#include "llvm/Analysis/DebugInfo.h"
#include "llvm/CodeGen/Analysis.h"
#include "llvm/CodeGen/MachineFunction.h"
-#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
-#include "llvm/CodeGen/MachineModuleInfo.h"
-#include "llvm/CodeGen/PseudoSourceValue.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/Target/TargetFrameLowering.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetMachine.h"
-#include "llvm/Target/TargetOptions.h"
#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
-#include "llvm/Function.h"
-#include "llvm/GlobalVariable.h"
#include "llvm/LLVMContext.h"
-#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
const TargetMachine &TM;
const TargetLowering &TLI;
SelectionDAG &DAG;
- CodeGenOpt::Level OptLevel;
// Libcall insertion helpers.
/// against each other, including inserted libcalls.
SDValue LastCALLSEQ_END;
- enum LegalizeAction {
- Legal, // The target natively supports this operation.
- Promote, // This operation should be executed in a larger type.
- Expand // Try to expand this to other ops, otherwise use a libcall.
- };
-
- /// ValueTypeActions - This is a bitvector that contains two bits for each
- /// value type, where the two bits correspond to the LegalizeAction enum.
- /// This can be queried with "getTypeAction(VT)".
- TargetLowering::ValueTypeActionImpl ValueTypeActions;
+ /// IsLegalizingCall - This member is used *only* for purposes of providing
+ /// helpful assertions that a libcall isn't created while another call is
+ /// being legalized (which could lead to non-serialized call sequences).
+ bool IsLegalizingCall;
/// LegalizedNodes - For nodes that are of legal width, and that have more
/// than one use, this map indicates what regularized operand to use. This
// If someone requests legalization of the new node, return itself.
if (From != To)
LegalizedNodes.insert(std::make_pair(To, To));
- }
-
-public:
- SelectionDAGLegalize(SelectionDAG &DAG, CodeGenOpt::Level ol);
- /// getTypeAction - Return how we should legalize values of this type, either
- /// it is already legal or we need to expand it into multiple registers of
- /// smaller integer type, or we need to promote it to a larger type.
- LegalizeAction getTypeAction(EVT VT) const {
- return (LegalizeAction)ValueTypeActions.getTypeAction(VT);
+ // Transfer SDDbgValues.
+ DAG.TransferDbgValues(From, To);
}
- /// isTypeLegal - Return true if this type is legal on this target.
- ///
- bool isTypeLegal(EVT VT) const {
- return getTypeAction(VT) == Legal;
- }
+public:
+ explicit SelectionDAGLegalize(SelectionDAG &DAG);
void LegalizeDAG();
private:
- /// LegalizeOp - We know that the specified value has a legal type.
- /// Recursively ensure that the operands have legal types, then return the
- /// result.
+ /// LegalizeOp - Return a legal replacement for the given operation, with
+ /// all legal operands.
SDValue LegalizeOp(SDValue O);
SDValue OptimizeFloatStore(StoreSDNode *ST);
DebugLoc dl);
SDValue ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned);
+ SDValue ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, const SDValue *Ops,
+ unsigned NumOps, bool isSigned, DebugLoc dl);
+
std::pair<SDValue, SDValue> ExpandChainLibCall(RTLIB::Libcall LC,
SDNode *Node, bool isSigned);
SDValue ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32,
RTLIB::Libcall Call_I32,
RTLIB::Libcall Call_I64,
RTLIB::Libcall Call_I128);
+ void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT, DebugLoc dl);
SDValue ExpandBUILD_VECTOR(SDNode *Node);
SDValue ExpandBitCount(unsigned Opc, SDValue Op, DebugLoc dl);
SDValue ExpandExtractFromVectorThroughStack(SDValue Op);
+ SDValue ExpandInsertToVectorThroughStack(SDValue Op);
SDValue ExpandVectorBuildThroughStack(SDNode* Node);
std::pair<SDValue, SDValue> ExpandAtomic(SDNode *Node);
return DAG.getVectorShuffle(NVT, dl, N1, N2, &NewMask[0]);
}
-SelectionDAGLegalize::SelectionDAGLegalize(SelectionDAG &dag,
- CodeGenOpt::Level ol)
+SelectionDAGLegalize::SelectionDAGLegalize(SelectionDAG &dag)
: TM(dag.getTarget()), TLI(dag.getTargetLoweringInfo()),
- DAG(dag), OptLevel(ol),
- ValueTypeActions(TLI.getValueTypeActions()) {
- assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE &&
- "Too many value types for ValueTypeActions to hold!");
+ DAG(dag) {
}
void SelectionDAGLegalize::LegalizeDAG() {
LastCALLSEQ_END = DAG.getEntryNode();
+ IsLegalizingCall = false;
// The legalize process is inherently a bottom-up recursive process (users
// legalize their uses before themselves). Given infinite stack space, we
// smaller type.
TLI.isLoadExtLegal(ISD::EXTLOAD, SVT) &&
TLI.ShouldShrinkFPConstant(OrigVT)) {
- const Type *SType = SVT.getTypeForEVT(*DAG.getContext());
+ Type *SType = SVT.getTypeForEVT(*DAG.getContext());
LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));
VT = SVT;
Extend = true;
SDValue CPIdx = DAG.getConstantPool(LLVMC, TLI.getPointerTy());
unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
if (Extend)
- return DAG.getExtLoad(ISD::EXTLOAD, OrigVT, dl,
+ return DAG.getExtLoad(ISD::EXTLOAD, dl, OrigVT,
DAG.getEntryNode(),
CPIdx, MachinePointerInfo::getConstantPool(),
VT, false, false, Alignment);
SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
return DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
ST->isVolatile(), ST->isNonTemporal(), Alignment);
- } else {
- // Do a (aligned) store to a stack slot, then copy from the stack slot
- // to the final destination using (unaligned) integer loads and stores.
- EVT StoredVT = ST->getMemoryVT();
- EVT RegVT =
- TLI.getRegisterType(*DAG.getContext(),
- EVT::getIntegerVT(*DAG.getContext(),
- StoredVT.getSizeInBits()));
- unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
- unsigned RegBytes = RegVT.getSizeInBits() / 8;
- unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
-
- // Make sure the stack slot is also aligned for the register type.
- SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
-
- // Perform the original store, only redirected to the stack slot.
- SDValue Store = DAG.getTruncStore(Chain, dl,
- Val, StackPtr, MachinePointerInfo(),
- StoredVT, false, false, 0);
- SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
- SmallVector<SDValue, 8> Stores;
- unsigned Offset = 0;
-
- // Do all but one copies using the full register width.
- for (unsigned i = 1; i < NumRegs; i++) {
- // Load one integer register's worth from the stack slot.
- SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr,
- MachinePointerInfo(),
- false, false, 0);
- // Store it to the final location. Remember the store.
- Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
- ST->getPointerInfo().getWithOffset(Offset),
- ST->isVolatile(), ST->isNonTemporal(),
- MinAlign(ST->getAlignment(), Offset)));
- // Increment the pointers.
