//
//===----------------------------------------------------------------------===//
-#include "llvm/Analysis/DebugInfo.h"
+#include "llvm/CodeGen/SelectionDAG.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallVector.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/TargetFrameInfo.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/DebugInfo.h"
+#include "llvm/IR/CallingConv.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/LLVMContext.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
-#include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/SmallVector.h"
-#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/Target/TargetFrameLowering.h"
+#include "llvm/Target/TargetLowering.h"
+#include "llvm/Target/TargetMachine.h"
using namespace llvm;
//===----------------------------------------------------------------------===//
/// will attempt merge setcc and brc instructions into brcc's.
///
namespace {
-class SelectionDAGLegalize {
+class SelectionDAGLegalize : public SelectionDAG::DAGUpdateListener {
const TargetMachine &TM;
const TargetLowering &TLI;
SelectionDAG &DAG;
- CodeGenOpt::Level OptLevel;
-
- // Libcall insertion helpers.
-
- /// LastCALLSEQ_END - This keeps track of the CALLSEQ_END node that has been
- /// legalized. We use this to ensure that calls are properly serialized
- /// against each other, including inserted libcalls.
- SDValue LastCALLSEQ_END;
-
- /// 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;
-
- 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;
+ /// LegalizePosition - The iterator for walking through the node list.
+ SelectionDAG::allnodes_iterator LegalizePosition;
- /// LegalizedNodes - For nodes that are of legal width, and that have more
- /// than one use, this map indicates what regularized operand to use. This
- /// allows us to avoid legalizing the same thing more than once.
- DenseMap<SDValue, SDValue> LegalizedNodes;
+ /// LegalizedNodes - The set of nodes which have already been legalized.
+ SmallPtrSet<SDNode *, 16> LegalizedNodes;
- void AddLegalizedOperand(SDValue From, SDValue To) {
- LegalizedNodes.insert(std::make_pair(From, To));
- // If someone requests legalization of the new node, return itself.
- if (From != To)
- LegalizedNodes.insert(std::make_pair(To, To));
- }
+ // Libcall insertion helpers.
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);
- }
-
- /// isTypeLegal - Return true if this type is legal on this target.
- ///
- bool isTypeLegal(EVT VT) const {
- return getTypeAction(VT) == Legal;
- }
+ 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.
- SDValue LegalizeOp(SDValue O);
+ /// LegalizeOp - Legalizes the given operation.
+ void LegalizeOp(SDNode *Node);
SDValue OptimizeFloatStore(StoreSDNode *ST);
+ void LegalizeLoadOps(SDNode *Node);
+ void LegalizeStoreOps(SDNode *Node);
+
/// PerformInsertVectorEltInMemory - Some target cannot handle a variable
/// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
/// is necessary to spill the vector being inserted into to memory, perform
/// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
SDValue ShuffleWithNarrowerEltType(EVT NVT, EVT VT, DebugLoc dl,
SDValue N1, SDValue N2,
- SmallVectorImpl<int> &Mask) const;
-
- bool LegalizeAllNodesNotLeadingTo(SDNode *N, SDNode *Dest,
- SmallPtrSet<SDNode*, 32> &NodesLeadingTo);
+ ArrayRef<int> Mask) const;
void LegalizeSetCCCondCode(EVT VT, SDValue &LHS, SDValue &RHS, SDValue &CC,
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_F64, RTLIB::Libcall Call_F80,
- RTLIB::Libcall Call_PPCF128);
+ RTLIB::Libcall Call_F128, RTLIB::Libcall Call_PPCF128);
SDValue ExpandIntLibCall(SDNode *Node, bool isSigned,
RTLIB::Libcall Call_I8,
RTLIB::Libcall Call_I16,
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);
+ SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP);
+
std::pair<SDValue, SDValue> ExpandAtomic(SDNode *Node);
- void ExpandNode(SDNode *Node, SmallVectorImpl<SDValue> &Results);
- void PromoteNode(SDNode *Node, SmallVectorImpl<SDValue> &Results);
+ void ExpandNode(SDNode *Node);
+ void PromoteNode(SDNode *Node);
+
+ void ForgetNode(SDNode *N) {
+ LegalizedNodes.erase(N);
+ if (LegalizePosition == SelectionDAG::allnodes_iterator(N))
+ ++LegalizePosition;
+ }
+
+public:
+ // DAGUpdateListener implementation.
+ virtual void NodeDeleted(SDNode *N, SDNode *E) {
+ ForgetNode(N);
+ }
+ virtual void NodeUpdated(SDNode *N) {}
+
+ // Node replacement helpers
+ void ReplacedNode(SDNode *N) {
+ if (N->use_empty()) {
+ DAG.RemoveDeadNode(N);
+ } else {
+ ForgetNode(N);
+ }
+ }
+ void ReplaceNode(SDNode *Old, SDNode *New) {
+ DAG.ReplaceAllUsesWith(Old, New);
+ ReplacedNode(Old);
+ }
+ void ReplaceNode(SDValue Old, SDValue New) {
+ DAG.ReplaceAllUsesWith(Old, New);
+ ReplacedNode(Old.getNode());
+ }
+ void ReplaceNode(SDNode *Old, const SDValue *New) {
+ DAG.ReplaceAllUsesWith(Old, New);
+ ReplacedNode(Old);
+ }
};
}
SDValue
SelectionDAGLegalize::ShuffleWithNarrowerEltType(EVT NVT, EVT VT, DebugLoc dl,
SDValue N1, SDValue N2,
- SmallVectorImpl<int> &Mask) const {
+ ArrayRef<int> Mask) const {
unsigned NumMaskElts = VT.getVectorNumElements();
unsigned NumDestElts = NVT.getVectorNumElements();
unsigned NumEltsGrowth = NumDestElts / NumMaskElts;
return DAG.getVectorShuffle(NVT, dl, N1, N2, &NewMask[0]);
}
-SelectionDAGLegalize::SelectionDAGLegalize(SelectionDAG &dag,
- CodeGenOpt::Level ol)
- : 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!");
+SelectionDAGLegalize::SelectionDAGLegalize(SelectionDAG &dag)
+ : SelectionDAG::DAGUpdateListener(dag),
+ TM(dag.getTarget()), TLI(dag.getTargetLoweringInfo()),
+ 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
- // could just start legalizing on the root and traverse the whole graph. In
- // practice however, this causes us to run out of stack space on large basic
- // blocks. To avoid this problem, compute an ordering of the nodes where each
- // node is only legalized after all of its operands are legalized.
DAG.AssignTopologicalOrder();
- for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
- E = prior(DAG.allnodes_end()); I != llvm::next(E); ++I)
- LegalizeOp(SDValue(I, 0));
-
- // Finally, it's possible the root changed. Get the new root.
- SDValue OldRoot = DAG.getRoot();
- assert(LegalizedNodes.count(OldRoot) && "Root didn't get legalized?");
- DAG.setRoot(LegalizedNodes[OldRoot]);
-
- LegalizedNodes.clear();
-
- // Remove dead nodes now.
- DAG.RemoveDeadNodes();
-}
-
-/// FindCallEndFromCallStart - Given a chained node that is part of a call
-/// sequence, find the CALLSEQ_END node that terminates the call sequence.
