diff --git a/llvm/lib/Transforms/ClearDFG/ClearDFG.cpp b/llvm/lib/Transforms/ClearDFG/ClearDFG.cpp
index 6d04074b96b1d5043f25c58f6be6ef018cc1edb7..07d7667fa044732293833382a9f57b05c8297509 100644
--- a/llvm/lib/Transforms/ClearDFG/ClearDFG.cpp
+++ b/llvm/lib/Transforms/ClearDFG/ClearDFG.cpp
@@ -74,6 +74,7 @@ public:
// Generate code for this internal node now. This way all the cloned
// functions for children exist.
deleteNode(N);
+ DEBUG(errs() << "\tDone - " << "\n");
//errs() << "DONE: Generating Code for Node (I) - " << N->getFuncPointer()->getName() << "\n";
}
diff --git a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
index 6926e6cfed65679b455152cffc311ae946dab314..59bdb622b93ef45a4d2b01ed415ff5e86b524470 100644
--- a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
@@ -147,6 +147,7 @@ private:
Constant* llvm_visc_ocl_clearContext;
Constant* llvm_visc_ocl_argument_scalar;
Constant* llvm_visc_ocl_argument_ptr;
+ Constant* llvm_visc_ocl_output_ptr;
Constant* llvm_visc_ocl_free;
Constant* llvm_visc_ocl_getOutput;
Constant* llvm_visc_ocl_executeNode;
@@ -245,6 +246,7 @@ void CGT_NVPTX::initRuntimeAPI() {
DECLARE(llvm_visc_ocl_clearContext);
DECLARE(llvm_visc_ocl_argument_scalar);
DECLARE(llvm_visc_ocl_argument_ptr);
+ DECLARE(llvm_visc_ocl_output_ptr);
DECLARE(llvm_visc_ocl_free);
DECLARE(llvm_visc_ocl_getOutput);
DECLARE(llvm_visc_ocl_executeNode);
@@ -332,7 +334,7 @@ void CGT_NVPTX::insertRuntimeCalls(DFInternalNode* N, Kernel* K, const Twine& Fi
// have those arguments)
// Add Index and Dim arguments except for the root node
- if(!N->isRoot())
+ if(!N->isRoot() && !N->getParent()->isChildGraphStreaming())
addIdxDimArgs(F_X86);
/* TODO: Use this code to verufy if this is a good pattern for PTX kernel
@@ -480,24 +482,23 @@ void CGT_NVPTX::insertRuntimeCalls(DFInternalNode* N, Kernel* K, const Twine& Fi
// Set output if struct is not an empty struct
StructType* OutputTy = K->KernelLeafNode->getOutputType();
- Value *outputSize, *d_Output;
+ std::vector<Value*> d_Outputs;
if(!OutputTy->isEmptyTy()) {
switchToTimer(visc_TimerID_COPY_PTR, RI);
// Not an empty struct
- unsigned outputIndex = KF->getFunctionType()->getNumParams();
- outputSize = ConstantExpr::getSizeOf(OutputTy);
- Value* setOutputArgs[] = {GraphID,
- Constant::getNullValue(Type::getInt8PtrTy(M.getContext())),
- ConstantInt::get(Type::getInt32Ty(M.getContext()),outputIndex),
- ConstantExpr::getSizeOf(OutputTy),
- False,
- True
- };
-
- d_Output = CallInst::Create(llvm_visc_ocl_argument_ptr,
- ArrayRef<Value*>(setOutputArgs, 6),
+ // Iterate over all elements of the struct and put them in
+ for(unsigned i=0; i < OutputTy->getNumElements(); i++) {
+ unsigned outputIndex = KF->getFunctionType()->getNumParams()+i;
+ Value* setOutputArgs[] = {GraphID,
+ ConstantInt::get(Type::getInt32Ty(M.getContext()),outputIndex),
+ ConstantExpr::getSizeOf(OutputTy->getElementType(i))};
+
+ CallInst* d_Output = CallInst::Create(llvm_visc_ocl_output_ptr,
+ ArrayRef<Value*>(setOutputArgs, 3),
"d_output."+KF->getName(),
RI);
+ d_Outputs.push_back(d_Output);
+ }
}
// Enqueue kernel
@@ -529,22 +530,30 @@ void CGT_NVPTX::insertRuntimeCalls(DFInternalNode* N, Kernel* K, const Twine& Fi
switchToTimer(visc_TimerID_READ_OUTPUT, RI);
// Read Output Struct if not empty
