CUDA: add CodeGen support for global variable address spaces.

Because in CUDA types do not have associated address spaces,
globals are declared in their "native" address space, and accessed
by bitcasting the pointer to address space 0.  This relies on address
space 0 being a unified address space.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@157167 91177308-0d34-0410-b5e6-96231b3b80d8

include/clang/Basic/AddressSpaces.h

@@ -29,6 +29,10 @@ enum ID {
   opencl_local,
   opencl_constant,
 
+  cuda_device,
+  cuda_constant,
+  cuda_shared,
+
   Last,
   Count = Last-Offset
 };

lib/AST/ASTContext.cpp

@@ -206,7 +206,10 @@ static const LangAS::Map *getAddressSpaceMap(const TargetInfo &T,
     static const unsigned FakeAddrSpaceMap[] = {
       1, // opencl_global
       2, // opencl_local
-      3  // opencl_constant
+      3, // opencl_constant
+      4, // cuda_device
+      5, // cuda_constant
+      6  // cuda_shared
     };
     return &FakeAddrSpaceMap;
   } else {

lib/Basic/Targets.cpp

@@ -949,7 +949,10 @@ namespace {
   static const unsigned PTXAddrSpaceMap[] = {
     0,    // opencl_global
     4,    // opencl_local
-    1     // opencl_constant
+    1,    // opencl_constant
+    0,    // cuda_device
+    1,    // cuda_constant
+    4,    // cuda_shared
   };
   class PTXTargetInfo : public TargetInfo {
     static const char * const GCCRegNames[];
@@ -3384,7 +3387,10 @@ namespace {
   static const unsigned TCEOpenCLAddrSpaceMap[] = {
       3, // opencl_global
       4, // opencl_local
-      5  // opencl_constant
+      5, // opencl_constant
+      0, // cuda_device
+      0, // cuda_constant
+      0  // cuda_shared
   };
 
   class TCETargetInfo : public TargetInfo{

lib/CodeGen/CodeGenModule.cpp

@@ -1172,11 +1172,12 @@ CodeGenModule::GetOrCreateLLVMGlobal(StringRef MangledName,
     DeferredDecls.erase(DDI);
   }
 
+  unsigned AddrSpace = GetGlobalVarAddressSpace(D, Ty->getAddressSpace());
   llvm::GlobalVariable *GV =
     new llvm::GlobalVariable(getModule(), Ty->getElementType(), false,
                              llvm::GlobalValue::ExternalLinkage,
                              0, MangledName, 0,
-                             false, Ty->getAddressSpace());
+                             false, AddrSpace);
 
   // Handle things which are present even on external declarations.
   if (D) {
@@ -1202,7 +1203,10 @@ CodeGenModule::GetOrCreateLLVMGlobal(StringRef MangledName,
     GV->setThreadLocal(D->isThreadSpecified());
   }
 
-  return GV;
+  if (AddrSpace != Ty->getAddressSpace())
+    return llvm::ConstantExpr::getBitCast(GV, Ty);
+  else
+    return GV;
 }
 
 
@@ -1487,6 +1491,20 @@ CodeGenModule::MaybeEmitGlobalStdInitializerListInitializer(const VarDecl *D,
   return llvmInit;
 }
 
+unsigned CodeGenModule::GetGlobalVarAddressSpace(const VarDecl *D,
+                                                 unsigned AddrSpace) {
+  if (LangOpts.CUDA && CodeGenOpts.CUDAIsDevice) {
+    if (D->hasAttr<CUDAConstantAttr>())
+      AddrSpace = getContext().getTargetAddressSpace(LangAS::cuda_constant);
+    else if (D->hasAttr<CUDASharedAttr>())
+      AddrSpace = getContext().getTargetAddressSpace(LangAS::cuda_shared);
+    else
+      AddrSpace = getContext().getTargetAddressSpace(LangAS::cuda_device);
+  }
+
+  return AddrSpace;
+}
+
 void CodeGenModule::EmitGlobalVarDefinition(const VarDecl *D) {
   llvm::Constant *Init = 0;
   QualType ASTTy = D->getType();
@@ -1566,7 +1584,7 @@ void CodeGenModule::EmitGlobalVarDefinition(const VarDecl *D) {
   if (GV == 0 ||
       GV->getType()->getElementType() != InitType ||
       GV->getType()->getAddressSpace() !=
-        getContext().getTargetAddressSpace(ASTTy)) {
+       GetGlobalVarAddressSpace(D, getContext().getTargetAddressSpace(ASTTy))) {
 
     // Move the old entry aside so that we'll create a new one.
     Entry->setName(StringRef());

lib/CodeGen/CodeGenModule.h

@@ -517,6 +517,12 @@ public:
   CreateOrReplaceCXXRuntimeVariable(StringRef Name, llvm::Type *Ty,
                                     llvm::GlobalValue::LinkageTypes Linkage);
 
+  /// GetGlobalVarAddressSpace - Return the address space of the underlying
+  /// global variable for D, as determined by its declaration.  Normally this
+  /// is the same as the address space of D's type, but in CUDA, address spaces
+  /// are associated with declarations, not types.
+  unsigned GetGlobalVarAddressSpace(const VarDecl *D, unsigned AddrSpace);
+
   /// GetAddrOfGlobalVar - Return the llvm::Constant for the address of the
   /// given global variable.  If Ty is non-null and if the global doesn't exist,
   /// then it will be greated with the specified type instead of whatever the

test/CodeGenCUDA/address-spaces.cu

@@ -0,0 +1,24 @@
+// RUN: %clang_cc1 -emit-llvm %s -o - -fcuda-is-device -triple ptx32-unknown-unknown | FileCheck %s
+
+#include "../SemaCUDA/cuda.h"
+
+// CHECK: @i = global
+__device__ int i;
+
+// CHECK: @j = addrspace(1) global
+__constant__ int j;
+
+// CHECK: @k = addrspace(4) global
+__shared__ int k;
+
+__device__ void foo() {
+  // CHECK: load i32* @i
+  i++;
+
+  // CHECK: load i32* bitcast (i32 addrspace(1)* @j to i32*)
+  j++;
+
+  // CHECK: load i32* bitcast (i32 addrspace(4)* @k to i32*)
+  k++;
+}
+