[AST] Produce ReturnStmt containing RecoveryExpr when type is wrong

[sammccall]

Summary

Previously we just drop the ReturnStmt and its argument from the AST,
which blocks analysis of broken code.

Fixes https://github.com/llvm/llvm-project/issues/39944

[sammccall]

Add evaluation test

[sammccall]

No changes to constant evaluation needed. This is expected as these patterns already exist today: return undef(); produces a RecoveryExpr and return; is preserved in the AST with no value even for non-void functions.

[hokein]

Thanks, this looks great.

clang/lib/Sema/SemaStmt.cpp

4037–4042

iniializers => inializers

clang/test/AST/ast-dump-recovery.cpp

365

nit: I would check the function-decl is invalid as well (this is important).

clang/include/clang/Sema/Sema.h

public:
 
   ExprResult
   PerformMoveOrCopyInitialization(const InitializedEntity &Entity,
                                   const NamedReturnInfo &NRInfo, Expr *Value,
                                   bool SupressSimplerImplicitMoves = false);
 
   StmtResult ActOnReturnStmt(SourceLocation ReturnLoc, Expr *RetValExp,
                              Scope *CurScope);
-  StmtResult BuildReturnStmt(SourceLocation ReturnLoc, Expr *RetValExp);
+  StmtResult BuildReturnStmt(SourceLocation ReturnLoc, Expr *RetValExp,
+                             bool AllowRecovery = false);
   StmtResult ActOnCapScopeReturnStmt(SourceLocation ReturnLoc, Expr *RetValExp,
                                      NamedReturnInfo &NRInfo,
                                      bool SupressSimplerImplicitMoves);
 
   StmtResult ActOnGCCAsmStmt(SourceLocation AsmLoc, bool IsSimple,
                              bool IsVolatile, unsigned NumOutputs,
                              unsigned NumInputs, IdentifierInfo **Names,
                              MultiExprArg Constraints, MultiExprArg Exprs,

clang/lib/Sema/SemaStmt.cpp

 Sema::ActOnReturnStmt(SourceLocation ReturnLoc, Expr *RetValExp,
                       Scope *CurScope) {
   // Correct typos, in case the containing function returns 'auto' and
   // RetValExp should determine the deduced type.
   ExprResult RetVal = CorrectDelayedTyposInExpr(
       RetValExp, nullptr, /*RecoverUncorrectedTypos=*/true);
   if (RetVal.isInvalid())
     return StmtError();
-  StmtResult R = BuildReturnStmt(ReturnLoc, RetVal.get());
+  StmtResult R =
+      BuildReturnStmt(ReturnLoc, RetVal.get(), /*AllowRecovery=*/true);
   if (R.isInvalid() || ExprEvalContexts.back().isDiscardedStatementContext())
     return R;
 
   if (VarDecl *VD =
       const_cast<VarDecl*>(cast<ReturnStmt>(R.get())->getNRVOCandidate())) {
     CurScope->addNRVOCandidate(VD);
   } else {
     CurScope->setNoNRVO();
if (!D || !S.SourceMgr.isInSystemHeader(D->getLocation()))
     return false;
   for (const DeclContext *DC = D->getDeclContext(); DC; DC = DC->getParent()) {
     if (DC->isStdNamespace())
       return true;
   }
   return false;
 }
 
-StmtResult Sema::BuildReturnStmt(SourceLocation ReturnLoc, Expr *RetValExp) {
+StmtResult Sema::BuildReturnStmt(SourceLocation ReturnLoc, Expr *RetValExp,
+                                 bool AllowRecovery) {
   // Check for unexpanded parameter packs.
   if (RetValExp && DiagnoseUnexpandedParameterPack(RetValExp))
     return StmtError();
 