-static SDNode *FindCallEndFromCallStart(SDNode *Node, int depth = 0) {
- if (Node->getOpcode() == ISD::CALLSEQ_START)
- depth++;
- else if (Node->getOpcode() == ISD::CALLSEQ_END) {
- depth--;
- if (depth == 0)
- return Node;
- }
- if (Node->use_empty())
- return 0; // No CallSeqEnd
-
- // The chain is usually at the end.
- SDValue TheChain(Node, Node->getNumValues()-1);
- if (TheChain.getValueType() != MVT::Other) {
- // Sometimes it's at the beginning.
- TheChain = SDValue(Node, 0);
- if (TheChain.getValueType() != MVT::Other) {
- // Otherwise, hunt for it.
- for (unsigned i = 1, e = Node->getNumValues(); i != e; ++i)
- if (Node->getValueType(i) == MVT::Other) {
- TheChain = SDValue(Node, i);
- break;
- }
-
- // Otherwise, we walked into a node without a chain.
- if (TheChain.getValueType() != MVT::Other)
- return 0;
+ // Visit all the nodes. We start in topological order, so that we see
+ // nodes with their original operands intact. Legalization can produce
+ // new nodes which may themselves need to be legalized. Iterate until all
+ // nodes have been legalized.
+ for (;;) {
+ bool AnyLegalized = false;
+ for (LegalizePosition = DAG.allnodes_end();
+ LegalizePosition != DAG.allnodes_begin(); ) {
+ --LegalizePosition;
+
+ SDNode *N = LegalizePosition;
+ if (LegalizedNodes.insert(N)) {
+ AnyLegalized = true;
+ LegalizeOp(N);
+ }
}
- }
-
- for (SDNode::use_iterator UI = Node->use_begin(),
- E = Node->use_end(); UI != E; ++UI) {
-
- // Make sure to only follow users of our token chain.
- SDNode *User = *UI;
- for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i)
- if (User->getOperand(i) == TheChain)
- if (SDNode *Result = FindCallEndFromCallStart(User, depth))
- return Result;
- }
- return 0;
-}
-
-/// FindCallStartFromCallEnd - Given a chained node that is part of a call
-/// sequence, find the CALLSEQ_START node that initiates the call sequence.
-static SDNode *FindCallStartFromCallEnd(SDNode *Node) {
- int nested = 0;
- assert(Node && "Didn't find callseq_start for a call??");
- while (Node->getOpcode() != ISD::CALLSEQ_START || nested) {
- Node = Node->getOperand(0).getNode();
- assert(Node->getOperand(0).getValueType() == MVT::Other &&
- "Node doesn't have a token chain argument!");
- switch (Node->getOpcode()) {
- default:
+ if (!AnyLegalized)
break;
- case ISD::CALLSEQ_START:
- if (!nested)
- return Node;
- nested--;
- break;
- case ISD::CALLSEQ_END:
- nested++;
- break;
- }
- }
- return 0;
-}
-
-/// LegalizeAllNodesNotLeadingTo - Recursively walk the uses of N, looking to
-/// see if any uses can reach Dest. If no dest operands can get to dest,
-/// legalize them, legalize ourself, and return false, otherwise, return true.
-///
-/// Keep track of the nodes we fine that actually do lead to Dest in
-/// NodesLeadingTo. This avoids retraversing them exponential number of times.
-///
-bool SelectionDAGLegalize::LegalizeAllNodesNotLeadingTo(SDNode *N, SDNode *Dest,
- SmallPtrSet<SDNode*, 32> &NodesLeadingTo) {
- if (N == Dest) return true; // N certainly leads to Dest :)
-
- // If we've already processed this node and it does lead to Dest, there is no
- // need to reprocess it.
- if (NodesLeadingTo.count(N)) return true;
-
- // If the first result of this node has been already legalized, then it cannot
- // reach N.
- if (LegalizedNodes.count(SDValue(N, 0))) return false;
- // Okay, this node has not already been legalized. Check and legalize all
- // operands. If none lead to Dest, then we can legalize this node.
- bool OperandsLeadToDest = false;
- for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
- OperandsLeadToDest |= // If an operand leads to Dest, so do we.
- LegalizeAllNodesNotLeadingTo(N->getOperand(i).getNode(), Dest,
- NodesLeadingTo);
-
- if (OperandsLeadToDest) {
- NodesLeadingTo.insert(N);
- return true;
}
- // Okay, this node looks safe, legalize it and return false.
- LegalizeOp(SDValue(N, 0));
- return false;
+ // Remove dead nodes now.
+ DAG.RemoveDeadNodes();
}
/// ExpandConstantFP - Expands the ConstantFP node to an integer constant or
/// a load from the constant pool.
-static SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP,
- SelectionDAG &DAG, const TargetLowering &TLI) {
+SDValue
+SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) {
bool Extend = false;
DebugLoc dl = CFP->getDebugLoc();
// 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,
- DAG.getEntryNode(),
- CPIdx, MachinePointerInfo::getConstantPool(),
- VT, false, false, Alignment);
- return DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx,
- MachinePointerInfo::getConstantPool(), false, false,
- Alignment);
+ if (Extend) {
+ SDValue Result =
+ DAG.getExtLoad(ISD::EXTLOAD, dl, OrigVT,
+ DAG.getEntryNode(),
+ CPIdx, MachinePointerInfo::getConstantPool(),
+ VT, false, false, Alignment);
+ return Result;
+ }
+ SDValue Result =
+ DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx,
+ MachinePointerInfo::getConstantPool(), false, false, false,
+ Alignment);
+ return Result;
}
/// ExpandUnalignedStore - Expands an unaligned store to 2 half-size stores.
-static
-SDValue ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG,
- const TargetLowering &TLI) {
+static void ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG,
+ const TargetLowering &TLI,
+ SelectionDAGLegalize *DAGLegalize) {
+ assert(ST->getAddressingMode() == ISD::UNINDEXED &&
+ "unaligned indexed stores not implemented!");
SDValue Chain = ST->getChain();
SDValue Ptr = ST->getBasePtr();
SDValue Val = ST->getValue();
// same size, then a (misaligned) int store.
// FIXME: Does not handle truncating floating point stores!
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);
- }
+ Result = DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
+ ST->isVolatile(), ST->isNonTemporal(), Alignment);
+ DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
+ return;
+ }
+ // 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();
+ MVT 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, 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.
+ SDValue Result =
+ DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
+ Stores.size());
+ DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
+ return;
}
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);
NewStoredVT, ST->isVolatile(), ST->isNonTemporal(),
Alignment);
- return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2);
+ SDValue Result =
+ DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2);
+ DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
}
/// ExpandUnalignedLoad - Expands an unaligned load to 2 half-size loads.
-static
-SDValue ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG,
- const TargetLowering &TLI) {
+static void
+ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG,
+ const TargetLowering &TLI,
+ SDValue &ValResult, SDValue &ChainResult) {
+ assert(LD->getAddressingMode() == ISD::UNINDEXED &&
+ "unaligned indexed loads not implemented!");
SDValue Chain = LD->getChain();
SDValue Ptr = LD->getBasePtr();
EVT VT = LD->getValueType(0);
DebugLoc dl = LD->getDebugLoc();
if (VT.isFloatingPoint() || VT.isVector()) {
EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
- if (TLI.isTypeLegal(intVT)) {
+ if (TLI.isTypeLegal(intVT) && TLI.isTypeLegal(LoadedVT)) {
// Expand to a (misaligned) integer load of the same size,
// then bitconvert to floating point or vector.
SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr, LD->getPointerInfo(),
LD->isVolatile(),
- LD->isNonTemporal(), LD->getAlignment());
+ LD->isNonTemporal(),
+ LD->isInvariant(), LD->getAlignment());
SDValue Result = DAG.getNode(ISD::BITCAST, dl, LoadedVT, newLoad);
- if (VT.isFloatingPoint() && LoadedVT != VT)
- Result = DAG.getNode(ISD::FP_EXTEND, dl, VT, Result);
+ if (LoadedVT != VT)
+ Result = DAG.getNode(VT.isFloatingPoint() ? ISD::FP_EXTEND :
+ ISD::ANY_EXTEND, dl, VT, Result);
- SDValue Ops[] = { Result, Chain };
- return DAG.getMergeValues(Ops, 2, dl);
+ ValResult = Result;
+ ChainResult = Chain;
+ return;
}
// Copy the value to a (aligned) stack slot using (unaligned) integer
// loads and stores, then do a (aligned) load from the stack slot.
- EVT RegVT = TLI.getRegisterType(*DAG.getContext(), intVT);
+ MVT RegVT = TLI.getRegisterType(*DAG.getContext(), intVT);
unsigned LoadedBytes = LoadedVT.getSizeInBits() / 8;
unsigned RegBytes = RegVT.getSizeInBits() / 8;
unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes;
SDValue Load = DAG.getLoad(RegVT, dl, Chain, Ptr,
LD->getPointerInfo().getWithOffset(Offset),
LD->isVolatile(), LD->isNonTemporal(),
+ LD->isInvariant(),
MinAlign(LD->getAlignment(), Offset));
// Follow the load with a store to the stack slot. Remember the store.
Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, StackPtr,
// 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.
- SDValue Ops[] = { Load, TF };
- return DAG.getMergeValues(Ops, 2, dl);
+ ValResult = Load;
+ ChainResult = TF;
+ return;
}
assert(LoadedVT.isInteger() && !LoadedVT.isVector() &&
"Unaligned load of unsupported type.");
// 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);
SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
Hi.getValue(1));
- SDValue Ops[] = { Result, TF };
- return DAG.getMergeValues(Ops, 2, dl);
+ ValResult = Result;
+ ChainResult = TF;
}
/// PerformInsertVectorEltInMemory - Some target cannot handle a variable
false, false, 0);
// Load the updated vector.
return DAG.getLoad(VT, dl, Ch, StackPtr,
- MachinePointerInfo::getFixedStack(SPFI), false, false, 0);
+ MachinePointerInfo::getFixedStack(SPFI), false, false,
+ false, 0);
}
// probably means that we need to integrate dag combiner and legalizer
// together.
// We generally can't do this one for long doubles.
- SDValue Tmp1 = ST->getChain();
- SDValue Tmp2 = ST->getBasePtr();
- SDValue Tmp3;
+ SDValue Chain = ST->getChain();
+ SDValue Ptr = ST->getBasePtr();
unsigned Alignment = ST->getAlignment();
bool isVolatile = ST->isVolatile();
bool isNonTemporal = ST->isNonTemporal();
DebugLoc dl = ST->getDebugLoc();
if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) {
if (CFP->getValueType(0) == MVT::f32 &&
- getTypeAction(MVT::i32) == Legal) {
- Tmp3 = DAG.getConstant(CFP->getValueAPF().
+ TLI.isTypeLegal(MVT::i32)) {
+ SDValue Con = DAG.getConstant(CFP->getValueAPF().
bitcastToAPInt().zextOrTrunc(32),
MVT::i32);
- return DAG.getStore(Tmp1, dl, Tmp3, Tmp2, ST->getPointerInfo(),
+ return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
isVolatile, isNonTemporal, Alignment);
}
if (CFP->getValueType(0) == MVT::f64) {
// If this target supports 64-bit registers, do a single 64-bit store.
- if (getTypeAction(MVT::i64) == Legal) {
- Tmp3 = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
+ if (TLI.isTypeLegal(MVT::i64)) {
+ SDValue Con = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
zextOrTrunc(64), MVT::i64);
- return DAG.getStore(Tmp1, dl, Tmp3, Tmp2, ST->getPointerInfo(),
+ return DAG.getStore(Chain, dl, Con, Ptr, 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.
SDValue Hi = DAG.getConstant(IntVal.lshr(32).trunc(32), MVT::i32);
if (TLI.isBigEndian()) std::swap(Lo, Hi);
- Lo = DAG.getStore(Tmp1, dl, Lo, Tmp2, ST->getPointerInfo(), isVolatile,
+ Lo = DAG.getStore(Chain, dl, Lo, Ptr, ST->getPointerInfo(), isVolatile,
isNonTemporal, Alignment);
- Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
DAG.getIntPtrConstant(4));
- Hi = DAG.getStore(Tmp1, dl, Hi, Tmp2,
+ Hi = DAG.getStore(Chain, dl, Hi, Ptr,
ST->getPointerInfo().getWithOffset(4),
isVolatile, isNonTemporal, MinAlign(Alignment, 4U));
}
}
}
- 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.
-SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) {
- if (Op.getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
- return Op;
+void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) {
+ StoreSDNode *ST = cast<StoreSDNode>(Node);
+ SDValue Chain = ST->getChain();
+ SDValue Ptr = ST->getBasePtr();
+ DebugLoc dl = Node->getDebugLoc();
- SDNode *Node = Op.getNode();
+ unsigned Alignment = ST->getAlignment();
+ bool isVolatile = ST->isVolatile();
+ bool isNonTemporal = ST->isNonTemporal();
+
+ if (!ST->isTruncatingStore()) {
+ if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
+ ReplaceNode(ST, OptStore);
+ return;
+ }
+
+ {
+ SDValue Value = ST->getValue();
+ MVT VT = Value.getSimpleValueType();
+ switch (TLI.getOperationAction(ISD::STORE, VT)) {
+ default: llvm_unreachable("This action is not supported yet!");
+ case TargetLowering::Legal:
+ // If this is an unaligned store and the target doesn't support it,
+ // expand it.
+ if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT())) {
+ Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
+ unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
+ if (ST->getAlignment() < ABIAlignment)
+ ExpandUnalignedStore(cast<StoreSDNode>(Node),
+ DAG, TLI, this);
+ }
+ break;
+ case TargetLowering::Custom: {
+ SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
+ if (Res.getNode())
+ ReplaceNode(SDValue(Node, 0), Res);
+ return;
+ }
+ case TargetLowering::Promote: {
+ MVT NVT = TLI.getTypeToPromoteTo(ISD::STORE, VT);
+ assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
+ "Can only promote stores to same size type");
+ Value = DAG.getNode(ISD::BITCAST, dl, NVT, Value);
+ SDValue Result =
+ DAG.getStore(Chain, dl, Value, Ptr,
+ ST->getPointerInfo(), isVolatile,
+ isNonTemporal, Alignment);
+ ReplaceNode(SDValue(Node, 0), Result);
+ break;
+ }
+ }
+ return;
+ }
+ } else {
+ SDValue Value = ST->getValue();
+
+ EVT StVT = ST->getMemoryVT();
+ unsigned StWidth = StVT.getSizeInBits();
+
+ if (StWidth != StVT.getStoreSizeInBits()) {
+ // Promote to a byte-sized store with upper bits zero if not
+ // storing an integral number of bytes. For example, promote
+ // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
+ EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
+ StVT.getStoreSizeInBits());
+ Value = DAG.getZeroExtendInReg(Value, dl, StVT);
+ SDValue Result =
+ DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
+ NVT, isVolatile, isNonTemporal, Alignment);
+ ReplaceNode(SDValue(Node, 0), Result);
+ } else if (StWidth & (StWidth - 1)) {
+ // If not storing a power-of-2 number of bits, expand as two stores.
