Default vaarg lowering should support indirect struct types.

[jyknight]

Summary

Fixes PR11517 for SPARC.

On most targets, clang lowers va_arg itself, eschewing the use of the
llvm vaarg instruction. This is necessary (at least for now) as the type
argument to the vaarg instruction cannot represent all the ABI
information that is needed to support complex calling conventions.

However, on targets with a simpler varrags ABIs, the LLVM instruction
can work just fine, and clang can simply lower to it. Unfortunately,
even on such targets, vaarg with a struct argument would fail, because
the default lowering to vaarg was naive: it didn't take into account the
ABI attribute computed by classifyArgumentType. In particular, for the
DefaultABIInfo, structs are supposed to be passed indirectly and so
llvm's vaarg instruction should be emitted with a pointer argument.

Now, vaarg instruction emission is able to use computed ABIArgInfo for
the provided argument type, which allows the default ABI support to work
for structs too.

I haven't touched the EmitVAArg implementation for PPC32_SVR4 or XCore,
although I believe both are now redundant, and could be switched over to
use the default implementation as well.

[dschuff]

This LGTM for PNaCl and WebAssembly.

It's possible that for wasm we may still end up doing our own thing for varargs instead of using the default, but this still looks like strictly an improvement, (and it makes bootstrapping a backend simpler).

[sunfish]

For WebAssembly, I'm expecting we'll switch to emitVoidPtrVAArg and eschew LLVM's vaarg instruction too, as other targets do. This exposes the expanded code to the full optimizer, which seems preferable.

cfe/trunk/lib/CodeGen/CGExprAgg.cpp

 void AggExprEmitter::VisitChooseExpr(const ChooseExpr *CE) {
   Visit(CE->getChosenSubExpr());
 }
 
 void AggExprEmitter::VisitVAArgExpr(VAArgExpr *VE) {
   Address ArgValue = Address::invalid();
   Address ArgPtr = CGF.EmitVAArg(VE, ArgValue);
 
+  // If EmitVAArg fails, emit an error.
   if (!ArgPtr.isValid()) {
-    // If EmitVAArg fails, we fall back to the LLVM instruction.
-    llvm::Value *Val = Builder.CreateVAArg(ArgValue.getPointer(),
-                                           CGF.ConvertType(VE->getType()));
-    if (!Dest.isIgnored())
-      Builder.CreateStore(Val, Dest.getAddress());
+    CGF.ErrorUnsupported(VE, "aggregate va_arg expression");
     return;
   }
 
   EmitFinalDestCopy(VE->getType(), CGF.MakeAddrLValue(ArgPtr, VE->getType()));
 }
 
 void AggExprEmitter::VisitCXXBindTemporaryExpr(CXXBindTemporaryExpr *E) {
   // Ensure that we have a slot, but if we already do, remember

cfe/trunk/lib/CodeGen/CGExprScalar.cpp

Value *ScalarExprEmitter::VisitVAArgExpr(VAArgExpr *VE) {
   if (Ty->isVariablyModifiedType())
     CGF.EmitVariablyModifiedType(Ty);
 
   Address ArgValue = Address::invalid();
   Address ArgPtr = CGF.EmitVAArg(VE, ArgValue);
 
   llvm::Type *ArgTy = ConvertType(VE->getType());
 
-  // If EmitVAArg fails, we fall back to the LLVM instruction.
-  if (!ArgPtr.isValid())
-    return Builder.CreateVAArg(ArgValue.getPointer(), ArgTy);
+  // If EmitVAArg fails, emit an error.
+  if (!ArgPtr.isValid()) {
+    CGF.ErrorUnsupported(VE, "va_arg expression");
+    return llvm::UndefValue::get(ArgTy);
+  }
 
   // FIXME Volatility.
   llvm::Value *Val = Builder.CreateLoad(ArgPtr);
 
   // If EmitVAArg promoted the type, we must truncate it.
   if (ArgTy != Val->getType()) {
     if (ArgTy->isPointerTy() && !Val->getType()->isPointerTy())
       Val = Builder.CreateIntToPtr(Val, ArgTy);

cfe/trunk/lib/CodeGen/TargetInfo.cpp

static bool canExpandIndirectArgument(QualType Ty, ASTContext &Context) {
   // Make sure there are not any holes in the struct.
   if (Size != Context.getTypeSize(Ty))
     return false;
 