- Offset += RegBytes;
- StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
- Increment);
- Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
- }
+ }
+ // Do a (aligned) store to a stack slot, then copy from the stack slot
+ // to the final destination using (unaligned) integer loads and stores.
+ EVT StoredVT = ST->getMemoryVT();
+ EVT RegVT =
+ TLI.getRegisterType(*DAG.getContext(),
+ EVT::getIntegerVT(*DAG.getContext(),
+ StoredVT.getSizeInBits()));
+ unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
+ unsigned RegBytes = RegVT.getSizeInBits() / 8;
+ unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
- // The last store may be partial. Do a truncating store. On big-endian
- // machines this requires an extending load from the stack slot to ensure
- // that the bits are in the right place.
- EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
- 8 * (StoredBytes - Offset));
+ // Make sure the stack slot is also aligned for the register type.
+ SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
- // Load from the stack slot.
- SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, RegVT, dl, Store, StackPtr,
- MachinePointerInfo(),
- MemVT, false, false, 0);
+ // Perform the original store, only redirected to the stack slot.
+ SDValue Store = DAG.getTruncStore(Chain, dl,
+ Val, StackPtr, MachinePointerInfo(),
+ StoredVT, false, false, 0);
+ SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
+ SmallVector<SDValue, 8> Stores;
+ unsigned Offset = 0;
- Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
- ST->getPointerInfo()
- .getWithOffset(Offset),
- MemVT, ST->isVolatile(),
- ST->isNonTemporal(),
- MinAlign(ST->getAlignment(), Offset)));
- // The order of the stores doesn't matter - say it with a TokenFactor.
- return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
- Stores.size());
+ // Do all but one copies using the full register width.
+ for (unsigned i = 1; i < NumRegs; i++) {
+ // Load one integer register's worth from the stack slot.
+ SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr,
+ MachinePointerInfo(),
+ false, false, 0);
+ // Store it to the final location. Remember the store.
+ Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
+ ST->getPointerInfo().getWithOffset(Offset),
+ ST->isVolatile(), ST->isNonTemporal(),
+ MinAlign(ST->getAlignment(), Offset)));
+ // Increment the pointers.
+ Offset += RegBytes;
+ StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
+ Increment);
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
}
+
+ // The last store may be partial. Do a truncating store. On big-endian
+ // machines this requires an extending load from the stack slot to ensure
+ // that the bits are in the right place.
+ EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
+ 8 * (StoredBytes - Offset));
+
+ // Load from the stack slot.
+ SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Store, StackPtr,
+ MachinePointerInfo(),
+ MemVT, false, false, 0);
+
+ Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
+ ST->getPointerInfo()
+ .getWithOffset(Offset),
+ MemVT, ST->isVolatile(),
+ ST->isNonTemporal(),
+ MinAlign(ST->getAlignment(), Offset)));
+ // The order of the stores doesn't matter - say it with a TokenFactor.
+ return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
+ Stores.size());
}
assert(ST->getMemoryVT().isInteger() &&
!ST->getMemoryVT().isVector() &&
int IncrementSize = NumBits / 8;
// Divide the stored value in two parts.
- SDValue ShiftAmount = DAG.getConstant(NumBits, TLI.getShiftAmountTy());
+ SDValue ShiftAmount = DAG.getConstant(NumBits,
+ TLI.getShiftAmountTy(Val.getValueType()));
SDValue Lo = Val;
SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
// The last copy may be partial. Do an extending load.
EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
8 * (LoadedBytes - Offset));
- SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, RegVT, dl, Chain, Ptr,
+ SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr,
LD->getPointerInfo().getWithOffset(Offset),
MemVT, LD->isVolatile(),
LD->isNonTemporal(),
Stores.size());
// Finally, perform the original load only redirected to the stack slot.
- Load = DAG.getExtLoad(LD->getExtensionType(), VT, dl, TF, StackBase,
+ Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
MachinePointerInfo(), LoadedVT, false, false, 0);
// Callers expect a MERGE_VALUES node.
// Load the value in two parts
SDValue Lo, Hi;
if (TLI.isLittleEndian()) {
- Lo = DAG.getExtLoad(ISD::ZEXTLOAD, VT, dl, Chain, Ptr, LD->getPointerInfo(),
+ Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(),
NewLoadedVT, LD->isVolatile(),
LD->isNonTemporal(), Alignment);
Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
DAG.getConstant(IncrementSize, TLI.getPointerTy()));
- Hi = DAG.getExtLoad(HiExtType, VT, dl, Chain, Ptr,
+ Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr,
LD->getPointerInfo().getWithOffset(IncrementSize),
NewLoadedVT, LD->isVolatile(),
LD->isNonTemporal(), MinAlign(Alignment,IncrementSize));
} else {
- Hi = DAG.getExtLoad(HiExtType, VT, dl, Chain, Ptr, LD->getPointerInfo(),
+ Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(),
NewLoadedVT, LD->isVolatile(),
LD->isNonTemporal(), Alignment);
Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
DAG.getConstant(IncrementSize, TLI.getPointerTy()));
- Lo = DAG.getExtLoad(ISD::ZEXTLOAD, VT, dl, Chain, Ptr,
+ Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr,
LD->getPointerInfo().getWithOffset(IncrementSize),
NewLoadedVT, LD->isVolatile(),
LD->isNonTemporal(), MinAlign(Alignment,IncrementSize));
}
// aggregate the two parts
- SDValue ShiftAmount = DAG.getConstant(NumBits, TLI.getShiftAmountTy());
+ SDValue ShiftAmount = DAG.getConstant(NumBits,
+ TLI.getShiftAmountTy(Hi.getValueType()));
SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount);
Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo);
DebugLoc dl = ST->getDebugLoc();
if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) {
if (CFP->getValueType(0) == MVT::f32 &&
- getTypeAction(MVT::i32) == Legal) {
+ TLI.isTypeLegal(MVT::i32)) {
Tmp3 = DAG.getConstant(CFP->getValueAPF().
bitcastToAPInt().zextOrTrunc(32),
MVT::i32);
if (CFP->getValueType(0) == MVT::f64) {
// If this target supports 64-bit registers, do a single 64-bit store.
- if (getTypeAction(MVT::i64) == Legal) {
+ if (TLI.isTypeLegal(MVT::i64)) {
Tmp3 = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
zextOrTrunc(64), MVT::i64);
return DAG.getStore(Tmp1, dl, Tmp3, Tmp2, ST->getPointerInfo(),
isVolatile, isNonTemporal, Alignment);
}
- if (getTypeAction(MVT::i32) == Legal && !ST->isVolatile()) {
+ if (TLI.isTypeLegal(MVT::i32) && !ST->isVolatile()) {
// Otherwise, if the target supports 32-bit registers, use 2 32-bit
// stores. If the target supports neither 32- nor 64-bits, this
// xform is certainly not worth it.
}
}
}
- return SDValue();
+ return SDValue(0, 0);
}
-/// LegalizeOp - We know that the specified value has a legal type, and
-/// that its operands are legal. Now ensure that the operation itself
-/// is legal, recursively ensuring that the operands' operations remain
-/// legal.
