GH-48728: [C++] Cache compiled regex matchers in string kernels

[HyukjinKwon]

Rationale for this change

String operations with regex patterns (match, replace, extract) were recompiling regex patterns on every invocation. This PR implements caching to compile once and reuse.

Benchmark shows roughly 36% performance improvement (2.52s -> 1.61s for 200 operations).

arrow/cpp/src/arrow/compute/kernels/scalar_string_ascii.cc

Line 1371 in 727106f

// TODO Cache matcher across invocations (for regex compilation)

arrow/cpp/src/arrow/compute/kernels/scalar_string_ascii.cc

Line 1381 in 727106f

// TODO Cache matcher across invocations (for regex compilation)

arrow/cpp/src/arrow/compute/kernels/scalar_string_ascii.cc

Line 1965 in 727106f

// TODO Cache replacer across invocations (for regex compilation)

arrow/cpp/src/arrow/compute/kernels/scalar_string_ascii.cc

Line 2218 in 727106f

// TODO cache this once per ExtractRegexOptions

What changes are included in this PR?

• Added CachedOptionsWrapper<T> template for kernel state with caching support
• Updated MatchSubstringState, ReplaceState, and ExtractRegexState to use caching
• Modified Exec() methods to call GetOrCreate<Matcher>() instead of direct Matcher::Make()

Are these changes tested?

Yes. All existing tests pass. Benchmark demonstrates measurable performance improvement when same pattern is used across multiple operations.

(Generated by ChatGPT)

Benchmark:

# Benchmark script: Compare WITH vs WITHOUT caching

# Step 1: Measure WITH caching (current implementation)
cd /.../arrow/cpp/build
/usr/bin/time -p ./debug/arrow-compute-scalar-type-test \
  --gtest_filter="TestStringKernels/0.MatchSubstringRegex" \
  --gtest_repeat=200 \
  --gtest_brief=1

# Step 2: Temporarily remove caching
cd /.../arrow/cpp
git stash push -m "Temp for benchmark" src/arrow/compute/kernels/scalar_string_ascii.cc

# Step 3: Rebuild WITHOUT caching
cd build
touch ../src/arrow/compute/kernels/scalar_string_ascii.cc
cmake --build .

# Step 4: Measure WITHOUT caching (reverted to old TODO code)
/usr/bin/time -p ./debug/arrow-compute-scalar-type-test \
  --gtest_filter="TestStringKernels/0.MatchSubstringRegex" \
  --gtest_repeat=200 \
  --gtest_brief=1

# Step 5: Restore caching
cd ..
git stash pop
cd build && touch ../src/arrow/compute/kernels/scalar_string_ascii.cc
cmake --build .

Results:

╔════════════════════════════════════════════════════════╗
║              BENCHMARK RESULTS                         ║
╠════════════════════════════════════════════════════════╣
║  WITHOUT Caching:      2.52 seconds                    ║
║  WITH Caching:         1.61 seconds                    ║
║  ─────────────────────────────────────                 ║
║  Time Saved:           0.91 seconds                    ║
║  Improvement:          36.1% FASTER                    ║
╚════════════════════════════════════════════════════════╝

Test Configuration:
  • Test: TestStringKernels/0.MatchSubstringRegex
  • Iterations: 200 repetitions
  • Pattern: Complex regex with groups/alternation
  • Per-operation: 12.6ms → 8.05ms (4.5ms saved)

Are there any user-facing changes?

No, this is an optiomization.

• GitHub Issue: [C++] Cache compiled regex matchers in string kernels #48728

[github-actions]

⚠️ GitHub issue #48728 has been automatically assigned in GitHub to PR creator.