#include "SparcISelLowering.h"
#include "SparcMachineFunctionInfo.h"
#include "SparcTargetMachine.h"
+#include "MCTargetDesc/SparcBaseInfo.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
{
assert (ArgFlags.isSRet());
- //Assign SRet argument
+ // Assign SRet argument.
State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
0,
LocVT, LocInfo));
static const uint16_t RegList[] = {
SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5
};
- //Try to get first reg
+ // Try to get first reg.
if (unsigned Reg = State.AllocateReg(RegList, 6)) {
State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
} else {
- //Assign whole thing in stack
+ // Assign whole thing in stack.
State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
State.AllocateStack(8,4),
LocVT, LocInfo));
return true;
}
- //Try to get second reg
+ // Try to get second reg.
if (unsigned Reg = State.AllocateReg(RegList, 6))
State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
else
#include "SparcGenCallingConv.inc"
+// The calling conventions in SparcCallingConv.td are described in terms of the
+// callee's register window. This function translates registers to the
+// corresponding caller window %o register.
+static unsigned toCallerWindow(unsigned Reg) {
+ assert(SP::I0 + 7 == SP::I7 && SP::O0 + 7 == SP::O7 && "Unexpected enum");
+ if (Reg >= SP::I0 && Reg <= SP::I7)
+ return Reg - SP::I0 + SP::O0;
+ return Reg;
+}
+
SDValue
SparcTargetLowering::LowerReturn(SDValue Chain,
CallingConv::ID CallConv, bool IsVarArg,
const SmallVectorImpl<ISD::OutputArg> &Outs,
const SmallVectorImpl<SDValue> &OutVals,
- DebugLoc DL, SelectionDAG &DAG) const {
+ SDLoc DL, SelectionDAG &DAG) const {
if (Subtarget->is64Bit())
return LowerReturn_64(Chain, CallConv, IsVarArg, Outs, OutVals, DL, DAG);
return LowerReturn_32(Chain, CallConv, IsVarArg, Outs, OutVals, DL, DAG);
CallingConv::ID CallConv, bool IsVarArg,
const SmallVectorImpl<ISD::OutputArg> &Outs,
const SmallVectorImpl<SDValue> &OutVals,
- DebugLoc DL, SelectionDAG &DAG) const {
+ SDLoc DL, SelectionDAG &DAG) const {
MachineFunction &MF = DAG.getMachineFunction();
// CCValAssign - represent the assignment of the return value to locations.
RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
}
- unsigned RetAddrOffset = 8; //Call Inst + Delay Slot
+ unsigned RetAddrOffset = 8; // Call Inst + Delay Slot
// If the function returns a struct, copy the SRetReturnReg to I0
if (MF.getFunction()->hasStructRetAttr()) {
SparcMachineFunctionInfo *SFI = MF.getInfo<SparcMachineFunctionInfo>();
CallingConv::ID CallConv, bool IsVarArg,
const SmallVectorImpl<ISD::OutputArg> &Outs,
const SmallVectorImpl<SDValue> &OutVals,
- DebugLoc DL, SelectionDAG &DAG) const {
+ SDLoc DL, SelectionDAG &DAG) const {
// CCValAssign - represent the assignment of the return value to locations.
SmallVector<CCValAssign, 16> RVLocs;
CallingConv::ID CallConv,
bool IsVarArg,
const SmallVectorImpl<ISD::InputArg> &Ins,
- DebugLoc DL,
+ SDLoc DL,
SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals) const {
if (Subtarget->is64Bit())
CallingConv::ID CallConv,
bool isVarArg,
const SmallVectorImpl<ISD::InputArg> &Ins,
- DebugLoc dl,
+ SDLoc dl,
SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals) const {
MachineFunction &MF = DAG.getMachineFunction();
CCValAssign &VA = ArgLocs[i];
if (i == 0 && Ins[i].Flags.isSRet()) {
- //Get SRet from [%fp+64]
+ // Get SRet from [%fp+64].
int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, 64, true);
SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
SDValue Arg = DAG.getLoad(MVT::i32, dl, Chain, FIPtr,
if (VA.needsCustom()) {
assert(VA.getValVT() == MVT::f64);
- //If it is double-word aligned, just load.
+ // If it is double-word aligned, just load.
if (Offset % 8 == 0) {
int FI = MF.getFrameInfo()->CreateFixedObject(8,
Offset,
}
if (MF.getFunction()->hasStructRetAttr()) {
- //Copy the SRet Argument to SRetReturnReg
+ // Copy the SRet Argument to SRetReturnReg.
SparcMachineFunctionInfo *SFI = MF.getInfo<SparcMachineFunctionInfo>();
unsigned Reg = SFI->getSRetReturnReg();
if (!Reg) {
CallingConv::ID CallConv,
bool IsVarArg,
const SmallVectorImpl<ISD::InputArg> &Ins,
- DebugLoc DL,
+ SDLoc DL,
SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals) const {
MachineFunction &MF = DAG.getMachineFunction();
getTargetMachine(), ArgLocs, *DAG.getContext());
CCInfo.AnalyzeFormalArguments(Ins, CC_Sparc64);
+ // The argument array begins at %fp+BIAS+128, after the register save area.
