Improve performance of strings count_matches utility with specific expression patterns

[davidwendt]

Description

Updates the logic in the cudf::strings::detail::count_matches() utility to use the faster positional::END_ONLY regex state engine path when input pattern does not include empty-matching instruction paths.
The empty-matching-only instructions require tracking the start/end position which the END_ONLY path does not record. These patterns are not common like where the entire pattern is just an anchor like "\b" and "$" or the pattern can match virtually nothing (i.e. return an empty string) like with "(?:)" or "a*" or "a{0,}".
The instruction set is analyzed for such a path and if it does not exist, the faster END_ONLY code logic is used.

The count_matches utility is used in count_re as well as split_re, findall and extract_all.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[davidwendt]

Benchmark results show up to 25% improvement for count_re

# count

## [0] NVIDIA RTX A6000

| max_width | num_rows | pattern |   Ref Time |   Cmp Time |          Diff |   %Diff |
|-----------|----------|---------|------------|------------|---------------|---------|
|    32     |  32768   |    0    | 169.300 us | 157.907 us |    -11.393 us |  -6.73% |
|    64     |  32768   |    0    | 326.756 us | 305.313 us |    -21.442 us |  -6.56% |
|    128    |  32768   |    0    | 648.013 us | 590.018 us |    -57.995 us |  -8.95% |
|    256    |  32768   |    0    |   1.250 ms |   1.141 ms |   -109.176 us |  -8.74% |
|    32     |  262144  |    0    | 648.595 us | 548.233 us |   -100.362 us | -15.47% |
|    64     |  262144  |    0    |   1.357 ms |   1.137 ms |   -220.500 us | -16.25% |
|    128    |  262144  |    0    |   2.953 ms |   2.344 ms |   -609.145 us | -20.63% |
|    256    |  262144  |    0    |   6.731 ms |   5.042 ms |  -1689.250 us | -25.10% |
|    32     | 2097152  |    0    |   4.519 ms |   3.846 ms |   -673.378 us | -14.90% |
|    64     | 2097152  |    0    |   9.556 ms |   7.984 ms |  -1571.139 us | -16.44% |
|    128    | 2097152  |    0    |  20.753 ms |  16.747 ms |  -4006.077 us | -19.30% |
|    256    | 2097152  |    0    |  46.701 ms |  35.662 ms | -11039.035 us | -23.64% |
|    32     |  32768   |    1    |  40.960 us |  40.610 us |     -0.350 us |  -0.85% |
|    64     |  32768   |    1    |  52.573 us |  52.247 us |     -0.326 us |  -0.62% |
|    128    |  32768   |    1    |  82.011 us |  81.694 us |     -0.317 us |  -0.39% |
|    256    |  32768   |    1    | 149.608 us | 148.300 us |     -1.307 us |  -0.87% |
|    32     |  262144  |    1    |  97.580 us |  92.489 us |     -5.091 us |  -5.22% |
|    64     |  262144  |    1    | 140.513 us | 132.223 us |     -8.290 us |  -5.90% |
|    128    |  262144  |    1    | 253.803 us | 242.245 us |    -11.558 us |  -4.55% |
|    256    |  262144  |    1    | 609.375 us | 577.544 us |    -31.831 us |  -5.22% |
|    32     | 2097152  |    1    | 548.561 us | 507.003 us |    -41.558 us |  -7.58% |
|    64     | 2097152  |    1    | 836.430 us | 802.318 us |    -34.112 us |  -4.08% |
|    128    | 2097152  |    1    |   1.665 ms |   1.602 ms |    -63.098 us |  -3.79% |
|    256    | 2097152  |    1    |   4.002 ms |   3.831 ms |   -170.964 us |  -4.27% |

and up to 21% for split_re

# split_re

## [0] NVIDIA RTX A6000

| max_width | num_rows |   Ref Time |   Cmp Time |          Diff |   %Diff |
|-----------|----------|------------|------------| --------------|---------|
|    32     |  32768   | 408.231 us | 404.176 us |     -4.055 us |  -0.99% |
|    64     |  32768   | 699.318 us | 720.646 us |     21.327 us |   3.05% |
|    128    |  32768   |   1.437 ms |   1.397 ms |    -40.356 us |  -2.81% |
|    256    |  32768   |   2.627 ms |   2.628 ms |      1.324 us |   0.05% |
|    32     |  262144  |   1.567 ms |   1.407 ms |   -159.413 us | -10.18% |
|    64     |  262144  |   3.141 ms |   2.785 ms |   -356.621 us | -11.35% |
|    128    |  262144  |   6.778 ms |   5.750 ms |  -1027.694 us | -15.16% |
|    256    |  262144  |  15.556 ms |  12.274 ms |  -3282.096 us | -21.10% |
|    32     | 2097152  |  10.744 ms |   9.664 ms |  -1079.461 us | -10.05% |
|    64     | 2097152  |  22.108 ms |  19.810 ms |  -2297.471 us | -10.39% |
|    128    | 2097152  |  48.239 ms |  41.677 ms |  -6562.017 us | -13.60% |
|    256    | 2097152  | 109.929 ms |  89.671 ms | -20258.460 us | -18.43% |

