#include "PTXMachineFunctionInfo.h"
#include "PTXRegisterInfo.h"
#include "PTXSubtarget.h"
+#include "llvm/Function.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
using namespace llvm;
-//===----------------------------------------------------------------------===//
-// Calling Convention Implementation
-//===----------------------------------------------------------------------===//
-
-#include "PTXGenCallingConv.inc"
-
//===----------------------------------------------------------------------===//
// TargetLowering Implementation
//===----------------------------------------------------------------------===//
setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
setMinFunctionAlignment(2);
+ // Let LLVM use loads/stores for all mem* operations
+ maxStoresPerMemcpy = 4096;
+ maxStoresPerMemmove = 4096;
+ maxStoresPerMemset = 4096;
+
////////////////////////////////////
/////////// Expansion //////////////
////////////////////////////////////
return "PTXISD::LOAD_PARAM";
case PTXISD::STORE_PARAM:
return "PTXISD::STORE_PARAM";
+ case PTXISD::READ_PARAM:
+ return "PTXISD::READ_PARAM";
+ case PTXISD::WRITE_PARAM:
+ return "PTXISD::WRITE_PARAM";
case PTXISD::EXIT:
return "PTXISD::EXIT";
case PTXISD::RET:
MachineFunction &MF = DAG.getMachineFunction();
const PTXSubtarget& ST = getTargetMachine().getSubtarget<PTXSubtarget>();
PTXMachineFunctionInfo *MFI = MF.getInfo<PTXMachineFunctionInfo>();
+ PTXParamManager &PM = MFI->getParamManager();
switch (CallConv) {
default:
if (MFI->isKernel() || ST.useParamSpaceForDeviceArgs()) {
// We just need to emit the proper LOAD_PARAM ISDs
for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
-
assert((!MFI->isKernel() || Ins[i].VT != MVT::i1) &&
"Kernels cannot take pred operands");
+ unsigned ParamSize = Ins[i].VT.getStoreSizeInBits();
+ unsigned Param = PM.addArgumentParam(ParamSize);
+ const std::string &ParamName = PM.getParamName(Param);
+ SDValue ParamValue = DAG.getTargetExternalSymbol(ParamName.c_str(),
+ MVT::Other);
SDValue ArgValue = DAG.getNode(PTXISD::LOAD_PARAM, dl, Ins[i].VT, Chain,
- DAG.getTargetConstant(i, MVT::i32));
+ ParamValue);
InVals.push_back(ArgValue);
-
- // Instead of storing a physical register in our argument list, we just
- // store the total size of the parameter, in bits. The ASM printer
- // knows how to process this.
- MFI->addArgReg(Ins[i].VT.getStoreSizeInBits());
}
}
else {
- // For device functions, we use the PTX calling convention to do register
- // assignments then create CopyFromReg ISDs for the allocated registers
-
- SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), ArgLocs,
- *DAG.getContext());
-
- CCInfo.AnalyzeFormalArguments(Ins, CC_PTX);
-
- for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
-
- CCValAssign& VA = ArgLocs[i];
- EVT RegVT = VA.getLocVT();
- TargetRegisterClass* TRC = 0;
-
- assert(VA.isRegLoc() && "CCValAssign must be RegLoc");
+ for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
+ EVT RegVT = Ins[i].VT;
+ TargetRegisterClass* TRC = getRegClassFor(RegVT);
+ unsigned RegType;
// Determine which register class we need
if (RegVT == MVT::i1) {
- TRC = PTX::RegPredRegisterClass;
+ RegType = PTXRegisterType::Pred;
}
else if (RegVT == MVT::i16) {
- TRC = PTX::RegI16RegisterClass;
+ RegType = PTXRegisterType::B16;
}
else if (RegVT == MVT::i32) {
- TRC = PTX::RegI32RegisterClass;
+ RegType = PTXRegisterType::B32;
}
else if (RegVT == MVT::i64) {
- TRC = PTX::RegI64RegisterClass;
+ RegType = PTXRegisterType::B64;
}
else if (RegVT == MVT::f32) {
- TRC = PTX::RegF32RegisterClass;
+ RegType = PTXRegisterType::F32;
}
else if (RegVT == MVT::f64) {
- TRC = PTX::RegF64RegisterClass;
+ RegType = PTXRegisterType::F64;
}
else {
llvm_unreachable("Unknown parameter type");
}
+ // Use a unique index in the instruction to prevent instruction folding.