+ assert(!StVT.isVector() && "Unsupported truncstore!");
+ unsigned RoundWidth = 1 << Log2_32(StWidth);
+ assert(RoundWidth < StWidth);
+ unsigned ExtraWidth = StWidth - RoundWidth;
+ assert(ExtraWidth < RoundWidth);
+ assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
+ "Store size not an integral number of bytes!");
+ EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
+ EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
+ SDValue Lo, Hi;
+ unsigned IncrementSize;
+
+ if (TLI.isLittleEndian()) {
+ // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16)
+ // Store the bottom RoundWidth bits.
+ Lo = DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
+ RoundVT,
+ isVolatile, isNonTemporal, Alignment);
+
+ // Store the remaining ExtraWidth bits.
+ IncrementSize = RoundWidth / 8;
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
+ DAG.getIntPtrConstant(IncrementSize));
+ Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
+ DAG.getConstant(RoundWidth,
+ TLI.getShiftAmountTy(Value.getValueType())));
+ Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr,
+ ST->getPointerInfo().getWithOffset(IncrementSize),
+ ExtraVT, isVolatile, isNonTemporal,
+ MinAlign(Alignment, IncrementSize));
+ } else {
+ // Big endian - avoid unaligned stores.
+ // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
+ // Store the top RoundWidth bits.
+ Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
+ DAG.getConstant(ExtraWidth,
+ TLI.getShiftAmountTy(Value.getValueType())));
+ Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, ST->getPointerInfo(),
+ RoundVT, isVolatile, isNonTemporal, Alignment);
+
+ // Store the remaining ExtraWidth bits.
+ IncrementSize = RoundWidth / 8;
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
+ DAG.getIntPtrConstant(IncrementSize));
+ Lo = DAG.getTruncStore(Chain, dl, Value, Ptr,
+ ST->getPointerInfo().getWithOffset(IncrementSize),
+ ExtraVT, isVolatile, isNonTemporal,
+ MinAlign(Alignment, IncrementSize));
+ }
+
+ // The order of the stores doesn't matter.
+ SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
+ ReplaceNode(SDValue(Node, 0), Result);
+ } else {
+ switch (TLI.getTruncStoreAction(ST->getValue().getSimpleValueType(),
+ StVT.getSimpleVT())) {
+ default: llvm_unreachable("This action is not supported yet!");
+ case TargetLowering::Legal:
+ // If this is an unaligned store and the target doesn't support it,
+ // expand it.
+ if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT())) {
+ Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
+ unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
+ if (ST->getAlignment() < ABIAlignment)
+ ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
+ }
+ break;
+ case TargetLowering::Custom: {
+ SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
+ if (Res.getNode())
+ ReplaceNode(SDValue(Node, 0), Res);
+ return;
+ }
+ case TargetLowering::Expand:
+ assert(!StVT.isVector() &&
+ "Vector Stores are handled in LegalizeVectorOps");
+
+ // TRUNCSTORE:i16 i32 -> STORE i16
+ assert(TLI.isTypeLegal(StVT) &&
+ "Do not know how to expand this store!");
+ Value = DAG.getNode(ISD::TRUNCATE, dl, StVT, Value);
+ SDValue Result =
+ DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
+ isVolatile, isNonTemporal, Alignment);
+ ReplaceNode(SDValue(Node, 0), Result);
+ break;
+ }
+ }
+ }
+}
+
+void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) {
+ LoadSDNode *LD = cast<LoadSDNode>(Node);
+ SDValue Chain = LD->getChain(); // The chain.
+ SDValue Ptr = LD->getBasePtr(); // The base pointer.
+ SDValue Value; // The value returned by the load op.
DebugLoc dl = Node->getDebugLoc();
+ ISD::LoadExtType ExtType = LD->getExtensionType();
+ if (ExtType == ISD::NON_EXTLOAD) {
+ MVT VT = Node->getSimpleValueType(0);
+ SDValue RVal = SDValue(Node, 0);
+ SDValue RChain = SDValue(Node, 1);
+
+ switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
+ default: llvm_unreachable("This action is not supported yet!");
+ case TargetLowering::Legal:
+ // If this is an unaligned load and the target doesn't support it,
+ // expand it.
+ if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT())) {
+ Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
+ unsigned ABIAlignment =
+ TLI.getDataLayout()->getABITypeAlignment(Ty);
+ if (LD->getAlignment() < ABIAlignment){
+ ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, RVal, RChain);
+ }
+ }
+ break;
+ case TargetLowering::Custom: {
+ SDValue Res = TLI.LowerOperation(RVal, DAG);
+ if (Res.getNode()) {
+ RVal = Res;
+ RChain = Res.getValue(1);
+ }
+ break;
+ }
+ case TargetLowering::Promote: {
+ MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
+ assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
+ "Can only promote loads to same size type");
+
+ SDValue Res = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getPointerInfo(),
+ LD->isVolatile(), LD->isNonTemporal(),
+ LD->isInvariant(), LD->getAlignment());
+ RVal = DAG.getNode(ISD::BITCAST, dl, VT, Res);
+ RChain = Res.getValue(1);
+ break;
+ }
+ }
+ if (RChain.getNode() != Node) {
+ assert(RVal.getNode() != Node && "Load must be completely replaced");
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), RVal);
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), RChain);
+ ReplacedNode(Node);
+ }
+ return;
+ }
+
+ EVT SrcVT = LD->getMemoryVT();
+ unsigned SrcWidth = SrcVT.getSizeInBits();
+ unsigned Alignment = LD->getAlignment();
+ bool isVolatile = LD->isVolatile();
+ bool isNonTemporal = LD->isNonTemporal();
+
+ if (SrcWidth != SrcVT.getStoreSizeInBits() &&
+ // Some targets pretend to have an i1 loading operation, and actually
+ // load an i8. This trick is correct for ZEXTLOAD because the top 7
+ // bits are guaranteed to be zero; it helps the optimizers understand
+ // that these bits are zero. It is also useful for EXTLOAD, since it
+ // tells the optimizers that those bits are undefined. It would be
+ // nice to have an effective generic way of getting these benefits...
+ // Until such a way is found, don't insist on promoting i1 here.
+ (SrcVT != MVT::i1 ||
+ TLI.getLoadExtAction(ExtType, MVT::i1) == TargetLowering::Promote)) {
+ // Promote to a byte-sized load if not loading an integral number of
+ // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
+ unsigned NewWidth = SrcVT.getStoreSizeInBits();
+ EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
+ SDValue Ch;
+
+ // The extra bits are guaranteed to be zero, since we stored them that
+ // way. A zext load from NVT thus automatically gives zext from SrcVT.
+
+ ISD::LoadExtType NewExtType =
+ ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
+
+ SDValue Result =
+ DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
+ Chain, Ptr, LD->getPointerInfo(),
+ NVT, isVolatile, isNonTemporal, Alignment);
+
+ Ch = Result.getValue(1); // The chain.
