diff --git a/llvm/include/llvm/IR/DFGraph.h b/llvm/include/llvm/IR/DFGraph.h
index d068301bc7d48e55fb78c2c10f7c10efbc797d4f..247154b6801abe7b7dffc1e90487ad872d832f9b 100644
--- a/llvm/include/llvm/IR/DFGraph.h
+++ b/llvm/include/llvm/IR/DFGraph.h
@@ -496,7 +496,6 @@ public:
}
void removeGenFuncForTarget(visc::Target T) {
- errs() << "Target tag = " << T << "\n";
switch (T) {
case visc::CPU_TARGET:
GenFuncs.CPUGenFunc = NULL;
diff --git a/llvm/include/llvm/SupportVISC/DFG2LLVM.h b/llvm/include/llvm/SupportVISC/DFG2LLVM.h
index a036d255c81aeec436b1add99ac0738ec3a46860..9a48405cd5d2631905e265a547167f79dc2eb681 100644
--- a/llvm/include/llvm/SupportVISC/DFG2LLVM.h
+++ b/llvm/include/llvm/SupportVISC/DFG2LLVM.h
@@ -17,6 +17,7 @@
#include "llvm/BuildDFG/BuildDFG.h"
#include "llvm/SupportVISC/VISCHint.h"
#include "llvm/SupportVISC/VISCTimer.h"
+#include "llvm/SupportVISC/VISCUtils.h"
using namespace llvm;
using namespace builddfg;
@@ -84,8 +85,10 @@ protected:
// Functions
Value* getStringPointer(const Twine& S, Instruction* InsertBefore, const Twine& Name = "");
- void addArgument(Function*, Type*, const Twine& Name = "");
- void addIdxDimArgs(Function* F);
+// void addArgument(Function*, Type*, const Twine& Name = "");
+ Function *addArgument(Function*, Type*, const Twine& Name = "");
+// void addIdxDimArgs(Function* F);
+ Function *addIdxDimArgs(Function* F);
std::vector<Value*> extractElements(Value*, std::vector<Type*>,
std::vector<std::string>, Instruction*);
Argument* getArgumentAt(Function* F, unsigned offset);
@@ -223,7 +226,29 @@ Value* CodeGenTraversal::getStringPointer(const Twine& S, Instruction* IB, const
}
// Add an argument of type Ty to the given function F
-void CodeGenTraversal::addArgument(Function* F, Type* Ty, const Twine& name) {
+//void CodeGenTraversal::addArgument(Function* F, Type* Ty, const Twine& name) {
+// // Add the argument to argument list
+// new Argument(Ty, name, F);
+//
+// // 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
+// FunctionType* FTy = FunctionType::get(F->getReturnType(), ArgTypes, F->isVarArg());
+// PointerType* PTy = PointerType::get(FTy, cast<PointerType>(F->getType())->getAddressSpace());
+//
+// // Change the function type
+// F->mutateType(PTy);
+//}
+
+// Creates a function with an additional argument of the specified type and
+// name. The previous function is not deleted.
+Function *CodeGenTraversal::addArgument(Function* F, Type* Ty, const Twine& name) {
// Add the argument to argument list
new Argument(Ty, name, F);
@@ -235,21 +260,43 @@ void CodeGenTraversal::addArgument(Function* F, Type* Ty, const Twine& name) {
}
// 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
+ // added arguments. So, we create a clone of this function with the correct
+ // type.
FunctionType* FTy = FunctionType::get(F->getReturnType(), ArgTypes, F->isVarArg());
- PointerType* PTy = PointerType::get(FTy, cast<PointerType>(F->getType())->getAddressSpace());
+ Function *newF = viscUtils::cloneFunction(F, FTy, false);
+
+ // Check if the function is used by a metadata node
+ if(F->isUsedByMetadata()) {
+ viscUtils::fixHintMetadata(*F->getParent(), F, newF);
+ }
- // Change the function type
- F->mutateType(PTy);
+ return newF;
}
// Change the argument list of function F to add index and limit arguments
-void CodeGenTraversal::addIdxDimArgs(Function* F) {
+//void CodeGenTraversal::addIdxDimArgs(Function* F) {
+// // Add Index and Dim arguments
+// std::string names[] = {"idx_x", "idx_y", "idx_z", "dim_x", "dim_y", "dim_z"};
+// for (int i = 0; i < 6; ++i) {
+// addArgument(F, Type::getInt32Ty(F->getContext()), names[i]);
+// }
+//}
+
+// Return new function with additional index and limit arguments.