+ const unsigned ArgArea = 128;
+
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
CCValAssign &VA = ArgLocs[i];
if (VA.isRegLoc()) {
assert(VA.isMemLoc());
// The CC_Sparc64_Full/Half functions compute stack offsets relative to the
// beginning of the arguments area at %fp+BIAS+128.
- unsigned Offset = VA.getLocMemOffset() + 128;
+ unsigned Offset = VA.getLocMemOffset() + ArgArea;
unsigned ValSize = VA.getValVT().getSizeInBits() / 8;
// Adjust offset for extended arguments, SPARC is big-endian.
// The caller will have written the full slot with extended bytes, but we
MachinePointerInfo::getFixedStack(FI),
false, false, false, 0));
}
+
+ if (!IsVarArg)
+ return Chain;
+
+ // This function takes variable arguments, some of which may have been passed
+ // in registers %i0-%i5. Variable floating point arguments are never passed
+ // in floating point registers. They go on %i0-%i5 or on the stack like
+ // integer arguments.
+ //
+ // The va_start intrinsic needs to know the offset to the first variable
+ // argument.
+ unsigned ArgOffset = CCInfo.getNextStackOffset();
+ SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>();
+ // Skip the 128 bytes of register save area.
+ FuncInfo->setVarArgsFrameOffset(ArgOffset + ArgArea +
+ Subtarget->getStackPointerBias());
+
+ // Save the variable arguments that were passed in registers.
+ // The caller is required to reserve stack space for 6 arguments regardless
+ // of how many arguments were actually passed.
+ SmallVector<SDValue, 8> OutChains;
+ for (; ArgOffset < 6*8; ArgOffset += 8) {
+ unsigned VReg = MF.addLiveIn(SP::I0 + ArgOffset/8, &SP::I64RegsRegClass);
+ SDValue VArg = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
+ int FI = MF.getFrameInfo()->CreateFixedObject(8, ArgOffset + ArgArea, true);
+ OutChains.push_back(DAG.getStore(Chain, DL, VArg,
+ DAG.getFrameIndex(FI, getPointerTy()),
+ MachinePointerInfo::getFixedStack(FI),
+ false, false, 0));
+ }
+
+ if (!OutChains.empty())
+ Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
+ &OutChains[0], OutChains.size());
+
return Chain;
}
SDValue
SparcTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
SmallVectorImpl<SDValue> &InVals) const {
+ if (Subtarget->is64Bit())
+ return LowerCall_64(CLI, InVals);
+ return LowerCall_32(CLI, InVals);
+}
+
+// Lower a call for the 32-bit ABI.
+SDValue
+SparcTargetLowering::LowerCall_32(TargetLowering::CallLoweringInfo &CLI,
+ SmallVectorImpl<SDValue> &InVals) const {
SelectionDAG &DAG = CLI.DAG;
- DebugLoc &dl = CLI.DL;
+ SDLoc &dl = CLI.DL;
SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
- //Create local copies for byval args.
+ // Create local copies for byval args.
SmallVector<SDValue, 8> ByValArgs;
for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
ISD::ArgFlagsTy Flags = Outs[i].Flags;
SDValue SizeNode = DAG.getConstant(Size, MVT::i32);
Chain = DAG.getMemcpy(Chain, dl, FIPtr, Arg, SizeNode, Align,
- false, //isVolatile,
- (Size <= 32), //AlwaysInline if size <= 32
+ false, // isVolatile,
+ (Size <= 32), // AlwaysInline if size <= 32
MachinePointerInfo(), MachinePointerInfo());
ByValArgs.push_back(FIPtr);
}
- Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(ArgsSize, true));
+ Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(ArgsSize, true),
+ dl);
SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
SmallVector<SDValue, 8> MemOpChains;
ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
- //Use local copy if it is a byval arg.
+ // Use local copy if it is a byval arg.
if (Flags.isByVal())
Arg = ByValArgs[byvalArgIdx++];
if (VA.isMemLoc()) {
unsigned Offset = VA.getLocMemOffset() + StackOffset;
- //if it is double-word aligned, just store.
+ // if it is double-word aligned, just store.
if (Offset % 8 == 0) {
SDValue StackPtr = DAG.getRegister(SP::O6, MVT::i32);
SDValue PtrOff = DAG.getIntPtrConstant(Offset);
if (NextVA.isRegLoc()) {
RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), Lo));
} else {
- //Store the low part in stack.
+ // Store the low part in stack.
unsigned Offset = NextVA.getLocMemOffset() + StackOffset;
SDValue StackPtr = DAG.getRegister(SP::O6, MVT::i32);
SDValue PtrOff = DAG.getIntPtrConstant(Offset);
// stuck together.
SDValue InFlag;
for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
- unsigned Reg = RegsToPass[i].first;
- // Remap I0->I7 -> O0->O7.