   return true;
 }
 
 namespace {
+Address EmitVAArgInstr(CodeGenFunction &CGF, Address VAListAddr, QualType Ty,
+                       const ABIArgInfo &AI) {
+  // This default implementation defers to the llvm backend's va_arg
+  // instruction. It can handle only passing arguments directly
+  // (typically only handled in the backend for primitive types), or
+  // aggregates passed indirectly by pointer (NOTE: if the "byval"
+  // flag has ABI impact in the callee, this implementation cannot
+  // work.)
+
+  // Only a few cases are covered here at the moment -- those needed
+  // by the default abi.
+  llvm::Value *Val;
+
+  if (AI.isIndirect()) {
+    assert(!AI.getPaddingType() &&
+           "Unepxected PaddingType seen in arginfo in generic VAArg emitter!");
+    assert(
+        !AI.getIndirectRealign() &&
+        "Unepxected IndirectRealign seen in arginfo in generic VAArg emitter!");
+
+    auto TyInfo = CGF.getContext().getTypeInfoInChars(Ty);
+    CharUnits TyAlignForABI = TyInfo.second;
+
+    llvm::Type *BaseTy =
+        llvm::PointerType::getUnqual(CGF.ConvertTypeForMem(Ty));
+    llvm::Value *Addr =
+        CGF.Builder.CreateVAArg(VAListAddr.getPointer(), BaseTy);
+    return Address(Addr, TyAlignForABI);
+  } else {
+    assert((AI.isDirect() || AI.isExtend()) &&
+           "Unexpected ArgInfo Kind in generic VAArg emitter!");
+
+    assert(!AI.getInReg() &&
+           "Unepxected InReg seen in arginfo in generic VAArg emitter!");
+    assert(!AI.getPaddingType() &&
+           "Unepxected PaddingType seen in arginfo in generic VAArg emitter!");
+    assert(!AI.getDirectOffset() &&
+           "Unepxected DirectOffset seen in arginfo in generic VAArg emitter!");
+    assert(!AI.getCoerceToType() &&
+           "Unepxected CoerceToType seen in arginfo in generic VAArg emitter!");
+
+    Address Temp = CGF.CreateMemTemp(Ty, "varet");
+    Val = CGF.Builder.CreateVAArg(VAListAddr.getPointer(), CGF.ConvertType(Ty));
+    CGF.Builder.CreateStore(Val, Temp);
+    return Temp;
+  }
+}
+
 /// DefaultABIInfo - The default implementation for ABI specific
 /// details. This implementation provides information which results in
 /// self-consistent and sensible LLVM IR generation, but does not
 /// conform to any particular ABI.
 class DefaultABIInfo : public ABIInfo {
 public:
   DefaultABIInfo(CodeGen::CodeGenTypes &CGT) : ABIInfo(CGT) {}
 
   ABIArgInfo classifyReturnType(QualType RetTy) const;
   ABIArgInfo classifyArgumentType(QualType RetTy) const;
 
   void computeInfo(CGFunctionInfo &FI) const override {
     if (!getCXXABI().classifyReturnType(FI))
       FI.getReturnInfo() = classifyReturnType(FI.getReturnType());
     for (auto &I : FI.arguments())
       I.info = classifyArgumentType(I.type);
   }
 
   Address EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
-                    QualType Ty) const override;
+                    QualType Ty) const override {
+    return EmitVAArgInstr(CGF, VAListAddr, Ty, classifyArgumentType(Ty));
+  }
 };
 
 class DefaultTargetCodeGenInfo : public TargetCodeGenInfo {
 public:
   DefaultTargetCodeGenInfo(CodeGen::CodeGenTypes &CGT)
     : TargetCodeGenInfo(new DefaultABIInfo(CGT)) {}
 };
 
-Address DefaultABIInfo::EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
-                                  QualType Ty) const {
-  return Address::invalid();
-}
-
 ABIArgInfo DefaultABIInfo::classifyArgumentType(QualType Ty) const {
   Ty = useFirstFieldIfTransparentUnion(Ty);
 
   if (isAggregateTypeForABI(Ty)) {
     // Records with non-trivial destructors/copy-constructors should not be
     // passed by value.
     if (CGCXXABI::RecordArgABI RAA = getRecordArgABI(Ty, getCXXABI()))
       return getNaturalAlignIndirect(Ty, RAA == CGCXXABI::RAA_DirectInMemory);
public:
   explicit WebAssemblyABIInfo(CodeGen::CodeGenTypes &CGT)
       : DefaultABIInfo(CGT) {}
 
 private:
   ABIArgInfo classifyReturnType(QualType RetTy) const;
   ABIArgInfo classifyArgumentType(QualType Ty) const;
 