+/// LegalizeOp - Return a legal replacement for the given operation, with
+/// all legal operands.
SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) {
if (Op.getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
return Op;
DebugLoc dl = Node->getDebugLoc();
for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
- assert(getTypeAction(Node->getValueType(i)) == Legal &&
+ assert(TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
+ TargetLowering::TypeLegal &&
"Unexpected illegal type!");
for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
- assert((isTypeLegal(Node->getOperand(i).getValueType()) ||
+ assert((TLI.getTypeAction(*DAG.getContext(),
+ Node->getOperand(i).getValueType()) ==
+ TargetLowering::TypeLegal ||
Node->getOperand(i).getOpcode() == ISD::TargetConstant) &&
"Unexpected illegal type!");
bool isCustom = false;
// Figure out the correct action; the way to query this varies by opcode
- TargetLowering::LegalizeAction Action;
+ TargetLowering::LegalizeAction Action = TargetLowering::Legal;
bool SimpleFinishLegalizing = true;
switch (Node->getOpcode()) {
case ISD::INTRINSIC_W_CHAIN:
Action = TLI.getOperationAction(Node->getOpcode(), InnerType);
break;
}
+ case ISD::ATOMIC_STORE: {
+ Action = TLI.getOperationAction(Node->getOpcode(),
+ Node->getOperand(2).getValueType());
+ break;
+ }
case ISD::SELECT_CC:
case ISD::SETCC:
case ISD::BR_CC: {
case ISD::BRCOND:
// Branches tweak the chain to include LastCALLSEQ_END
Ops[0] = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ops[0],
- LastCALLSEQ_END);
+ LastCALLSEQ_END);
Ops[0] = LegalizeOp(Ops[0]);
LastCALLSEQ_END = DAG.getEntryNode();
break;
// Legalizing shifts/rotates requires adjusting the shift amount
// to the appropriate width.
if (!Ops[1].getValueType().isVector())
- Ops[1] = LegalizeOp(DAG.getShiftAmountOperand(Ops[1]));
+ Ops[1] = LegalizeOp(DAG.getShiftAmountOperand(Ops[0].getValueType(),
+ Ops[1]));
break;
case ISD::SRL_PARTS:
case ISD::SRA_PARTS:
// Legalizing shifts/rotates requires adjusting the shift amount
// to the appropriate width.
if (!Ops[2].getValueType().isVector())
- Ops[2] = LegalizeOp(DAG.getShiftAmountOperand(Ops[2]));
+ Ops[2] = LegalizeOp(DAG.getShiftAmountOperand(Ops[0].getValueType(),
+ Ops[2]));
break;
}
}
break;
case ISD::CALLSEQ_START: {
- static int depth = 0;
SDNode *CallEnd = FindCallEndFromCallStart(Node);
// Recursively Legalize all of the inputs of the call end that do not lead
// Merge in the last call to ensure that this call starts after the last
// call ended.
- if (LastCALLSEQ_END.getOpcode() != ISD::EntryToken && depth == 0) {
+ if (LastCALLSEQ_END.getOpcode() != ISD::EntryToken) {
Tmp1 = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
Tmp1, LastCALLSEQ_END);
Tmp1 = LegalizeOp(Tmp1);
// sequence have been legalized, legalize the call itself. During this
// process, no libcalls can/will be inserted, guaranteeing that no calls
// can overlap.
-
- SDValue Saved_LastCALLSEQ_END = LastCALLSEQ_END ;
+ assert(!IsLegalizingCall && "Inconsistent sequentialization of calls!");
// Note that we are selecting this call!
LastCALLSEQ_END = SDValue(CallEnd, 0);
+ IsLegalizingCall = true;
- depth++;
// Legalize the call, starting from the CALLSEQ_END.
LegalizeOp(LastCALLSEQ_END);
- depth--;
- assert(depth >= 0 && "Un-matched CALLSEQ_START?");
- if (depth > 0)
- LastCALLSEQ_END = Saved_LastCALLSEQ_END;
+ assert(!IsLegalizingCall && "CALLSEQ_END should have cleared this!");
return Result;
}
case ISD::CALLSEQ_END:
Result.getResNo());
}
}
+ assert(IsLegalizingCall && "Call sequence imbalance between start/end?");
// This finishes up call legalization.
+ IsLegalizingCall = false;
+
// If the CALLSEQ_END node has a flag, remember that we legalized it.
AddLegalizedOperand(SDValue(Node, 0), Result.getValue(0));
if (Node->getNumValues() == 2)
// If this is an unaligned load and the target doesn't support it,
// expand it.
if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT())) {
- const Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
+ Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
unsigned ABIAlignment = TLI.getTargetData()->getABITypeAlignment(Ty);
if (LD->getAlignment() < ABIAlignment){
Result = ExpandUnalignedLoad(cast<LoadSDNode>(Result.getNode()),
ISD::LoadExtType NewExtType =
ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
- Result = DAG.getExtLoad(NewExtType, Node->getValueType(0), dl,
+ Result = DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
Tmp1, Tmp2, LD->getPointerInfo(),
NVT, isVolatile, isNonTemporal, Alignment);
if (TLI.isLittleEndian()) {
// EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
// Load the bottom RoundWidth bits.
- Lo = DAG.getExtLoad(ISD::ZEXTLOAD, Node->getValueType(0), dl,
+ Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
Tmp1, Tmp2,
LD->getPointerInfo(), RoundVT, isVolatile,
isNonTemporal, Alignment);
IncrementSize = RoundWidth / 8;
Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
DAG.getIntPtrConstant(IncrementSize));
- Hi = DAG.getExtLoad(ExtType, Node->getValueType(0), dl, Tmp1, Tmp2,
+ Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Tmp1, Tmp2,
LD->getPointerInfo().getWithOffset(IncrementSize),
ExtraVT, isVolatile, isNonTemporal,
MinAlign(Alignment, IncrementSize));
// Move the top bits to the right place.
Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
- DAG.getConstant(RoundWidth, TLI.getShiftAmountTy()));
+ DAG.getConstant(RoundWidth,
+ TLI.getShiftAmountTy(Hi.getValueType())));
// Join the hi and lo parts.
Result = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
// Big endian - avoid unaligned loads.
// EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
// Load the top RoundWidth bits.
- Hi = DAG.getExtLoad(ExtType, Node->getValueType(0), dl, Tmp1, Tmp2,
+ Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Tmp1, Tmp2,
LD->getPointerInfo(), RoundVT, isVolatile,
isNonTemporal, Alignment);
Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
DAG.getIntPtrConstant(IncrementSize));
Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
- Node->getValueType(0), dl, Tmp1, Tmp2,
+ dl, Node->getValueType(0), Tmp1, Tmp2,
LD->getPointerInfo().getWithOffset(IncrementSize),
ExtraVT, isVolatile, isNonTemporal,
MinAlign(Alignment, IncrementSize));
// Move the top bits to the right place.
Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
- DAG.getConstant(ExtraWidth, TLI.getShiftAmountTy()));
+ DAG.getConstant(ExtraWidth,
+ TLI.getShiftAmountTy(Hi.getValueType())));
// Join the hi and lo parts.