+// The original function is removed from the module and erased.
+Function *CodeGenTraversal::addIdxDimArgs(Function* F) {
+ errs() << "Function Type: " << *F->getFunctionType() << "\n";
// Add Index and Dim arguments
std::string names[] = {"idx_x", "idx_y", "idx_z", "dim_x", "dim_y", "dim_z"};
+ Function *newF;
for (int i = 0; i < 6; ++i) {
- addArgument(F, Type::getInt32Ty(F->getContext()), names[i]);
+ newF = addArgument(F, Type::getInt64Ty(F->getContext()), names[i]);
+ F->replaceAllUsesWith(UndefValue::get(F->getType()));
+ F->eraseFromParent();
+ F = newF;
}
+ errs() << "Function Type after adding args: " << *newF->getFunctionType() << "\n";
+ return newF;
}
// Extract elements from an aggregate value. TyList contains the type of each
diff --git a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
index bb83544969c3fa88fb0dd10557cae734ced28786..845e5a7f9f50fcd9cf5eaa8455b5d913ab401a1c 100644
--- a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
@@ -354,8 +354,7 @@ void CGT_NVPTX::insertRuntimeCalls(DFInternalNode* N, Kernel* K, const Twine& Fi
for (Function::const_arg_iterator i = F->arg_begin(), e = F->arg_end();
i != e; ++i) {
dest_iterator->setName(i->getName()); // Copy the name over...
- // Add mapping to VMap and increment dest iterator
- VMap[&*i] = &*dest_iterator;
+ // Increment dest iterator
++dest_iterator;
}
@@ -364,17 +363,29 @@ void CGT_NVPTX::insertRuntimeCalls(DFInternalNode* N, Kernel* K, const Twine& Fi
ReturnInst* RI = ReturnInst::Create(M.getContext(),
UndefValue::get(F_X86->getReturnType()), BB);
- //Add the generated function info to DFNode
-// N->setGenFunc(F_X86, visc::CPU_TARGET);
- N->addGenFunc(F_X86, visc::GPU_TARGET, true);
-
// FIXME: Adding Index and Dim arguments are probably not required except
// for consistency purpose (DFG2LLVM_X86 does assume that all leaf nodes do
// have those arguments)
// Add Index and Dim arguments except for the root node
if(!N->isRoot() && !N->getParent()->isChildGraphStreaming())
- addIdxDimArgs(F_X86);
+ F_X86 = addIdxDimArgs(F_X86);
+
+ BB = &*F_X86->begin();
+ RI = cast<ReturnInst>(BB->getTerminator());
+
+ //Add the generated function info to DFNode
+// N->setGenFunc(F_X86, visc::CPU_TARGET);
+ N->addGenFunc(F_X86, visc::GPU_TARGET, true);
+
+ // Loop over the arguments, to create the VMap
+ dest_iterator = F_X86->arg_begin();
+ for (Function::const_arg_iterator i = F->arg_begin(), e = F->arg_end();
+ i != e; ++i) {
+ // Add mapping to VMap and increment dest iterator
+ VMap[&*i] = &*dest_iterator;
+ ++dest_iterator;
+ }
/* TODO: Use this code to verufy if this is a good pattern for PTX kernel
diff --git a/llvm/lib/Transforms/DFG2LLVM_SPIR/DFG2LLVM_SPIR.cpp b/llvm/lib/Transforms/DFG2LLVM_SPIR/DFG2LLVM_SPIR.cpp
index f31e75931c4a997d8c43716e299faf323ac48477..d62d0561ac47c384c21835608a512d3c8246a022 100644
--- a/llvm/lib/Transforms/DFG2LLVM_SPIR/DFG2LLVM_SPIR.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_SPIR/DFG2LLVM_SPIR.cpp
@@ -369,7 +369,7 @@ void CGT_SPIR::insertRuntimeCalls(DFInternalNode* N, Kernel* K, const Twine& Fil
// Add Index and Dim arguments except for the root node
if(!N->isRoot() && !N->getParent()->isChildGraphStreaming())
- addIdxDimArgs(F_X86);
+ F_X86 = addIdxDimArgs(F_X86);