   // DefaultABIInfo's classifyReturnType and classifyArgumentType are
-  // non-virtual, but computeInfo is virtual, so we overload that.
+  // non-virtual, but computeInfo and EmitVAArg is virtual, so we
+  // overload them.
   void computeInfo(CGFunctionInfo &FI) const override {
     if (!getCXXABI().classifyReturnType(FI))
       FI.getReturnInfo() = classifyReturnType(FI.getReturnType());
     for (auto &Arg : FI.arguments())
       Arg.info = classifyArgumentType(Arg.type);
   }
 
   Address EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
if (!getCXXABI().classifyReturnType(FI))
     FI.getReturnInfo() = classifyReturnType(FI.getReturnType());
 
   for (auto &I : FI.arguments())
     I.info = classifyArgumentType(I.type);
 }
 
 Address PNaClABIInfo::EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
                                 QualType Ty) const {
-  return Address::invalid();
+  // The PNaCL ABI is a bit odd, in that varargs don't use normal
+  // function classification. Structs get passed directly for varargs
+  // functions, through a rewriting transform in
+  // pnacl-llvm/lib/Transforms/NaCl/ExpandVarArgs.cpp, which allows
+  // this target to actually support a va_arg instructions with an
+  // aggregate type, unlike other targets.
+  return EmitVAArgInstr(CGF, VAListAddr, Ty, ABIArgInfo::getDirect());
 }
 
 /// \brief Classify argument of given type \p Ty.
 ABIArgInfo PNaClABIInfo::classifyArgumentType(QualType Ty) const {
   if (isAggregateTypeForABI(Ty)) {
     if (CGCXXABI::RecordArgABI RAA = getRecordArgABI(Ty, getCXXABI()))
       return getNaturalAlignIndirect(Ty, RAA == CGCXXABI::RAA_DirectInMemory);
     return getNaturalAlignIndirect(Ty);
public:
   }
 
   bool initDwarfEHRegSizeTable(CodeGen::CodeGenFunction &CGF,
                                llvm::Value *Address) const override;
 };
 
 }
 
+// TODO: this implementation is now likely redundant with
+// DefaultABIInfo::EmitVAArg.
 Address PPC32_SVR4_ABIInfo::EmitVAArg(CodeGenFunction &CGF, Address VAList,
                                       QualType Ty) const {
   const unsigned OverflowLimit = 8;
   if (const ComplexType *CTy = Ty->getAs<ComplexType>()) {
     // TODO: Implement this. For now ignore.
     (void)CTy;
-    return Address::invalid();
+    return Address::invalid(); // FIXME?
   }
 
   // struct __va_list_tag {
   //   unsigned char gpr;
   //   unsigned char fpr;
   //   unsigned short reserved;
   //   void *overflow_arg_area;
   //   void *reg_save_area;
PPC32TargetCodeGenInfo::initDwarfEHRegSizeTable(CodeGen::CodeGenFunction &CGF,
 
   return false;
 }
 
 // PowerPC-64
 
 namespace {
 /// PPC64_SVR4_ABIInfo - The 64-bit PowerPC ELF (SVR4) ABI information.
-class PPC64_SVR4_ABIInfo : public DefaultABIInfo {
+class PPC64_SVR4_ABIInfo : public ABIInfo {
 public:
   enum ABIKind {
     ELFv1 = 0,
     ELFv2
   };
 
 private:
   static const unsigned GPRBits = 64;
private:
   }
 
   bool IsQPXVectorTy(QualType Ty) const {
     return IsQPXVectorTy(Ty.getTypePtr());
   }
 
 public:
   PPC64_SVR4_ABIInfo(CodeGen::CodeGenTypes &CGT, ABIKind Kind, bool HasQPX)
-    : DefaultABIInfo(CGT), Kind(Kind), HasQPX(HasQPX) {}
+      : ABIInfo(CGT), Kind(Kind), HasQPX(HasQPX) {}
 
   bool isPromotableTypeForABI(QualType Ty) const;
   CharUnits getParamTypeAlignment(QualType Ty) const;
 
   ABIArgInfo classifyReturnType(QualType RetTy) const;
   ABIArgInfo classifyArgumentType(QualType Ty) const;
 
   bool isHomogeneousAggregateBaseType(QualType Ty) const override;

 }
 
 Address AArch64ABIInfo::EmitDarwinVAArg(Address VAListAddr, QualType Ty,
                                         CodeGenFunction &CGF) const {
   // The backend's lowering doesn't support va_arg for aggregates or
   // illegal vector types.  Lower VAArg here for these cases and use
   // the LLVM va_arg instruction for everything else.
   if (!isAggregateTypeForABI(Ty) && !isIllegalVectorType(Ty))
-    return Address::invalid();
+    return EmitVAArgInstr(CGF, VAListAddr, Ty, ABIArgInfo::getDirect());
 
