From 7b1d63e4065823b7dbef3001dcd602142d22b50e Mon Sep 17 00:00:00 2001
From: Akash Kothari <akashk4@tyler.cs.illinois.edu>
Date: Mon, 13 Jan 2020 08:15:00 -0600
Subject: [PATCH] Deprecating __visc__node in GenVISC
---
hpvm/include/GenVISC/GenVISC.h | 3 -
hpvm/lib/Transforms/GenVISC/GenVISC.cpp | 743 +-----------------------
2 files changed, 1 insertion(+), 745 deletions(-)
diff --git a/hpvm/include/GenVISC/GenVISC.h b/hpvm/include/GenVISC/GenVISC.h
index ef1ca3469c..585af33953 100644
--- a/hpvm/include/GenVISC/GenVISC.h
+++ b/hpvm/include/GenVISC/GenVISC.h
@@ -44,9 +44,6 @@ public:
// Functions
virtual bool runOnModule(Module &M);
- void generateTest(CallInst* CI);
- Function* genKernel(Function* KernelFunction, StructType* RetTy);
- void genHost(CallInst*, Function*, unsigned, unsigned, unsigned, unsigned, StructType*);
};
} // End of namespace
diff --git a/hpvm/lib/Transforms/GenVISC/GenVISC.cpp b/hpvm/lib/Transforms/GenVISC/GenVISC.cpp
index e19f084992..9cce997f34 100644
--- a/hpvm/lib/Transforms/GenVISC/GenVISC.cpp
+++ b/hpvm/lib/Transforms/GenVISC/GenVISC.cpp
@@ -154,7 +154,6 @@ IS_VISC_CALL(cos)
IS_VISC_CALL(init)
-IS_VISC_CALL(node)
IS_VISC_CALL(cleanup)
IS_VISC_CALL(wait)
IS_VISC_CALL(trackMemory)
@@ -170,48 +169,6 @@ static unsigned getNumericValue(Value* V) {
return cast<ConstantInt>(V)->getZExtValue();
}
-
-
-// Add <numArgs> to the argument list of Function <F>. The names for these arguments
-// should be put in the string array <names>. Ideally the length of <names>
-// array should be numArgs. But, even when the length is not numArgs the
-// arguments would be added correctly. The names however would not be as
-// intuitive.
-static Function* addArgs(Function* F, unsigned numArgs, std::string names[]) {
- if(numArgs == 0) return F; // Return if no arguments are to be added.
-
- // Create the argument type list with added argument types
- std::vector<Type*> ArgTypes;
- for(Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
- ai != ae; ++ai) {
- ArgTypes.push_back(ai->getType());
- }
- // Adding new arguments to the function argument list, would not change the
- // function type. We need to change the type of this function to reflect the
- // added arguments
- for(unsigned i = 0; i < numArgs; ++i) {
-// ArgTypes.push_back(Type::getInt32Ty(F->getContext()));
- ArgTypes.push_back(Type::getInt64Ty(F->getContext()));
- }
- FunctionType* newFT = FunctionType::get(F->getReturnType(), ArgTypes, F->isVarArg());
-
- // Change the function type
- Function* newF = cloneFunction(F, newFT, false);
-
- // Add names to the extra arguments to the Function argument list
- unsigned numOldArgs = F->getFunctionType()->getNumParams();
- for(Function::arg_iterator ai = newF->arg_begin(), ae = newF->arg_end();
- ai != ae; ++ai) {
- if (ai->getArgNo() < numOldArgs)
- continue;
- ai->setName(names[(ai->getArgNo() - numOldArgs) % names->size()]);
- }
-
- replaceNodeFunctionInIR(*F->getParent(), F, newF);
- return newF;
-}
-
-
// Take the __visc__return instruction and generate code for combining the
// values being returned into a struct and returning it.
// The first operand is the number of returned values
@@ -255,130 +212,6 @@ static Value* genCodeForReturn(CallInst* CI) {
return IV;
}
-// The visc launch intrinsic requires all the input parameters to the kernel
-// function be placed in contiguous memory and pointer to that input be passed
-// as the second argument to the launch intrinsic. This generates code to bring
-// together all the input and dimension arguments in one packed struct
-// <InStruct>. First pack the arguments to the kernel function and then add the
-// dimension arguments depending on the hierarchy of DFG user wants to generate.
-static void marshallArguments(unsigned levels, unsigned numArgs, unsigned argOffset, unsigned numDims, unsigned dimOffset, Value* InStruct, CallInst* CI, Function* KernelF) {
- DEBUG(errs() << "Kernel Function = " << KernelF->getName() << "\n");
-
- // Get module context and i32 0 constant, as they would be frequently used in
- // this function.