/* TODO: Use this code to verufy if this is a good pattern for OCL kernel
diff --git a/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp b/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
index 8794f423e03af521ab44562301924ba106454e1a..d55640dc6f875777ab3fb3fbe75763a4dd6d016d 100644
--- a/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
@@ -949,9 +949,9 @@ void CGT_X86::invokeChild_X86(DFNode* C, Function* F_X86,
Args.push_back(getInValueAt(C, i, F_X86, IB));
}
- Value* I32Zero = ConstantInt::get(Type::getInt32Ty(F_X86->getContext()), 0);
+ Value* I64Zero = ConstantInt::get(Type::getInt64Ty(F_X86->getContext()), 0);
for(unsigned j=0; j<6; j++)
- Args.push_back(I32Zero);
+ Args.push_back(I64Zero);
errs() << "Function type: " << *CF_X86->getType() << "\n";
errs() << "Function type: " << *CF->getType() << "\n";
@@ -971,7 +971,7 @@ void CGT_X86::invokeChild_X86(DFNode* C, Function* F_X86,
std::string varNames[3] = {"x", "y", "z"};
unsigned numArgs = CI->getNumArgOperands();
for(unsigned j=0; j < C->getNumOfDim(); j++) {
- Value* indexLimit;
+ Value* indexLimit = NULL;
// Limit can either be a constant or an arguement of the internal node.
// In case of constant we can use that constant value directly in the
// new F_X86 function. In case of an argument, we need to get the mapped
@@ -1300,7 +1300,7 @@ void CGT_X86::codeGen(DFInternalNode* N) {
if (!(C->hasX86GenFuncForTarget(visc::CPU_TARGET))) {
errs() << "No CPU x86 version for child node "
<< C->getFuncPointer()->getName()
- << " . Skip code gen for parent node "
+ << "\n Skip code gen for parent node "
<< N->getFuncPointer()->getName() << "\n";
codeGen = false;
}
@@ -1318,34 +1318,42 @@ void CGT_X86::codeGen(DFInternalNode* N) {
// Create new function with the same type
F_X86 = Function::Create(F->getFunctionType(), F->getLinkage(), F->getName(), &M);
- errs() << "--------------" << F->getName() << "\n";
+
// Loop over the arguments, copying the names of arguments over.
Function::arg_iterator dest_iterator = F_X86->arg_begin();
- assert(false && "Got here\n");
for (Function::const_arg_iterator i = F->arg_begin(), e = F->arg_end();
i != e; ++i) {
dest_iterator->setName(i->getName()); // Copy the name over...
- // Add mapping to VMap and increment dest iterator
- VMap[&*i] = &*dest_iterator;
+ // Increment dest iterator
++dest_iterator;
}
-
- assert(false && "Got here\n");
// Add a basic block to this empty function
BasicBlock *BB = BasicBlock::Create(F_X86->getContext(), "entry", F_X86);
ReturnInst* RI = ReturnInst::Create(F_X86->getContext(),
UndefValue::get(F_X86->getReturnType()), BB);
- //Add generated function info to DFNode
-// N->setGenFunc(F_X86, visc::CPU_TARGET);
- N->addGenFunc(F_X86, visc::CPU_TARGET, true);
-
// Add Index and Dim arguments except for the root node and the child graph of
// parent node is not streaming
if(!N->isRoot() && !N->getParent()->isChildGraphStreaming())
- addIdxDimArgs(F_X86);
+ F_X86 = addIdxDimArgs(F_X86);
+
+ BB = &*F_X86->begin();
+ RI = cast<ReturnInst>(BB->getTerminator());
+ //Add generated function info to DFNode
+// N->setGenFunc(F_X86, visc::CPU_TARGET);
+ N->addGenFunc(F_X86, visc::CPU_TARGET, true);
+
+ // Loop over the arguments, to create the VMap.