+ // Yes, this is a hack.
+ SDValue Index = DAG.getTargetConstant(i, MVT::i32);
unsigned Reg = MF.getRegInfo().createVirtualRegister(TRC);
- MF.getRegInfo().addLiveIn(VA.getLocReg(), Reg);
+ SDValue ArgValue = DAG.getNode(PTXISD::READ_PARAM, dl, RegVT, Chain,
+ Index);
- SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
InVals.push_back(ArgValue);
- MFI->addArgReg(VA.getLocReg());
+ MFI->addRegister(Reg, RegType, PTXRegisterSpace::Argument);
}
}
assert(Outs.size() == 0 && "Kernel must return void.");
return DAG.getNode(PTXISD::EXIT, dl, MVT::Other, Chain);
case CallingConv::PTX_Device:
- //assert(Outs.size() <= 1 && "Can at most return one value.");
+ assert(Outs.size() <= 1 && "Can at most return one value.");
break;
}
MachineFunction& MF = DAG.getMachineFunction();
PTXMachineFunctionInfo *MFI = MF.getInfo<PTXMachineFunctionInfo>();
+ PTXParamManager &PM = MFI->getParamManager();
SDValue Flag;
+ const PTXSubtarget& ST = getTargetMachine().getSubtarget<PTXSubtarget>();
- // Even though we could use the .param space for return arguments for
- // device functions if SM >= 2.0 and the number of return arguments is
- // only 1, we just always use registers since this makes the codegen
- // easier.
- SmallVector<CCValAssign, 16> RVLocs;
- CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
- getTargetMachine(), RVLocs, *DAG.getContext());
-
- CCInfo.AnalyzeReturn(Outs, RetCC_PTX);
-
- for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
- CCValAssign& VA = RVLocs[i];
-
- assert(VA.isRegLoc() && "CCValAssign must be RegLoc");
+ if (ST.useParamSpaceForDeviceArgs()) {
+ assert(Outs.size() < 2 && "Device functions can return at most one value");
+
+ if (Outs.size() == 1) {
+ unsigned ParamSize = OutVals[0].getValueType().getSizeInBits();
+ unsigned Param = PM.addReturnParam(ParamSize);
+ const std::string &ParamName = PM.getParamName(Param);
+ SDValue ParamValue = DAG.getTargetExternalSymbol(ParamName.c_str(),
+ MVT::Other);
+ Chain = DAG.getNode(PTXISD::STORE_PARAM, dl, MVT::Other, Chain,
+ ParamValue, OutVals[0]);
+ }
+ } else {
+ for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
+ EVT RegVT = Outs[i].VT;
+ TargetRegisterClass* TRC = 0;
+ unsigned RegType;
- unsigned Reg = VA.getLocReg();
+ // Determine which register class we need
+ if (RegVT == MVT::i1) {
+ TRC = PTX::RegPredRegisterClass;
+ RegType = PTXRegisterType::Pred;
+ }
+ else if (RegVT == MVT::i16) {
+ TRC = PTX::RegI16RegisterClass;
+ RegType = PTXRegisterType::B16;
+ }
+ else if (RegVT == MVT::i32) {
+ TRC = PTX::RegI32RegisterClass;
+ RegType = PTXRegisterType::B32;
+ }
+ else if (RegVT == MVT::i64) {
+ TRC = PTX::RegI64RegisterClass;
+ RegType = PTXRegisterType::B64;
+ }
+ else if (RegVT == MVT::f32) {
+ TRC = PTX::RegF32RegisterClass;
+ RegType = PTXRegisterType::F32;
+ }
+ else if (RegVT == MVT::f64) {
+ TRC = PTX::RegF64RegisterClass;
+ RegType = PTXRegisterType::F64;
+ }
+ else {
+ llvm_unreachable("Unknown parameter type");
+ }
- DAG.getMachineFunction().getRegInfo().addLiveOut(Reg);
+ unsigned Reg = MF.getRegInfo().createVirtualRegister(TRC);
