diff --git a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
index 48d65ec3b4113146ac4e79be681d4e1fc06c221a..681249231e630d09d0212a689ce61350fbc94b90 100644
--- a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
@@ -65,6 +65,7 @@ namespace {
private:
//Member variables
Module &M;
+ Module &KernelM;
BuildDFG &DFG;
DFNode * KernelLaunchNode;
@@ -88,6 +89,8 @@ namespace {
//Functions
+ std::string getKernelsModuleName(Module &M);
+ void writeKernelsModule();
void transformFunctionToVoid(Function* F);
void initRuntimeAPI();
void addIdxDimArgs(Function* F);
@@ -100,8 +103,44 @@ namespace {
void codeGen(DFLeafNode* N);
public:
+
// Constructor
- CodeGenTraversal(Module &_M, BuildDFG &_DFG) : M(_M), DFG(_DFG) { }
+ CodeGenTraversal(Module &_M, BuildDFG &_DFG) : M(_M), DFG(_DFG), KernelM(*CloneModule(&_M)) {
+
+ // Copying instead of creating new, in order to preserve required info (metadata)
+
+ // Remove functions, global variables and aliases
+ std::vector<GlobalVariable*> gvv = std::vector<GlobalVariable*>();
+ for (Module::global_iterator mi = KernelM.global_begin(),
+ me = KernelM.global_end(); (mi != me); ++mi) {
+ GlobalVariable* gv = &*mi;
+ gvv.push_back(gv);
+ }
+ for (std::vector<GlobalVariable*>::iterator vi = gvv.begin(); vi != gvv.end(); ++vi)
+ (*vi)->eraseFromParent();
+
+ std::vector<Function*> fv = std::vector<Function*>();
+ for (Module::iterator mi = KernelM.begin(),
+ me = KernelM.end(); (mi != me); ++mi) {
+ Function* f = &*mi;
+ fv.push_back(f);
+ }
+ for (std::vector<Function*>::iterator vi = fv.begin(); vi != fv.end(); ++vi)
+ (*vi)->eraseFromParent();
+
+ std::vector<GlobalAlias*> av = std::vector<GlobalAlias*>();
+ for (Module::alias_iterator mi = KernelM.alias_begin(),
+ me = KernelM.alias_end(); (mi != me); ++mi) {
+ GlobalAlias* a = &*mi;
+ av.push_back(a);
+ }
+ for (std::vector<GlobalAlias*>::iterator vi = av.begin(); vi != av.end(); ++vi)
+ (*vi)->eraseFromParent();
+
+ changeDataLayout(KernelM);
+ changeTargetTriple(KernelM);
+
+ }
virtual void visit(DFInternalNode* N) {
for(DFGraph::children_iterator i = N->getChildGraph()->begin(),
@@ -557,8 +596,8 @@ namespace {
ValueToValueMapTy VMap;
F_nvptx = CloneFunction(F, VMap, true);
- // Insert the cloned function into the module
- M.getFunctionList().push_back(F_nvptx);
+ // Insert the cloned function into the kernels module
+ KernelM.getFunctionList().push_back(F_nvptx);
DEBUG(errs() << *F_nvptx->getType());
DEBUG(errs() << *F_nvptx);
@@ -617,6 +656,7 @@ namespace {
ArgDFNode = Leaf_HandleToDFNodeMap[ArgII];
int numOfDim = ArgDFNode->getNumOfDim();
DEBUG(errs() << "\t Got node dimension : " << numOfDim << "\n");
+// IntegerType* IntTy = Type::getInt32Ty(KernelM.getContext());
IntegerType* IntTy = Type::getInt32Ty(getGlobalContext());
ConstantInt* numOfDimConstant = ConstantInt::getSigned(IntTy, (int64_t) numOfDim);
@@ -648,7 +688,7 @@ namespace {
// Argument of the function to be called
ConstantInt * DimConstant =
- ConstantInt::get(Type::getInt32Ty(getGlobalContext()), dim);
+ ConstantInt::get(Type::getInt32Ty(getGlobalContext()) /*KernelM.getContext()*/ , dim);
ArrayRef<Value *> Args(DimConstant);
// The following is to find which function to call
@@ -662,29 +702,29 @@ namespace {
// launch, so we need to specify a global id
FunctionType* FT =
- FunctionType::get(Type::getInt32Ty(getGlobalContext()),
- std::vector<Type*>(1, Type::getInt32Ty(getGlobalContext())),
+ FunctionType::get(Type::getInt32Ty(getGlobalContext() /*KernelM.getContext()*/),
+ std::vector<Type*>(1, Type::getInt32Ty(getGlobalContext() /*KernelM.getContext()*/)),
false);
OpenCLFunction = cast<Function>
- (M.getOrInsertFunction(StringRef("get_global_id"), FT));
+ (KernelM.getOrInsertFunction(StringRef("get_global_id"), FT));
} else if (Leaf_HandleToDFNodeMap[ArgII] == Leaf_HandleToDFNodeMap[II]) {
// We are asking for this node's id with respect to its parent
// this is a local id call
FunctionType* FT =
- FunctionType::get(Type::getInt32Ty(getGlobalContext()),
- std::vector<Type*>(1, Type::getInt32Ty(getGlobalContext())),