Result = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
// If this is an unaligned load and the target doesn't support it,
// expand it.
if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT())) {
- const Type *Ty =
+ Type *Ty =
LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
unsigned ABIAlignment =
TLI.getTargetData()->getABITypeAlignment(Ty);
}
break;
case TargetLowering::Expand:
- if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) && isTypeLegal(SrcVT)) {
+ if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) && TLI.isTypeLegal(SrcVT)) {
SDValue Load = DAG.getLoad(SrcVT, dl, Tmp1, Tmp2,
LD->getPointerInfo(),
LD->isVolatile(), LD->isNonTemporal(),
Tmp2 = LegalizeOp(Load.getValue(1));
break;
}
+
+ // If this is a promoted vector load, and the vector element types are
+ // legal, then scalarize it.
+ if (ExtType == ISD::EXTLOAD && SrcVT.isVector() &&
+ TLI.isTypeLegal(Node->getValueType(0).getScalarType())) {
+ SmallVector<SDValue, 8> LoadVals;
+ SmallVector<SDValue, 8> LoadChains;
+ unsigned NumElem = SrcVT.getVectorNumElements();
+ unsigned Stride = SrcVT.getScalarType().getSizeInBits()/8;
+
+ for (unsigned Idx=0; Idx<NumElem; Idx++) {
+ Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
+ DAG.getIntPtrConstant(Stride));
+ SDValue ScalarLoad = DAG.getExtLoad(ISD::EXTLOAD, dl,
+ Node->getValueType(0).getScalarType(),
+ Tmp1, Tmp2, LD->getPointerInfo().getWithOffset(Idx * Stride),
+ SrcVT.getScalarType(),
+ LD->isVolatile(), LD->isNonTemporal(),
+ LD->getAlignment());
+
+ LoadVals.push_back(ScalarLoad.getValue(0));
+ LoadChains.push_back(ScalarLoad.getValue(1));
+ }
+ Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+ &LoadChains[0], LoadChains.size());
+ SDValue ValRes = DAG.getNode(ISD::BUILD_VECTOR, dl,
+ Node->getValueType(0), &LoadVals[0], LoadVals.size());
+
+ Tmp1 = LegalizeOp(ValRes); // Relegalize new nodes.
+ Tmp2 = LegalizeOp(Result.getValue(0)); // Relegalize new nodes.
+ break;
+ }
+
+ // If this is a promoted vector load, and the vector element types are
+ // illegal, create the promoted vector from bitcasted segments.
+ if (ExtType == ISD::EXTLOAD && SrcVT.isVector()) {
+ EVT MemElemTy = Node->getValueType(0).getScalarType();
+ EVT SrcSclrTy = SrcVT.getScalarType();
+ unsigned SizeRatio =
+ (MemElemTy.getSizeInBits() / SrcSclrTy.getSizeInBits());
+
+ SmallVector<SDValue, 8> LoadVals;
+ SmallVector<SDValue, 8> LoadChains;
+ unsigned NumElem = SrcVT.getVectorNumElements();
+ unsigned Stride = SrcVT.getScalarType().getSizeInBits()/8;
+
+ for (unsigned Idx=0; Idx<NumElem; Idx++) {
+ Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
+ DAG.getIntPtrConstant(Stride));
+ SDValue ScalarLoad = DAG.getExtLoad(ISD::EXTLOAD, dl,
+ SrcVT.getScalarType(),
+ Tmp1, Tmp2, LD->getPointerInfo().getWithOffset(Idx * Stride),
+ SrcVT.getScalarType(),
+ LD->isVolatile(), LD->isNonTemporal(),
+ LD->getAlignment());
+ if (TLI.isBigEndian()) {
+ // MSB (which is garbage, comes first)
+ LoadVals.push_back(ScalarLoad.getValue(0));
+ for (unsigned i = 0; i<SizeRatio-1; ++i)
+ LoadVals.push_back(DAG.getUNDEF(SrcVT.getScalarType()));
+ } else {
+ // LSB (which is data, comes first)
+ for (unsigned i = 0; i<SizeRatio-1; ++i)
+ LoadVals.push_back(DAG.getUNDEF(SrcVT.getScalarType()));
+ LoadVals.push_back(ScalarLoad.getValue(0));
+ }
+ LoadChains.push_back(ScalarLoad.getValue(1));
+ }
+
+ Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+ &LoadChains[0], LoadChains.size());
+ EVT TempWideVector = EVT::getVectorVT(*DAG.getContext(),
+ SrcVT.getScalarType(), NumElem*SizeRatio);
+ SDValue ValRes = DAG.getNode(ISD::BUILD_VECTOR, dl,
+ TempWideVector, &LoadVals[0], LoadVals.size());
+
+ // Cast to the correct type
+ ValRes = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), ValRes);
+
+ Tmp1 = LegalizeOp(ValRes); // Relegalize new nodes.
+ Tmp2 = LegalizeOp(Result.getValue(0)); // Relegalize new nodes.
+ break;
+
+ }
+
// FIXME: This does not work for vectors on most targets. Sign- and
// zero-extend operations are currently folded into extending loads,
// whether they are legal or not, and then we end up here without any
"EXTLOAD should always be supported!");
// Turn the unsupported load into an EXTLOAD followed by an explicit
// zero/sign extend inreg.
- Result = DAG.getExtLoad(ISD::EXTLOAD, Node->getValueType(0), dl,
+ Result = DAG.getExtLoad(ISD::EXTLOAD, dl, Node->getValueType(0),
Tmp1, Tmp2, LD->getPointerInfo(), SrcVT,
LD->isVolatile(), LD->isNonTemporal(),
LD->getAlignment());
// If this is an unaligned store and the target doesn't support it,
// expand it.
if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT())) {
- const Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
+ Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
unsigned ABIAlignment= TLI.getTargetData()->getABITypeAlignment(Ty);
if (ST->getAlignment() < ABIAlignment)
Result = ExpandUnalignedStore(cast<StoreSDNode>(Result.getNode()),
Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
DAG.getIntPtrConstant(IncrementSize));
Hi = DAG.getNode(ISD::SRL, dl, Tmp3.getValueType(), Tmp3,
- DAG.getConstant(RoundWidth, TLI.getShiftAmountTy()));
+ DAG.getConstant(RoundWidth,
+ TLI.getShiftAmountTy(Tmp3.getValueType())));
Hi = DAG.getTruncStore(Tmp1, dl, Hi, Tmp2,
ST->getPointerInfo().getWithOffset(IncrementSize),
ExtraVT, isVolatile, isNonTemporal,
// TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
// Store the top RoundWidth bits.