- if (Reg >= SP::I0 && Reg <= SP::I7)
- Reg = Reg-SP::I0+SP::O0;
-
+ unsigned Reg = toCallerWindow(RegsToPass[i].first);
Chain = DAG.getCopyToReg(Chain, dl, Reg, RegsToPass[i].second, InFlag);
InFlag = Chain.getValue(1);
}
Ops.push_back(Callee);
if (hasStructRetAttr)
Ops.push_back(DAG.getTargetConstant(SRetArgSize, MVT::i32));
- for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
- unsigned Reg = RegsToPass[i].first;
- if (Reg >= SP::I0 && Reg <= SP::I7)
- Reg = Reg-SP::I0+SP::O0;
-
- Ops.push_back(DAG.getRegister(Reg, RegsToPass[i].second.getValueType()));
- }
+ for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
+ Ops.push_back(DAG.getRegister(toCallerWindow(RegsToPass[i].first),
+ RegsToPass[i].second.getValueType()));
if (InFlag.getNode())
Ops.push_back(InFlag);
InFlag = Chain.getValue(1);
Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(ArgsSize, true),
- DAG.getIntPtrConstant(0, true), InFlag);
+ DAG.getIntPtrConstant(0, true), InFlag, dl);
InFlag = Chain.getValue(1);
// Assign locations to each value returned by this call.
// Copy all of the result registers out of their specified physreg.
for (unsigned i = 0; i != RVLocs.size(); ++i) {
- unsigned Reg = RVLocs[i].getLocReg();
-
- // Remap I0->I7 -> O0->O7.
- if (Reg >= SP::I0 && Reg <= SP::I7)
- Reg = Reg-SP::I0+SP::O0;
-
- Chain = DAG.getCopyFromReg(Chain, dl, Reg,
+ Chain = DAG.getCopyFromReg(Chain, dl, toCallerWindow(RVLocs[i].getLocReg()),
RVLocs[i].getValVT(), InFlag).getValue(1);
InFlag = Chain.getValue(2);
InVals.push_back(Chain.getValue(0));
return getDataLayout()->getTypeAllocSize(ElementTy);
}
+
+// Fixup floating point arguments in the ... part of a varargs call.
+//
+// The SPARC v9 ABI requires that floating point arguments are treated the same
+// as integers when calling a varargs function. This does not apply to the
+// fixed arguments that are part of the function's prototype.
+//
+// This function post-processes a CCValAssign array created by
+// AnalyzeCallOperands().
+static void fixupVariableFloatArgs(SmallVectorImpl<CCValAssign> &ArgLocs,
+ ArrayRef<ISD::OutputArg> Outs) {
+ for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
+ const CCValAssign &VA = ArgLocs[i];
+ // FIXME: What about f32 arguments? C promotes them to f64 when calling
+ // varargs functions.
+ if (!VA.isRegLoc() || VA.getLocVT() != MVT::f64)
+ continue;
+ // The fixed arguments to a varargs function still go in FP registers.
+ if (Outs[VA.getValNo()].IsFixed)
+ continue;
+
+ // This floating point argument should be reassigned.
+ CCValAssign NewVA;
+
+ // Determine the offset into the argument array.
+ unsigned Offset = 8 * (VA.getLocReg() - SP::D0);
+ assert(Offset < 16*8 && "Offset out of range, bad register enum?");
+
+ if (Offset < 6*8) {
+ // This argument should go in %i0-%i5.
+ unsigned IReg = SP::I0 + Offset/8;
+ // Full register, just bitconvert into i64.
+ NewVA = CCValAssign::getReg(VA.getValNo(), VA.getValVT(),
+ IReg, MVT::i64, CCValAssign::BCvt);
+ } else {
+ // This needs to go to memory, we're out of integer registers.
+ NewVA = CCValAssign::getMem(VA.getValNo(), VA.getValVT(),
+ Offset, VA.getLocVT(), VA.getLocInfo());
+ }
+ ArgLocs[i] = NewVA;
+ }
+}
+
+// Lower a call for the 64-bit ABI.
+SDValue
+SparcTargetLowering::LowerCall_64(TargetLowering::CallLoweringInfo &CLI,
+ SmallVectorImpl<SDValue> &InVals) const {
+ SelectionDAG &DAG = CLI.DAG;
+ SDLoc DL = CLI.DL;
+ SDValue Chain = CLI.Chain;
+
+ // Analyze operands of the call, assigning locations to each operand.
+ SmallVector<CCValAssign, 16> ArgLocs;
+ CCState CCInfo(CLI.CallConv, CLI.IsVarArg, DAG.getMachineFunction(),
+ DAG.getTarget(), ArgLocs, *DAG.getContext());
+ CCInfo.AnalyzeCallOperands(CLI.Outs, CC_Sparc64);
+
+ // Get the size of the outgoing arguments stack space requirement.
+ // The stack offset computed by CC_Sparc64 includes all arguments.
+ // Called functions expect 6 argument words to exist in the stack frame, used
+ // or not.
+ unsigned ArgsSize = std::max(6*8u, CCInfo.getNextStackOffset());
+
+ // Keep stack frames 16-byte aligned.
+ ArgsSize = RoundUpToAlignment(ArgsSize, 16);
+
+ // Varargs calls require special treatment.
+ if (CLI.IsVarArg)
+ fixupVariableFloatArgs(ArgLocs, CLI.Outs);
+
+ // Adjust the stack pointer to make room for the arguments.