if(!OutputTy->isEmptyTy()) {
- Value* GetOutputArgs[] = {GraphID,
+ std::vector<Value*>h_Outputs;
+ Value* KernelOutput = UndefValue::get(OutputTy);
+ for(unsigned i=0; i < OutputTy->getNumElements(); i++) {
+ Value* GetOutputArgs[] = {GraphID,
Constant::getNullValue(Type::getInt8PtrTy(M.getContext())),
- d_Output,
- outputSize
+ d_Outputs[i],
+ ConstantExpr::getSizeOf(OutputTy->getElementType(i))
};
- CallInst* h_Output = CallInst::Create(llvm_visc_ocl_getOutput,
+ CallInst* h_Output = CallInst::Create(llvm_visc_ocl_getOutput,
ArrayRef<Value*>(GetOutputArgs, 4),
"h_output."+KF->getName()+".addr",
RI);
- // Read each device pointer listed in output struct
- // Load the output struct
- CastInst* BI = BitCastInst::CreatePointerCast(h_Output, KF->getReturnType()->getPointerTo(), "output.ptr", RI);
- Value* KernelOutput = new LoadInst(BI, "output."+KF->getName(), RI);
+ // Read each device pointer listed in output struct
+ // Load the output struct
+ CastInst* BI = BitCastInst::CreatePointerCast(h_Output,
+ OutputTy->getElementType(i)->getPointerTo(), "output.ptr", RI);
+
+ Value* OutputElement = new LoadInst(BI, "output."+KF->getName(), RI);
+ KernelOutput = InsertValueInst::Create(KernelOutput, OutputElement, ArrayRef<unsigned>(i),
+ KF->getName()+"output", RI);
+ }
OutputMap[K->KernelLeafNode] = KernelOutput;
}
-
+
// Read all the pointer arguments which had side effects i.e., had out
// attribute
DEBUG(errs() << "Output Pointers : " << OutputPointers.size() << "\n");
@@ -642,7 +651,14 @@ void CGT_NVPTX::insertRuntimeCalls(DFInternalNode* N, Kernel* K, const Twine& Fi
// Right now, only targeting the one level case. In general, device functions
// can return values so we don't need to change them
void CGT_NVPTX::codeGen(DFInternalNode* N) {
-
+ errs () << "Inside node: " << N->getFuncPointer()->getName() << "\n";
+ if(KernelLaunchNode == NULL)
+ errs () << "No kernel launch node\n";
+ else {
+ errs() << "KernelLaunchNode is not null: "<< KernelLaunchNode<<"\n";
+ errs () << "KernelLaunchNode: " << KernelLaunchNode->getFuncPointer()->getName() << "\n";
+ }
+
if (!KernelLaunchNode) {
DEBUG(errs() << "No code generated (host code for kernel launch complete).\n");
return;
@@ -721,6 +737,13 @@ void CGT_NVPTX::codeGen(DFLeafNode* N) {
DEBUG(errs() << "Skipping dummy node\n");
return;
}
+
+ // Generate code only if it has the right hint
+ if(!checkPreferredTarget(N, visc::GPU_TARGET)) {
+ errs() << "Skipping node: "<< N->getFuncPointer()->getName() << "\n";
+ return;
+ }
+
// Checking which node is the kernel launch
DFNode* PNode = N->getParent();
@@ -1185,23 +1208,16 @@ void CGT_NVPTX::transformFunctionToVoid(Function* F) {
else {
// The struct has return values, thus needs to be converted to parameter
- int initialNumParams = F->arg_size();
-
- Type* ArgType = FRetTy->getPointerTo(GENERIC_ADDRSPACE);
- new Argument(ArgType, "ret_struct_ptr", F);
- DEBUG(errs() << "\tCreated parameter\n");
-
- // Find where the new parameter is in the header
- Function::arg_iterator ai, ae;
- int check = 0;
- for (ai = F->arg_begin(), ae = F->arg_end();
- ai != ae; ++ai) {
- if (ai->getName().equals("ret_struct_ptr")) break;
- check++;
+ // Iterate over all element types of return struct and add arguments to the
+ // function
+ std::vector<Argument*> Args;