[davidwendt]

/ok to test

[davidwendt]

/ok to test

[davidwendt]

/ok to test

[davidwendt]

/ok to test

[davidwendt]

/ok to test

[gforsyth]

/ok to test

[coderabbitai]

📝 Walkthrough

Walkthrough

Detects when a regex can match an empty string, exposes that flag on device, adapts the counting kernel to select a positional policy accordingly, and adds a test validating empty-match counting for several patterns.

Changes

Empty Match Detection and Handling

Layer / File(s)	Summary
Match flags contract definition cpp/src/strings/regex/regcomp.h	Defines match_flags enum with NONE and EMPTY_MATCH values, and declares reprog::compute_match_flags() const method.
Host-side empty-match detection algorithm cpp/src/strings/regex/regcomp.cpp	Implements compute_match_flags() using depth-first graph traversal with visited-set tracking to identify non-consuming instruction paths and determine if empty matches are possible.
Device-side empty-match state and accessor cpp/src/strings/regex/regex.cuh	Adds _empty_match_possible member and is_empty_match_possible() accessor to reprog_device.
Device initialization of empty-match flag cpp/src/strings/regex/regexec.cpp	Initializes reprog_device::_empty_match_possible from host compute_match_flags() during device program creation.
Kernel template instantiation based on empty-match flag cpp/src/strings/count_matches.cu	Templates count_fn on positional policy, uses prog.find<P>(), and conditionally instantiates BEGIN_END or END_ONLY kernel variants based on the empty-match predicate.
Empty-match counting validation test cpp/tests/strings/contains_tests.cpp	Adds CountEmptyMatching test validating count_re for patterns that can and cannot match empty strings, including anchors and boundary assertions.

🎯 3 (Moderate) | ⏱️ ~20 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 42.86% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately summarizes the main change: improving count_matches performance by using a faster regex execution path for patterns without empty-matching instruction paths.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description check	✅ Passed	The pull request description clearly explains the purpose of the changes: optimizing count_matches() to use the faster positional::END_ONLY path when patterns don't contain empty-matching instruction paths, with specific examples and context about affected utilities.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 3

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@cpp/src/strings/regex/regcomp.cpp`:
- Around line 1217-1243: The OR branch logic in reprog::compute_match_flags
(inside the check_paths lambda) incorrectly uses logical OR causing a
consuming-path result if either branch consumes; change the OR handling to
require both branches to be consuming by replacing the current `||` behavior
with logical AND so that the OR case returns self(self, inst.u2.left_id,
visited) && self(self, inst.u1.right_id, visited); this ensures patterns where
one branch reaches END without consuming (e.g., a*, a?, etc.) are detected as
empty-match paths.

In `@cpp/tests/strings/contains_tests.cpp`:
- Around line 411-415: The test loop is using the loop variable pattern but is
hardcoded to regex_program::create("a*"); change it so regex_program::create
uses the loop variable pattern (the variable named pattern) so each entry from
patterns is compiled and passed to cudf::strings::count_re(sv, *prog); ensure
you still compare *results to expected and keep the existing variables sv,
expected and function calls unchanged.
- Around line 408-410: The test currently hardcodes
cudf::strings::regex_program::create("a*") instead of using the loop variable
pattern, so change the call to use pattern (e.g.,
cudf::strings::regex_program::create(pattern)) in the contains test (where the
regex is constructed) and update expectations: for patterns that match empty
everywhere ("a*", "X?", "b{0,}", "()", "(?:)", "[a-z]*") keep expected
{6,6,1,4}; for "^" and "$" use {1,1,1,1}; for "^$" use {0,0,1,0}; and set
appropriate counts for "\\b" and "\\B" (or split the loop into separate
TEST_CASEs per pattern) so the assertion compares the result against the correct
per-pattern expected vector instead of a single hardcoded expected.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: b6057e77-3a65-4626-876d-d6cc754e656f

📥 Commits

Reviewing files that changed from the base of the PR and between 57e27e7 and 4c791a3.

📒 Files selected for processing (6)

• cpp/src/strings/count_matches.cu
• cpp/src/strings/regex/regcomp.cpp
• cpp/src/strings/regex/regcomp.h
• cpp/src/strings/regex/regex.cuh
• cpp/src/strings/regex/regexec.cpp
• cpp/tests/strings/contains_tests.cpp

[davidwendt]

/merge