tr: add ASCII range translation fast path

[parasol-aser]

What

Adds a narrow fast path for bytewise ASCII range translations such as:

tr 'a-z' 'A-Z'

The change detects translation tables that modify one contiguous ASCII range by a constant wrapping delta, then processes that range with an AVX2 range compare plus masked add on x86/x86_64 hosts that support AVX2. Other translations continue to use the existing single-byte or table-lookup paths, and non-AVX2 hosts use the scalar fallback.

Why

Before this change, tr 'a-z' 'A-Z' mapped every byte through a scalar 256-byte translation table. This is a common case, and it can be handled more directly by checking whether each byte falls within the translated ASCII range and adding the fixed delta.

Measurements

Environment:

• CPU: Intel(R) Xeon(R) Platinum 8272CL CPU @ 2.60GHz
• OS: Linux x86_64
• Rust: rustc 1.92.0
• Candidate branch: perf/P002
• Baseline commit: 4b5a2af7a916910bfeaf46b298a963d8a038565a
• hyperfine was not installed, so this used /usr/bin/time, 2 warmups, and 12 measured runs.

Input was corpus/large_text.txt repeated 16 times, 1,342,178,256 bytes total.

/usr/bin/time -f '%e %M' ./runs/P002-rerun/bin/tr-baseline 'a-z' 'A-Z' < runs/P002-rerun/input/large_text_x16.txt > /dev/null
/usr/bin/time -f '%e %M' ./runs/P002-rerun/bin/tr-p002 'a-z' 'A-Z' < runs/P002-rerun/input/large_text_x16.txt > /dev/null
/usr/bin/time -f '%e %M' /usr/bin/tr 'a-z' 'A-Z' < runs/P002-rerun/input/large_text_x16.txt > /dev/null

implementation	mean wall time	stddev	throughput
uutils baseline	1.068 s	0.129 s	1198.1 MiB/s
uutils candidate	0.421 s	0.029 s	3041.6 MiB/s
GNU tr	1.251 s	0.209 s	1023.3 MiB/s

The candidate is 2.54x faster than the uutils baseline on this workload, a 60.6% wall-time reduction. The earlier 80 MiB pipeline benchmark also showed a 53.8% reduction, from 0.065 s to 0.030 s.

Correctness

For the 1.3 GB input, baseline uutils, candidate uutils, and GNU tr produced the same transformed output SHA256:

6f2d6cb371ca0b423a90a5690ee7f6dac0be6a7d889f308ff5b15f2957e853db

Tests

cargo test --release --test tests -- --nocapture --test-threads=1 test_tr::test_ascii_range_translate_alignment_boundaries
cargo clippy --release -p uu_tr -- -D warnings
cargo fmt --check --package uu_tr

The regression test covers 0, 1, 31, 32, and 33 byte inputs around the AVX2 lane width, all byte values, and a UTF-8/non-ASCII boundary case, with GNU parity when GNU tr is available.

Caveats

The speedup is from the AVX2 path on this x86_64 host. Non-AVX2 targets use the scalar fallback and should be behavior-preserving, but I did not benchmark those targets here.

[github-actions]

GNU testsuite comparison:

Skip an intermittent issue tests/date/date-locale-hour (fails in this run but passes in the 'main' branch)
Note: The gnu test tests/cut/bounded-memory is now being skipped but was previously passing.
Note: The gnu test tests/printf/printf-surprise is now being skipped but was previously passing.

[oech3]

You need to add spell-checker: ignore