+ for (unsigned i=0; i<FRetTy->getNumElements(); i++) {
+ Argument* RetArg = new Argument(FRetTy->getElementType(i)->getPointerTo(), "ret_arg", F);
+ Args.push_back(RetArg);
+ DEBUG(errs() << "\tCreated parameter: " << *RetArg << "\n");
}
- // DEBUG(errs() << "\tcheck = " << check << "\tinitialNumParams = " << initialNumParams << "\n");
- assert(check == initialNumParams);
+ Function::arg_iterator ai, ae;
DEBUG(errs() << "\tReplacing Return statements\n");
// Replace return statements with extractValue and store instructions
@@ -1209,11 +1225,15 @@ void CGT_NVPTX::transformFunctionToVoid(Function* F) {
rie = RItoRemove.end(); rii != rie; ++rii) {
ReturnInst* RI = (*rii);
Value* RetVal = RI->getReturnValue();
+ for(unsigned i = 0; i < Args.size(); i++) {
+ ExtractValueInst* EI = ExtractValueInst::Create(RetVal, ArrayRef<unsigned>(i),
+ Args[i]->getName()+".val", RI);
+ new StoreInst(EI, Args[i], RI);
+ }
// assert(RetVal && "Return value should not be null at this point");
// StructType* RetType = cast<StructType>(RetVal->getType());
// assert(RetType && "Return type is not a struct");
- new StoreInst(RetVal, &(*ai), RI);
ReturnInst::Create((F->getContext()), 0, RI);
RI->eraseFromParent();
@@ -1237,7 +1257,6 @@ void CGT_NVPTX::transformFunctionToVoid(Function* F) {
// Change the function type
F->mutateType(PTy);
-
}
/******************************************************************************
diff --git a/llvm/lib/Transforms/DFG2LLVM_SPIR/DFG2LLVM_SPIR.cpp b/llvm/lib/Transforms/DFG2LLVM_SPIR/DFG2LLVM_SPIR.cpp
index a05fe6080262ccf4657650c6169900ac4e35d6f9..51a40baa94bae48dc4fe4b5994d3bd43f07e8e4c 100644
--- a/llvm/lib/Transforms/DFG2LLVM_SPIR/DFG2LLVM_SPIR.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_SPIR/DFG2LLVM_SPIR.cpp
@@ -151,6 +151,7 @@ private:
Constant* llvm_visc_ocl_clearContext;
Constant* llvm_visc_ocl_argument_scalar;
Constant* llvm_visc_ocl_argument_ptr;
+ Constant* llvm_visc_ocl_output_ptr;
Constant* llvm_visc_ocl_free;
Constant* llvm_visc_ocl_getOutput;
Constant* llvm_visc_ocl_executeNode;
@@ -177,6 +178,7 @@ public:
// Constructor
CGT_SPIR(Module &_M, BuildDFG &_DFG) : CodeGenTraversal(_M, _DFG), KernelM(*CloneModule(&_M)) {
+ KernelLaunchNode = NULL;
init();
initRuntimeAPI();
@@ -250,6 +252,7 @@ void CGT_SPIR::initRuntimeAPI() {
DECLARE(llvm_visc_ocl_clearContext);
DECLARE(llvm_visc_ocl_argument_scalar);
DECLARE(llvm_visc_ocl_argument_ptr);
+ DECLARE(llvm_visc_ocl_output_ptr);
DECLARE(llvm_visc_ocl_free);
DECLARE(llvm_visc_ocl_getOutput);
DECLARE(llvm_visc_ocl_executeNode);
@@ -337,7 +340,7 @@ void CGT_SPIR::insertRuntimeCalls(DFInternalNode* N, Kernel* K, const Twine& Fil
// have those arguments)
// Add Index and Dim arguments except for the root node
- if(!N->isRoot())
+ if(!N->isRoot() && !N->getParent()->isChildGraphStreaming())
addIdxDimArgs(F_X86);
/* TODO: Use this code to verufy if this is a good pattern for OCL kernel
@@ -485,24 +488,23 @@ void CGT_SPIR::insertRuntimeCalls(DFInternalNode* N, Kernel* K, const Twine& Fil
// Set output if struct is not an empty struct
StructType* OutputTy = K->KernelLeafNode->getOutputType();
- Value *outputSize, *d_Output;
+ std::vector<Value*> d_Outputs;
if(!OutputTy->isEmptyTy()) {
switchToTimer(visc_TimerID_COPY_PTR, RI);
// Not an empty struct
