Use musttail for variadic method thunks when possible

[rnk]

Summary

This avoids cloning variadic virtual methods when the target supports
musttail and the return type is not covariant. I think we never
implemented this previously because it doesn't handle the covariant
case. But, in the MS ABI, there are some cases where vtable thunks must
be emitted even when the variadic method defintion is not available, so
it looks like we need to implement this. Do it for both ABIs, since it's
a nice size improvement and simplification for Itanium.

Fixes PR43173.

[efriedma]

Do we have test coverage for a variadic, covariant thunk for a function without a definition? I don't think there's any way for us to actually emit that, but we should make sure the error message is right.

I'm a little concerned that using musttail thunks with the Itanium ABI will flush out bugs that have been hiding because we have less test coverage on Windows. But it's probably the right thing to do.

clang/lib/CodeGen/CGVTables.cpp

206

This is fine, I guess, but how is it related?

[rnk]

In D67028#1653439, @efriedma wrote:

Do we have test coverage for a variadic, covariant thunk for a function without a definition? I don't think there's any way for us to actually emit that, but we should make sure the error message is right.

That appears to have been PR18098, so we have a test for it in thunks.cpp. The way the Itanium C++ ABI works, thunks are always emitted as weak_odr in TUs that provide the implementation, so emitting them with the vtable is just an optimization. Basically, clang never *has* to emit a thunk when the definition isn't present, it can always choose to simply declare it, and that's what we do here: https://github.com/llvm/llvm-project/blob/7c8952197b86790b31731d34d559281840916e1f/clang/lib/CodeGen/CGVTables.cpp#L558

In the MS ABI, deriving a new class may require the creation of new thunks for methods that were not overridden, so we can't use the same trick.

I think my two new tests are redundant with tests in thunks.cpp, so perhaps I should just add some new RUN lines there.

I'm a little concerned that using musttail thunks with the Itanium ABI will flush out bugs that have been hiding because we have less test coverage on Windows. But it's probably the right thing to do.

True, it's a risk. One other thing worth mentioning is that the IR cloning doesn't actually work in the presence of labels-as-values, so we are improving conformance in that extra, extra corner case.

clang/lib/CodeGen/CGVTables.cpp

206

The MS ABI implementation of performThisAdjustment returns a pointer that needs to be cast. The covariant test exercises this codepath in the MS ABI, and we didn't do that before. I guess the Itanium one does the cast internally, but the MS ABI impl has this comment:

// Don't need to bitcast back, the call CodeGen will handle this.
return V;

I probably removed the cast when trying to clean up our generated IR.

[efriedma]

In the MS ABI, deriving a new class may require the creation of new thunks for methods that were not overridden, so we can't use the same trick.

Yes. MSVC emits an error message "covariant returns with multiple or virtual inheritance not supported for varargs functions" in cases like the following. It looks like the equivalent isn't implemented in clang? (Is that the llvm_unreachable in this patch?)

[efriedma]

Oops, meant to actually include the testcase in my last comment:

struct V1 { };
struct V2 : virtual V1 { };
struct A {

virtual V1 *f(int,...);

};
struct B : A {

virtual void b();
virtual V2 *f(int,...);

};
B b;

[rnk]

I think what I said applies to your test case. Basically, in the Itanium C++ ABI, virtual method definitions provide all their thunks as weak_odr. We only emit thunks referenced by vtables as an optimization, and they are marked available_externally. In your test case, we hit the early return that I linked to, so we don't try to clone, and we don't need to emit an error.

[efriedma]

In your test case, we hit the early return that I linked to, so we don't try to clone, and we don't need to emit an error.

Yes, that's what happens with the Itanium ABI; what happens with the MS ABI?

[rnk]

• merge into thunks.cpp

[rnk]

In D67028#1659881, @efriedma wrote:

In your test case, we hit the early return that I linked to, so we don't try to clone, and we don't need to emit an error.

Yes, that's what happens with the Itanium ABI; what happens with the MS ABI?

I see, it doesn't work. There's also the case of the MS ABI for ISAs that don't support musttail, although that's less interesting.

[rnk]

• emit an error if we try to clone without a definition

[rnk]

clang/test/CodeGenCXX/ms-thunks-variadic-return.cpp

9 ↗

(On Diff #218995)

I do love the optimism of the CGM.ErrorUnsupported diagnostic: "cannot compile this ${unimplementable_feature} yet".