Hi = DAG.getNode(ISD::SRL, dl, Tmp3.getValueType(), Tmp3,
- DAG.getConstant(ExtraWidth, TLI.getShiftAmountTy()));
+ DAG.getConstant(ExtraWidth,
+ TLI.getShiftAmountTy(Tmp3.getValueType())));
Hi = DAG.getTruncStore(Tmp1, dl, Hi, Tmp2, ST->getPointerInfo(),
RoundVT, isVolatile, isNonTemporal, Alignment);
// If this is an unaligned store and the target doesn't support it,
// expand it.
if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT())) {
- const Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
+ Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
unsigned ABIAlignment= TLI.getTargetData()->getABITypeAlignment(Ty);
if (ST->getAlignment() < ABIAlignment)
Result = ExpandUnalignedStore(cast<StoreSDNode>(Result.getNode()),
case TargetLowering::Custom:
Result = TLI.LowerOperation(Result, DAG);
break;
- case Expand:
+ case TargetLowering::Expand:
+
+ EVT WideScalarVT = Tmp3.getValueType().getScalarType();
+ EVT NarrowScalarVT = StVT.getScalarType();
+
+ // The Store type is illegal, must scalarize the vector store.
+ SmallVector<SDValue, 8> Stores;
+ bool ScalarLegal = TLI.isTypeLegal(WideScalarVT);
+ if (!TLI.isTypeLegal(StVT) && StVT.isVector() && ScalarLegal) {
+ unsigned NumElem = StVT.getVectorNumElements();
+
+ unsigned ScalarSize = StVT.getScalarType().getSizeInBits();
+ // Round odd types to the next pow of two.
+ if (!isPowerOf2_32(ScalarSize))
+ ScalarSize = NextPowerOf2(ScalarSize);
+ // Types smaller than 8 bits are promoted to 8 bits.
+ ScalarSize = std::max<unsigned>(ScalarSize, 8);
+ // Store stride
+ unsigned Stride = ScalarSize/8;
+ assert(isPowerOf2_32(Stride) && "Stride must be a power of two");
+
+ for (unsigned Idx=0; Idx<NumElem; Idx++) {
+ SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
+ WideScalarVT, Tmp3, DAG.getIntPtrConstant(Idx));
+
+
+ EVT NVT = EVT::getIntegerVT(*DAG.getContext(), ScalarSize);
+
+ Ex = DAG.getNode(ISD::TRUNCATE, dl, NVT, Ex);
+ Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
+ DAG.getIntPtrConstant(Stride));
+ SDValue Store = DAG.getStore(Tmp1, dl, Ex, Tmp2,
+ ST->getPointerInfo().getWithOffset(Idx*Stride),
+ isVolatile, isNonTemporal, Alignment);
+ Stores.push_back(Store);
+ }
+ Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+ &Stores[0], Stores.size());
+ break;
+ }
+
+ // The Store type is illegal, must scalarize the vector store.
+ // However, the scalar type is illegal. Must bitcast the result
+ // and store it in smaller parts.
+ if (!TLI.isTypeLegal(StVT) && StVT.isVector()) {
+ unsigned WideNumElem = StVT.getVectorNumElements();
+ unsigned Stride = NarrowScalarVT.getSizeInBits()/8;
+
+ unsigned SizeRatio =
+ (WideScalarVT.getSizeInBits() / NarrowScalarVT.getSizeInBits());
+
+ EVT CastValueVT = EVT::getVectorVT(*DAG.getContext(), NarrowScalarVT,
+ SizeRatio*WideNumElem);
+
+ // Cast the wide elem vector to wider vec with smaller elem type.
+ // Example <2 x i64> -> <4 x i32>
+ Tmp3 = DAG.getNode(ISD::BITCAST, dl, CastValueVT, Tmp3);
+
+ for (unsigned Idx=0; Idx<WideNumElem*SizeRatio; Idx++) {
+ // Extract elment i
+ SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
+ NarrowScalarVT, Tmp3, DAG.getIntPtrConstant(Idx));
+ // bump pointer.
+ Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
+ DAG.getIntPtrConstant(Stride));
+
+ // Store if, this element is:
+ // - First element on big endian, or
+ // - Last element on little endian
+ if (( TLI.isBigEndian() && (Idx%SizeRatio == 0)) ||
+ ((!TLI.isBigEndian() && (Idx%SizeRatio == SizeRatio-1)))) {
+ SDValue Store = DAG.getStore(Tmp1, dl, Ex, Tmp2,
+ ST->getPointerInfo().getWithOffset(Idx*Stride),
+ isVolatile, isNonTemporal, Alignment);
+ Stores.push_back(Store);
+ }
+ }
+ Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+ &Stores[0], Stores.size());
+ break;
+ }
+
+
// TRUNCSTORE:i16 i32 -> STORE i16
- assert(isTypeLegal(StVT) && "Do not know how to expand this store!");
+ assert(TLI.isTypeLegal(StVT) && "Do not know how to expand this store!");
Tmp3 = DAG.getNode(ISD::TRUNCATE, dl, StVT, Tmp3);
Result = DAG.getStore(Tmp1, dl, Tmp3, Tmp2, ST->getPointerInfo(),
isVolatile, isNonTemporal, Alignment);
if (Op.getValueType().isVector())
return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr,MachinePointerInfo(),
false, false, 0);
- return DAG.getExtLoad(ISD::EXTLOAD, Op.getValueType(), dl, Ch, StackPtr,
+ return DAG.getExtLoad(ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr,
MachinePointerInfo(),
Vec.getValueType().getVectorElementType(),
false, false, 0);
}
+SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
+ assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
+
+ SDValue Vec = Op.getOperand(0);
+ SDValue Part = Op.getOperand(1);
+ SDValue Idx = Op.getOperand(2);
+ DebugLoc dl = Op.getDebugLoc();
+
+ // Store the value to a temporary stack slot, then LOAD the returned part.
+
+ SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
+ int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
+ MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
+
+ // First store the whole vector.
+ SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo,
+ false, false, 0);
+
+ // Then store the inserted part.
+
+ // Add the offset to the index.
+ unsigned EltSize =
+ Vec.getValueType().getVectorElementType().getSizeInBits()/8;
+
+ Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
+ DAG.getConstant(EltSize, Idx.getValueType()));
+
+ if (Idx.getValueType().bitsGT(TLI.getPointerTy()))
+ Idx = DAG.getNode(ISD::TRUNCATE, dl, TLI.getPointerTy(), Idx);
+ else
+ Idx = DAG.getNode(ISD::ZERO_EXTEND, dl, TLI.getPointerTy(), Idx);
+
+ SDValue SubStackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx,
+ StackPtr);
+
+ // Store the subvector.
+ Ch = DAG.getStore(DAG.getEntryNode(), dl, Part, SubStackPtr,
+ MachinePointerInfo(), false, false, 0);
+
+ // Finally, load the updated vector.
+ return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo,
+ false, false, 0);
+}
+
SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
// We can't handle this case efficiently. Allocate a sufficiently
// aligned object on the stack, store each element into it, then load
SDValue SignBit;
EVT FloatVT = Tmp2.getValueType();
EVT IVT = EVT::getIntegerVT(*DAG.getContext(), FloatVT.getSizeInBits());
- if (isTypeLegal(IVT)) {
+ if (TLI.isTypeLegal(IVT)) {
// Convert to an integer with the same sign bit.
