[Clang] Allow devirtualization involving array subscripts with constant indices when the pointee type is known

[AZero13]

By C++1z [expr.add]/6, if the array element type and the pointee type are not similar, behavior is undefined.

[llvmbot]

@llvm/pr-subscribers-clang

Author: AZero13 (AZero13)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/130528.diff

1 Files Affected:

• (modified) clang/lib/AST/DeclCXX.cpp (+22-1)

diff --git a/clang/lib/AST/DeclCXX.cpp b/clang/lib/AST/DeclCXX.cpp
index 7eff776882629..ecba340f0425a 100644
--- a/clang/lib/AST/DeclCXX.cpp
+++ b/clang/lib/AST/DeclCXX.cpp
@@ -1689,7 +1689,7 @@ static NamedDecl* getLambdaCallOperatorHelper(const CXXRecordDecl &RD) {
   //     auto L = [](T) { }; // But A's call operator would want A's here.
   //   }
   //
-  // Walk the call operator’s redecl chain to find the one that belongs
+  // Walk the call operator's redecl chain to find the one that belongs
   // to this module.
   //
   // TODO: We need to fix this properly (see
@@ -2465,6 +2465,27 @@ CXXMethodDecl *CXXMethodDecl::getDevirtualizedMethod(const Expr *Base,
   if (Base->isPRValue() && Base->getType()->isRecordType())
     return this;
 
+  // Handle array subscripts with constant indices when the pointee type is known
+  if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(Base)) {
+    QualType BaseTy = ASE->getBase()->getType();
+    
+    // Check if it's a pointer to a record type
+    if (BaseTy->isPointerType() && 
+        BaseTy->getPointeeType()->isRecordType()) {
+      // For C++17 and later, we can devirtualize array access beyond p[0]
+      // According to [expr.add]/6, if the array element type and the pointee
+      // type are not similar, behavior is undefined, so we can assume they are
+      // the same type
+      const ASTContext &Context = getParent()->getASTContext();
+      const LangOptions &LangOpts = Context.getLangOpts();
+      if (LangOpts.CPlusPlus17 && 
+          ASE->getIdx()->isIntegerConstantExpr(Context)) {
+        // It's a constant index, so it's safe to devirtualize
+        return this;
+      }
+    }
+  }
+
   // If we don't even know what we would call, we can't devirtualize.
   const CXXRecordDecl *BestDynamicDecl = Base->getBestDynamicClassType();
   if (!BestDynamicDecl)

[github-actions]

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

[erichkeane]

This patch is very lacking on details, please elaborate in the commit message (that is, the github summary). Also, this needs a release note.

[erichkeane]

I don't get the comment on 89: Why can we virtualize this but not that one?

[erichkeane]

Unrelated change.

[erichkeane]

I don't see the 'beyond [0]' in that standardeeze, can you better clarify it?

Also, where are you checking for >0 here? It seems like you'd do this in every case? What is making this skip the 0 case?