CodeGen: Emit string literal in constant address space

[yaxunl]

Summary

Some targets have constant address space (e.g. amdgcn). For them string literal should be
emitted in constant address space then casted to default address space.

[rjmccall]

The part about string literals looks fine, but:

lib/CodeGen/CGDecl.cpp

1375

I don't understand why a patch about string literals is changing auto variable emission.

[yaxunl]

lib/CodeGen/CGDecl.cpp

1375

It is a bug about alloca revealed by the lit test

char l_array[] = "l_array";

Loc contains the alloca casted to default address space, therefore it needs to be casted back to alloca address space here, otherwise CreateBitCast returns invalid bitcast. Unlike lifetime.start, memcpy does not require alloca address space, so an alternative fix is to let BP take address space of Loc.

[rjmccall]

lib/CodeGen/CGDecl.cpp

1375

Yeah, I think using the address space of Loc is more appropriate.

[yaxunl]

Revised by John's comments. Also refactored to extract common code.

[rjmccall]

lib/CodeGen/CGDecl.cpp

1373

CGM.Int8Ty exists.

lib/CodeGen/CodeGenModule.cpp

3077

This should really just be a static function in this file.

[yaxunl]

Revised by John's comments.

[rjmccall]

Thanks, LGTM.

lib/CodeGen/CGDecl.cpp

   // If this is a simple aggregate initialization, we can optimize it
   // in various ways.
   bool isVolatile = type.isVolatileQualified();
 
   llvm::Value *SizeVal =
     llvm::ConstantInt::get(IntPtrTy,
                            getContext().getTypeSizeInChars(type).getQuantity());
 
-  llvm::Type *BP = AllocaInt8PtrTy;
+  llvm::Type *BP = llvm::Type::getInt8Ty(CGM.getLLVMContext())
+                       ->getPointerTo(Loc.getAddressSpace());
   if (Loc.getType() != BP)
     Loc = Builder.CreateBitCast(Loc, BP);
 
   // If the initializer is all or mostly zeros, codegen with memset then do
   // a few stores afterward.
   if (shouldUseMemSetPlusStoresToInitialize(constant,
                 CGM.getDataLayout().getTypeAllocSize(constant->getType()))) {
     Builder.CreateMemSet(Loc, llvm::ConstantInt::get(Int8Ty, 0), SizeVal,
                          isVolatile);
     // Zero and undef don't require a stores.
     if (!constant->isNullValue() && !isa<llvm::UndefValue>(constant)) {
       Loc = Builder.CreateBitCast(Loc,
         constant->getType()->getPointerTo(Loc.getAddressSpace()));
       emitStoresForInitAfterMemset(constant, Loc.getPointer(),
                                    isVolatile, Builder);
     }
   } else {
     // Otherwise, create a temporary global with the initializer then
     // memcpy from the global to the alloca.
     std::string Name = getStaticDeclName(CGM, D);
-    unsigned AS = 0;
-    if (getLangOpts().OpenCL) {
-      AS = CGM.getContext().getTargetAddressSpace(LangAS::opencl_constant);
-      BP = llvm::PointerType::getInt8PtrTy(getLLVMContext(), AS);
-    }
+    unsigned AS = CGM.getContext().getTargetAddressSpace(
+        CGM.getStringLiteralAddressSpace());
+    BP = llvm::PointerType::getInt8PtrTy(getLLVMContext(), AS);
+
     llvm::GlobalVariable *GV =
       new llvm::GlobalVariable(CGM.getModule(), constant->getType(), true,
                                llvm::GlobalValue::PrivateLinkage,
                                constant, Name, nullptr,
                                llvm::GlobalValue::NotThreadLocal, AS);
     GV->setAlignment(Loc.getAlignment().getQuantity());
     GV->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
 

lib/CodeGen/CodeGenModule.h

public:
   /// would be. If IsForDefinition is true, it is guaranteed that an actual
   /// global with type Ty will be returned, not conversion of a variable with
   /// the same mangled name but some other type.
   llvm::Constant *GetAddrOfGlobalVar(const VarDecl *D,
                                      llvm::Type *Ty = nullptr,
                                      ForDefinition_t IsForDefinition
                                        = NotForDefinition);
 
+  /// Return the AST address space of string literal, which is used to emit
+  /// the string literal as global variable in LLVM IR.
+  /// Note: This is not necessarily the address space of the string literal
+  /// in AST. For address space agnostic language, e.g. C++, string literal
+  /// in AST is always in default address space.
+  LangAS getStringLiteralAddressSpace() const;
+
+  /// Cast the string literal global variable to default address space when
+  /// necessary.
+  llvm::Constant *
+  castStringLiteralToDefaultAddressSpace(llvm::GlobalVariable *GV);
+
   /// Return the address of the given function. If Ty is non-null, then this
   /// function will use the specified type if it has to create it.
   llvm::Constant *GetAddrOfFunction(GlobalDecl GD, llvm::Type *Ty = nullptr,
                                     bool ForVTable = false,
                                     bool DontDefer = false,
                                     ForDefinition_t IsForDefinition
                                       = NotForDefinition);
 