+ // FIXME: Use hasReservedCallFrame to avoid %sp adjustments around all calls
+ // with more than 6 arguments.
+ Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(ArgsSize, true),
+ DL);
+
+ // Collect the set of registers to pass to the function and their values.
+ // This will be emitted as a sequence of CopyToReg nodes glued to the call
+ // instruction.
+ SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
+
+ // Collect chains from all the memory opeations that copy arguments to the
+ // stack. They must follow the stack pointer adjustment above and precede the
+ // call instruction itself.
+ SmallVector<SDValue, 8> MemOpChains;
+
+ for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
+ const CCValAssign &VA = ArgLocs[i];
+ SDValue Arg = CLI.OutVals[i];
+
+ // Promote the value if needed.
+ switch (VA.getLocInfo()) {
+ default:
+ llvm_unreachable("Unknown location info!");
+ case CCValAssign::Full:
+ break;
+ case CCValAssign::SExt:
+ Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
+ break;
+ case CCValAssign::ZExt:
+ Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
+ break;
+ case CCValAssign::AExt:
+ Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
+ break;
+ case CCValAssign::BCvt:
+ Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
+ break;
+ }
+
+ if (VA.isRegLoc()) {
+ // The custom bit on an i32 return value indicates that it should be
+ // passed in the high bits of the register.
+ if (VA.getValVT() == MVT::i32 && VA.needsCustom()) {
+ Arg = DAG.getNode(ISD::SHL, DL, MVT::i64, Arg,
+ DAG.getConstant(32, MVT::i32));
+
+ // The next value may go in the low bits of the same register.
+ // Handle both at once.
+ if (i+1 < ArgLocs.size() && ArgLocs[i+1].isRegLoc() &&
+ ArgLocs[i+1].getLocReg() == VA.getLocReg()) {
+ SDValue NV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64,
+ CLI.OutVals[i+1]);
+ Arg = DAG.getNode(ISD::OR, DL, MVT::i64, Arg, NV);
+ // Skip the next value, it's already done.
+ ++i;
+ }
+ }
+ RegsToPass.push_back(std::make_pair(toCallerWindow(VA.getLocReg()), Arg));
+ continue;
+ }
+
+ assert(VA.isMemLoc());
+
+ // Create a store off the stack pointer for this argument.
+ SDValue StackPtr = DAG.getRegister(SP::O6, getPointerTy());
+ // The argument area starts at %fp+BIAS+128 in the callee frame,
+ // %sp+BIAS+128 in ours.
+ SDValue PtrOff = DAG.getIntPtrConstant(VA.getLocMemOffset() +
+ Subtarget->getStackPointerBias() +
+ 128);
+ PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
+ MemOpChains.push_back(DAG.getStore(Chain, DL, Arg, PtrOff,
+ MachinePointerInfo(),
+ false, false, 0));
+ }
+
+ // Emit all stores, make sure they occur before the call.
+ if (!MemOpChains.empty())
+ Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
+ &MemOpChains[0], MemOpChains.size());
+
+ // Build a sequence of CopyToReg nodes glued together with token chain and
+ // glue operands which copy the outgoing args into registers. The InGlue is
+ // necessary since all emitted instructions must be stuck together in order
+ // to pass the live physical registers.
+ SDValue InGlue;
+ for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
+ Chain = DAG.getCopyToReg(Chain, DL,
+ RegsToPass[i].first, RegsToPass[i].second, InGlue);
+ InGlue = Chain.getValue(1);
+ }
+
+ // If the callee is a GlobalAddress node (quite common, every direct call is)
+ // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
+ // Likewise ExternalSymbol -> TargetExternalSymbol.
+ SDValue Callee = CLI.Callee;
+ if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
+ Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, getPointerTy());
+ else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee))
+ Callee = DAG.getTargetExternalSymbol(E->getSymbol(), getPointerTy());
+
+ // Build the operands for the call instruction itself.
+ SmallVector<SDValue, 8> Ops;
+ Ops.push_back(Chain);
+ Ops.push_back(Callee);
+ for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
+ Ops.push_back(DAG.getRegister(RegsToPass[i].first,
+ RegsToPass[i].second.getValueType()));
+
+ // Make sure the CopyToReg nodes are glued to the call instruction which
+ // consumes the registers.
+ if (InGlue.getNode())
+ Ops.push_back(InGlue);
+
+ // Now the call itself.
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
+ Chain = DAG.getNode(SPISD::CALL, DL, NodeTys, &Ops[0], Ops.size());
+ InGlue = Chain.getValue(1);
+
+ // Revert the stack pointer immediately after the call.
+ Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(ArgsSize, true),
+ DAG.getIntPtrConstant(0, true), InGlue, DL);
+ InGlue = Chain.getValue(1);
+
+ // Now extract the return values. This is more or less the same as
+ // LowerFormalArguments_64.
+
+ // Assign locations to each value returned by this call.
+ SmallVector<CCValAssign, 16> RVLocs;
+ CCState RVInfo(CLI.CallConv, CLI.IsVarArg, DAG.getMachineFunction(),
+ DAG.getTarget(), RVLocs, *DAG.getContext());
+ RVInfo.AnalyzeCallResult(CLI.Ins, CC_Sparc64);
+
+ // Copy all of the result registers out of their specified physreg.