+
+ if (ExtType == ISD::SEXTLOAD)
+ // Having the top bits zero doesn't help when sign extending.
+ Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
+ Result.getValueType(),
+ Result, DAG.getValueType(SrcVT));
+ else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
+ // All the top bits are guaranteed to be zero - inform the optimizers.
+ Result = DAG.getNode(ISD::AssertZext, dl,
+ Result.getValueType(), Result,
+ DAG.getValueType(SrcVT));
+
+ Value = Result;
+ Chain = Ch;
+ } else if (SrcWidth & (SrcWidth - 1)) {
+ // If not loading a power-of-2 number of bits, expand as two loads.
+ assert(!SrcVT.isVector() && "Unsupported extload!");
+ unsigned RoundWidth = 1 << Log2_32(SrcWidth);
+ assert(RoundWidth < SrcWidth);
+ unsigned ExtraWidth = SrcWidth - RoundWidth;
+ assert(ExtraWidth < RoundWidth);
+ assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
+ "Load size not an integral number of bytes!");
+ EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
+ EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
+ SDValue Lo, Hi, Ch;
+ unsigned IncrementSize;
+
+ if (TLI.isLittleEndian()) {
+ // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
+ // Load the bottom RoundWidth bits.
+ Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
+ Chain, Ptr,
+ LD->getPointerInfo(), RoundVT, isVolatile,
+ isNonTemporal, Alignment);
+
+ // Load the remaining ExtraWidth bits.
+ IncrementSize = RoundWidth / 8;
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
+ DAG.getIntPtrConstant(IncrementSize));
+ Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
+ LD->getPointerInfo().getWithOffset(IncrementSize),
+ ExtraVT, isVolatile, isNonTemporal,
+ MinAlign(Alignment, IncrementSize));
+
+ // Build a factor node to remember that this load is independent of
+ // the other one.
+ Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
+ Hi.getValue(1));
+
+ // Move the top bits to the right place.
+ Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
+ DAG.getConstant(RoundWidth,
+ TLI.getShiftAmountTy(Hi.getValueType())));
+
+ // Join the hi and lo parts.
+ Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
+ } else {
+ // Big endian - avoid unaligned loads.
+ // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
+ // Load the top RoundWidth bits.
+ Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
+ LD->getPointerInfo(), RoundVT, isVolatile,
+ isNonTemporal, Alignment);
+
+ // Load the remaining ExtraWidth bits.
+ IncrementSize = RoundWidth / 8;
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
+ DAG.getIntPtrConstant(IncrementSize));
+ Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
+ dl, Node->getValueType(0), Chain, Ptr,
+ LD->getPointerInfo().getWithOffset(IncrementSize),
+ ExtraVT, isVolatile, isNonTemporal,
+ MinAlign(Alignment, IncrementSize));
+
+ // Build a factor node to remember that this load is independent of
+ // the other one.
+ Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
+ Hi.getValue(1));
+
+ // Move the top bits to the right place.
+ Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
+ DAG.getConstant(ExtraWidth,
+ TLI.getShiftAmountTy(Hi.getValueType())));
+
+ // Join the hi and lo parts.
+ Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
+ }
+
+ Chain = Ch;
+ } else {
+ bool isCustom = false;
+ switch (TLI.getLoadExtAction(ExtType, SrcVT.getSimpleVT())) {
+ default: llvm_unreachable("This action is not supported yet!");
+ case TargetLowering::Custom:
+ isCustom = true;
+ // FALLTHROUGH
+ case TargetLowering::Legal: {
+ Value = SDValue(Node, 0);
+ Chain = SDValue(Node, 1);
+
+ if (isCustom) {
+ SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
+ if (Res.getNode()) {
+ Value = Res;
+ Chain = Res.getValue(1);
+ }
+ } else {
+ // If this is an unaligned load and the target doesn't support it,
+ // expand it.
+ if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT())) {
+ Type *Ty =
+ LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
+ unsigned ABIAlignment =
+ TLI.getDataLayout()->getABITypeAlignment(Ty);
+ if (LD->getAlignment() < ABIAlignment){
+ ExpandUnalignedLoad(cast<LoadSDNode>(Node),
+ DAG, TLI, Value, Chain);
+ }
+ }
+ }
+ break;
+ }
+ case TargetLowering::Expand:
+ if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) && TLI.isTypeLegal(SrcVT)) {
+ SDValue Load = DAG.getLoad(SrcVT, dl, Chain, Ptr,
+ LD->getPointerInfo(),
+ LD->isVolatile(), LD->isNonTemporal(),
+ LD->isInvariant(), LD->getAlignment());
+ unsigned ExtendOp;
+ switch (ExtType) {
+ case ISD::EXTLOAD:
+ ExtendOp = (SrcVT.isFloatingPoint() ?
+ ISD::FP_EXTEND : ISD::ANY_EXTEND);
+ break;
+ case ISD::SEXTLOAD: ExtendOp = ISD::SIGN_EXTEND; break;
+ case ISD::ZEXTLOAD: ExtendOp = ISD::ZERO_EXTEND; break;
+ default: llvm_unreachable("Unexpected extend load type!");
+ }
+ Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
+ Chain = Load.getValue(1);
+ break;
+ }
+
+ assert(!SrcVT.isVector() &&
+ "Vector Loads are handled in LegalizeVectorOps");
+
+ // 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
+ // support for legalizing them.
+ assert(ExtType != ISD::EXTLOAD &&
+ "EXTLOAD should always be supported!");
+ // Turn the unsupported load into an EXTLOAD followed by an explicit
+ // zero/sign extend inreg.
+ SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl, Node->getValueType(0),
+ Chain, Ptr, LD->getPointerInfo(), SrcVT,
+ LD->isVolatile(), LD->isNonTemporal(),
+ LD->getAlignment());
+ SDValue ValRes;
+ if (ExtType == ISD::SEXTLOAD)
+ ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
+ Result.getValueType(),
+ Result, DAG.getValueType(SrcVT));
+ else
+ ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT.getScalarType());
+ Value = ValRes;
+ Chain = Result.getValue(1);
+ break;
+ }
+ }
+
+ // Since loads produce two values, make sure to remember that we legalized
+ // both of them.
+ if (Chain.getNode() != Node) {
+ assert(Value.getNode() != Node && "Load must be completely replaced");
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value);
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
+ ReplacedNode(Node);
+ }
+}
+
+/// LegalizeOp - Return a legal replacement for the given operation, with
+/// all legal operands.
+void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
+ if (Node->getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
+ return;
+
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!");
- // Note that LegalizeOp may be reentered even from single-use nodes, which
- // means that we always must cache transformed nodes.
- DenseMap<SDValue, SDValue>::iterator I = LegalizedNodes.find(Op);
- if (I != LegalizedNodes.end()) return I->second;
-
- SDValue Tmp1, Tmp2, Tmp3, Tmp4;
- SDValue Result = Op;
- 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:
case ISD::INTRINSIC_WO_CHAIN:
case ISD::INTRINSIC_VOID:
- case ISD::VAARG:
case ISD::STACKSAVE:
Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
break;
+ case ISD::VAARG:
+ Action = TLI.getOperationAction(Node->getOpcode(),
+ Node->getValueType(0));
+ if (Action != TargetLowering::Promote)
+ Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
+ break;
case ISD::SINT_TO_FP:
case ISD::UINT_TO_FP:
case ISD::EXTRACT_VECTOR_ELT:
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: {
unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
Node->getOpcode() == ISD::SETCC ? 2 : 1;
unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
- EVT OpVT = Node->getOperand(CompareOperand).getValueType();
+ MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType();
ISD::CondCode CCCode =
cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
Action = TLI.getCondCodeAction(CCCode, OpVT);
case ISD::FRAME_TO_ARGS_OFFSET:
case ISD::EH_SJLJ_SETJMP:
case ISD::EH_SJLJ_LONGJMP:
- case ISD::EH_SJLJ_DISPATCHSETUP:
// These operations lie about being legal: when they claim to be legal,
// they should actually be expanded.
Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
if (Action == TargetLowering::Legal)
Action = TargetLowering::Expand;
break;
- case ISD::TRAMPOLINE:
+ case ISD::INIT_TRAMPOLINE:
+ case ISD::ADJUST_TRAMPOLINE:
case ISD::FRAMEADDR:
case ISD::RETURNADDR:
// These operations lie about being legal: when they claim to be legal,
if (Action == TargetLowering::Legal)
Action = TargetLowering::Custom;
break;
- case ISD::BUILD_VECTOR:
- // A weird case: legalization for BUILD_VECTOR never legalizes the
- // operands!
- // FIXME: This really sucks... changing it isn't semantically incorrect,
- // but it massively pessimizes the code for floating-point BUILD_VECTORs
- // because ConstantFP operands get legalized into constant pool loads
- // before the BUILD_VECTOR code can see them. It doesn't usually bite,
- // though, because BUILD_VECTORS usually get lowered into other nodes
- // which get legalized properly.
- SimpleFinishLegalizing = false;
+ case ISD::DEBUGTRAP:
+ Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
+ if (Action == TargetLowering::Expand) {
+ // replace ISD::DEBUGTRAP with ISD::TRAP
+ SDValue NewVal;
+ NewVal = DAG.getNode(ISD::TRAP, Node->getDebugLoc(), Node->getVTList(),
+ Node->getOperand(0));
+ ReplaceNode(Node, NewVal.getNode());
+ LegalizeOp(NewVal.getNode());
+ return;
+ }
break;
+
default:
if (Node->getOpcode() >= ISD::BUILTIN_OP_END) {
Action = TargetLowering::Legal;
} else {
Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
- }
- break;
- }
-
- if (SimpleFinishLegalizing) {
- SmallVector<SDValue, 8> Ops, ResultVals;
- for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
- Ops.push_back(LegalizeOp(Node->getOperand(i)));
- switch (Node->getOpcode()) {
- default: break;
- case ISD::BR:
- case ISD::BRIND:
- case ISD::BR_JT:
- 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);
- Ops[0] = LegalizeOp(Ops[0]);
- LastCALLSEQ_END = DAG.getEntryNode();
- break;
- case ISD::SHL:
- case ISD::SRL:
- case ISD::SRA:
- case ISD::ROTL:
- case ISD::ROTR:
- // Legalizing shifts/rotates requires adjusting the shift amount
- // to the appropriate width.
- if (!Ops[1].getValueType().isVector())
- Ops[1] = LegalizeOp(DAG.getShiftAmountOperand(Ops[1]));
- break;
- case ISD::SRL_PARTS:
- case ISD::SRA_PARTS:
- case ISD::SHL_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]));
- break;
- }
-
- Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(), Ops.data(),
- Ops.size()), 0);
- switch (Action) {
- case TargetLowering::Legal:
- for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
- ResultVals.push_back(Result.getValue(i));
- break;
- case TargetLowering::Custom:
- // FIXME: The handling for custom lowering with multiple results is
- // a complete mess.
- Tmp1 = TLI.LowerOperation(Result, DAG);
- if (Tmp1.getNode()) {
- for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i) {
- if (e == 1)
- ResultVals.push_back(Tmp1);
- else
- ResultVals.push_back(Tmp1.getValue(i));
- }
- break;
- }
-
- // FALL THROUGH
- case TargetLowering::Expand:
- ExpandNode(Result.getNode(), ResultVals);
- break;
- case TargetLowering::Promote:
- PromoteNode(Result.getNode(), ResultVals);
- break;
- }
- if (!ResultVals.empty()) {
- for (unsigned i = 0, e = ResultVals.size(); i != e; ++i) {
- if (ResultVals[i] != SDValue(Node, i))
- ResultVals[i] = LegalizeOp(ResultVals[i]);
- AddLegalizedOperand(SDValue(Node, i), ResultVals[i]);
- }
- return ResultVals[Op.getResNo()];
- }
- }
-
- switch (Node->getOpcode()) {
- default:
-#ifndef NDEBUG
- dbgs() << "NODE: ";
- Node->dump( &DAG);
- dbgs() << "\n";
-#endif
- assert(0 && "Do not know how to legalize this operator!");
-
- case ISD::BUILD_VECTOR:
- switch (TLI.getOperationAction(ISD::BUILD_VECTOR, Node->getValueType(0))) {
- default: assert(0 && "This action is not supported yet!");
- case TargetLowering::Custom:
- Tmp3 = TLI.LowerOperation(Result, DAG);
- if (Tmp3.getNode()) {
- Result = Tmp3;
- break;
- }
- // FALLTHROUGH
- case TargetLowering::Expand:
- Result = ExpandBUILD_VECTOR(Result.getNode());
- break;
- }
- break;
- case ISD::CALLSEQ_START: {
- SDNode *CallEnd = FindCallEndFromCallStart(Node);
-
- // Recursively Legalize all of the inputs of the call end that do not lead
- // to this call start. This ensures that any libcalls that need be inserted
- // are inserted *before* the CALLSEQ_START.
- {SmallPtrSet<SDNode*, 32> NodesLeadingTo;
- for (unsigned i = 0, e = CallEnd->getNumOperands(); i != e; ++i)
- LegalizeAllNodesNotLeadingTo(CallEnd->getOperand(i).getNode(), Node,
- NodesLeadingTo);
- }
-
- // Now that we have legalized all of the inputs (which may have inserted
- // libcalls), create the new CALLSEQ_START node.
- Tmp1 = LegalizeOp(Node->getOperand(0)); // Legalize the chain.
-
- // Merge in the last call to ensure that this call starts after the last
- // call ended.
- if (LastCALLSEQ_END.getOpcode() != ISD::EntryToken) {
- Tmp1 = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
- Tmp1, LastCALLSEQ_END);
- Tmp1 = LegalizeOp(Tmp1);
- }
-
- // Do not try to legalize the target-specific arguments (#1+).
- if (Tmp1 != Node->getOperand(0)) {
- SmallVector<SDValue, 8> Ops(Node->op_begin(), Node->op_end());
- Ops[0] = Tmp1;
- Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(), &Ops[0],
- Ops.size()), Result.getResNo());
- }
-
- // Remember that the CALLSEQ_START is legalized.
- AddLegalizedOperand(Op.getValue(0), Result);
- if (Node->getNumValues() == 2) // If this has a flag result, remember it.
- AddLegalizedOperand(Op.getValue(1), Result.getValue(1));
-
- // Now that the callseq_start and all of the non-call nodes above this call
- // sequence have been legalized, legalize the call itself. During this
- // process, no libcalls can/will be inserted, guaranteeing that no calls
- // can overlap.
- assert(!IsLegalizingCall && "Inconsistent sequentialization of calls!");
- // Note that we are selecting this call!