   CharUnits SlotSize = CharUnits::fromQuantity(8);
 
   // Empty records are ignored for parameter passing purposes.
   if (isEmptyRecord(getContext(), Ty, true)) {
     Address Addr(CGF.Builder.CreateLoad(VAListAddr, "ap.cur"), SlotSize);
     Addr = CGF.Builder.CreateElementBitCast(Addr, CGF.ConvertTypeForMem(Ty));
     return Addr;
public:
   XCoreTargetCodeGenInfo(CodeGenTypes &CGT)
     :TargetCodeGenInfo(new XCoreABIInfo(CGT)) {}
   void emitTargetMD(const Decl *D, llvm::GlobalValue *GV,
                     CodeGen::CodeGenModule &M) const override;
 };
 
 } // End anonymous namespace.
 
+// TODO: this implementation is likely now redundant with the default
+// EmitVAArg.
 Address XCoreABIInfo::EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
                                 QualType Ty) const {
   CGBuilderTy &Builder = CGF.Builder;
 
   // Get the VAList.
   CharUnits SlotSize = CharUnits::fromQuantity(4);
   Address AP(Builder.CreateLoad(VAListAddr), SlotSize);
 

cfe/trunk/test/CodeGen/le32-vaarg.c

 // RUN: %clang_cc1 -triple le32-unknown-nacl -emit-llvm -o - %s | FileCheck %s
 #include <stdarg.h>
 
 int get_int(va_list *args) {
   return va_arg(*args, int);
 }
 // CHECK: define i32 @get_int
 // CHECK: [[RESULT:%[a-z_0-9]+]] = va_arg {{.*}}, i32{{$}}
-// CHECK: ret i32 [[RESULT]]
+// CHECK: store i32 [[RESULT]], i32* [[LOC:%[a-z_0-9]+]]
+// CHECK: [[RESULT2:%[a-z_0-9]+]] = load i32, i32* [[LOC]]
+// CHECK: ret i32 [[RESULT2]]
 
 struct Foo {
   int x;
 };
 
 struct Foo dest;
 
 void get_struct(va_list *args) {
   dest = va_arg(*args, struct Foo);
 }
 // CHECK: define void @get_struct
 // CHECK: [[RESULT:%[a-z_0-9]+]] = va_arg {{.*}}, %struct.Foo{{$}}
-// CHECK: store %struct.Foo [[RESULT]], %struct.Foo* @dest
+// CHECK: store %struct.Foo [[RESULT]], %struct.Foo* [[LOC:%[a-z_0-9]+]]
+// CHECK: [[LOC2:%[a-z_0-9]+]] = bitcast {{.*}} [[LOC]] to i8*
+// CHECK: call void @llvm.memcpy{{.*}}@dest{{.*}}, i8* [[LOC2]]
 
 void skip_struct(va_list *args) {
   va_arg(*args, struct Foo);
 }
 // CHECK: define void @skip_struct
 // CHECK: va_arg {{.*}}, %struct.Foo{{$}}

cfe/trunk/test/CodeGen/sparc-vaarg.c

+// RUN: %clang_cc1 -triple sparc -emit-llvm -o - %s | FileCheck %s
+#include <stdarg.h>
+
+// CHECK-LABEL: define i32 @get_int
+// CHECK: [[RESULT:%[a-z_0-9]+]] = va_arg {{.*}}, i32{{$}}
+// CHECK: store i32 [[RESULT]], i32* [[LOC:%[a-z_0-9]+]]
+// CHECK: [[RESULT2:%[a-z_0-9]+]] = load i32, i32* [[LOC]]
+// CHECK: ret i32 [[RESULT2]]
+int get_int(va_list *args) {
+  return va_arg(*args, int);
+}
+
+struct Foo {
+  int x;
+};
+
+struct Foo dest;
+
+// CHECK-LABEL: define void @get_struct
+// CHECK: [[RESULT:%[a-z_0-9]+]] = va_arg {{.*}}, %struct.Foo*{{$}}
+// CHECK: [[RESULT2:%[a-z_0-9]+]] = bitcast {{.*}} [[RESULT]] to i8*
+// CHECK: call void @llvm.memcpy{{.*}}@dest{{.*}}, i8* [[RESULT2]]
+void get_struct(va_list *args) {
+ dest = va_arg(*args, struct Foo);
+}
+
+enum E { Foo_one = 1 };
+
+enum E enum_dest;
+
+// CHECK-LABEL: define void @get_enum
+// CHECK: va_arg i8** {{.*}}, i32
+void get_enum(va_list *args) {
+  enum_dest = va_arg(*args, enum E);
+}