- LLVMContext& Ctx = CI->getParent()->getContext();
- Constant* IntZero = ConstantInt::get(Type::getInt32Ty(Ctx), 0);
-
- // Find the arguments to be passed to kernel function and pack them in a
- // struct. Specifically first generate a GEP instruction to find the correct
- // memory location in InStruct and then generate Store instruction to store
- // the argument in that location.
- Function::arg_iterator ai = KernelF->arg_begin();
- Function::arg_iterator ae = KernelF->arg_end();
-
- for(unsigned i = 0; i < numArgs && ai != ae; i++, ai++) {
- Value* arg = CI->getArgOperand(i+argOffset);
- DEBUG(errs() << "Argument: " << ai->getName() << "\n");
- DEBUG(errs() << "Passing: " << *arg << "\n");
- // Create constant int (i)
- Constant* Int_i = ConstantInt::get(Type::getInt32Ty(Ctx), i);
- // Get Element pointer instruction
- Value* GEPIndices[] = { IntZero, Int_i };
- GetElementPtrInst* GEP = GetElementPtrInst::Create(nullptr, InStruct,
- ArrayRef<Value*>(GEPIndices, 2),
- InStruct->getName()+"."+ai->getName(),
- CI);
- // Store instruction
- if(GEP->getType()->getPointerElementType() != arg->getType()) {
- // Arguments type might not match with the kernel function definition
- // One reason might be because of default argument promotions, where all
- // arguments of type float are always promoted to double and types char,
- // short int are promoted to int.
- // LLVM 4.0 also promotes pointers to i8*. In case both are pointer types,
- // we just issue a warning and cast it to appropriate type
- if(arg->getType() == Type::getDoubleTy(Ctx)) {
- DEBUG(errs() << "Cast from " << *arg->getType() << " To " <<
- *GEP->getType()->getPointerElementType() << "\n");
- CastInst* CastI = BitCastInst::CreateFPCast(arg,
- GEP->getType()->getPointerElementType(), GEP->getName()+".cast",
- CI);
- new StoreInst(CastI, GEP, CI);
- } else if (arg->getType() == Type::getInt32Ty(Ctx)) {
- CastInst* CastI = BitCastInst::CreateIntegerCast(arg,
- GEP->getType()->getPointerElementType(), false,
- GEP->getName()+".cast", CI);
- new StoreInst(CastI, GEP, CI);
- } else if (arg->getType()->isPointerTy() && GEP->getType()->getPointerElementType()->isPointerTy()) {
- errs() << "WARNING: Argument type mismatch between kernel and __visc__node call. Forcing cast\n";
- CastInst* CastI = CastInst::CreatePointerCast(arg,
- GEP->getType()->getPointerElementType(), GEP->getName()+".cast",
- CI);
- new StoreInst(CastI, GEP, CI);
- } else {
- errs() << "Error: Mismatch in argument types\n";
- errs() << "__visc__node call: " << *CI << "\n";
- errs() << "Argument: " << *arg << "\n";
- errs() << "Expected: " << *ai << "\n";
- llvm_unreachable("Mismatch in argument types of kernel function and __visc__node call");
- }
- } else {
- new StoreInst(arg, GEP, CI);
- }
- }
-
- // Based on the hierarchy of the DFG we want, we need to pass the dimension
- // for each level. The number of dimensions we need to pass to the launch
- // intrinsic is the product of the number of levels and dimesions at each
- // level.
- // Marshall dim arguments
- DEBUG(errs() << *CI << "\n");
- std::string names[] = {"dimX", "dimY", "dimZ"};
- for(unsigned i=0; i< numDims*levels; i++) {
- Value* arg = CI->getArgOperand(i+dimOffset);
- DEBUG(errs() << "Passing: " << *arg << "\n");
- // Create constant int (i)
- Constant* Int_i = ConstantInt::get(Type::getInt32Ty(Ctx), i+numArgs);
- // Get Element pointer instruction
- Value* GEPIndices[] = { IntZero, Int_i };
- GetElementPtrInst* GEP = GetElementPtrInst::Create(nullptr, InStruct,
- ArrayRef<Value*>(GEPIndices, 2),
- InStruct->getName()+"."+names[i%numDims]+Twine(i/levels),
- CI);
- // Store instruction
- DEBUG(errs() << *arg << " " << *GEP << "\n");
- StoreInst* SI = new StoreInst(arg, GEP, CI);
- DEBUG(errs() << *SI << "\n");
-
- }
-}
-
-// Returns vector of all wait instructions, waiting on the passed graphID value
-static std::vector<CallInst*>* getWaitList(Value* GraphID) {
- DEBUG(errs() << "Getting Uses of: " << *GraphID << "\n");
- std::vector<CallInst*>* WaitList = new std::vector<CallInst*>();
- // It must have been loaded from memory somewhere
- for(Value::user_iterator ui = GraphID->user_begin(),
- ue = GraphID->user_end(); ui!=ue; ++ui) {
- if(CallInst* waitI = dyn_cast<CallInst>(*ui)) {
- DEBUG(errs() << "Use: " << *waitI << "\n");
- assert(isVISCCall_wait(waitI)
- && "GraphID can only be used by __visc__wait call");
- WaitList->push_back(waitI);
- }
- //else if (PHINode* PN = dyn_cast<PHINode>(*ui)){
- //errs() << "Found PhiNode use of graphID\n";
- //std::vector<CallInst*>* phiWaitList = getWaitList(PN);
- //WaitList->insert(WaitList->end(), phiWaitList->begin(), phiWaitList->end());
- //free(phiWaitList);
- //}
- else {
- DEBUG(errs() << *(*ui) << "\n");
- llvm_unreachable("Error: Operation on Graph ID not supported!\n");
- }
- }
- return WaitList;
-}
-
// Analyse the attribute call for this function. Add the in and out
// attributes to pointer parameters.
