[SpecialCaseList] Filtering Globs with matching prefix and suffix

[vitalybuka]

This commit enhances the SpecialCaseList::GlobMatcher to filter globs
more efficiently by considering both prefixes and suffixes.

Previously, the GlobMatcher used a RadixTree to store globs based
on their prefixes. This allowed for quick lookup of potential matches
by matching the query string's prefix against the stored prefixes.
However, for globs with common prefixes but different suffixes,
unnecessary glob matching attempts could still occur.

This change introduces a nested RadixTree structure:
PrefixSuffixToGlob: RadixTree<Prefix, RadixTree<Suffix, Globs>>.
Now, when a query string is matched, it first finds matching prefixes,
and then within those prefix matches, it further filters by matching
the reversed suffix of the query string against the reversed suffixes
of the globs. This significantly reduces the number of Glob::match
calls, especially for large special case lists with many globs sharing
common prefixes but differing in their suffixes.

According to SpecialCaseListBM:

Lookup benchmarks (significant improvements):

OVERALL_GEOMEAN                       -0.5815

Lookup *suffix and prefix*suffix like benchmarks (huge improvements):

OVERALL_GEOMEAN                       -0.9316

https://gist.github.com/vitalybuka/e586751902760ced6beefcdf0d7b26fd

[llvmbot]

@llvm/pr-subscribers-llvm-support

Author: Vitaly Buka (vitalybuka)

Changes

This commit enhances the SpecialCaseList::GlobMatcher to filter globs
more efficiently by considering both prefixes and suffixes.

Previously, the GlobMatcher used a RadixTree to store globs based
on their prefixes. This allowed for quick lookup of potential matches
by matching the query string's prefix against the stored prefixes.
However, for globs with common prefixes but different suffixes,
unnecessary glob matching attempts could still occur.

This change introduces a nested RadixTree structure:
PrefixSuffixToGlob: RadixTree<Prefix, RadixTree<Suffix, Globs>>.
Now, when a query string is matched, it first finds matching prefixes,
and then within those prefix matches, it further filters by matching
the reversed suffix of the query string against the reversed suffixes
of the globs. This significantly reduces the number of Glob::match
calls, especially for large special case lists with many globs sharing
common prefixes but differing in their suffixes.

---

Full diff: https://github.com/llvm/llvm-project/pull/164543.diff

2 Files Affected:

• (modified) llvm/include/llvm/Support/SpecialCaseList.h (+3-2)
• (modified) llvm/lib/Support/SpecialCaseList.cpp (+11-7)

diff --git a/llvm/include/llvm/Support/SpecialCaseList.h b/llvm/include/llvm/Support/SpecialCaseList.h
index 16f309329a0b5..471f8e779fa24 100644
--- a/llvm/include/llvm/Support/SpecialCaseList.h
+++ b/llvm/include/llvm/Support/SpecialCaseList.h
@@ -167,8 +167,9 @@ class SpecialCaseList {
     std::vector<GlobMatcher::Glob> Globs;
 
     RadixTree<iterator_range<StringRef::const_iterator>,
-              SmallVector<const GlobMatcher::Glob *, 1>>
-        PrefixToGlob;
+              RadixTree<iterator_range<StringRef::const_reverse_iterator>,
+                        SmallVector<const GlobMatcher::Glob *, 1>>>
+        PrefixSuffixToGlob;
   };
 
   /// Represents a set of patterns and their line numbers
diff --git a/llvm/lib/Support/SpecialCaseList.cpp b/llvm/lib/Support/SpecialCaseList.cpp
index 2a86cc37b6000..9bd1c199695d1 100644
--- a/llvm/lib/Support/SpecialCaseList.cpp
+++ b/llvm/lib/Support/SpecialCaseList.cpp
@@ -92,8 +92,10 @@ void SpecialCaseList::GlobMatcher::preprocess(bool BySize) {
 
   for (auto &G : Globs) {
     StringRef Prefix = G.Pattern.prefix();
+    StringRef Suffix = G.Pattern.suffix();
 
-    auto &V = PrefixToGlob.emplace(Prefix).first->second;
+    auto &SToGlob = PrefixSuffixToGlob.emplace(Prefix).first->second;
+    auto &V = SToGlob.emplace(reverse(Suffix)).first->second;
     V.emplace_back(&G);
   }
 }
@@ -101,12 +103,14 @@ void SpecialCaseList::GlobMatcher::preprocess(bool BySize) {
 void SpecialCaseList::GlobMatcher::match(
     StringRef Query,
     llvm::function_ref<void(StringRef Rule, unsigned LineNo)> Cb) const {
-  if (!PrefixToGlob.empty()) {
-    for (const auto &[_, V] : PrefixToGlob.find_prefixes(Query)) {
-      for (const auto *G : reverse(V)) {
-        if (G->Pattern.match(Query)) {
-          Cb(G->Name, G->LineNo);
-          break;
+  if (!PrefixSuffixToGlob.empty()) {
+    for (const auto &[_, SToGlob] : PrefixSuffixToGlob.find_prefixes(Query)) {
+      for (const auto &[_, V] : SToGlob.find_prefixes(reverse(Query))) {
+        for (const auto *G : reverse(V)) {
+          if (G->Pattern.match(Query)) {
+            Cb(G->Name, G->LineNo);
+            break;
+          }
         }
       }
     }

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[Copilot]

Pull Request Overview

This PR optimizes the SpecialCaseList::GlobMatcher by introducing a nested RadixTree structure that filters globs using both prefix and suffix matching, reducing unnecessary glob match attempts and improving lookup performance by ~58% overall and ~93% for patterns with suffixes.

Key Changes:

• Introduced nested RadixTree structure mapping prefixes to suffix-based RadixTrees
• Modified preprocess() to build the nested structure using both prefix and reversed suffix
• Updated match() to perform two-level filtering: prefix matching followed by reversed suffix matching

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
llvm/lib/Support/SpecialCaseList.cpp	Implements nested RadixTree lookup logic with prefix and suffix matching
llvm/include/llvm/Support/SpecialCaseList.h	Updates data structure from single RadixTree to nested RadixTree for prefix-suffix indexing