SignBit = DAG.getNode(ISD::BITCAST, dl, IVT, Tmp2);
} else {
assert(BitShift < LoadTy.getSizeInBits() && "Pointer advanced wrong?");
if (BitShift)
SignBit = DAG.getNode(ISD::SHL, dl, LoadTy, SignBit,
- DAG.getConstant(BitShift,TLI.getShiftAmountTy()));
+ DAG.getConstant(BitShift,
+ TLI.getShiftAmountTy(SignBit.getValueType())));
}
}
// Now get the sign bit proper, by seeing whether the value is negative.
unsigned SrcSize = SrcOp.getValueType().getSizeInBits();
unsigned SlotSize = SlotVT.getSizeInBits();
unsigned DestSize = DestVT.getSizeInBits();
- const Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
+ Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
unsigned DestAlign = TLI.getTargetData()->getPrefTypeAlignment(DestType);
// Emit a store to the stack slot. Use a truncstore if the input value is
false, false, DestAlign);
assert(SlotSize < DestSize && "Unknown extension!");
- return DAG.getExtLoad(ISD::EXTLOAD, DestVT, dl, Store, FIPtr,
+ return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr,
PtrInfo, SlotVT, false, false, DestAlign);
}
}
} else {
assert(Node->getOperand(i).getOpcode() == ISD::UNDEF);
- const Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
+ Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
CV.push_back(UndefValue::get(OpNTy));
}
}
// and leave the Hi part unset.
SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
bool isSigned) {
+ assert(!IsLegalizingCall && "Cannot overlap legalization of calls!");
// The input chain to this libcall is the entry node of the function.
// Legalizing the call will automatically add the previous call to the
// dependence.
TargetLowering::ArgListEntry Entry;
for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
EVT ArgVT = Node->getOperand(i).getValueType();
- const Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
+ Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
Entry.isSExt = isSigned;
Entry.isZExt = !isSigned;
TLI.getPointerTy());
// Splice the libcall in wherever FindInputOutputChains tells us to.
- const Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
+ Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
// isTailCall may be true since the callee does not reference caller stack
// frame. Check if it's in the right position.
return CallInfo.first;
}
+/// ExpandLibCall - Generate a libcall taking the given operands as arguments
+/// and returning a result of type RetVT.
+SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
+ const SDValue *Ops, unsigned NumOps,
+ bool isSigned, DebugLoc dl) {
+ TargetLowering::ArgListTy Args;
+ Args.reserve(NumOps);
+
+ TargetLowering::ArgListEntry Entry;
+ for (unsigned i = 0; i != NumOps; ++i) {
+ Entry.Node = Ops[i];
+ Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
+ Entry.isSExt = isSigned;
+ Entry.isZExt = !isSigned;
+ Args.push_back(Entry);
+ }
+ SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
+ TLI.getPointerTy());
+
+ Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
+ std::pair<SDValue,SDValue> CallInfo =
+ TLI.LowerCallTo(DAG.getEntryNode(), RetTy, isSigned, !isSigned, false,
+ false, 0, TLI.getLibcallCallingConv(LC), false,
+ /*isReturnValueUsed=*/true,
+ Callee, Args, DAG, dl);
+
+ // Legalize the call sequence, starting with the chain. This will advance
+ // the LastCALLSEQ_END to the legalized version of the CALLSEQ_END node that
+ // was added by LowerCallTo (guaranteeing proper serialization of calls).
+ LegalizeOp(CallInfo.second);
+
+ return CallInfo.first;
+}
+
// ExpandChainLibCall - Expand a node into a call to a libcall. Similar to
// ExpandLibCall except that the first operand is the in-chain.
std::pair<SDValue, SDValue>
SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
SDNode *Node,
bool isSigned) {
+ assert(!IsLegalizingCall && "Cannot overlap legalization of calls!");
SDValue InChain = Node->getOperand(0);
TargetLowering::ArgListTy Args;
TargetLowering::ArgListEntry Entry;
for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) {
EVT ArgVT = Node->getOperand(i).getValueType();
- const Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
+ Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
Entry.Node = Node->getOperand(i);
Entry.Ty = ArgTy;
Entry.isSExt = isSigned;
TLI.getPointerTy());
// Splice the libcall in wherever FindInputOutputChains tells us to.
- const Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
+ Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
std::pair<SDValue, SDValue> CallInfo =
TLI.LowerCallTo(InChain, RetTy, isSigned, !isSigned, false, false,
0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
return ExpandLibCall(LC, Node, isSigned);
}
+/// isDivRemLibcallAvailable - Return true if divmod libcall is available.
+static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
+ const TargetLowering &TLI) {
+ RTLIB::Libcall LC;
+ switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
+ default: assert(0 && "Unexpected request for libcall!");
+ case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
+ case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
+ case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
+ case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
+ case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
+ }
+
+ return TLI.getLibcallName(LC) != 0;
+}
+
+/// UseDivRem - Only issue divrem libcall if both quotient and remainder are
+/// needed.
+static bool UseDivRem(SDNode *Node, bool isSigned, bool isDIV) {
+ unsigned OtherOpcode = 0;
+ if (isSigned)
+ OtherOpcode = isDIV ? ISD::SREM : ISD::SDIV;
+ else
+ OtherOpcode = isDIV ? ISD::UREM : ISD::UDIV;
+
+ SDValue Op0 = Node->getOperand(0);
+ SDValue Op1 = Node->getOperand(1);
+ for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
+ UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
+ SDNode *User = *UI;
+ if (User == Node)
+ continue;
+ if (User->getOpcode() == OtherOpcode &&
+ User->getOperand(0) == Op0 &&
+ User->getOperand(1) == Op1)
+ return true;
+ }
+ return false;
+}
+
+/// ExpandDivRemLibCall - Issue libcalls to __{u}divmod to compute div / rem
+/// pairs.
+void
+SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
+ SmallVectorImpl<SDValue> &Results) {
+ unsigned Opcode = Node->getOpcode();
+ bool isSigned = Opcode == ISD::SDIVREM;
+
+ RTLIB::Libcall LC;
+ switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
+ default: assert(0 && "Unexpected request for libcall!");
+ case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
+ case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
+ case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
+ case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
+ case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
+ }
+
+ // The input chain to this libcall is the entry node of the function.
+ // Legalizing the call will automatically add the previous call to the
+ // dependence.
+ SDValue InChain = DAG.getEntryNode();
+
+ EVT RetVT = Node->getValueType(0);
+ Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
+
+ TargetLowering::ArgListTy Args;
+ TargetLowering::ArgListEntry Entry;
+ for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
+ EVT ArgVT = Node->getOperand(i).getValueType();
+ Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
+ Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
+ Entry.isSExt = isSigned;
+ Entry.isZExt = !isSigned;
+ Args.push_back(Entry);
+ }
+
+ // Also pass the return address of the remainder.
+ SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
+ Entry.Node = FIPtr;
+ Entry.Ty = RetTy->getPointerTo();
+ Entry.isSExt = isSigned;
+ Entry.isZExt = !isSigned;
+ Args.push_back(Entry);
+
+ SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
+ TLI.getPointerTy());
+
+ // Splice the libcall in wherever FindInputOutputChains tells us to.