- unsigned outputIndex = KF->getFunctionType()->getNumParams();
- outputSize = ConstantExpr::getSizeOf(OutputTy);
- Value* setOutputArgs[] = {GraphID,
- Constant::getNullValue(Type::getInt8PtrTy(M.getContext())),
- ConstantInt::get(Type::getInt32Ty(M.getContext()),outputIndex),
- ConstantExpr::getSizeOf(OutputTy),
- False,
- True
- };
-
- d_Output = CallInst::Create(llvm_visc_ocl_argument_ptr,
- ArrayRef<Value*>(setOutputArgs, 6),
+ // Iterate over all elements of the struct and put them in
+ for(unsigned i=0; i < OutputTy->getNumElements(); i++) {
+ unsigned outputIndex = KF->getFunctionType()->getNumParams()+i;
+ Value* setOutputArgs[] = {GraphID,
+ ConstantInt::get(Type::getInt32Ty(M.getContext()),outputIndex),
+ ConstantExpr::getSizeOf(OutputTy->getElementType(i))};
+
+ CallInst* d_Output = CallInst::Create(llvm_visc_ocl_output_ptr,
+ ArrayRef<Value*>(setOutputArgs, 3),
"d_output."+KF->getName(),
RI);
+ d_Outputs.push_back(d_Output);
+ }
}
// Enqueue kernel
@@ -534,19 +536,27 @@ void CGT_SPIR::insertRuntimeCalls(DFInternalNode* N, Kernel* K, const Twine& Fil
switchToTimer(visc_TimerID_READ_OUTPUT, RI);
// Read Output Struct if not empty
if(!OutputTy->isEmptyTy()) {
- Value* GetOutputArgs[] = {GraphID,
+ std::vector<Value*>h_Outputs;
+ Value* KernelOutput = UndefValue::get(OutputTy);
+ for(unsigned i=0; i < OutputTy->getNumElements(); i++) {
+ Value* GetOutputArgs[] = {GraphID,
Constant::getNullValue(Type::getInt8PtrTy(M.getContext())),
- d_Output,
- outputSize
+ d_Outputs[i],
+ ConstantExpr::getSizeOf(OutputTy->getElementType(i))
};
- CallInst* h_Output = CallInst::Create(llvm_visc_ocl_getOutput,
+ CallInst* h_Output = CallInst::Create(llvm_visc_ocl_getOutput,
ArrayRef<Value*>(GetOutputArgs, 4),
"h_output."+KF->getName()+".addr",
RI);
- // Read each device pointer listed in output struct
- // Load the output struct
- CastInst* BI = BitCastInst::CreatePointerCast(h_Output, KF->getReturnType()->getPointerTo(), "output.ptr", RI);
- Value* KernelOutput = new LoadInst(BI, "output."+KF->getName(), RI);
+ // Read each device pointer listed in output struct
+ // Load the output struct
+ CastInst* BI = BitCastInst::CreatePointerCast(h_Output,
+ OutputTy->getElementType(i)->getPointerTo(), "output.ptr", RI);
+
+ Value* OutputElement = new LoadInst(BI, "output."+KF->getName(), RI);
+ KernelOutput = InsertValueInst::Create(KernelOutput, OutputElement, ArrayRef<unsigned>(i),
+ KF->getName()+"output", RI);
+ }
OutputMap[K->KernelLeafNode] = KernelOutput;
}
@@ -647,6 +657,14 @@ void CGT_SPIR::insertRuntimeCalls(DFInternalNode* N, Kernel* K, const Twine& Fil
// Right now, only targeting the one level case. In general, device functions
// can return values so we don't need to change them
void CGT_SPIR::codeGen(DFInternalNode* N) {
+ errs () << "Inside node: " << N->getFuncPointer()->getName() << "\n";
+ if(KernelLaunchNode == NULL)
+ errs () << "No kernel launch node\n";
+ else {
+ errs() << "KernelLaunchNode is not null: "<< KernelLaunchNode<<"\n";
+ errs () << "KernelLaunchNode: " << KernelLaunchNode->getFuncPointer()->getName() << "\n";
+ }
+
if (!KernelLaunchNode) {
DEBUG(errs() << "No code generated (host code for kernel launch complete).\n");
@@ -720,6 +738,31 @@ void CGT_SPIR::codeGen(DFInternalNode* N) {
}