+ dest_iterator = F_X86->arg_begin();
+ for (Function::const_arg_iterator i = F->arg_begin(), e = F->arg_end();
+ i != e; ++i) {
+ // Add mapping and increment dest iterator
+ VMap[&*i] = &*dest_iterator;
+ ++dest_iterator;
+ }
+
// Iterate over children in topological order
for(DFGraph::children_iterator ci = N->getChildGraph()->begin(),
ce = N->getChildGraph()->end(); ci != ce; ++ci) {
@@ -1675,14 +1683,14 @@ void CGT_X86::codeGen(DFLeafNode* N) {
// Insert the cloned function into the module
M.getFunctionList().push_back(F_X86);
- // Add generated function info to DFNode
-// N->setGenFunc(F_X86, visc::CPU_TARGET);
- N->addGenFunc(F_X86, visc::CPU_TARGET, true);
-
// Add the new argument to the argument list. Add arguments only if the cild
// graph of parent node is not streaming
if(!N->getParent()->isChildGraphStreaming())
- addIdxDimArgs(F_X86);
+ F_X86 = addIdxDimArgs(F_X86);
+
+ // Add generated function info to DFNode
+// N->setGenFunc(F_X86, visc::CPU_TARGET);
+ N->addGenFunc(F_X86, visc::CPU_TARGET, true);
// Go through the arguments, and any pointer arguments with in attribute need
// to have x86_argument_ptr call to get the x86 ptr of the argument
diff --git a/llvm/projects/visc-rt/visc-rt.cpp b/llvm/projects/visc-rt/visc-rt.cpp
index e61802d25e2b5ccd4affa7044deb8712feaf9c03..8d52e471652ab5ca17e30ca9328f067ae1ab9942 100644
--- a/llvm/projects/visc-rt/visc-rt.cpp
+++ b/llvm/projects/visc-rt/visc-rt.cpp
@@ -99,7 +99,7 @@ void llvm_visc_x86_dstack_pop() {
//DEBUG(cout << "DStack size = " << DStack.size() << flush << "\n");
}
-unsigned llvm_visc_x86_getDimLimit(unsigned level, unsigned dim) {
+uint64_t llvm_visc_x86_getDimLimit(unsigned level, unsigned dim) {
//DEBUG(cout << "Request limit for dim " << dim << " of ancestor " << level <<flush << "\n");
//unsigned size = DStack.size();
//DEBUG(cout << "\t Return: " << DStack[size-level-1].getDimLimit(dim) <<flush << "\n");
@@ -107,7 +107,7 @@ unsigned llvm_visc_x86_getDimLimit(unsigned level, unsigned dim) {
return 0;
}
-unsigned llvm_visc_x86_getDimInstance(unsigned level, unsigned dim) {
+uint64_t llvm_visc_x86_getDimInstance(unsigned level, unsigned dim) {
//DEBUG(cout << "Request instance id for dim " << dim << " of ancestor " << level <<flush << "\n");
//unsigned size = DStack.size();
//DEBUG(cout << "\t Return: " << DStack[size-level-1].getDimInstance(dim) <<flush << "\n");
diff --git a/llvm/projects/visc-rt/visc-rt.h b/llvm/projects/visc-rt/visc-rt.h
index 20cc6e35a6f0b7802d4e662b61f5c6e2bf086149..69392671fb6a244922c75a6c11c82b2405787732 100644
--- a/llvm/projects/visc-rt/visc-rt.h
+++ b/llvm/projects/visc-rt/visc-rt.h
@@ -71,8 +71,8 @@ class DFGDepth {
void llvm_visc_x86_dstack_push(unsigned n, unsigned limitX = 0, unsigned iX = 0,
unsigned limitY = 0, unsigned iY = 0, unsigned limitZ = 0, unsigned iZ = 0);
void llvm_visc_x86_dstack_pop();
-unsigned llvm_visc_x86_getDimLimit(unsigned level, unsigned dim);
-unsigned llvm_visc_x86_getDimInstance(unsigned level, unsigned dim);
+uint64_t llvm_visc_x86_getDimLimit(unsigned level, unsigned dim);
+uint64_t llvm_visc_x86_getDimInstance(unsigned level, unsigned dim);
/********************* Memory Tracker **********************************/