+ for (unsigned i = 0; i != RVLocs.size(); ++i) {
+ CCValAssign &VA = RVLocs[i];
+ unsigned Reg = toCallerWindow(VA.getLocReg());
+
+ // When returning 'inreg {i32, i32 }', two consecutive i32 arguments can
+ // reside in the same register in the high and low bits. Reuse the
+ // CopyFromReg previous node to avoid duplicate copies.
+ SDValue RV;
+ if (RegisterSDNode *SrcReg = dyn_cast<RegisterSDNode>(Chain.getOperand(1)))
+ if (SrcReg->getReg() == Reg && Chain->getOpcode() == ISD::CopyFromReg)
+ RV = Chain.getValue(0);
+
+ // But usually we'll create a new CopyFromReg for a different register.
+ if (!RV.getNode()) {
+ RV = DAG.getCopyFromReg(Chain, DL, Reg, RVLocs[i].getLocVT(), InGlue);
+ Chain = RV.getValue(1);
+ InGlue = Chain.getValue(2);
+ }
+
+ // Get the high bits for i32 struct elements.
+ if (VA.getValVT() == MVT::i32 && VA.needsCustom())
+ RV = DAG.getNode(ISD::SRL, DL, VA.getLocVT(), RV,
+ DAG.getConstant(32, MVT::i32));
+
+ // The callee promoted the return value, so insert an Assert?ext SDNode so
+ // we won't promote the value again in this function.
+ switch (VA.getLocInfo()) {
+ case CCValAssign::SExt:
+ RV = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), RV,
+ DAG.getValueType(VA.getValVT()));
+ break;
+ case CCValAssign::ZExt:
+ RV = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), RV,
+ DAG.getValueType(VA.getValVT()));
+ break;
+ default:
+ break;
+ }
+
+ // Truncate the register down to the return value type.
+ if (VA.isExtInLoc())
+ RV = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), RV);
+
+ InVals.push_back(RV);
+ }
+
+ return Chain;
+}
+
//===----------------------------------------------------------------------===//
// TargetLowering Implementation
//===----------------------------------------------------------------------===//
setTruncStoreAction(MVT::f64, MVT::f32, Expand);
// Custom legalize GlobalAddress nodes into LO/HI parts.
- setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
- setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
- setOperationAction(ISD::ConstantPool , MVT::i32, Custom);
+ setOperationAction(ISD::GlobalAddress, getPointerTy(), Custom);
+ setOperationAction(ISD::GlobalTLSAddress, getPointerTy(), Custom);
+ setOperationAction(ISD::ConstantPool, getPointerTy(), Custom);
+ setOperationAction(ISD::BlockAddress, getPointerTy(), Custom);
// Sparc doesn't have sext_inreg, replace them with shl/sra
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
if (Subtarget->is64Bit()) {
+ setOperationAction(ISD::BITCAST, MVT::f64, Expand);
+ setOperationAction(ISD::BITCAST, MVT::i64, Expand);
+ setOperationAction(ISD::SELECT, MVT::i64, Expand);
+ setOperationAction(ISD::SETCC, MVT::i64, Expand);
setOperationAction(ISD::BR_CC, MVT::i64, Custom);
setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
}
// FIXME: There are instructions available for ATOMIC_FENCE
// on SparcV8 and later.
- setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
+ if (!Subtarget->isV9()) {
+ // SparcV8 does not have FNEGD and FABSD.
+ setOperationAction(ISD::FNEG, MVT::f64, Custom);
+ setOperationAction(ISD::FABS, MVT::f64, Custom);
+ }
+
setOperationAction(ISD::FSIN , MVT::f64, Expand);
setOperationAction(ISD::FCOS , MVT::f64, Expand);
setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
setStackPointerRegisterToSaveRestore(SP::O6);
- if (TM.getSubtarget<SparcSubtarget>().isV9())
+ if (Subtarget->isV9())
setOperationAction(ISD::CTPOP, MVT::i32, Legal);
setMinFunctionAlignment(2);
/// isMaskedValueZeroForTargetNode - Return true if 'Op & Mask' is known to
/// be zero. Op is expected to be a target specific node. Used by DAG
/// combiner.
-void SparcTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
- APInt &KnownZero,
- APInt &KnownOne,
- const SelectionDAG &DAG,
- unsigned Depth) const {
+void SparcTargetLowering::computeMaskedBitsForTargetNode
+ (const SDValue Op,
+ APInt &KnownZero,
+ APInt &KnownOne,
+ const SelectionDAG &DAG,
+ unsigned Depth) const {
APInt KnownZero2, KnownOne2;
KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
}
}
+// Convert to a target node and set target flags.