+ DebugLoc dl = Node->getDebugLoc();
+ std::pair<SDValue, SDValue> CallInfo =
+ TLI.LowerCallTo(InChain, RetTy, isSigned, !isSigned, false, false,
+ 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
+ /*isReturnValueUsed=*/true, Callee, Args, DAG, dl);
+
+ // Legalize the call sequence, starting with the chain. This will advance
+ // the LastCALLSEQ to the legalized version of the CALLSEQ_END node that
+ // was added by LowerCallTo (guaranteeing proper serialization of calls).
+ LegalizeOp(CallInfo.second);
+
+ // Remainder is loaded back from the stack frame.
+ SDValue Rem = DAG.getLoad(RetVT, dl, LastCALLSEQ_END, FIPtr,
+ MachinePointerInfo(), false, false, 0);
+ Results.push_back(CallInfo.first);
+ Results.push_back(Rem);
+}
+
/// ExpandLegalINT_TO_FP - This function is responsible for legalizing a
/// INT_TO_FP operation of the specified operand when the target requests that
/// we expand it. At this point, we know that the result and operand types are
if (!isSigned) {
SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
- SDValue ShiftConst = DAG.getConstant(1, TLI.getShiftAmountTy());
+ SDValue ShiftConst =
+ DAG.getConstant(1, TLI.getShiftAmountTy(Op0.getValueType()));
SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
SDValue AndConst = DAG.getConstant(1, MVT::i64);
SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
}
// Otherwise, implement the fully general conversion.
- EVT SHVT = TLI.getShiftAmountTy();
SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
DAG.getConstant(UINT64_C(0xfffffffffffff800), MVT::i64));
Op0, DAG.getConstant(UINT64_C(0x0020000000000000), MVT::i64),
ISD::SETUGE);
SDValue Sel2 = DAG.getNode(ISD::SELECT, dl, MVT::i64, Ge, Sel, Op0);
+ EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType());
SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
DAG.getConstant(32, SHVT));
false, false, Alignment);
else {
FudgeInReg =
- LegalizeOp(DAG.getExtLoad(ISD::EXTLOAD, DestVT, dl,
+ LegalizeOp(DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
DAG.getEntryNode(), CPIdx,
MachinePointerInfo::getConstantPool(),
MVT::f32, false, false, Alignment));
///
SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, DebugLoc dl) {
EVT VT = Op.getValueType();
- EVT SHVT = TLI.getShiftAmountTy();
+ EVT SHVT = TLI.getShiftAmountTy(VT);
SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
switch (VT.getSimpleVT().SimpleTy) {
default: assert(0 && "Unhandled Expand type in BSWAP!");
default: assert(0 && "Cannot expand this yet!");
case ISD::CTPOP: {
EVT VT = Op.getValueType();
- EVT ShVT = TLI.getShiftAmountTy();
+ EVT ShVT = TLI.getShiftAmountTy(VT);
unsigned Len = VT.getSizeInBits();
assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
Op = DAG.getNode(ISD::SRL, dl, VT,
DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
DAG.getConstant(Len - 8, ShVT));
-
+
return Op;
}
case ISD::CTLZ: {
//
// but see also: http://www.hackersdelight.org/HDcode/nlz.cc
EVT VT = Op.getValueType();
- EVT ShVT = TLI.getShiftAmountTy();
+ EVT ShVT = TLI.getShiftAmountTy(VT);
unsigned len = VT.getSizeInBits();
for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
SDValue Tmp3 = DAG.getConstant(1ULL << i, ShVT);
Results.push_back(DAG.getConstant(0, MVT::i32));
Results.push_back(Node->getOperand(0));
break;
+ case ISD::ATOMIC_FENCE:
case ISD::MEMBARRIER: {
// If the target didn't lower this, lower it to '__sync_synchronize()' call
+ // FIXME: handle "fence singlethread" more efficiently.
TargetLowering::ArgListTy Args;
std::pair<SDValue, SDValue> CallResult =
TLI.LowerCallTo(Node->getOperand(0), Type::getVoidTy(*DAG.getContext()),
Results.push_back(CallResult.second);
break;
}
+ case ISD::ATOMIC_LOAD: {
+ // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP.
+ SDValue Zero = DAG.getConstant(0, cast<AtomicSDNode>(Node)->getMemoryVT());
+ SDValue Swap = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl,
+ cast<AtomicSDNode>(Node)->getMemoryVT(),
+ Node->getOperand(0),
+ Node->getOperand(1), Zero, Zero,
+ cast<AtomicSDNode>(Node)->getMemOperand(),
+ cast<AtomicSDNode>(Node)->getOrdering(),
+ cast<AtomicSDNode>(Node)->getSynchScope());
+ Results.push_back(Swap.getValue(0));
+ Results.push_back(Swap.getValue(1));
+ break;
+ }
+ case ISD::ATOMIC_STORE: {
+ // There is no libcall for atomic store; fake it with ATOMIC_SWAP.
+ SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
+ cast<AtomicSDNode>(Node)->getMemoryVT(),
+ Node->getOperand(0),
+ Node->getOperand(1), Node->getOperand(2),
+ cast<AtomicSDNode>(Node)->getMemOperand(),
+ cast<AtomicSDNode>(Node)->getOrdering(),
+ cast<AtomicSDNode>(Node)->getSynchScope());
+ Results.push_back(Swap.getValue(1));
+ break;
+ }
// By default, atomic intrinsics are marked Legal and lowered. Targets
// which don't support them directly, however, may want libcalls, in which
// case they mark them Expand, and we get here.
// SAR. However, it is doubtful that any exist.
EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
EVT VT = Node->getValueType(0);
- EVT ShiftAmountTy = TLI.getShiftAmountTy();
+ EVT ShiftAmountTy = TLI.getShiftAmountTy(VT);
if (VT.isVector())
ShiftAmountTy = VT;
unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
}
case ISD::FP_ROUND_INREG: {
// The only way we can lower this is to turn it into a TRUNCSTORE,
- // EXTLOAD pair, targetting a temporary location (a stack slot).
+ // EXTLOAD pair, targeting a temporary location (a stack slot).
