diff --git a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
index 768742b53e4ad9628e040147b80a0cd039269674..ac93e5f79727523b671c03f0bd9078a42727e794 100644
--- a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
@@ -1242,7 +1242,7 @@ void CGT_NVPTX::codeGen(DFLeafNode* N) {
int parentLevel = ParentDFNode->getLevel();
int parentReplFactor = ParentDFNode->getNumOfDim();
- if (!parentLevel || !parentReplFactor) {
+ if ((N == ArgDFNode) && (!parentLevel || !parentReplFactor)) {
// We only have one level in the hierarchy or the parent node is not
// replicated. This indicates that the parent node is the kernel
// launch, so the instances are global_size (gridDim x blockDim)
diff --git a/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp b/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
index 5f9f6cb12fc28f64ca1b2ba72c54b27035cdf3f5..75280bd50d674be701b7b59f36f75125ceed23e0 100644
--- a/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
@@ -223,6 +223,12 @@ std::vector<IntrinsicInst*>* CGT_X86::getUseList(Value* GraphID) {
if(IntrinsicInst* waitI = dyn_cast<IntrinsicInst>(*ui)) {
UseList->push_back(waitI);
}
+ //else if (PHINode* PN = dyn_cast<PHINode>(*ui)){
+ //errs() << "Found PhiNode use of graphID\n";
+ //std::vector<IntrinsicInst*>* phiUseList = getUseList(PN);
+ //UseList->insert(UseList->end(), phiUseList->begin(), phiUseList->end());
+ //free(phiUseList);
+ //}
else {
llvm_unreachable("Error: Operation on Graph ID not supported!\n");
}
diff --git a/llvm/lib/Transforms/GenVISC/GenVISC.cpp b/llvm/lib/Transforms/GenVISC/GenVISC.cpp
index 8957ebe2c8a5b099e4f930a90e748ededb780fae..1deb088a2d7e4e98ec91ff2a483d6c3b92d7ecc9 100644
--- a/llvm/lib/Transforms/GenVISC/GenVISC.cpp
+++ b/llvm/lib/Transforms/GenVISC/GenVISC.cpp
@@ -52,29 +52,43 @@ static void ReplaceCallWithIntrinsic(Instruction* I, Intrinsic::ID IntrinsicID,
// Find the correct intrinsic call
Module* M = CI->getParent()->getParent()->getParent();
- Function* F = Intrinsic::getDeclaration(M, IntrinsicID);
- FunctionType* FTy = F->getFunctionType();
- DEBUG(errs() << *F << "\n");
-
- // Create argument list
- assert(CI->getNumArgOperands() == FTy->getNumParams()
- && "Number of arguments of call do not match with Intrinsic");
+ Function* F;
+ std::vector<Type*> ArgTypes;
std::vector<Value*> args;
- for(unsigned i=0; i < CI->getNumArgOperands(); i++) {
- Value* V = CI->getArgOperand(i);
- // Either the type should match or both should be of pointer type
- assert(V->getType() == FTy->getParamType(i) ||
- (V->getType()->isPointerTy() && FTy->getParamType(i)->isPointerTy())
- && "Dummy function call argument does not match with Intrinsic argument!");
- // If the types do not match, then both must be pointer type and pointer
- // cast needs to be performed
- if(V->getType() != FTy->getParamType(i)) {
- V = CastInst::CreatePointerCast(V, FTy->getParamType(i), "", CI);
+ if(Intrinsic::isOverloaded(IntrinsicID)) {
+ // This is an overloaded intrinsic. The types must exactly match. Get the
+ // argument types
+ for(unsigned i=0; i < CI->getNumArgOperands(); i++) {
+ ArgTypes.push_back(CI->getArgOperand(i)->getType());
+ args.push_back(CI->getArgOperand(i));
+ }
+ F = Intrinsic::getDeclaration(M, IntrinsicID, ArgTypes);
+ DEBUG(errs() << *F << "\n");
+ }
+ else { // Non-overloaded intrinsic
+ F = Intrinsic::getDeclaration(M, IntrinsicID);
+ FunctionType* FTy = F->getFunctionType();
+ DEBUG(errs() << *F << "\n");
+
+ // Create argument list
+ assert(CI->getNumArgOperands() == FTy->getNumParams()
+ && "Number of arguments of call do not match with Intrinsic");
+ for(unsigned i=0; i < CI->getNumArgOperands(); i++) {
+ Value* V = CI->getArgOperand(i);
+ // Either the type should match or both should be of pointer type
+ assert(V->getType() == FTy->getParamType(i) ||
+ (V->getType()->isPointerTy() && FTy->getParamType(i)->isPointerTy())
+ && "Dummy function call argument does not match with Intrinsic argument!");
+ // If the types do not match, then both must be pointer type and pointer
+ // cast needs to be performed
+ if(V->getType() != FTy->getParamType(i)) {
+ V = CastInst::CreatePointerCast(V, FTy->getParamType(i), "", CI);
+ }
+ args.push_back(V);
}
- args.push_back(V);
}
// Insert call instruction
- CallInst* Inst = CallInst::Create(F, args, CI->getName(), CI);
+ CallInst* Inst = CallInst::Create(F, args, F->getReturnType()->isVoidTy()? "" : CI->getName(), CI);
DEBUG(errs() << "\tSubstitute with: " << *Inst << "\n");
@@ -118,6 +132,11 @@ IS_VISC_CALL(atomic_max)
IS_VISC_CALL(atomic_and)
IS_VISC_CALL(atomic_or)
IS_VISC_CALL(atomic_xor)
+// Misc Fn
+IS_VISC_CALL(floor)
+IS_VISC_CALL(rsqrt)
+IS_VISC_CALL(sqrt)
+
IS_VISC_CALL(init)
IS_VISC_CALL(node)
@@ -336,6 +355,12 @@ static std::vector<CallInst*>* getWaitList(Value* GraphID) {
&& "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");
@@ -976,7 +1001,7 @@ bool GenVISC::runOnModule(Module &M) {
CI->getArgOperand(2), CI->getArgOperand(3)
};
CallInst* CreateNode3DInst = CallInst::Create(CreateNode3DF,
- ArrayRef<Value*>(CreateNode3DArgs, 2),
+ ArrayRef<Value*>(CreateNode3DArgs, 4),
graphFunc->getName()+".node", CI);
DEBUG(errs() << "Found visc createNode3D call: " << *CI << "\n");
DEBUG(errs() << "\tSubstitute with: " << *CreateNode3DInst << "\n");
@@ -1146,6 +1171,15 @@ bool GenVISC::runOnModule(Module &M) {
if (isVISCCall_atomic_xor(I)) {
ReplaceCallWithIntrinsic(I, Intrinsic::visc_atomic_xor, &toBeErased);
}
+ if (isVISCCall_floor(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::nvvm_floor_f, &toBeErased);
+ }
+ if (isVISCCall_rsqrt(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::nvvm_rsqrt_approx_f, &toBeErased);
+ }
+ if (isVISCCall_sqrt(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::nvvm_sqrt_f, &toBeErased);
+ }
}
}