- Chain = DAG.getCopyToReg(Chain, dl, Reg, OutVals[i], Flag);
+ SDValue Copy = DAG.getCopyToReg(Chain, dl, Reg, OutVals[i]/*, Flag*/);
+ SDValue OutReg = DAG.getRegister(Reg, RegVT);
- // Guarantee that all emitted copies are stuck together,
- // avoiding something bad
- Flag = Chain.getValue(1);
+ Chain = DAG.getNode(PTXISD::WRITE_PARAM, dl, MVT::Other, Copy, OutReg);
- MFI->addRetReg(Reg);
+ MFI->addRegister(Reg, RegType, PTXRegisterSpace::Return);
+ }
}
if (Flag.getNode() == 0) {
SmallVectorImpl<SDValue> &InVals) const {
MachineFunction& MF = DAG.getMachineFunction();
- PTXMachineFunctionInfo *MFI = MF.getInfo<PTXMachineFunctionInfo>();
-
+ PTXMachineFunctionInfo *PTXMFI = MF.getInfo<PTXMachineFunctionInfo>();
+ PTXParamManager &PM = PTXMFI->getParamManager();
+ MachineFrameInfo *MFI = MF.getFrameInfo();
+
assert(getTargetMachine().getSubtarget<PTXSubtarget>().callsAreHandled() &&
"Calls are not handled for the target device");
- // Is there a more "LLVM"-way to create a variable-length array of values?
- SDValue* ops = new SDValue[OutVals.size() + 2];
+ // Identify the callee function
+ const GlobalValue *GV = cast<GlobalAddressSDNode>(Callee)->getGlobal();
+ const Function *function = cast<Function>(GV);
+
+ // allow non-device calls only for printf
+ bool isPrintf = function->getName() == "printf" || function->getName() == "puts";
+
+ assert((isPrintf || function->getCallingConv() == CallingConv::PTX_Device) &&
+ "PTX function calls must be to PTX device functions");
+
+ unsigned outSize = isPrintf ? 2 : Outs.size();
+
+ std::vector<SDValue> Ops;
+ // The layout of the ops will be [Chain, #Ins, Ins, Callee, #Outs, Outs]
+ Ops.resize(outSize + Ins.size() + 4);
+
+ Ops[0] = Chain;
+
+ // Identify the callee function
+ Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
+ Ops[Ins.size()+2] = Callee;
+
+ // #Outs
+ Ops[Ins.size()+3] = DAG.getTargetConstant(outSize, MVT::i32);
+
+ if (isPrintf) {
+ // first argument is the address of the global string variable in memory
+ unsigned Param0 = PM.addLocalParam(getPointerTy().getSizeInBits());
+ SDValue ParamValue0 = DAG.getTargetExternalSymbol(PM.getParamName(Param0).c_str(),
+ MVT::Other);
+ Chain = DAG.getNode(PTXISD::STORE_PARAM, dl, MVT::Other, Chain,
+ ParamValue0, OutVals[0]);
+ Ops[Ins.size()+4] = ParamValue0;
+
+ // alignment is the maximum size of all the arguments
+ unsigned alignment = 0;
+ for (unsigned i = 1; i < OutVals.size(); ++i) {
+ alignment = std::max(alignment,
+ OutVals[i].getValueType().getSizeInBits());
+ }
- ops[0] = Chain;
+ // size is the alignment multiplied by the number of arguments
+ unsigned size = alignment * (OutVals.size() - 1);
+
+ // second argument is the address of the stack object (unless no arguments)
+ unsigned Param1 = PM.addLocalParam(getPointerTy().getSizeInBits());
+ SDValue ParamValue1 = DAG.getTargetExternalSymbol(PM.getParamName(Param1).c_str(),
+ MVT::Other);
+ Ops[Ins.size()+5] = ParamValue1;
+
+ if (size > 0)
+ {
+ // create a local stack object to store the arguments
+ unsigned StackObject = MFI->CreateStackObject(size / 8, alignment / 8, false);