// NOTE: there is a choice here between constantly creating new stack
// slots and always reusing the same one. We currently always create
case ISD::EXTRACT_SUBVECTOR:
Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
break;
+ case ISD::INSERT_SUBVECTOR:
+ Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
+ break;
case ISD::CONCAT_VECTORS: {
Results.push_back(ExpandVectorBuildThroughStack(Node));
break;
EVT VT = Node->getValueType(0);
EVT EltVT = VT.getVectorElementType();
- if (getTypeAction(EltVT) == Promote)
+ if (!TLI.isTypeLegal(EltVT))
EltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
unsigned NumElems = VT.getVectorNumElements();
SmallVector<SDValue, 8> Ops;
// 1 -> Hi
Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0),
DAG.getConstant(OpTy.getSizeInBits()/2,
- TLI.getShiftAmountTy()));
+ TLI.getShiftAmountTy(Node->getOperand(0).getValueType())));
Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1);
} else {
// 0 -> Lo
Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
RTLIB::REM_F80, RTLIB::REM_PPCF128));
break;
+ case ISD::FMA:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
+ RTLIB::FMA_F80, RTLIB::FMA_PPCF128));
+ break;
case ISD::FP16_TO_FP32:
Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
break;
unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
Tmp2 = Node->getOperand(0);
Tmp3 = Node->getOperand(1);
- if (TLI.isOperationLegalOrCustom(DivRemOpc, VT)) {
+ if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
+ (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
+ UseDivRem(Node, isSigned, false))) {
Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
} else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
// X % Y -> X-X/Y*Y
Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
- } else if (isSigned) {
+ } else if (isSigned)
Tmp1 = ExpandIntLibCall(Node, true,
RTLIB::SREM_I8,
RTLIB::SREM_I16, RTLIB::SREM_I32,
RTLIB::SREM_I64, RTLIB::SREM_I128);
- } else {
+ else
Tmp1 = ExpandIntLibCall(Node, false,
RTLIB::UREM_I8,
RTLIB::UREM_I16, RTLIB::UREM_I32,
RTLIB::UREM_I64, RTLIB::UREM_I128);
- }
Results.push_back(Tmp1);
break;
}
unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
EVT VT = Node->getValueType(0);
SDVTList VTs = DAG.getVTList(VT, VT);
- if (TLI.isOperationLegalOrCustom(DivRemOpc, VT))
+ if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
+ (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
+ UseDivRem(Node, isSigned, true)))
Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
Node->getOperand(1));
else if (isSigned)
Results.push_back(Tmp1.getValue(1));
break;
}
+ case ISD::SDIVREM:
+ case ISD::UDIVREM:
+ // Expand into divrem libcall
+ ExpandDivRemLibCall(Node, Results);
+ break;
case ISD::MUL: {
EVT VT = Node->getValueType(0);
SDVTList VTs = DAG.getVTList(VT, VT);
case ISD::UMULO:
case ISD::SMULO: {
EVT VT = Node->getValueType(0);
+ EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
SDValue LHS = Node->getOperand(0);
SDValue RHS = Node->getOperand(1);
SDValue BottomHalf;
TopHalf = BottomHalf.getValue(1);
} else if (TLI.isTypeLegal(EVT::getIntegerVT(*DAG.getContext(),
VT.getSizeInBits() * 2))) {
- EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
Tmp1 = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS);
// have a libcall big enough.
// Also, we can fall back to a division in some cases, but that's a big
// performance hit in the general case.
- EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
if (WideVT == MVT::i16)
LC = RTLIB::MUL_I16;
else if (WideVT == MVT::i128)
LC = RTLIB::MUL_I128;
assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!");
- LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
- RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
-
- SDValue Ret = ExpandLibCall(LC, Node, isSigned);
- BottomHalf = DAG.getNode(ISD::TRUNCATE, dl, VT, Ret);
- TopHalf = DAG.getNode(ISD::SRL, dl, Ret.getValueType(), Ret,
- DAG.getConstant(VT.getSizeInBits(), TLI.getPointerTy()));
- TopHalf = DAG.getNode(ISD::TRUNCATE, dl, VT, TopHalf);
+
+ // The high part is obtained by SRA'ing all but one of the bits of low
+ // part.
+ unsigned LoSize = VT.getSizeInBits();
+ SDValue HiLHS = DAG.getNode(ISD::SRA, dl, VT, RHS,
+ DAG.getConstant(LoSize-1, TLI.getPointerTy()));
+ SDValue HiRHS = DAG.getNode(ISD::SRA, dl, VT, LHS,
+ DAG.getConstant(LoSize-1, TLI.getPointerTy()));
+
+ // Here we're passing the 2 arguments explicitly as 4 arguments that are
+ // pre-lowered to the correct types. This all depends upon WideVT not
+ // being a legal type for the architecture and thus has to be split to
+ // two arguments.
+ SDValue Args[] = { LHS, HiLHS, RHS, HiRHS };
+ SDValue Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
+ BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
+ DAG.getIntPtrConstant(0));
+ TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
+ DAG.getIntPtrConstant(1));
}
+
if (isSigned) {
- Tmp1 = DAG.getConstant(VT.getSizeInBits() - 1, TLI.getShiftAmountTy());
+ Tmp1 = DAG.getConstant(VT.getSizeInBits() - 1,
+ TLI.getShiftAmountTy(BottomHalf.getValueType()));
Tmp1 = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, Tmp1);
TopHalf = DAG.getSetCC(dl, TLI.getSetCCResultType(VT), TopHalf, Tmp1,
ISD::SETNE);
Tmp2 = DAG.getNode(ISD::ANY_EXTEND, dl, PairTy, Node->getOperand(1));
Tmp2 = DAG.getNode(ISD::SHL, dl, PairTy, Tmp2,
DAG.getConstant(PairTy.getSizeInBits()/2,
- TLI.getShiftAmountTy()));
+ TLI.getShiftAmountTy(PairTy)));
Results.push_back(DAG.getNode(ISD::OR, dl, PairTy, Tmp1, Tmp2));
break;
}
SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
- SDValue LD = DAG.getExtLoad(ISD::SEXTLOAD, PTy, dl, Chain, Addr,
+ SDValue LD = DAG.getExtLoad(ISD::SEXTLOAD, dl, PTy, Chain, Addr,
MachinePointerInfo::getJumpTable(), MemVT,
false, false, 0);
Addr = LD;
Tmp2.getOperand(0), Tmp2.getOperand(1),
Node->getOperand(2));
} else {
+ // We test only the i1 bit. Skip the AND if UNDEF.
+ Tmp3 = (Tmp2.getOpcode() == ISD::UNDEF) ? Tmp2 :
+ DAG.getNode(ISD::AND, dl, Tmp2.getValueType(), Tmp2,
+ DAG.getConstant(1, Tmp2.getValueType()));
Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, Tmp1,
- DAG.getCondCode(ISD::SETNE), Tmp2,
- DAG.getConstant(0, Tmp2.getValueType()),
+ DAG.getCondCode(ISD::SETNE), Tmp3,
+ DAG.getConstant(0, Tmp3.getValueType()),
Node->getOperand(2));
}
Results.push_back(Tmp1);
Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
Tmp1 = DAG.getNode(ISD::BSWAP, dl, NVT, Tmp1);
Tmp1 = DAG.getNode(ISD::SRL, dl, NVT, Tmp1,
- DAG.getConstant(DiffBits, TLI.getShiftAmountTy()));
+ DAG.getConstant(DiffBits, TLI.getShiftAmountTy(NVT)));
Results.push_back(Tmp1);
break;
}
// SelectionDAG::Legalize - This is the entry point for the file.
//
-void SelectionDAG::Legalize(CodeGenOpt::Level OptLevel) {
+void SelectionDAG::Legalize() {
/// run - This is the main entry point to this class.
///
- SelectionDAGLegalize(*this, OptLevel).LegalizeDAG();
+ SelectionDAGLegalize(*this).LegalizeDAG();
}
-