I suppose it could be done if we standardized a new ABI for variadic virtual methods with pointer-like return values, so that they emit the main implementation under a new symbol name that takes a va_list and then the variadic one thunks to it. =P

[efriedma]

LGTM

clang/lib/CodeGen/CGVTables.cpp

CodeGenFunction::GenerateVarArgsThunk(llvm::Function *Fn,
   const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>();
   QualType ResultType = FPT->getReturnType();
 
   // Get the original function
   assert(FnInfo.isVariadic());
   llvm::Type *Ty = CGM.getTypes().GetFunctionType(FnInfo);
   llvm::Value *Callee = CGM.GetAddrOfFunction(GD, Ty, /*ForVTable=*/true);
   llvm::Function *BaseFn = cast<llvm::Function>(Callee);
+  assert(!BaseFn->isDeclaration() && "cannot clone undefined variadic method");
 
   // Clone to thunk.
   llvm::ValueToValueMapTy VMap;
 
   // We are cloning a function while some Metadata nodes are still unresolved.
   // Ensure that the value mapper does not encounter any of them.
   resolveTopLevelMetadata(BaseFn, VMap);
   llvm::Function *NewFn = llvm::CloneFunction(BaseFn, VMap);
llvm::BasicBlock::iterator ThisStore =
                I.getOperand(0) == ThisPtr.getPointer();
       });
   assert(ThisStore != EntryBB->end() &&
          "Store of this should be in entry block?");
   // Adjust "this", if necessary.
   Builder.SetInsertPoint(&*ThisStore);
   llvm::Value *AdjustedThisPtr =
       CGM.getCXXABI().performThisAdjustment(*this, ThisPtr, Thunk.This);
+  AdjustedThisPtr = Builder.CreateBitCast(AdjustedThisPtr,
+                                          ThisStore->getOperand(0)->getType());
   ThisStore->setOperand(0, AdjustedThisPtr);
 
   if (!Thunk.Return.isEmpty()) {
     // Fix up the returned value, if necessary.
     for (llvm::BasicBlock &BB : *Fn) {
       llvm::Instruction *T = BB.getTerminator();
       if (isa<llvm::ReturnInst>(T)) {
         RValue RV = RValue::get(T->getOperand(0));
void CodeGenFunction::EmitCallAndReturnForThunk(llvm::FunctionCallee Callee,
   const CXXMethodDecl *MD = cast<CXXMethodDecl>(CurGD.getDecl());
 
   // Adjust the 'this' pointer if necessary
   llvm::Value *AdjustedThisPtr =
     Thunk ? CGM.getCXXABI().performThisAdjustment(
                           *this, LoadCXXThisAddress(), Thunk->This)
           : LoadCXXThis();
 
-  if (CurFnInfo->usesInAlloca() || IsUnprototyped) {
-    // We don't handle return adjusting thunks, because they require us to call
-    // the copy constructor.  For now, fall through and pretend the return
-    // adjustment was empty so we don't crash.
+  // If perfect forwarding is required a variadic method, a method using
+  // inalloca, or an unprototyped thunk, use musttail. Emit an error if this
+  // thunk requires a return adjustment, since that is impossible with musttail.
+  if (CurFnInfo->usesInAlloca() || CurFnInfo->isVariadic() || IsUnprototyped) {
     if (Thunk && !Thunk->Return.isEmpty()) {
       if (IsUnprototyped)
         CGM.ErrorUnsupported(
             MD, "return-adjusting thunk with incomplete parameter type");
+      else if (CurFnInfo->isVariadic())
+        llvm_unreachable("shouldn't try to emit musttail return-adjusting "
+                         "thunks for variadic functions");
       else
         CGM.ErrorUnsupported(
             MD, "non-trivial argument copy for return-adjusting thunk");
     }
     EmitMustTailThunk(CurGD, AdjustedThisPtr, Callee);
     return;
   }
 
llvm::Constant *CodeGenVTables::maybeEmitThunk(GlobalDecl GD,
   // that the return type is meaningless. These thunks can be used to call
   // functions with differing return types, and the caller is required to cast
   // the prototype appropriately to extract the correct value.
   if (IsUnprototyped)
     ThunkFn->addFnAttr("thunk");
 
   CGM.SetLLVMFunctionAttributesForDefinition(GD.getDecl(), ThunkFn);
 