static void handleVISCAttributes(Function* F, CallInst* CI) {
@@ -420,380 +253,6 @@ static void handleVISCAttributes(Function* F, CallInst* CI) {
DEBUG(errs() << "Kernel after adding In/Out VISC attributes:\n" << *F << "\n");
}
-// Recursively generate internal nodes for all the levels. Node at each level
-// will create the appropriate instances of the child node at that level using
-// the visc createNode intrinsic, and pass on the remaining dimensions to the
-// child node.
-static Function* genInternalNode(Function* KernelF, unsigned level,
- unsigned numArgs, unsigned numDims, unsigned dimOffset, CallInst* CI) {
- // Create new function with the same type
- Module* module = KernelF->getParent();
- Function* ChildNodeF;
-
- // Recursively generate node for lower level
- if(level > 1) {
- ChildNodeF = genInternalNode(KernelF, level-1, numArgs, numDims, dimOffset, CI);
- addHint(ChildNodeF, getPreferredTarget(KernelF));
-// Internal nodes always get a CPU hint. If code geneation for them is not
-// needed and can be skipped, this is handled by the accelerator backends
-// addHint(ChildNodeF, visc::CPU_TARGET);
- } else {
- ChildNodeF = KernelF;
- }
-
- // Generate Internal node for current level
- Function* InternalF = Function::Create(ChildNodeF->getFunctionType(),
- ChildNodeF->getLinkage(),
- KernelF->getName()+"Internal_level"+Twine(level),
- module);
- // Create a basic block in this function
- BasicBlock *BB = BasicBlock::Create(InternalF->getContext(), "entry", InternalF);
- ReturnInst* RI = ReturnInst::Create(InternalF->getContext(),
- UndefValue::get(InternalF->getReturnType()), BB);
- // Copy correct attributes
- InternalF->setAttributes(ChildNodeF->getAttributes());
- // Loop over the arguments, copying the names of arguments over.
- Function::arg_iterator dest_iterator = InternalF->arg_begin();
- for (Function::const_arg_iterator i = ChildNodeF->arg_begin(), e = ChildNodeF->arg_end();
- i != e; ++i, ++dest_iterator) {
- DEBUG(errs() << "Copying argument: " << i->getName() << "\n");
- dest_iterator->setName(i->getName()); // Copy the name over...
- DEBUG(errs() << "New Argument: " << *dest_iterator << "\n");
- }
-
- // Add extra dimesnion arguments
- std::string dimNames[] = {"dimX", "dimY", "dimZ"};
- DEBUG(errs() << "Adding extra args to function Function:\n" << *InternalF << "\n");
- InternalF = addArgs(InternalF, numDims, dimNames);
- // update RI
- RI = cast<ReturnInst>(InternalF->getEntryBlock().getTerminator());
- DEBUG(errs() << "After Adding extra args to function Function:\n" << *InternalF << "\n");
-
- // Insert createNode intrinsic
- // First generate constant expression to bitcast the function pointer to
- // internal node to i8*
- Value* NodeF = ConstantExpr::getPointerCast(ChildNodeF, Type::getInt8PtrTy(module->getContext()));
-
- // Use args vectors to get the arguments for visc createNode
- // intrinsic
- std::vector<Value*> args;
-
- // Push the i8* pointer to internal node into the args vector
- args.push_back(NodeF);
-
- // Traverse the argument list of internal node function in reverse to get the
- // dimesnions to be used to create instances of child node at this level
- Function::arg_iterator ai = InternalF->arg_end();
- for(unsigned i=0; i<numDims; i++, ai--);
- DEBUG(errs() << "Iterator at: " << *ai << "\n");
-
- // ai now points to the first dimension argument to be passed to the
- // createNode intrinsic. Follow it to push the dim argument into
- // the args vector
- for(unsigned i=0; i < numDims; i++, ai++) {
- args.push_back(&*ai);
- }
-
- // Based on the number of dimensions choose the appropriate visc createNode
- // intrinsic
- DEBUG(errs() << "Number of dims = " << numDims << "\n");
- Intrinsic::ID createNodeXD;