+//static bool checkPreferredTarget(DFNode* N, visc::Target T) {
+ //Function* F = N->getFuncPointer();
+ //Module* M = F->getParent();
+ //NamedMDNode* HintNode;
+ //switch (T) {
+ //case visc::GPU_TARGET:
+ //HintNode = M->getOrInsertNamedMetadata("visc_hint_gpu");
+ //break;
+ //case visc::SPIR_TARGET:
+ //HintNode = M->getOrInsertNamedMetadata("visc_hint_spir");
+ //break;
+ //case visc::CPU_TARGET:
+ //HintNode = M->getOrInsertNamedMetadata("visc_hint_cpu");
+ //break;
+ //default:
+ //llvm_unreachable("Target Not supported yet!");
+ //}
+ //for (unsigned i = 0; i < HintNode->getNumOperands(); i++) {
+ //MDNode* MetaNode = HintNode->getOperand(i);
+ //if(F == MetaNode->getOperand(0))
+ //return true;
+ //}
+ //return false;
+//}
+
void CGT_SPIR::codeGen(DFLeafNode* N) {
// Skip code generation if it is a dummy node
@@ -728,6 +771,12 @@ void CGT_SPIR::codeGen(DFLeafNode* N) {
return;
}
+ // Generate code only if it has the right hint
+ if(!checkPreferredTarget(N, visc::SPIR_TARGET)) {
+ errs() << "Skipping node: "<< N->getFuncPointer()->getName() << "\n";
+ return;
+ }
+
// Checking which node is the kernel launch
DFNode* PNode = N->getParent();
int pLevel = PNode->getLevel();
@@ -739,6 +788,7 @@ void CGT_SPIR::codeGen(DFLeafNode* N) {
// (2) Parent does not have multiple instances
if (!pLevel || !pReplFactor) {
KernelLaunchNode = PNode;
+ errs() << "Setting Kernel Launch Node\n";
kernel = new Kernel(NULL,
N,
N->getInArgMap(),
@@ -1203,23 +1253,16 @@ void CGT_SPIR::transformFunctionToVoid(Function* F) {
else {
// The struct has return values, thus needs to be converted to parameter
- int initialNumParams = F->arg_size();
-
- Type* ArgType = FRetTy->getPointerTo(GENERIC_ADDRSPACE);
- new Argument(ArgType, "ret_struct_ptr", F);
- DEBUG(errs() << "\tCreated parameter\n");
-
- // Find where the new parameter is in the header
- Function::arg_iterator ai, ae;
- int check = 0;
- for (ai = F->arg_begin(), ae = F->arg_end();
- ai != ae; ++ai) {
- if (ai->getName().equals("ret_struct_ptr")) break;
- check++;
+ // Iterate over all element types of return struct and add arguments to the
+ // function
+ std::vector<Argument*> Args;
+ for (unsigned i=0; i<FRetTy->getNumElements(); i++) {
+ Argument* RetArg = new Argument(FRetTy->getElementType(i)->getPointerTo(), "ret_arg", F);
+ Args.push_back(RetArg);
+ DEBUG(errs() << "\tCreated parameter: " << *RetArg << "\n");
}
- // DEBUG(errs() << "\tcheck = " << check << "\tinitialNumParams = " << initialNumParams << "\n");
- assert(check == initialNumParams);
+ Function::arg_iterator ai, ae;
DEBUG(errs() << "\tReplacing Return statements\n");
// Replace return statements with extractValue and store instructions
@@ -1227,11 +1270,15 @@ void CGT_SPIR::transformFunctionToVoid(Function* F) {
rie = RItoRemove.end(); rii != rie; ++rii) {
ReturnInst* RI = (*rii);
Value* RetVal = RI->getReturnValue();
+ for(unsigned i = 0; i < Args.size(); i++) {
+ ExtractValueInst* EI = ExtractValueInst::Create(RetVal, ArrayRef<unsigned>(i),
+ Args[i]->getName()+".val", RI);
+ new StoreInst(EI, Args[i], RI);
+ }
// assert(RetVal && "Return value should not be null at this point");
// StructType* RetType = cast<StructType>(RetVal->getType());
// assert(RetType && "Return type is not a struct");
- new StoreInst(RetVal, &(*ai), RI);