   // HACK: We suppress simpler implicit move here in msvc compatibility mode
   // just as a temporary work around, as the MSVC STL has issues with
   // this change.
   bool SupressSimplerImplicitMoves =
StmtResult Sema::BuildReturnStmt(SourceLocation ReturnLoc, Expr *RetValExp,
   // FIXME: Add a flag to the ScopeInfo to indicate whether we're performing
   // deduction.
   if (getLangOpts().CPlusPlus14) {
     if (AutoType *AT = FnRetType->getContainedAutoType()) {
       FunctionDecl *FD = cast<FunctionDecl>(CurContext);
       // If we've already decided this function is invalid, e.g. because
       // we saw a `return` whose expression had an error, don't keep
       // trying to deduce its return type.
-      if (FD->isInvalidDecl())
-        return StmtError();
-      if (DeduceFunctionTypeFromReturnExpr(FD, ReturnLoc, RetValExp, AT)) {
+      // (Some return values may be needlessly wrapped in RecoveryExpr).
+      if (FD->isInvalidDecl() ||
+          DeduceFunctionTypeFromReturnExpr(FD, ReturnLoc, RetValExp, AT)) {
         FD->setInvalidDecl();
+        if (!AllowRecovery)
+          return StmtError();
+        // The deduction failure is diagnosed and marked, try to recover.
+        if (RetValExp) {
+          // Wrap return value with a recovery expression of the previous type.
+          // If no deduction yet, use DependentTy.
+          auto Recovery = CreateRecoveryExpr(
+              RetValExp->getBeginLoc(), RetValExp->getEndLoc(), RetValExp,
+              AT->isDeduced() ? FnRetType : QualType());
+          if (Recovery.isInvalid())
             return StmtError();
+          RetValExp = Recovery.get();
+        } else {
+          // Nothing to do: a ReturnStmt with no value is fine recovery.
+        }
       } else {
         FnRetType = FD->getReturnType();
       }
     }
   }
   const VarDecl *NRVOCandidate = getCopyElisionCandidate(NRInfo, FnRetType);
 
   bool HasDependentReturnType = FnRetType->isDependentType();
 
   ReturnStmt *Result = nullptr;
   if (FnRetType->isVoidType()) {
     if (RetValExp) {
-      if (isa<InitListExpr>(RetValExp)) {
+      if (auto *ILE = dyn_cast<InitListExpr>(RetValExp)) {
         // We simply never allow init lists as the return value of void
         // functions. This is compatible because this was never allowed before,
         // so there's no legacy code to deal with.
         NamedDecl *CurDecl = getCurFunctionOrMethodDecl();
         int FunctionKind = 0;
         if (isa<ObjCMethodDecl>(CurDecl))
           FunctionKind = 1;
         else if (isa<CXXConstructorDecl>(CurDecl))
           FunctionKind = 2;
         else if (isa<CXXDestructorDecl>(CurDecl))
           FunctionKind = 3;
 
         Diag(ReturnLoc, diag::err_return_init_list)
             << CurDecl << FunctionKind << RetValExp->getSourceRange();
 
-        // Drop the expression.
-        RetValExp = nullptr;
+        // Preserve the initializers in the AST.
+        RetValExp = AllowRecovery
+                        ? CreateRecoveryExpr(ILE->getLBraceLoc(),
+                                             ILE->getRBraceLoc(), ILE->inits())
+                              .get()
+                        : nullptr;
       } else if (!RetValExp->isTypeDependent()) {
         // C99 6.8.6.4p1 (ext_ since GCC warns)
         unsigned D = diag::ext_return_has_expr;
         if (RetValExp->getType()->isVoidType()) {
           NamedDecl *CurDecl = getCurFunctionOrMethodDecl();
           if (isa<CXXConstructorDecl>(CurDecl) ||
               isa<CXXDestructorDecl>(CurDecl))
             D = diag::err_ctor_dtor_returns_void;
if (FnRetType->isVoidType()) {
     // In C++ the return statement is handled via a copy initialization,
     // the C version of which boils down to CheckSingleAssignmentConstraints.
     if (!HasDependentReturnType && !RetValExp->isTypeDependent()) {
       // we have a non-void function with an expression, continue checking
       InitializedEntity Entity =
           InitializedEntity::InitializeResult(ReturnLoc, RetType);
       ExprResult Res = PerformMoveOrCopyInitialization(
           Entity, NRInfo, RetValExp, SupressSimplerImplicitMoves);
+      if (Res.isInvalid() && AllowRecovery)
+        Res = CreateRecoveryExpr(RetValExp->getBeginLoc(),
+                                 RetValExp->getEndLoc(), RetValExp, RetType);
       if (Res.isInvalid()) {
         // FIXME: Clean up temporaries here anyway?
         return StmtError();
       }
       RetValExp = Res.getAs<Expr>();
 