- switch(numDims) {
- case 0:
- createNodeXD = Intrinsic::visc_createNode;
- break;
- case 1:
- createNodeXD = Intrinsic::visc_createNode1D;
- break;
- case 2:
- createNodeXD = Intrinsic::visc_createNode2D;
- break;
- case 3:
- createNodeXD = Intrinsic::visc_createNode3D;
- break;
- default:
- llvm_unreachable("Invalid number of dimensions!");
- break;
- };
-
- // Generate the visc createNode intrinsic, using the args vector as parameter
- Function* CreateNodeF = Intrinsic::getDeclaration(module, createNodeXD);
- DEBUG(errs() << "Function chosen:\n" << *CreateNodeF << "\n");
- CallInst *CreateNodeCall = CallInst::Create(CreateNodeF, args, ChildNodeF->getName()+".node", RI);
- DEBUG(errs() << "Generate call: " << *CreateNodeCall << "\n");
-
- // Generate Bind intrinsics
- Function* bindInputF = Intrinsic::getDeclaration(module, Intrinsic::visc_bind_input);
- DEBUG(errs() << "Generating input binding:\n" << *bindInputF << "\n");
- for(unsigned i=0; i < ChildNodeF->arg_size(); i++) {
- std::vector<Value*> bindArgs;
- bindArgs.push_back(CreateNodeCall);
- bindArgs.push_back(ConstantInt::get(Type::getInt32Ty(module->getContext()), i));
- bindArgs.push_back(ConstantInt::get(Type::getInt32Ty(module->getContext()), i));
- bindArgs.push_back(ConstantInt::getFalse(module->getContext()));
- CallInst* bindInputCall = CallInst::Create(bindInputF, bindArgs, "", RI);
- DEBUG(errs() << *bindInputCall << "\n");
- }
-
- // Print the generated internal node for debugging
- DEBUG(errs() << "Generated Function:\n" << *InternalF << "\n");
-
- return InternalF;
-}
-
-// Change the OpenCL query function calls with visc intrinsics in function F.
-static void replaceOpenCLCallsWithVISCIntrinsics(Function *F) {
- Module* module = F->getParent();
- std::vector<CallInst *> IItoRemove;
-
- // Get first instruction
- inst_iterator i = inst_begin(F);
- Instruction *FI = &(*i);
-
- // Insert getNode intrinsic
- Intrinsic::ID getNodeID = Intrinsic::visc_getNode;
- Function* GetNodeF = Intrinsic::getDeclaration(module, getNodeID);
- std::vector<Value*> args;
- CallInst *GetNodeCall = CallInst::Create(GetNodeF, args, F->getName()+".node", FI);
- DEBUG(errs() << "Generate getNode intrinsic: " << *GetNodeCall << "\n");
-
- // Insert getParentNode intrinsic
- Intrinsic::ID getParentNodeID = Intrinsic::visc_getParentNode;
- Function* GetParentNodeF = Intrinsic::getDeclaration(module, getParentNodeID);
- args.push_back(GetNodeCall);
- CallInst *GetParentNodeCall = CallInst::Create(GetParentNodeF, args, F->getName()+".parentNode", FI);
- DEBUG(errs() << "Generate getParentNode intrinsic: " << *GetParentNodeCall << "\n");
-
- // Iterate through all instructions
- for (inst_iterator i = inst_begin(F), e = inst_end(F); i != e; ++i) {
- Instruction *I = &(*i);
- CallInst *CI;
-
- // Find OpenCL function calls
- if ((CI = dyn_cast<CallInst>(I))) {
- if ((CI->getCalledValue()->stripPointerCasts()->getName()).equals("get_global_id")) {
- DEBUG(errs() << "Found get_global_id call: " << *CI << "\n");
- CallSite OpenCLCallSite(CI);
- Value *arg0 = OpenCLCallSite.getArgument(0);
- // Find the intrinsic function to be called
- unsigned dim = getNumericValue(arg0);
- Intrinsic::ID getNodeInstanceID;
- Intrinsic::ID getNumNodeInstancesID;
- switch (dim) {
- case 0:
- getNodeInstanceID = Intrinsic::visc_getNodeInstanceID_x;
- getNumNodeInstancesID = Intrinsic::visc_getNumNodeInstances_x;
- break;
- case 1:
- getNodeInstanceID = Intrinsic::visc_getNodeInstanceID_y;
- getNumNodeInstancesID = Intrinsic::visc_getNumNodeInstances_y;
- break;
- case 2:
- getNodeInstanceID = Intrinsic::visc_getNodeInstanceID_z;
- getNumNodeInstancesID = Intrinsic::visc_getNumNodeInstances_z;
- break;
- default:
- assert(false && "Invalid dimension from valid OpenCL source!");
- break;
- }
-
-
- // Creating getNodeInstanceID intrinsic for parent node
- ArrayRef<Value *> Args0(GetParentNodeCall);
- Function* GetNodeInstanceIDF = Intrinsic::getDeclaration(module, getNodeInstanceID);
- CallInst* ParentIDIntrinsic = CallInst::Create(GetNodeInstanceIDF, Args0, "", CI);
-
- // Creating getNumNodeInstances intrinsic for this node
- ArrayRef<Value *> Args1(GetNodeCall);
- Function* GetNumNodeInstancesF = Intrinsic::getDeclaration(module, getNumNodeInstancesID);
- CallInst* InstancesIntrinsic = CallInst::Create(GetNumNodeInstancesF, Args1, "", CI);
- // Creating mul instruction
- BinaryOperator* MulInst = BinaryOperator::Create(Instruction::Mul,
- ParentIDIntrinsic,
- InstancesIntrinsic,
- "", CI);
- // Creating getNodeInstanceID intrinsic for this node
- CallInst* LocalIDIntrinsic = CallInst::Create(GetNodeInstanceIDF, Args1, "", CI);
- // Creating add instruction
- BinaryOperator* AddInst = BinaryOperator::Create(Instruction::Add,
- MulInst,
- LocalIDIntrinsic,
- "", CI);
- CI->replaceAllUsesWith(AddInst);
- IItoRemove.push_back(CI);
- }
- if ((CI->getCalledValue()->stripPointerCasts()->getName()).equals("get_local_id")) {
- DEBUG(errs() << "Found get_local_id call: " << *CI << "\n");
- // Value *arg0 = CI->getOperand(0);
- CallSite OpenCLCallSite(CI);
- Value *arg0 = OpenCLCallSite.getArgument(0);
-
- // Argument of the function to be called
- ArrayRef<Value *> Args(GetNodeCall);
-
- // Find the intrinsic function to be called
- unsigned dim = getNumericValue(arg0);
- Intrinsic::ID getNodeInstanceID;
- switch (dim) {
- case 0:
- getNodeInstanceID = Intrinsic::visc_getNodeInstanceID_x;
- break;
- case 1:
- getNodeInstanceID = Intrinsic::visc_getNodeInstanceID_y;
- break;
- case 2:
- getNodeInstanceID = Intrinsic::visc_getNodeInstanceID_z;
- break;
- default:
- assert(false && "Invalid dimension from valid OpenCL source!");
- break;
- }
- Function* GetNodeInstanceIDF = Intrinsic::getDeclaration(module, getNodeInstanceID);
- CallInst* VI = CallInst::Create(GetNodeInstanceIDF, Args, "", CI);
- CI->replaceAllUsesWith(VI);
- IItoRemove.push_back(CI);
- }
- if ((CI->getCalledValue()->stripPointerCasts()->getName()).equals("get_group_id")) {
- DEBUG(errs() << "Found get_group_id call: " << *CI << "\n");
- // Value *arg0 = CI->getOperand(0);
- CallSite OpenCLCallSite(CI);
- Value *arg0 = OpenCLCallSite.getArgument(0);
-
- // Argument of the function to be called
- ArrayRef<Value *> Args(GetParentNodeCall);
-
- // Find the intrinsic function to be called
- unsigned dim = getNumericValue(arg0);
- Intrinsic::ID getNodeInstanceID;
- switch (dim) {
- case 0:
- getNodeInstanceID = Intrinsic::visc_getNodeInstanceID_x;
- break;
- case 1:
- getNodeInstanceID = Intrinsic::visc_getNodeInstanceID_y;
- break;
- case 2:
- getNodeInstanceID = Intrinsic::visc_getNodeInstanceID_z;
- break;
- default:
- assert(false && "Invalid dimension from valid OpenCL source!");
- break;
- }
- Function* GetNodeInstanceIDF = Intrinsic::getDeclaration(module, getNodeInstanceID);
- CallInst* VI = CallInst::Create(GetNodeInstanceIDF, Args, "", CI);
- CI->replaceAllUsesWith(VI);
- IItoRemove.push_back(CI);
- }
- if ((CI->getCalledValue()->stripPointerCasts()->getName()).equals("get_global_size")) {
- DEBUG(errs() << "Found get_global_size call: " << *CI << "\n");
- CallSite OpenCLCallSite(CI);
- Value *arg0 = OpenCLCallSite.getArgument(0);
- // Find the intrinsic function to be called
- unsigned dim = getNumericValue(arg0);
- Intrinsic::ID getNumNodeInstancesID;
- switch (dim) {
- case 0:
- getNumNodeInstancesID = Intrinsic::visc_getNumNodeInstances_x;
- break;
- case 1:
- getNumNodeInstancesID = Intrinsic::visc_getNumNodeInstances_y;
- break;