ReturnInst::Create((F->getContext()), 0, RI);
RI->eraseFromParent();
diff --git a/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp b/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
index 98f654d3da748e1250a9a36ac18d8be6a3d1cb0e..5f9f6cb12fc28f64ca1b2ba72c54b27035cdf3f5 100644
--- a/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
@@ -837,7 +837,7 @@ Value* CGT_X86::getInValueAt(DFNode* Child, unsigned i, Function* ParentF_X86, I
// argument from argument list of this internal node
Value* inputVal;
if(SrcDF->isEntryNode()) {
- inputVal = getArgumentAt(ParentF_X86, i);
+ inputVal = getArgumentAt(ParentF_X86, E->getSourcePosition());
DEBUG(errs() << "Argument "<< i<< " = " << *inputVal << "\n");
}
else {
@@ -881,6 +881,12 @@ void CGT_X86::invokeChild_X86(DFNode* C, Function* F_X86,
for(unsigned j=0; j<6; j++)
Args.push_back(I32Zero);
+ errs() << "Function type: " << *CF_X86->getType() << "\n";
+ errs() << "Function type: " << *CF->getType() << "\n";
+ errs() << "Arguments: " << Args.size() << "\n";
+ for(unsigned i=0; i < Args.size(); i++)
+ errs() << *Args[i]->getType() << " ";
+ errs() << "\n";
// Call the F_X86 function associated with this node
CallInst* CI = CallInst::Create(CF_X86, Args,
CF_X86->getName()+"_output",
diff --git a/llvm/lib/Transforms/GenVISC/GenVISC.cpp b/llvm/lib/Transforms/GenVISC/GenVISC.cpp
index 8ad0602703745e0a2c8b78fbda3a6b410d493408..3aa7735ee36720574b589d578cdd303abbf4b194 100644
--- a/llvm/lib/Transforms/GenVISC/GenVISC.cpp
+++ b/llvm/lib/Transforms/GenVISC/GenVISC.cpp
@@ -120,7 +120,7 @@ static void addHint(Function* F, visc::Target T) {
HintNode = M->getOrInsertNamedMetadata("visc_hint_gpu");
}
else {
- HintNode = M->getOrInsertNamedMetadata("visc_hint_cpu");
+ HintNode = M->getOrInsertNamedMetadata("visc_hint_spir");
}
// Create a node for the function and add it to the hint node
@@ -826,11 +826,14 @@ bool GenVISC::runOnModule(Module &M) {
Function* LaunchF = Intrinsic::getDeclaration(&M, Intrinsic::visc_launch);
DEBUG(errs() << *LaunchF << "\n");
// Get i8* cast to function pointer
- Function* graphFunc = cast<Function>(CI->getArgOperand(0));
+ Function* graphFunc = cast<Function>(CI->getArgOperand(1));
Constant* F = ConstantExpr::getPointerCast(graphFunc, Type::getInt8PtrTy(Ctx));
+ ConstantInt* Op = cast<ConstantInt>(CI->getArgOperand(0));
+ Value* isStreaming = Op->isZero()? ConstantInt::getFalse(Ctx)
+ : ConstantInt::getTrue(Ctx);
- Value* LaunchArgs[] = {F, CI->getArgOperand(1), ConstantInt::getTrue(Ctx)};
+ Value* LaunchArgs[] = {F, CI->getArgOperand(2), isStreaming};
CallInst* LaunchInst = CallInst::Create(LaunchF,
ArrayRef<Value*>(LaunchArgs, 3),
"graphID", CI);
@@ -1038,9 +1041,12 @@ bool GenVISC::runOnModule(Module &M) {
}
// Erase the __visc__node calls
- DEBUG(errs() << "Erase Statements:\n");
+ DEBUG(errs() << "Erase " << toBeErased.size() << " Statements:\n");
for(auto I: toBeErased) {
DEBUG(errs() << *I << "\n");
+ }
+ for(auto I: toBeErased) {
+ DEBUG(errs() << "\tErasing " << *I << "\n");
I->eraseFromParent();
}
@@ -1125,7 +1131,7 @@ void GenVISC::genKernel(Function* KernelF, CallInst* CI, StructType* RetTy) {
Instruction *I = &(*i);
if(isVISCattributesCall(I)) {
handleVISCAttributes(KernelF, cast<CallInst>(I));
- I->eraseFromParent();
+ //I->eraseFromParent();
break;
}
}