+SDValue SparcTargetLowering::withTargetFlags(SDValue Op, unsigned TF,
+ SelectionDAG &DAG) const {
+ if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op))
+ return DAG.getTargetGlobalAddress(GA->getGlobal(),
+ SDLoc(GA),
+ GA->getValueType(0),
+ GA->getOffset(), TF);
+
+ if (const ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op))
+ return DAG.getTargetConstantPool(CP->getConstVal(),
+ CP->getValueType(0),
+ CP->getAlignment(),
+ CP->getOffset(), TF);
+
+ if (const BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(Op))
+ return DAG.getTargetBlockAddress(BA->getBlockAddress(),
+ Op.getValueType(),
+ 0,
+ TF);
+
+ if (const ExternalSymbolSDNode *ES = dyn_cast<ExternalSymbolSDNode>(Op))
+ return DAG.getTargetExternalSymbol(ES->getSymbol(),
+ ES->getValueType(0), TF);
+
+ llvm_unreachable("Unhandled address SDNode");
+}
+
+// Split Op into high and low parts according to HiTF and LoTF.
+// Return an ADD node combining the parts.
+SDValue SparcTargetLowering::makeHiLoPair(SDValue Op,
+ unsigned HiTF, unsigned LoTF,
+ SelectionDAG &DAG) const {
+ SDLoc DL(Op);
+ EVT VT = Op.getValueType();
+ SDValue Hi = DAG.getNode(SPISD::Hi, DL, VT, withTargetFlags(Op, HiTF, DAG));
+ SDValue Lo = DAG.getNode(SPISD::Lo, DL, VT, withTargetFlags(Op, LoTF, DAG));
+ return DAG.getNode(ISD::ADD, DL, VT, Hi, Lo);
+}
+
+// Build SDNodes for producing an address from a GlobalAddress, ConstantPool,
+// or ExternalSymbol SDNode.
+SDValue SparcTargetLowering::makeAddress(SDValue Op, SelectionDAG &DAG) const {
+ SDLoc DL(Op);
+ EVT VT = getPointerTy();
+
+ // Handle PIC mode first.
+ if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
+ // This is the pic32 code model, the GOT is known to be smaller than 4GB.
+ SDValue HiLo = makeHiLoPair(Op, SPII::MO_HI, SPII::MO_LO, DAG);
+ SDValue GlobalBase = DAG.getNode(SPISD::GLOBAL_BASE_REG, DL, VT);
+ SDValue AbsAddr = DAG.getNode(ISD::ADD, DL, VT, GlobalBase, HiLo);
+ return DAG.getLoad(VT, DL, DAG.getEntryNode(), AbsAddr,
+ MachinePointerInfo::getGOT(), false, false, false, 0);
+ }
+
+ // This is one of the absolute code models.
+ switch(getTargetMachine().getCodeModel()) {
+ default:
+ llvm_unreachable("Unsupported absolute code model");
+ case CodeModel::Small:
+ // abs32.
+ return makeHiLoPair(Op, SPII::MO_HI, SPII::MO_LO, DAG);
+ case CodeModel::Medium: {
+ // abs44.
+ SDValue H44 = makeHiLoPair(Op, SPII::MO_H44, SPII::MO_M44, DAG);
+ H44 = DAG.getNode(ISD::SHL, DL, VT, H44, DAG.getConstant(12, MVT::i32));
+ SDValue L44 = withTargetFlags(Op, SPII::MO_L44, DAG);
+ L44 = DAG.getNode(SPISD::Lo, DL, VT, L44);
+ return DAG.getNode(ISD::ADD, DL, VT, H44, L44);
+ }
+ case CodeModel::Large: {
+ // abs64.
+ SDValue Hi = makeHiLoPair(Op, SPII::MO_HH, SPII::MO_HM, DAG);
+ Hi = DAG.getNode(ISD::SHL, DL, VT, Hi, DAG.getConstant(32, MVT::i32));
+ SDValue Lo = makeHiLoPair(Op, SPII::MO_HI, SPII::MO_LO, DAG);
+ return DAG.getNode(ISD::ADD, DL, VT, Hi, Lo);
+ }
+ }
+}
+
SDValue SparcTargetLowering::LowerGlobalAddress(SDValue Op,
SelectionDAG &DAG) const {
- const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
- // FIXME there isn't really any debug info here
- DebugLoc dl = Op.getDebugLoc();
- SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32);
- SDValue Hi = DAG.getNode(SPISD::Hi, dl, MVT::i32, GA);
- SDValue Lo = DAG.getNode(SPISD::Lo, dl, MVT::i32, GA);
-
- if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
- return DAG.getNode(ISD::ADD, dl, MVT::i32, Lo, Hi);
-
- SDValue GlobalBase = DAG.getNode(SPISD::GLOBAL_BASE_REG, dl,
- getPointerTy());
- SDValue RelAddr = DAG.getNode(ISD::ADD, dl, MVT::i32, Lo, Hi);
- SDValue AbsAddr = DAG.getNode(ISD::ADD, dl, MVT::i32,
- GlobalBase, RelAddr);
- return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
- AbsAddr, MachinePointerInfo(), false, false, false, 0);
+ return makeAddress(Op, DAG);
}
SDValue SparcTargetLowering::LowerConstantPool(SDValue Op,
SelectionDAG &DAG) const {
- ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
- // FIXME there isn't really any debug info here
- DebugLoc dl = Op.getDebugLoc();
- const Constant *C = N->getConstVal();
- SDValue CP = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment());
- SDValue Hi = DAG.getNode(SPISD::Hi, dl, MVT::i32, CP);
- SDValue Lo = DAG.getNode(SPISD::Lo, dl, MVT::i32, CP);
- if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
- return DAG.getNode(ISD::ADD, dl, MVT::i32, Lo, Hi);
-
- SDValue GlobalBase = DAG.getNode(SPISD::GLOBAL_BASE_REG, dl,
- getPointerTy());
- SDValue RelAddr = DAG.getNode(ISD::ADD, dl, MVT::i32, Lo, Hi);
- SDValue AbsAddr = DAG.getNode(ISD::ADD, dl, MVT::i32,
- GlobalBase, RelAddr);
- return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
- AbsAddr, MachinePointerInfo(), false, false, false, 0);
+ return makeAddress(Op, DAG);
+}
+
+SDValue SparcTargetLowering::LowerBlockAddress(SDValue Op,
+ SelectionDAG &DAG) const {
+ return makeAddress(Op, DAG);
}
static SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc dl(Op);
// Convert the fp value to integer in an FP register.