- case 2:
- getNumNodeInstancesID = Intrinsic::visc_getNumNodeInstances_z;
- break;
- default:
- assert(false && "Invalid dimension from valid OpenCL source!");
- break;
- }
-
-
- // Creating getNumNodeInstances intrinsic for parent node
- ArrayRef<Value *> Args0(GetParentNodeCall);
- Function* GetNumNodeInstancesF = Intrinsic::getDeclaration(module, getNumNodeInstancesID);
- CallInst* ParentInstancesIntrinsic = CallInst::Create(GetNumNodeInstancesF, Args0, "", CI);
- // Creating getNumNodeInstances intrinsic for this node
- ArrayRef<Value *> Args1(GetNodeCall);
- CallInst* InstancesIntrinsic = CallInst::Create(GetNumNodeInstancesF, Args1, "", CI);
- // Creating mul instruction
- BinaryOperator* MulInst = BinaryOperator::Create(Instruction::Mul,
- ParentInstancesIntrinsic,
- InstancesIntrinsic,
- "", CI);
- CI->replaceAllUsesWith(MulInst);
- IItoRemove.push_back(CI);
-
- }
- if ((CI->getCalledValue()->stripPointerCasts()->getName()).equals("get_local_size")) {
- DEBUG(errs() << "Found get_local_size call: " << *CI << "\n");
- CallSite OpenCLCallSite(CI);
- Value *arg0 = OpenCLCallSite.getArgument(0);
-
- // Argument of the function to be called
- ArrayRef<Value *> Args(GetNodeCall);
-
- // Find the intrinsic function to be called
- unsigned dim = getNumericValue(arg0);
- Intrinsic::ID getNumNodeInstancesID;
- switch (dim) {
- case 0:
- getNumNodeInstancesID = Intrinsic::visc_getNumNodeInstances_x;
- break;
- case 1:
- getNumNodeInstancesID = Intrinsic::visc_getNumNodeInstances_y;
- break;
- case 2:
- getNumNodeInstancesID = Intrinsic::visc_getNumNodeInstances_z;
- break;
- default:
- assert(false && "Invalid dimension from valid OpenCL source!");
- break;
- }
- Function* GetNumNodeInstancesF = Intrinsic::getDeclaration(module, getNumNodeInstancesID);
- CallInst* VI = CallInst::Create(GetNumNodeInstancesF, Args, "", CI);
- CI->replaceAllUsesWith(VI);
- IItoRemove.push_back(CI);
- }
- if ((CI->getCalledValue()->stripPointerCasts()->getName()).equals("get_num_groups")) {
- DEBUG(errs() << "Found get_num_groups call: " << *CI << "\n");
- CallSite OpenCLCallSite(CI);
- Value *arg0 = OpenCLCallSite.getArgument(0);
-
- // Argument of the function to be called
- ArrayRef<Value *> Args(GetParentNodeCall);
-
- // Find the intrinsic function to be called
- unsigned dim = getNumericValue(arg0);
- Intrinsic::ID getNumNodeInstancesID;
- switch (dim) {
- case 0:
- getNumNodeInstancesID = Intrinsic::visc_getNumNodeInstances_x;
- break;
- case 1:
- getNumNodeInstancesID = Intrinsic::visc_getNumNodeInstances_y;
- break;
- case 2:
- getNumNodeInstancesID = Intrinsic::visc_getNumNodeInstances_z;
- break;
- default:
- assert(false && "Invalid dimension from valid OpenCL source!");
- break;
- }
- Function* GetNumNodeInstancesF = Intrinsic::getDeclaration(module, getNumNodeInstancesID);
- CallInst* VI = CallInst::Create(GetNumNodeInstancesF, Args, "", CI);
- CI->replaceAllUsesWith(VI);
- IItoRemove.push_back(CI);
- }
- }
- }
-
-//for (std::vector<CallInst *>::reverse_iterator ri = IItoRemove.rbegin(),
- // re = IItoRemove.rend(); ri != re; ++ri)
- for (CallInst *CI : reverse(IItoRemove))
- CI->eraseFromParent();
-
-}
-
-
// Public Functions of GenVISC pass
bool GenVISC::runOnModule(Module &M) {
errs() << "\nGENVISC PASS\n";
@@ -879,16 +338,7 @@ bool GenVISC::runOnModule(Module &M) {
CallInst* CI = cast<CallInst>(I);
LLVMContext& Ctx = CI->getContext();
- // If __visc__node call found, generate the test case
-
- if(isVISCCall_node(I)) {
- errs() << "Found visc node call in Function: " << f->getName() << "\n";
- assert(CI->getNumArgOperands() >= 5
- && "__visc__node call should have atleast 5 arguments!");
- generateTest(CI);
- // Place this call in the list of instructions to be erased.