       // If we have a related result type, we need to implicitly
       // convert back to the formal result type.  We can't pretend to

clang/test/AST/ast-dump-recovery.cpp

class DelegatingInit {
     DelegatingInit(double) : DelegatingInit(invalid) {}
     // CHECK:      CXXConstructorDecl {{.*}} DelegatingInit 'void (double)'
     // CHECK-NEXT: |-ParmVarDecl
     // CHECK-NEXT: |-CXXCtorInitializer 'DelegatingInit'
     // CHECK-NEXT: | `-ParenListExpr
     // CHECK-NEXT: |   `-RecoveryExpr {{.*}} '<dependent type>'
   };
 }
+
+float *brokenReturn() {
+  // CHECK:      FunctionDecl {{.*}} brokenReturn
+  return 42;
+  // CHECK:      ReturnStmt
+  // CHECK-NEXT: `-RecoveryExpr {{.*}} 'float *'
+  // CHECK-NEXT:   `-IntegerLiteral {{.*}} 'int' 42
+}
+
+// Return deduction treats the first, second *and* third differently!
+auto *brokenDeducedReturn(int *x, float *y, double *z) {
+  // CHECK:      FunctionDecl {{.*}} invalid brokenDeducedReturn
+  if (x) return x;
+  // CHECK:      ReturnStmt
+  // CHECK-NEXT: `-ImplicitCastExpr {{.*}} <LValueToRValue>
+  // CHECK-NEXT:   `-DeclRefExpr {{.*}} 'x' 'int *'
+  if (y) return y;
+  // CHECK:      ReturnStmt
+  // CHECK-NEXT: `-RecoveryExpr {{.*}} 'int *'
+  // CHECK-NEXT:   `-DeclRefExpr {{.*}} 'y' 'float *'
+  if (z) return z;
+  // CHECK:      ReturnStmt
+  // CHECK-NEXT: `-RecoveryExpr {{.*}} 'int *'
+  // CHECK-NEXT:   `-DeclRefExpr {{.*}} 'z' 'double *'
+  return x;
+  // Unfortunate: we wrap a valid return in RecoveryExpr.
+  // This is to avoid running deduction again after it failed once.
+  // CHECK:      ReturnStmt
+  // CHECK-NEXT: `-RecoveryExpr {{.*}} 'int *'
+  // CHECK-NEXT:   `-DeclRefExpr {{.*}} 'x' 'int *'
+}
+
+void returnInitListFromVoid() {
+  // CHECK:      FunctionDecl {{.*}} returnInitListFromVoid
+  return {7,8};
+  // CHECK:      ReturnStmt
+  // CHECK-NEXT: `-RecoveryExpr {{.*}} '<dependent type>'
+  // CHECK-NEXT:   |-IntegerLiteral {{.*}} 'int' 7
+  // CHECK-NEXT:   `-IntegerLiteral {{.*}} 'int' 8
+}

clang/test/SemaCXX/constant-expression-cxx11.cpp

 namespace Lifetime {
   void f() {
     constexpr int &n = n; // expected-error {{constant expression}} expected-note {{use of reference outside its lifetime}} expected-warning {{not yet bound to a value}}
     constexpr int m = m; // expected-error {{constant expression}} expected-note {{read of object outside its lifetime}}
   }
 