- LastCALLSEQ_END = SDValue(CallEnd, 0);
- IsLegalizingCall = true;
-
- // Legalize the call, starting from the CALLSEQ_END.
- LegalizeOp(LastCALLSEQ_END);
- assert(!IsLegalizingCall && "CALLSEQ_END should have cleared this!");
- return Result;
- }
- case ISD::CALLSEQ_END:
- // If the CALLSEQ_START node hasn't been legalized first, legalize it. This
- // will cause this node to be legalized as well as handling libcalls right.
- if (LastCALLSEQ_END.getNode() != Node) {
- LegalizeOp(SDValue(FindCallStartFromCallEnd(Node), 0));
- DenseMap<SDValue, SDValue>::iterator I = LegalizedNodes.find(Op);
- assert(I != LegalizedNodes.end() &&
- "Legalizing the call start should have legalized this node!");
- return I->second;
- }
-
- // Otherwise, the call start has been legalized and everything is going
- // according to plan. Just legalize ourselves normally here.
- Tmp1 = LegalizeOp(Node->getOperand(0)); // Legalize the chain.
- // Do not try to legalize the target-specific arguments (#1+), except for
- // an optional flag input.
- if (Node->getOperand(Node->getNumOperands()-1).getValueType() != MVT::Glue){
- if (Tmp1 != Node->getOperand(0)) {
- SmallVector<SDValue, 8> Ops(Node->op_begin(), Node->op_end());
- Ops[0] = Tmp1;
- Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(),
- &Ops[0], Ops.size()),
- Result.getResNo());
- }
- } else {
- Tmp2 = LegalizeOp(Node->getOperand(Node->getNumOperands()-1));
- if (Tmp1 != Node->getOperand(0) ||
- Tmp2 != Node->getOperand(Node->getNumOperands()-1)) {
- SmallVector<SDValue, 8> Ops(Node->op_begin(), Node->op_end());
- Ops[0] = Tmp1;
- Ops.back() = Tmp2;
- Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(),
- &Ops[0], Ops.size()),
- 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)
- AddLegalizedOperand(SDValue(Node, 1), Result.getValue(1));
- return Result.getValue(Op.getResNo());
- case ISD::LOAD: {
- LoadSDNode *LD = cast<LoadSDNode>(Node);
- Tmp1 = LegalizeOp(LD->getChain()); // Legalize the chain.
- Tmp2 = LegalizeOp(LD->getBasePtr()); // Legalize the base pointer.
-
- ISD::LoadExtType ExtType = LD->getExtensionType();
- if (ExtType == ISD::NON_EXTLOAD) {
- EVT VT = Node->getValueType(0);
- Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(),
- Tmp1, Tmp2, LD->getOffset()),
- Result.getResNo());
- Tmp3 = Result.getValue(0);
- Tmp4 = Result.getValue(1);
-
- switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
- default: assert(0 && "This action is not supported yet!");
- case TargetLowering::Legal:
- // 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());
- unsigned ABIAlignment = TLI.getTargetData()->getABITypeAlignment(Ty);
- if (LD->getAlignment() < ABIAlignment){
- Result = ExpandUnalignedLoad(cast<LoadSDNode>(Result.getNode()),
- DAG, TLI);
- Tmp3 = Result.getOperand(0);
- Tmp4 = Result.getOperand(1);
- Tmp3 = LegalizeOp(Tmp3);
- Tmp4 = LegalizeOp(Tmp4);
- }
- }
- break;
- case TargetLowering::Custom:
- Tmp1 = TLI.LowerOperation(Tmp3, DAG);
- if (Tmp1.getNode()) {
- Tmp3 = LegalizeOp(Tmp1);
- Tmp4 = LegalizeOp(Tmp1.getValue(1));
- }
- break;
- case TargetLowering::Promote: {
- // Only promote a load of vector type to another.
- assert(VT.isVector() && "Cannot promote this load!");
- // Change base type to a different vector type.
- EVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
-
- Tmp1 = DAG.getLoad(NVT, dl, Tmp1, Tmp2, LD->getPointerInfo(),
- LD->isVolatile(), LD->isNonTemporal(),
- LD->getAlignment());
- Tmp3 = LegalizeOp(DAG.getNode(ISD::BITCAST, dl, VT, Tmp1));
- Tmp4 = LegalizeOp(Tmp1.getValue(1));
- break;
- }
- }
- // Since loads produce two values, make sure to remember that we
- // legalized both of them.
- AddLegalizedOperand(SDValue(Node, 0), Tmp3);
- AddLegalizedOperand(SDValue(Node, 1), Tmp4);
- return Op.getResNo() ? Tmp4 : Tmp3;
- }
-
- EVT SrcVT = LD->getMemoryVT();
- unsigned SrcWidth = SrcVT.getSizeInBits();
- unsigned Alignment = LD->getAlignment();
- bool isVolatile = LD->isVolatile();
- bool isNonTemporal = LD->isNonTemporal();
-
- if (SrcWidth != SrcVT.getStoreSizeInBits() &&
- // Some targets pretend to have an i1 loading operation, and actually
- // load an i8. This trick is correct for ZEXTLOAD because the top 7
- // bits are guaranteed to be zero; it helps the optimizers understand
- // that these bits are zero. It is also useful for EXTLOAD, since it
- // tells the optimizers that those bits are undefined. It would be
- // nice to have an effective generic way of getting these benefits...
- // Until such a way is found, don't insist on promoting i1 here.
- (SrcVT != MVT::i1 ||
- TLI.getLoadExtAction(ExtType, MVT::i1) == TargetLowering::Promote)) {
- // Promote to a byte-sized load if not loading an integral number of
- // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
- unsigned NewWidth = SrcVT.getStoreSizeInBits();
- EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
- SDValue Ch;
-
- // The extra bits are guaranteed to be zero, since we stored them that
- // way. A zext load from NVT thus automatically gives zext from SrcVT.
-
- ISD::LoadExtType NewExtType =
- ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
-
- Result = DAG.getExtLoad(NewExtType, Node->getValueType(0), dl,
- Tmp1, Tmp2, LD->getPointerInfo(),
- NVT, isVolatile, isNonTemporal, Alignment);
-
- Ch = Result.getValue(1); // The chain.
-
- if (ExtType == ISD::SEXTLOAD)
- // Having the top bits zero doesn't help when sign extending.
- Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
- Result.getValueType(),
- Result, DAG.getValueType(SrcVT));
- else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
- // All the top bits are guaranteed to be zero - inform the optimizers.
- Result = DAG.getNode(ISD::AssertZext, dl,
- Result.getValueType(), Result,
- DAG.getValueType(SrcVT));
-
- Tmp1 = LegalizeOp(Result);
- Tmp2 = LegalizeOp(Ch);
- } else if (SrcWidth & (SrcWidth - 1)) {
- // If not loading a power-of-2 number of bits, expand as two loads.
- assert(!SrcVT.isVector() && "Unsupported extload!");
- unsigned RoundWidth = 1 << Log2_32(SrcWidth);
- assert(RoundWidth < SrcWidth);
- unsigned ExtraWidth = SrcWidth - RoundWidth;
- assert(ExtraWidth < RoundWidth);
- assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
- "Load size not an integral number of bytes!");
- EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
- EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
- SDValue Lo, Hi, Ch;
- unsigned IncrementSize;
-
- 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,
- Tmp1, Tmp2,
- LD->getPointerInfo(), RoundVT, isVolatile,
- isNonTemporal, Alignment);
-
- // Load the remaining ExtraWidth bits.