+ SDValue FrameIndex = DAG.getFrameIndex(StackObject, getPointerTy());
+
+ // store each of the arguments to the stack in turn
+ for (unsigned int i = 1; i != OutVals.size(); i++) {
+ SDValue FrameAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FrameIndex, DAG.getTargetConstant((i - 1) * 8, getPointerTy()));
+ Chain = DAG.getStore(Chain, dl, OutVals[i], FrameAddr,
+ MachinePointerInfo(),
+ false, false, 0);
+ }
- if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
- const GlobalValue *GV = G->getGlobal();
- Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
- ops[1] = Callee;
- } else {
- assert(false && "Function must be a GlobalAddressSDNode");
- }
+ // copy the address of the local frame index to get the address in non-local space
+ SDValue genericAddr = DAG.getNode(PTXISD::COPY_ADDRESS, dl, getPointerTy(), FrameIndex);
- for (unsigned i = 0; i != OutVals.size(); ++i) {
- unsigned Size = OutVals[i].getValueType().getSizeInBits();
- SDValue Index = DAG.getTargetConstant(MFI->getNextParam(Size), MVT::i32);
- Chain = DAG.getNode(PTXISD::STORE_PARAM, dl, MVT::Other, Chain,
- Index, OutVals[i]);
- ops[i+2] = Index;
+ // store this address in the second argument
+ Chain = DAG.getNode(PTXISD::STORE_PARAM, dl, MVT::Other, Chain, ParamValue1, genericAddr);
+ }
+ }
+ else
+ {
+ // Generate STORE_PARAM nodes for each function argument. In PTX, function
+ // arguments are explicitly stored into .param variables and passed as
+ // arguments. There is no register/stack-based calling convention in PTX.
+ for (unsigned i = 0; i != OutVals.size(); ++i) {
+ unsigned Size = OutVals[i].getValueType().getSizeInBits();
+ unsigned Param = PM.addLocalParam(Size);
+ const std::string &ParamName = PM.getParamName(Param);
+ SDValue ParamValue = DAG.getTargetExternalSymbol(ParamName.c_str(),
+ MVT::Other);
+ Chain = DAG.getNode(PTXISD::STORE_PARAM, dl, MVT::Other, Chain,
+ ParamValue, OutVals[i]);
+ Ops[i+Ins.size()+4] = ParamValue;
+ }
+ }
+
+ std::vector<SDValue> InParams;
+
+ // Generate list of .param variables to hold the return value(s).
+ Ops[1] = DAG.getTargetConstant(Ins.size(), MVT::i32);
+ for (unsigned i = 0; i < Ins.size(); ++i) {
+ unsigned Size = Ins[i].VT.getStoreSizeInBits();
+ unsigned Param = PM.addLocalParam(Size);
+ const std::string &ParamName = PM.getParamName(Param);
+ SDValue ParamValue = DAG.getTargetExternalSymbol(ParamName.c_str(),
+ MVT::Other);
+ Ops[i+2] = ParamValue;
+ InParams.push_back(ParamValue);
}
- ops[0] = Chain;
+ Ops[0] = Chain;
- Chain = DAG.getNode(PTXISD::CALL, dl, MVT::Other, ops, OutVals.size()+2);
+ // Create the CALL node.
+ Chain = DAG.getNode(PTXISD::CALL, dl, MVT::Other, &Ops[0], Ops.size());
- delete [] ops;
+ // Create the LOAD_PARAM nodes that retrieve the function return value(s).
+ for (unsigned i = 0; i < Ins.size(); ++i) {
+ SDValue Load = DAG.getNode(PTXISD::LOAD_PARAM, dl, Ins[i].VT, Chain,
+ InParams[i]);
+ InVals.push_back(Load);
+ }
return Chain;
}
+
+unsigned PTXTargetLowering::getNumRegisters(LLVMContext &Context, EVT VT) {
+ // All arguments consist of one "register," regardless of the type.
+ return 1;
+}
+