commit 4dc34ebf2a0716bf77ba110dab6777a3fc4397dd
author Peter Collingbourne <peter@pcc.me.uk> Sun May 20 21:08:35 2012 +0000
committer Peter Collingbourne <peter@pcc.me.uk> Sun May 20 21:08:35 2012 +0000
tree 3024562d1a661ae95d8f596d2182c63e5b9592ef
parent e5f8372b5dccf5894f0263d018599c14d8f31273

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

lib/CodeGen/CodeGenModule.cpp

diff --git a/lib/CodeGen/CodeGenModule.cpp b/lib/CodeGen/CodeGenModule.cpp
index f2dda5d..0da493f 100644
--- a/lib/CodeGen/CodeGenModule.cpp
+++ b/lib/CodeGen/CodeGenModule.cpp
@@ -1172,11 +1172,12 @@
     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 @@
     GV->setThreadLocal(D->isThreadSpecified());
   }
 
-  return GV;
+  if (AddrSpace != Ty->getAddressSpace())
+    return llvm::ConstantExpr::getBitCast(GV, Ty);
+  else
+    return GV;
 }
 
 
@@ -1487,6 +1491,20 @@
   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 @@
   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

diff --git a/lib/CodeGen/CodeGenModule.h b/lib/CodeGen/CodeGenModule.h
index bde03a7..8f3bd78 100644
--- a/lib/CodeGen/CodeGenModule.h
+++ b/lib/CodeGen/CodeGenModule.h
@@ -517,6 +517,12 @@
   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