- toBeErased.push_back(CI);
- }
+
if(isVISCCall_init(I)) {
ReplaceCallWithIntrinsic(I, Intrinsic::visc_init, &toBeErased);
}
@@ -1343,197 +793,6 @@ Value* GenVISC::getStringPointer(const Twine& S, Instruction* IB, const Twine& N
return SPtr;
}
-
-
-// Generate the test case using the dummy __visc__node call CI
-// First parse the arguments to find the kernel function, num of levels,
-// dimensions, arguments, inputs and outputs. Pass this information to genKernel
-// and genInternalNode functions to generate the test case.
-void GenVISC::generateTest(CallInst* CI) {
- // Parse the dummy function call here
- LLVMContext& Ctx = CI->getParent()->getContext();
-
- unsigned offset = 1; // argument at offset 1 is the number of dimensions
- // Find number of arguments
- assert(CI->getNumArgOperands() > offset
- && "Too few arguments for __visc__node call!");
- unsigned levels = getNumericValue(CI->getArgOperand(offset));
- errs() << "\tNum of levels = " << levels << "\n";
-
- // Find number of dimensions
- offset += 1;
- assert(CI->getNumArgOperands() > offset
- && "Too few arguments for __visc__node call!");
- unsigned numDims = getNumericValue(CI->getOperand(offset));
- errs() << "\tNum of dimensions = " << numDims << "\n";
-
-
- // Find number of arguments
- offset += numDims*levels + 1; // skip the dimesnions
- assert(CI->getNumArgOperands() > offset
- && "Too few arguments for __visc__node call!");
- unsigned numArgs = getNumericValue(CI->getArgOperand(offset));
- errs() << "\tNum of kernel arguments = " << numArgs << "\n";
-
- // Find number of outputs
- offset += numArgs + 1; // skip the kernel arguments
- assert(CI->getNumArgOperands() > offset
- && "Too few arguments for __visc__node call!");
- unsigned numOutputs = getNumericValue(CI->getArgOperand(offset));
- errs() << "\tNum of kernel outputs = " << numOutputs << "\n";
-
- // Find return struct type
- assert(numOutputs == 0 && "Not handled case where number of outputs is non-zero!");
- // This is always zero. One should look at the number of struct elements of
- // kernel function
- StructType* RetTy = StructType::create(Ctx, None, "rtype");
-
- Function* KernelF = genKernel(cast<Function>(CI->getArgOperand(0)->stripPointerCasts()), RetTy);
- genHost(CI, KernelF, levels, numDims, numArgs, numOutputs, RetTy);
-}
-
-
-
-// Make all the required changes to the kernel function. This would include
-// changing the function signature by adding any extra arguments required.
-// Changing the return type. Changing all the OpenCL query intrinsics with the
-// visc intrinsics.
-Function* GenVISC::genKernel(Function* KernelF, StructType* RetTy) {
- // Make changes to kernel here
- DEBUG(errs() << "Modifying Node Function: " << KernelF->getName() << "\n");
-
- // Find dummy __visc__attribute call in this function and add visc attributes
- // in/out to pointer arguments
- for (inst_iterator i = inst_begin(KernelF), e = inst_end(KernelF); i != e; ++i) {
- Instruction *I = &(*i);
- if(isVISCCall_attributes(I)) {
- handleVISCAttributes(KernelF, cast<CallInst>(I));
- //I->eraseFromParent();
- break;
- }
- }
-
- // Change arguments and types
- // Create the argument type list with added argument types
- //Function::ArgumentListType& argList = KernelF->getArgumentList();
- std::vector<Type*> argTypes;
- // Insert an i32 argument after every pointer argument. However adding an
- // argument does not change the attribute list of function and so the
- // arguments need to be shifted accordingly.