lib/CodeGen/CodeGenModule.cpp

       return LangAS::cuda_shared;
     else
       return LangAS::cuda_device;
   }
 
   return getTargetCodeGenInfo().getGlobalVarAddressSpace(*this, D);
 }
 
+LangAS CodeGenModule::getStringLiteralAddressSpace() const {
+  // OpenCL v1.2 s6.5.3: a string literal is in the constant address space.
+  if (LangOpts.OpenCL)
+    return LangAS::opencl_constant;
+  if (auto AS = getTarget().getConstantAddressSpace())
+    return AS.getValue();
+  return LangAS::Default;
+}
+
+// In address space agnostic languages, string literals are in default address
+// space in AST. However, certain targets (e.g. amdgcn) request them to be
+// emitted in constant address space in LLVM IR. To be consistent with other
+// parts of AST, string literal global variables in constant address space
+// need to be casted to default address space before being put into address
+// map and referenced by other part of CodeGen.
+// In OpenCL, string literals are in constant address space in AST, therefore
+// they should not be casted to default address space.
+llvm::Constant *CodeGenModule::castStringLiteralToDefaultAddressSpace(
+    llvm::GlobalVariable *GV) {
+  llvm::Constant *Cast = GV;
+  if (!LangOpts.OpenCL) {
+    if (auto AS = getTarget().getConstantAddressSpace()) {
+      if (AS != LangAS::Default)
+        Cast = getTargetCodeGenInfo().performAddrSpaceCast(
+            *this, GV, AS.getValue(), LangAS::Default,
+            GV->getValueType()->getPointerTo(
+                getContext().getTargetAddressSpace(LangAS::Default)));
+    }
+  }
+  return Cast;
+}
+
 template<typename SomeDecl>
 void CodeGenModule::MaybeHandleStaticInExternC(const SomeDecl *D,
                                                llvm::GlobalValue *GV) {
   if (!getLangOpts().CPlusPlus)
     return;
 
   // Must have 'used' attribute, or else inline assembly can't rely on
   // the name existing.

   Elements.resize(NumElements);
   return llvm::ConstantDataArray::get(VMContext, Elements);
 }
 
 static llvm::GlobalVariable *
 GenerateStringLiteral(llvm::Constant *C, llvm::GlobalValue::LinkageTypes LT,
                       CodeGenModule &CGM, StringRef GlobalName,
                       CharUnits Alignment) {
-  // OpenCL v1.2 s6.5.3: a string literal is in the constant address space.
-  unsigned AddrSpace = 0;
-  if (CGM.getLangOpts().OpenCL)
-    AddrSpace = CGM.getContext().getTargetAddressSpace(LangAS::opencl_constant);
+  unsigned AddrSpace = CGM.getContext().getTargetAddressSpace(
+      CGM.getStringLiteralAddressSpace());
 
   llvm::Module &M = CGM.getModule();
   // Create a global variable for this string
   auto *GV = new llvm::GlobalVariable(
       M, C->getType(), !CGM.getLangOpts().WritableStrings, LT, C, GlobalName,
       nullptr, llvm::GlobalVariable::NotThreadLocal, AddrSpace);
   GV->setAlignment(Alignment.getQuantity());
   GV->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);

   }
 
   auto GV = GenerateStringLiteral(C, LT, *this, GlobalVariableName, Alignment);
   if (Entry)
     *Entry = GV;
 
   SanitizerMD->reportGlobalToASan(GV, S->getStrTokenLoc(0), "<string literal>",
                                   QualType());
-  return ConstantAddress(GV, Alignment);
+
+  return ConstantAddress(castStringLiteralToDefaultAddressSpace(GV), Alignment);
 }
 
 /// GetAddrOfConstantStringFromObjCEncode - Return a pointer to a constant
 /// array for the given ObjCEncodeExpr node.
 ConstantAddress
 CodeGenModule::GetAddrOfConstantStringFromObjCEncode(const ObjCEncodeExpr *E) {
   std::string Str;
   getContext().getObjCEncodingForType(E->getEncodedType(), Str);

   // Get the default prefix if a name wasn't specified.
   if (!GlobalName)
     GlobalName = ".str";
   // Create a global variable for this.
   auto GV = GenerateStringLiteral(C, llvm::GlobalValue::PrivateLinkage, *this,
                                   GlobalName, Alignment);
   if (Entry)
     *Entry = GV;
-  return ConstantAddress(GV, Alignment);
+
+  return ConstantAddress(castStringLiteralToDefaultAddressSpace(GV), Alignment);
 }
 
 ConstantAddress CodeGenModule::GetAddrOfGlobalTemporary(
     const MaterializeTemporaryExpr *E, const Expr *Init) {
   assert((E->getStorageDuration() == SD_Static ||
           E->getStorageDuration() == SD_Thread) && "not a global temporary");
   const auto *VD = cast<VarDecl>(E->getExtendingDecl());
 

test/CodeGenCXX/amdgcn-string-literal.cpp

+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -emit-llvm %s -o - | FileCheck %s
+
+// CHECK: @.str = private unnamed_addr addrspace(4) constant [6 x i8] c"g_str\00", align 1
+// CHECK: @g_str = addrspace(1) global i8* addrspacecast (i8 addrspace(4)* getelementptr inbounds ([6 x i8], [6 x i8] addrspace(4)* @.str, i32 0, i32 0) to i8*), align 8
+// CHECK: @g_array = addrspace(1) global [8 x i8] c"g_array\00", align 1
+// CHECK: @.str.1 = private unnamed_addr addrspace(4) constant [6 x i8] c"l_str\00", align 1
+// CHECK: @_ZZ1fvE7l_array = private unnamed_addr addrspace(4) constant [8 x i8] c"l_array\00", align 1
+
+const char* g_str = "g_str";
+char g_array[] = "g_array";
+
+void g(const char* p);
+
+// CHECK-LABEL: define void @_Z1fv()
+void f() {
+  const char* l_str = "l_str";
+  
+  // CHECK: call void @llvm.memcpy.p0i8.p4i8.i64
+  char l_array[] = "l_array";
+
+  g(g_str);
+  g(g_array);
+  g(l_str);
+  g(l_array);
+
+  const char* p = g_str;
+  g(p);
+}