   constexpr int &get(int &&n) { return n; }
   // cxx2b-error@-1 {{non-const lvalue reference to type 'int' cannot bind to a temporary of type 'int'}}
-  // cxx2b-error@-2 {{no return statement in constexpr function}} See PR40598
   constexpr int &&get_rv(int &&n) { return static_cast<int&&>(n); }
   struct S {
     int &&r;
     int &s;
     int t;
-    constexpr S() : r(get_rv(0)), s(get(0)), t(r) {} // expected-note {{read of object outside its lifetime}}
-    constexpr S(int) : r(get_rv(0)), s(get(0)), t(s) {}
-    // cxx2b-warning@-1 {{reference 's' is not yet bound to a value when used here}}
-    // cxx2b-note@-2    {{read of uninitialized object is not allowed in a constant expression}}
-    // cxx11_20-note@-3 {{read of object outside its lifetime}}
+    constexpr S() : r(get_rv(0)), s(get(0)), t(r) {} // cxx11_20-note {{read of object outside its lifetime}}
+    constexpr S(int) : r(get_rv(0)), s(get(0)), t(s) {} // cxx11_20-note {{read of object outside its lifetime}}
   };
-  constexpr int k1 = S().t; // expected-error {{constant expression}} expected-note {{in call}}
-  constexpr int k2 = S(0).t; // expected-error {{constant expression}} expected-note {{in call}}
+  constexpr int k1 = S().t; // expected-error {{constant expression}} cxx11_20-note {{in call}}
+  constexpr int k2 = S(0).t; // expected-error {{constant expression}} cxx11_20-note {{in call}}
 
   struct Q {
     int n = 0;
     constexpr int f() const { return 0; }
   };
   constexpr Q *out_of_lifetime(Q q) { return &q; } // expected-warning {{address of stack}}
   constexpr int k3 = out_of_lifetime({})->n; // expected-error {{constant expression}} expected-note {{read of object outside its lifetime}}
   constexpr int k4 = out_of_lifetime({})->f(); // expected-error {{constant expression}} expected-note {{member call on object outside its lifetime}}

clang/test/SemaCXX/constant-expression-cxx14.cpp

namespace VirtualFromBase {
   constexpr X<X<S2>> xxs2;
   constexpr X<S2> *q = const_cast<X<X<S2>>*>(&xxs2);
   static_assert(q->f() == sizeof(X<S2>), ""); // cxx14-error {{constant expression}} cxx14-note {{virtual function}}
 }
 
 namespace Lifetime {
   constexpr int &get(int &&r) { return r; }
   // cxx2b-error@-1 {{non-const lvalue reference to type 'int' cannot bind to a temporary of type 'int'}}
-  // cxx2b-error@-2 {{no return statement in constexpr function}} See PR40598
   constexpr int f() {
     int &r = get(123);
     return r;
-    // cxx2b-note@-1 {{use of reference outside its lifetime is not allowed in a constant expression}}
-    // cxx14_20-note@-2 {{read of object outside its lifetime}}
+    // cxx14_20-note@-1 {{read of object outside its lifetime}}
   }
-  static_assert(f() == 123, ""); // expected-error {{constant expression}} expected-note {{in call}}
+  static_assert(f() == 123, ""); // expected-error {{constant expression}} cxx14_20-note {{in call}}
 
   constexpr int g() {
     int *p = 0;
     {
       int n = 0;
       p = &n;
       n = 42;
     }

clang/test/SemaCXX/constexpr-function-recovery-crash.cpp

 
 constexpr int test10() { return undef(); } // expected-error {{use of undeclared identifier 'undef'}}
 static_assert(test10() <= 1, "should not crash"); // expected-error {{static_assert expression is not an integral constant expression}}
 
 struct X {} array[] = {undef()}; // expected-error {{use of undeclared identifier 'undef'}}
 constexpr void test11() {
   for (X& e : array) {}
 }
+
+constexpr int test12() { return "wrong"; } // expected-error {{cannot initialize return object of type 'int'}}
+constexpr int force12 = test12();          // expected-error {{must be initialized by a constant}}