- //bool shiftAttr = false;
- for(Function::arg_iterator ai = KernelF->arg_begin(), ae = KernelF->arg_end();
- ai != ae; ++ai) {
-
- argTypes.push_back(ai->getType());
- if(ai->getType()->isPointerTy()) {
- // If it is a pointer argument, add an i64 type next
- argTypes.push_back(Type::getInt64Ty(KernelF->getContext()));
- }
-
- }
- // Adding new arguments to the function argument list, would not change the
- // function type. We need to change the type of this function to reflect the
- // added arguments
- FunctionType* newFT = FunctionType::get(RetTy, argTypes, KernelF->isVarArg());
-
- // Change the function type
- SmallVector<ReturnInst*, 8> Returns;
- Function* newKernelF = cloneFunction(KernelF, newFT, true, &Returns);
- DEBUG(errs() << *newKernelF << "\n");
-
- // Replace ret void instruction with ret %RetTy undef
- for(auto RI: Returns) {
- DEBUG(errs() << "Found return inst: "<< *RI << "\n");
- ReturnInst* newRI = ReturnInst::Create(KernelF->getContext(), UndefValue::get(RetTy));
- ReplaceInstWithInst(RI, newRI);
- }
-
- replaceNodeFunctionInIR(*KernelF->getParent(), KernelF, newKernelF);
- // Replace opencl query intrinsics with visc query intrinsics
- replaceOpenCLCallsWithVISCIntrinsics(newKernelF);
- return newKernelF;
-}
-
-// Generate the code replacing the dummy __visc__node call with visc launch
-// intrinsic and also generate the internal nodes required at each level
-// depending on the hierarchy of DFG needed. This would also involve marhsalling
-// all the input arguments to the kernel function in memory. Replaceing CI with
-// launch intrinsic, and all the dummy __visc__wait calls with the visc wait
-// intrinsic.
-void GenVISC::genHost(CallInst* CI, Function* KernelF, unsigned levels, unsigned numDims, unsigned numArgs, unsigned numOutputs, StructType* RetTy) {
- // Make host code changes here
- DEBUG(errs() << "Modifying Host code for __visc__node call site: " << *CI << "\n");
- DEBUG(errs() << "Kernel Function: " << KernelF->getName() << "\n");
- LLVMContext& Ctx = CI->getParent()->getContext();
-
- // Create a root funtion which has this as internal node
- Function* Root = genInternalNode(KernelF, levels, numArgs, numDims, 3, CI);
-
- // Add hint to compile root for CPU. This is always true.
- addHint(Root, visc::CPU_TARGET);
-
- // Generate argument struct type (All arguments followed by return struct type)
- std::vector<Type*> ArgList;
- unsigned offset = numDims*levels + 2 + 1 + 1;
- for(Function::arg_iterator ai=KernelF->arg_begin(), ae=KernelF->arg_end();
- ai!=ae; ai++) {
- Type* Ty = ai->getType();
- ArgList.push_back(Ty);
- }
- // Add the dimesnions arguments
- for(unsigned i=0; i<numDims*levels; i++) {
-// ArgList.push_back(Type::getInt32Ty(Ctx));
- ArgList.push_back(Type::getInt64Ty(Ctx));
- }
- ArgList.push_back(RetTy);
- StructType* ArgStructTy = StructType::create(ArgList, "struct.arg", true);
- DEBUG(errs() << *ArgStructTy << "\n");
-
- switchToTimer(visc_TimerID_ARG_PACK, CI);
- // Insert alloca inst for this argument struct type
- AllocaInst* AI = new AllocaInst(ArgStructTy, 0, "in.addr", CI);
-
- // Marshall all input arguments and dimension arguments into argument struct
- // type
- marshallArguments(levels, numArgs, offset, numDims, 3, AI, CI, KernelF);
-
- // Type cast argument struct to i8*
- CastInst* BI = BitCastInst::CreatePointerCast(AI,
- Type::getInt8PtrTy(Ctx),
- "args",
- CI);
-
- switchToTimer(visc_TimerID_NONE, CI);
-
- // Bitcast Root function to i8*
- Constant* Root_i8ptr = ConstantExpr::getPointerCast(Root, Type::getInt8PtrTy(Ctx));
- // Replace CI with launch call to a Root function
- Function* LaunchF = Intrinsic::getDeclaration(Root->getParent(), Intrinsic::visc_launch);
- DEBUG(errs() << "Intrinsic for launch: " << *LaunchF << "\n");
-
- Value* LaunchInstArgs[] = {Root_i8ptr, BI, ConstantInt::getFalse(Ctx)};
- CallInst* LaunchInst = CallInst::Create(LaunchF,
- ArrayRef<Value*>(LaunchInstArgs,3),
- "graph"+Root->getName(), CI);
- //ReplaceInstWithInst(LI, LaunchInst);
-
- DEBUG(errs() << *LaunchInst << "\n");
- // Add wait call
- // Replace all wait instructions with visc wait intrinsic instructions
- Function* WaitF = Intrinsic::getDeclaration(Root->getParent(), Intrinsic::visc_wait);
- std::vector<CallInst*>* WaitList = getWaitList(CI);
- for(unsigned i=0; i < WaitList->size(); ++i) {
- CallInst* waitCall = WaitList->at(i);
- CallInst* waitInst = CallInst::Create(WaitF,
- ArrayRef<Value*>(LaunchInst),
- "", CI);
- DEBUG(errs() << *waitInst << "\n");
- waitCall->eraseFromParent();
- }
-
- // Get result (optional)
-}
-
void GenVISC::initializeTimerSet(Instruction* InsertBefore) {
Value* TimerSetAddr;
StoreInst* SI;
--
GitLab