[NFC][clang-tidy]improve performance for misc-unused-using-decls check

[HerrCai0907]

UnusedUsingDeclsCheck::removeFromFoundDecls will be called with high frequency. At current time it will check every Context.
This patch adds a cache to reduce algorithm complexity.

[llvmbot]

@llvm/pr-subscribers-clang-tidy

Author: Congcong Cai (HerrCai0907)

Changes

UnusedUsingDeclsCheck::removeFromFoundDecls will be called with high frequency. At current time it will check every Context.
This patch adds a cache to reduce algorithm complexity.

---

Full diff: https://github.com/llvm/llvm-project/pull/78231.diff

2 Files Affected:

• (modified) clang-tools-extra/clang-tidy/misc/UnusedUsingDeclsCheck.cpp (+10-4)
• (modified) clang-tools-extra/clang-tidy/misc/UnusedUsingDeclsCheck.h (+1)

diff --git a/clang-tools-extra/clang-tidy/misc/UnusedUsingDeclsCheck.cpp b/clang-tools-extra/clang-tidy/misc/UnusedUsingDeclsCheck.cpp
index 051375263e53c3..59e487bab31195 100644
--- a/clang-tools-extra/clang-tidy/misc/UnusedUsingDeclsCheck.cpp
+++ b/clang-tools-extra/clang-tidy/misc/UnusedUsingDeclsCheck.cpp
@@ -117,8 +117,10 @@ void UnusedUsingDeclsCheck::check(const MatchFinder::MatchResult &Result) {
             /*SkipTrailingWhitespaceAndNewLine=*/true));
     for (const auto *UsingShadow : Using->shadows()) {
       const auto *TargetDecl = UsingShadow->getTargetDecl()->getCanonicalDecl();
-      if (shouldCheckDecl(TargetDecl))
+      if (shouldCheckDecl(TargetDecl)) {
         Context.UsingTargetDecls.insert(TargetDecl);
+        UsingTargetDeclsCache.insert(TargetDecl);
+      }
     }
     if (!Context.UsingTargetDecls.empty())
       Contexts.push_back(Context);
@@ -201,13 +203,16 @@ void UnusedUsingDeclsCheck::check(const MatchFinder::MatchResult &Result) {
 void UnusedUsingDeclsCheck::removeFromFoundDecls(const Decl *D) {
   if (!D)
     return;
+  const Decl *CanonicalDecl = D->getCanonicalDecl();
+  if (!UsingTargetDeclsCache.contains(CanonicalDecl))
+    return;
   // FIXME: Currently, we don't handle the using-decls being used in different
   // scopes (such as different namespaces, different functions). Instead of
   // giving an incorrect message, we mark all of them as used.
-  //
-  // FIXME: Use a more efficient way to find a matching context.
   for (auto &Context : Contexts) {
-    if (Context.UsingTargetDecls.contains(D->getCanonicalDecl()))
+    if (Context.IsUsed)
+      continue;
+    if (Context.UsingTargetDecls.contains(CanonicalDecl))
       Context.IsUsed = true;
   }
 }
@@ -224,6 +229,7 @@ void UnusedUsingDeclsCheck::onEndOfTranslationUnit() {
     }
   }
   Contexts.clear();
+  UsingTargetDeclsCache.clear();
 }
 
 } // namespace clang::tidy::misc
diff --git a/clang-tools-extra/clang-tidy/misc/UnusedUsingDeclsCheck.h b/clang-tools-extra/clang-tidy/misc/UnusedUsingDeclsCheck.h
index 498b3ffd2678c3..7bdaf12e8aecee 100644
--- a/clang-tools-extra/clang-tidy/misc/UnusedUsingDeclsCheck.h
+++ b/clang-tools-extra/clang-tidy/misc/UnusedUsingDeclsCheck.h
@@ -49,6 +49,7 @@ class UnusedUsingDeclsCheck : public ClangTidyCheck {
   };
 
   std::vector<UsingDeclContext> Contexts;
+  llvm::SmallPtrSet<const Decl *, 32> UsingTargetDeclsCache;
 
   StringRef RawStringHeaderFileExtensions;
   FileExtensionsSet HeaderFileExtensions;

[PiotrZSL]

LGTM.
There is open issue for performance of this check.
Consider mentioning this in release notes as some generic performance improvment.
Do you have any benchmarks ?

[hokein]

Thanks, the code looks good. Do you have any performance data after this change?

[HerrCai0907]

Thanks, the code looks good. Do you have any performance data after this change?

I don't run clang-tidy bench for some large project. But I do some test for some auto-generated cpp files which has lots of using, and the execution time reduce from 0.2s to 0.18s in my MAC M1 Pro.

[HerrCai0907]

LGTM. There is open issue for performance of this check. Consider mentioning this in release notes as some generic performance improvement.

I have also seen this issue. But after optimization, this check is still slow. I guess registering too many pattern causes performance issue but I cannot find a good solution to optimize it.