assert(Op.getValueType() == MVT::i32);
Op = DAG.getNode(SPISD::FTOI, dl, MVT::f32, Op.getOperand(0));
}
static SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc dl(Op);
assert(Op.getOperand(0).getValueType() == MVT::i32);
SDValue Tmp = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
// Convert the int value to FP in an FP register.
SDValue LHS = Op.getOperand(2);
SDValue RHS = Op.getOperand(3);
SDValue Dest = Op.getOperand(4);
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc dl(Op);
unsigned Opc, SPCC = ~0U;
// If this is a br_cc of a "setcc", and if the setcc got lowered into
// Get the condition flag.
SDValue CompareFlag;
if (LHS.getValueType().isInteger()) {
- EVT VTs[] = { LHS.getValueType(), MVT::Glue };
- SDValue Ops[2] = { LHS, RHS };
- CompareFlag = DAG.getNode(SPISD::CMPICC, dl, VTs, Ops, 2).getValue(1);
+ CompareFlag = DAG.getNode(SPISD::CMPICC, dl, MVT::Glue, LHS, RHS);
if (SPCC == ~0U) SPCC = IntCondCCodeToICC(CC);
// 32-bit compares use the icc flags, 64-bit uses the xcc flags.
Opc = LHS.getValueType() == MVT::i32 ? SPISD::BRICC : SPISD::BRXCC;
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
SDValue TrueVal = Op.getOperand(2);
SDValue FalseVal = Op.getOperand(3);
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc dl(Op);
unsigned Opc, SPCC = ~0U;
// If this is a select_cc of a "setcc", and if the setcc got lowered into
SDValue CompareFlag;
if (LHS.getValueType().isInteger()) {
- // subcc returns a value
- EVT VTs[] = { LHS.getValueType(), MVT::Glue };
- SDValue Ops[2] = { LHS, RHS };
- CompareFlag = DAG.getNode(SPISD::CMPICC, dl, VTs, Ops, 2).getValue(1);
+ CompareFlag = DAG.getNode(SPISD::CMPICC, dl, MVT::Glue, LHS, RHS);
Opc = LHS.getValueType() == MVT::i32 ?
SPISD::SELECT_ICC : SPISD::SELECT_XCC;
if (SPCC == ~0U) SPCC = IntCondCCodeToICC(CC);
MachineFunction &MF = DAG.getMachineFunction();
SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>();
+ // Need frame address to find the address of VarArgsFrameIndex.
+ MF.getFrameInfo()->setFrameAddressIsTaken(true);
+
// vastart just stores the address of the VarArgsFrameIndex slot into the
// memory location argument.
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc DL(Op);
SDValue Offset =
- DAG.getNode(ISD::ADD, dl, MVT::i32,
- DAG.getRegister(SP::I6, MVT::i32),
- DAG.getConstant(FuncInfo->getVarArgsFrameOffset(),
- MVT::i32));
+ DAG.getNode(ISD::ADD, DL, TLI.getPointerTy(),
+ DAG.getRegister(SP::I6, TLI.getPointerTy()),
+ DAG.getIntPtrConstant(FuncInfo->getVarArgsFrameOffset()));
const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
- return DAG.getStore(Op.getOperand(0), dl, Offset, Op.getOperand(1),
+ return DAG.getStore(Op.getOperand(0), DL, Offset, Op.getOperand(1),
MachinePointerInfo(SV), false, false, 0);
}
EVT VT = Node->getValueType(0);
SDValue InChain = Node->getOperand(0);
SDValue VAListPtr = Node->getOperand(1);
+ EVT PtrVT = VAListPtr.getValueType();
const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
- DebugLoc dl = Node->getDebugLoc();
- SDValue VAList = DAG.getLoad(MVT::i32, dl, InChain, VAListPtr,
+ SDLoc DL(Node);
+ SDValue VAList = DAG.getLoad(PtrVT, DL, InChain, VAListPtr,
MachinePointerInfo(SV), false, false, false, 0);
- // Increment the pointer, VAList, to the next vaarg
- SDValue NextPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, VAList,
- DAG.getConstant(VT.getSizeInBits()/8,
- MVT::i32));
- // Store the incremented VAList to the legalized pointer
- InChain = DAG.getStore(VAList.getValue(1), dl, NextPtr,
+ // Increment the pointer, VAList, to the next vaarg.