+  // Thunks for variadic methods are special because in general variadic
+  // arguments cannot be perferctly forwarded. In the general case, clang
+  // implements such thunks by cloning the original function body. However, for
+  // thunks with no return adjustment on targets that support musttail, we can
+  // use musttail to perfectly forward the variadic arguments.
+  bool ShouldCloneVarArgs = false;
   if (!IsUnprototyped && ThunkFn->isVarArg()) {
-    // Varargs thunks are special; we can't just generate a call because
-    // we can't copy the varargs.  Our implementation is rather
-    // expensive/sucky at the moment, so don't generate the thunk unless
-    // we have to.
-    // FIXME: Do something better here; GenerateVarArgsThunk is extremely ugly.
+    ShouldCloneVarArgs = true;
+    if (TI.Return.isEmpty()) {
+      switch (CGM.getTriple().getArch()) {
+      case llvm::Triple::x86_64:
+      case llvm::Triple::x86:
+      case llvm::Triple::aarch64:
+        ShouldCloneVarArgs = false;
+        break;
+      default:
+        break;
+      }
+    }
+  }
+
+  if (ShouldCloneVarArgs) {
     if (UseAvailableExternallyLinkage)
       return ThunkFn;
-    ThunkFn = CodeGenFunction(CGM).GenerateVarArgsThunk(ThunkFn, FnInfo, GD,
-                                                        TI);
+    ThunkFn =
+        CodeGenFunction(CGM).GenerateVarArgsThunk(ThunkFn, FnInfo, GD, TI);
   } else {
     // Normal thunk body generation.
     CodeGenFunction(CGM).generateThunk(ThunkFn, FnInfo, GD, TI, IsUnprototyped);
   }
 
   setThunkProperties(CGM, TI, ThunkFn, ForVTable, GD);
   return ThunkFn;
 }

clang/test/CodeGenCXX/linetable-virtual-variadic.cpp

-// RUN: %clang_cc1 -triple x86_64-apple-darwin -emit-llvm -debug-info-kind=line-tables-only %s -o - | FileCheck %s
-// RUN: %clang_cc1 -triple x86_64-apple-darwin -emit-llvm -debug-info-kind=line-directives-only %s -o - | FileCheck %s
+// Sparc64 is used because AArch64 and X86_64 would both use musttail.
+// RUN: %clang_cc1 -triple sparc64-linux-gnu -emit-llvm -debug-info-kind=line-tables-only %s -o - | FileCheck %s
+// RUN: %clang_cc1 -triple sparc64-linux-gnu -emit-llvm -debug-info-kind=line-directives-only %s -o - | FileCheck %s
 // Crasher for PR22929.
 class Base {
   virtual void VariadicFunction(...);
 };
 
 class Derived : public virtual Base {
   virtual void VariadicFunction(...);
 };

clang/test/CodeGenCXX/thunks-variadic-covariant.cpp

+// RUN: %clang_cc1 %s -triple=x86_64-windows-msvc -munwind-tables -emit-llvm -o - -O1 -disable-llvm-passes | FileCheck --check-prefix=MSVC %s
+// RUN: %clang_cc1 %s -triple=x86_64-windows-gnu -munwind-tables -emit-llvm -o - -O1 -disable-llvm-passes | FileCheck --check-prefix=ITANIUM %s
+
+struct A {
+  virtual A *f(int x, ...);
+};
+struct B {
+  virtual B *f(int x, ...);
+};
+struct C : A, B {
+  virtual void c();
+  virtual C *f(int x, ...);
+};
+C *makeC();
+C *C::f(int x, ...) { return makeC(); }
+C c;
+
+// MSVC-LABEL: define dso_local %struct.C* @"?f@C@@UEAAPEAU1@HZZ"
+// MSVC-SAME: (%struct.C* %this, i32 %x, ...)
+// MSVC: call %struct.C* @"?makeC@@YAPEAUC@@XZ"()
+
+// MSVC-LABEL: define weak_odr dso_local %struct.C* @"?f@C@@W7EAAPEAUB@@HZZ"
+// MSVC-SAME: (%struct.C* %this, i32 %x, ...)
+// MSVC: call %struct.C* @"?makeC@@YAPEAUC@@XZ"()
+// MSVC: getelementptr inbounds i8, i8* {{.*}}, i32 8
+// MSVC: ret %struct.C* %7
+
+// MSVC-LABEL: define linkonce_odr dso_local %struct.C* @"?f@C@@W7EAAPEAU1@HZZ"
+// MSVC-SAME: (%struct.C* %this, i32 %x, ...)
+// MSVC: getelementptr i8, i8* %{{[^,]*}}, i32 -8
+// MSVC: musttail call %struct.C* (%struct.C*, i32, ...) @"?f@C@@UEAAPEAU1@HZZ"(%struct.C* %{{[^,]*}}, i32 %{{[^,]*}}, ...)
+
+
+// ITANIUM-LABEL: define dso_local %struct.C* @_ZN1C1fEiz
+// ITANIUM-SAME: (%struct.C* %this, i32 %x, ...)
+// ITANIUM: call %struct.C* @_Z5makeCv()
+
+// ITANIUM-LABEL: define dso_local %struct.C* @_ZTchn8_h8_N1C1fEiz
+// ITANIUM-SAME: (%struct.C* %this, i32 %x, ...)
+// ITANIUM: call %struct.C* @_Z5makeCv()
+// ITANIUM: getelementptr inbounds i8, i8* %{{[^,]*}}, i64 8

