sequence):

is_range_indexer (int64 — the no-op-indexer fast path in take/reset_index/merge/reindex):

n	old	new	speedup
1,000	642 ns	503 ns	1.28x
10,000	2.96 µs	1.75 µs	1.69x
100,000	26.3 µs	14.0 µs	1.87x
1,000,000	260 µs	139 µs	1.87x
10,000,000	2.63 ms	1.40 ms	1.87x

is_sequence_range (int64, step=3 — Index construction):

n	old	new	speedup
1,000	505 ns	379 ns	1.33x
10,000	2.94 µs	1.78 µs	1.65x
100,000	27.2 µs	15.7 µs	1.74x
1,000,000	272 µs	157 µs	1.74x
10,000,000	2.71 ms	1.56 ms	1.74x

Early-exit (short-circuit) path: 0.99–1.02x across all sizes for both — the unroll doesn't cost anything when the scan breaks early.

Test plan

• Exhaustive correctness across int8/16/32/64, every break position, steps {1, 2, -1, 3}, plus length-mismatch / empty edge cases
• pandas/tests/libs/test_lib.py + downstream suites (ranges, sort_values, reset_index, merge, strings)

No whatsnew entry, following the has_nans/all_nans precedent (GH-65192) — these are internal scan helpers and the per-op effect on end-user timings is below the noise floor.

🤖 Generated with Claude Code