+ SDValue NextPtr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
+ DAG.getIntPtrConstant(VT.getSizeInBits()/8));
+ // Store the incremented VAList to the legalized pointer.
+ InChain = DAG.getStore(VAList.getValue(1), DL, NextPtr,
VAListPtr, MachinePointerInfo(SV), false, false, 0);
- // Load the actual argument out of the pointer VAList, unless this is an
- // f64 load.
- if (VT != MVT::f64)
- return DAG.getLoad(VT, dl, InChain, VAList, MachinePointerInfo(),
- false, false, false, 0);
-
- // Otherwise, load it as i64, then do a bitconvert.
- SDValue V = DAG.getLoad(MVT::i64, dl, InChain, VAList, MachinePointerInfo(),
- false, false, false, 0);
-
- // Bit-Convert the value to f64.
- SDValue Ops[2] = {
- DAG.getNode(ISD::BITCAST, dl, MVT::f64, V),
- V.getValue(1)
- };
- return DAG.getMergeValues(Ops, 2, dl);
+ // Load the actual argument out of the pointer VAList.
+ // We can't count on greater alignment than the word size.
+ return DAG.getLoad(VT, DL, InChain, VAList, MachinePointerInfo(),
+ false, false, false,
+ std::min(PtrVT.getSizeInBits(), VT.getSizeInBits())/8);
}
static SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
SDValue Chain = Op.getOperand(0); // Legalize the chain.
SDValue Size = Op.getOperand(1); // Legalize the size.
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc dl(Op);
unsigned SPReg = SP::O6;
SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, MVT::i32);
static SDValue getFLUSHW(SDValue Op, SelectionDAG &DAG) {
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc dl(Op);
SDValue Chain = DAG.getNode(SPISD::FLUSHW,
dl, MVT::Other, DAG.getEntryNode());
return Chain;
MFI->setFrameAddressIsTaken(true);
EVT VT = Op.getValueType();
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc dl(Op);
unsigned FrameReg = SP::I6;
uint64_t depth = Op.getConstantOperandVal(0);
MFI->setReturnAddressIsTaken(true);
EVT VT = Op.getValueType();
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc dl(Op);
unsigned RetReg = SP::I7;
uint64_t depth = Op.getConstantOperandVal(0);
if (depth == 0)
RetAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, RetReg, VT);
else {
+ // Need frame address to find return address of the caller.
+ MFI->setFrameAddressIsTaken(true);
+
// flush first to make sure the windowed registers' values are in stack
SDValue Chain = getFLUSHW(Op, DAG);
RetAddr = DAG.getCopyFromReg(Chain, dl, SP::I6, VT);
return RetAddr;
}
+static SDValue LowerF64Op(SDValue Op, SelectionDAG &DAG)
+{
+ SDLoc dl(Op);
+
+ assert(Op.getValueType() == MVT::f64 && "LowerF64Op called on non-double!");
+ assert(Op.getOpcode() == ISD::FNEG || Op.getOpcode() == ISD::FABS);
+
+ // Lower fneg/fabs on f64 to fneg/fabs on f32.
+ // fneg f64 => fneg f32:sub_even, fmov f32:sub_odd.
+ // fabs f64 => fabs f32:sub_even, fmov f32:sub_odd.
+
+ SDValue SrcReg64 = Op.getOperand(0);
+ SDValue Hi32 = DAG.getTargetExtractSubreg(SP::sub_even, dl, MVT::f32,
+ SrcReg64);
+ SDValue Lo32 = DAG.getTargetExtractSubreg(SP::sub_odd, dl, MVT::f32,
+ SrcReg64);
+
+ Hi32 = DAG.getNode(Op.getOpcode(), dl, MVT::f32, Hi32);
+
+ SDValue DstReg64 = SDValue(DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF,
+ dl, MVT::f64), 0);
+ DstReg64 = DAG.getTargetInsertSubreg(SP::sub_even, dl, MVT::f64,
+ DstReg64, Hi32);
+ DstReg64 = DAG.getTargetInsertSubreg(SP::sub_odd, dl, MVT::f64,
+ DstReg64, Lo32);
+ return DstReg64;
+}
+
SDValue SparcTargetLowering::
LowerOperation(SDValue Op, SelectionDAG &DAG) const {
switch (Op.getOpcode()) {
default: llvm_unreachable("Should not custom lower this!");
+
+ case ISD::FNEG:
+ case ISD::FABS: return LowerF64Op(Op, DAG);
+
case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
case ISD::GlobalTLSAddress:
llvm_unreachable("TLS not implemented for Sparc.");
case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
+ case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
std::pair<unsigned, const TargetRegisterClass*>
SparcTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
- EVT VT) const {
+ MVT VT) const {
if (Constraint.size() == 1) {
switch (Constraint[0]) {
case 'r':
projects / oota-llvm.git / blobdiff