[clang] [clang][bytecode] Fix CXXConstructExpr for multidim arrays (PR #164760)

[Timm Baeder via cfe-commits]

================
@@ -3274,34 +3274,43 @@ bool Compiler<Emitter>::VisitCXXConstructExpr(const CXXConstructExpr *E) {
   }
 
   if (T->isArrayType()) {
-    const ConstantArrayType *CAT =
-        Ctx.getASTContext().getAsConstantArrayType(E->getType());
-    if (!CAT)
-      return false;
-
-    size_t NumElems = CAT->getZExtSize();
     const Function *Func = getFunction(E->getConstructor());
     if (!Func)
       return false;
 
-    // FIXME(perf): We're calling the constructor once per array element here,
-    //   in the old intepreter we had a special-case for trivial constructors.
-    for (size_t I = 0; I != NumElems; ++I) {
-      if (!this->emitConstUint64(I, E))
-        return false;
-      if (!this->emitArrayElemPtrUint64(E))
-        return false;
+    if (!this->emitDupPtr(E))
+      return false;
 
-      // Constructor arguments.
-      for (const auto *Arg : E->arguments()) {
-        if (!this->visit(Arg))
-          return false;
+    std::function<bool(QualType)> initArrayDimension;
+    initArrayDimension = [&](QualType T) -> bool {
+      if (!T->isArrayType()) {
+        // Constructor arguments.
+        for (const auto *Arg : E->arguments()) {
+          if (!this->visit(Arg))
----------------
tbaederr wrote:

The `return false` case is hit when direct expression evaluation fails or when we hit the size limit of the bytecode emitter. It's not tested in the here attached test case I think, but at least the expression case is tested in an existing test.

https://github.com/llvm/llvm-project/pull/164760