clang/test/CodeGenCXX/thunks-variadic.cpp

+// RUN: %clang_cc1 %s -triple=x86_64-windows-msvc -munwind-tables -emit-llvm -o - -O1 -disable-llvm-passes | FileCheck --check-prefix=MSVC %s
+// RUN: %clang_cc1 %s -triple=x86_64-windows-gnu -munwind-tables -emit-llvm -o - -O1 -disable-llvm-passes | FileCheck --check-prefix=ITANIUM %s
+
+namespace test1 {
+class a {
+  virtual void b(const char *, ...);
+};
+class c {
+  virtual void b(const char *, ...);
+};
+class e : c, a {
+  void b(const char *, ...) override;
+} d;
+
+// MSVC-LABEL: define linkonce_odr dso_local void @"?b@e@test1@@G7EAAXPEBDZZ"
+// MSVC-SAME: (%"class.test1::e"* %this, i8* %[[ARG:[^,]+]], ...)
+// MSVC: getelementptr i8, i8* %{{.*}}, i32 -8
+// MSVC: musttail call void (%"class.test1::e"*, i8*, ...) @"?b@e@test1@@EEAAXPEBDZZ"(%"class.test1::e"* %{{.*}}, i8* %[[ARG]], ...)
+
+// ITANIUM-LABEL: define available_externally dso_local void @_ZThn8_N5test11e1bEPKcz
+// ITANIUM-SAME: (%"class.test1::e"* %this, i8* %{{.*}}, ...)
+// ITANIUM: getelementptr inbounds i8, i8* %{{.*}}, i64 -8
+// ITANIUM: musttail call void (%"class.test1::e"*, i8*, ...) @_ZN5test11e1bEPKcz(%"class.test1::e"* %{{.*}}, i8* %{{.*}}, ...)
+}

clang/test/CodeGenCXX/thunks.cpp

-// RUN: %clang_cc1 %s -triple=x86_64-pc-linux-gnu -munwind-tables -emit-llvm -o - | FileCheck --check-prefix=CHECK --check-prefix=CHECK-NONOPT %s
-// RUN: %clang_cc1 %s -triple=x86_64-pc-linux-gnu -debug-info-kind=standalone -dwarf-version=5 -munwind-tables -emit-llvm -o - | FileCheck --check-prefix=CHECK --check-prefix=CHECK-NONOPT --check-prefix=CHECK-DBG %s
-// RUN: %clang_cc1 %s -triple=x86_64-pc-linux-gnu -munwind-tables -emit-llvm -o - -O1 -disable-llvm-passes | FileCheck --check-prefix=CHECK --check-prefix=CHECK-OPT %s
+// Sparc64 is used because AArch64 and X86_64 would both use musttail.
+// RUN: %clang_cc1 %s -triple=sparc64-pc-linux-gnu -munwind-tables -emit-llvm -o - | FileCheck --check-prefix=CHECK --check-prefix=CHECK-NONOPT %s
+// RUN: %clang_cc1 %s -triple=sparc64-pc-linux-gnu -debug-info-kind=standalone -dwarf-version=5 -munwind-tables -emit-llvm -o - | FileCheck --check-prefix=CHECK --check-prefix=CHECK-NONOPT --check-prefix=CHECK-DBG %s
+// RUN: %clang_cc1 %s -triple=sparc64-pc-linux-gnu -munwind-tables -emit-llvm -o - -O1 -disable-llvm-passes | FileCheck --check-prefix=CHECK --check-prefix=CHECK-OPT %s
 
 namespace Test1 {
 
 // Check that we emit a non-virtual thunk for C::f.
 
 struct A {
   virtual void f();
 };