+ FunctionType::get(Type::getInt32Ty(getGlobalContext() /*KernelM.getContext()*/),
+ std::vector<Type*>(1, Type::getInt32Ty(getGlobalContext() /*KernelM.getContext()*/)),
false);
OpenCLFunction = cast<Function>
- (M.getOrInsertFunction(StringRef("get_local_id"), FT));
+ (KernelM.getOrInsertFunction(StringRef("get_local_id"), FT));
} else if (Leaf_HandleToDFNodeMap[ArgII] == N->getParent()) {
// We are asking for this node's parent's id with respect to its
// parent: this is a group id call
FunctionType* FT =
- FunctionType::get(Type::getInt32Ty(getGlobalContext()),
- std::vector<Type*>(1, Type::getInt32Ty(getGlobalContext())),
+ FunctionType::get(Type::getInt32Ty(getGlobalContext() /*KernelM.getContext()*/),
+ std::vector<Type*>(1, Type::getInt32Ty(getGlobalContext() /*KernelM.getContext()*/)),
false);
OpenCLFunction = cast<Function>
- (M.getOrInsertFunction(StringRef("get_group_id"), FT));
+ (KernelM.getOrInsertFunction(StringRef("get_group_id"), FT));
} else {
assert(false && "Unable to translate this intrinsic");
}
@@ -723,7 +763,7 @@ namespace {
// Argument of the function to be called
ConstantInt * DimConstant =
- ConstantInt::get(Type::getInt32Ty(getGlobalContext()), dim);
+ ConstantInt::get(Type::getInt32Ty(getGlobalContext() /*KernelM.getContext()*/), dim);
ArrayRef<Value *> Args(DimConstant);
// The following is to find which function to call
@@ -736,29 +776,29 @@ namespace {
// replicated. This indicates that the parent node is the kernel
// launch, so the instances are global_size (gridDim x blockDim)
FunctionType* FT =
- FunctionType::get(Type::getInt32Ty(getGlobalContext()),
- std::vector<Type*>(1, Type::getInt32Ty(getGlobalContext())),
+ FunctionType::get(Type::getInt32Ty(getGlobalContext() /*KernelM.getContext()*/),
+ std::vector<Type*>(1, Type::getInt32Ty(getGlobalContext() /*KernelM.getContext()*/)),
false);
OpenCLFunction = cast<Function>
- (M.getOrInsertFunction(StringRef("get_global_size"), FT));
+ (KernelM.getOrInsertFunction(StringRef("get_global_size"), FT));
} else if (Leaf_HandleToDFNodeMap[ArgII] == Leaf_HandleToDFNodeMap[II]) {
// We are asking for this node's instances
// this is a local size (block dim) call
FunctionType* FT =
- FunctionType::get(Type::getInt32Ty(getGlobalContext()),
- std::vector<Type*>(1, Type::getInt32Ty(getGlobalContext())),
+ FunctionType::get(Type::getInt32Ty(getGlobalContext() /*KernelM.getContext()*/),
+ std::vector<Type*>(1, Type::getInt32Ty(getGlobalContext() /*KernelM.getContext()*/)),
false);
OpenCLFunction = cast<Function>
- (M.getOrInsertFunction(StringRef("get_local_size"), FT));
+ (KernelM.getOrInsertFunction(StringRef("get_local_size"), FT));
} else if (Leaf_HandleToDFNodeMap[ArgII] == N->getParent()) {
// We are asking for this node's parent's instances
// this is a (global_size/local_size) (grid dim) call
FunctionType* FT =
- FunctionType::get(Type::getInt32Ty(getGlobalContext()),
- std::vector<Type*>(1, Type::getInt32Ty(getGlobalContext())),
+ FunctionType::get(Type::getInt32Ty(getGlobalContext() /*KernelM.getContext()*/),
+ std::vector<Type*>(1, Type::getInt32Ty(getGlobalContext() /*KernelM.getContext()*/)),
false);
OpenCLFunction = cast<Function>
- (M.getOrInsertFunction(StringRef("get_num_groups"), FT));
+ (KernelM.getOrInsertFunction(StringRef("get_num_groups"), FT));
} else {
assert(false && "Unable to translate this intrinsic");
}
@@ -815,6 +855,26 @@ namespace {
return true;
}
+ std::string CodeGenTraversal::getKernelsModuleName(Module &M) {
+ std::string mid = M.getModuleIdentifier();
+ return mid.append("_kernels.ll");
+ }
+
+ void CodeGenTraversal::writeKernelsModule() {
+
+ //TODO: Add metadata
+
+ char* ErrorMessage = NULL;
+ LLVMModuleRef KernelMRef = wrap(&KernelM);
+ LLVMPrintModuleToFile(KernelMRef,
+ getKernelsModuleName(M).c_str(),
+ &ErrorMessage);
+ if (ErrorMessage) {
+ LLVMDisposeMessage(ErrorMessage);
+ }
+ LLVMDisposeModule(KernelMRef);
+ }
+
void CodeGenTraversal::transformFunctionToVoid(Function* F) {
// FIXME: Maybe do that using the Node?