- 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,
- LD->getPointerInfo().getWithOffset(IncrementSize),
- ExtraVT, isVolatile, isNonTemporal,
- MinAlign(Alignment, IncrementSize));
-
- // Build a factor node to remember that this load is independent of
- // the other one.
- Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
- Hi.getValue(1));
-
- // Move the top bits to the right place.
- Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
- DAG.getConstant(RoundWidth, TLI.getShiftAmountTy()));
-
- // Join the hi and lo parts.
- Result = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
- } else {
- // 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,
- LD->getPointerInfo(), RoundVT, isVolatile,
- isNonTemporal, Alignment);
-
- // Load the remaining ExtraWidth bits.
- IncrementSize = RoundWidth / 8;
- Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
- DAG.getIntPtrConstant(IncrementSize));
- Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
- Node->getValueType(0), dl, Tmp1, Tmp2,
- LD->getPointerInfo().getWithOffset(IncrementSize),
- ExtraVT, isVolatile, isNonTemporal,
- MinAlign(Alignment, IncrementSize));
-
- // Build a factor node to remember that this load is independent of
- // the other one.
- Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
- Hi.getValue(1));
-
- // Move the top bits to the right place.
- Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
- DAG.getConstant(ExtraWidth, TLI.getShiftAmountTy()));
-
- // Join the hi and lo parts.
- Result = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
- }
-
- Tmp1 = LegalizeOp(Result);
- Tmp2 = LegalizeOp(Ch);
- } else {
- switch (TLI.getLoadExtAction(ExtType, SrcVT)) {
- default: assert(0 && "This action is not supported yet!");
- case TargetLowering::Custom:
- isCustom = true;
- // FALLTHROUGH
- case TargetLowering::Legal:
- Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(),
- Tmp1, Tmp2, LD->getOffset()),
- Result.getResNo());
- Tmp1 = Result.getValue(0);
- Tmp2 = Result.getValue(1);
-
- if (isCustom) {
- Tmp3 = TLI.LowerOperation(Result, DAG);
- if (Tmp3.getNode()) {
- Tmp1 = LegalizeOp(Tmp3);
- Tmp2 = LegalizeOp(Tmp3.getValue(1));
- }
- } else {
- // 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());
- unsigned ABIAlignment =
- TLI.getTargetData()->getABITypeAlignment(Ty);
- if (LD->getAlignment() < ABIAlignment){
- Result = ExpandUnalignedLoad(cast<LoadSDNode>(Result.getNode()),
- DAG, TLI);
- Tmp1 = Result.getOperand(0);
- Tmp2 = Result.getOperand(1);
- Tmp1 = LegalizeOp(Tmp1);
- Tmp2 = LegalizeOp(Tmp2);
- }
- }
- }
- break;
- case TargetLowering::Expand:
- if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) && isTypeLegal(SrcVT)) {
- SDValue Load = DAG.getLoad(SrcVT, dl, Tmp1, Tmp2,
- LD->getPointerInfo(),
- LD->isVolatile(), LD->isNonTemporal(),
- LD->getAlignment());
- unsigned ExtendOp;
- switch (ExtType) {
- case ISD::EXTLOAD:
- ExtendOp = (SrcVT.isFloatingPoint() ?
- ISD::FP_EXTEND : ISD::ANY_EXTEND);
- break;
- case ISD::SEXTLOAD: ExtendOp = ISD::SIGN_EXTEND; break;
- case ISD::ZEXTLOAD: ExtendOp = ISD::ZERO_EXTEND; break;
- default: llvm_unreachable("Unexpected extend load type!");
- }
- Result = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
- Tmp1 = LegalizeOp(Result); // Relegalize new nodes.
- Tmp2 = LegalizeOp(Load.getValue(1));
- 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
- // support for legalizing them.
- assert(ExtType != ISD::EXTLOAD &&
- "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,
- Tmp1, Tmp2, LD->getPointerInfo(), SrcVT,
- LD->isVolatile(), LD->isNonTemporal(),
- LD->getAlignment());
- SDValue ValRes;
- if (ExtType == ISD::SEXTLOAD)
- ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
- Result.getValueType(),
- Result, DAG.getValueType(SrcVT));
- else
- ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT);
- Tmp1 = LegalizeOp(ValRes); // Relegalize new nodes.
- Tmp2 = LegalizeOp(Result.getValue(1)); // Relegalize new nodes.
- break;
- }
- }
-
- // Since loads produce two values, make sure to remember that we legalized
- // both of them.
- AddLegalizedOperand(SDValue(Node, 0), Tmp1);
- AddLegalizedOperand(SDValue(Node, 1), Tmp2);
- return Op.getResNo() ? Tmp2 : Tmp1;
- }
- case ISD::STORE: {
- StoreSDNode *ST = cast<StoreSDNode>(Node);
- Tmp1 = LegalizeOp(ST->getChain()); // Legalize the chain.
- Tmp2 = LegalizeOp(ST->getBasePtr()); // Legalize the pointer.
- unsigned Alignment = ST->getAlignment();
- bool isVolatile = ST->isVolatile();
- bool isNonTemporal = ST->isNonTemporal();
-
- if (!ST->isTruncatingStore()) {
- if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
- Result = SDValue(OptStore, 0);
- break;
- }
-
- {
- Tmp3 = LegalizeOp(ST->getValue());
- Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(),
- Tmp1, Tmp3, Tmp2,
- ST->getOffset()),
- Result.getResNo());
-
- EVT VT = Tmp3.getValueType();
- switch (TLI.getOperationAction(ISD::STORE, VT)) {
- default: assert(0 && "This action is not supported yet!");
- case TargetLowering::Legal:
- // 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());
- unsigned ABIAlignment= TLI.getTargetData()->getABITypeAlignment(Ty);
- if (ST->getAlignment() < ABIAlignment)
- Result = ExpandUnalignedStore(cast<StoreSDNode>(Result.getNode()),
- DAG, TLI);
- }
- break;
- case TargetLowering::Custom:
- Tmp1 = TLI.LowerOperation(Result, DAG);
- if (Tmp1.getNode()) Result = Tmp1;
- break;
- case TargetLowering::Promote:
- assert(VT.isVector() && "Unknown legal promote case!");
- Tmp3 = DAG.getNode(ISD::BITCAST, dl,
- TLI.getTypeToPromoteTo(ISD::STORE, VT), Tmp3);
- Result = DAG.getStore(Tmp1, dl, Tmp3, Tmp2,
- ST->getPointerInfo(), isVolatile,
- isNonTemporal, Alignment);
- break;
- }
- break;
- }
- } else {
- Tmp3 = LegalizeOp(ST->getValue());
-
- EVT StVT = ST->getMemoryVT();
- unsigned StWidth = StVT.getSizeInBits();
-
- if (StWidth != StVT.getStoreSizeInBits()) {
- // Promote to a byte-sized store with upper bits zero if not
- // storing an integral number of bytes. For example, promote
- // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
- EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
- StVT.getStoreSizeInBits());
- Tmp3 = DAG.getZeroExtendInReg(Tmp3, dl, StVT);
- Result = DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2, ST->getPointerInfo(),
- NVT, isVolatile, isNonTemporal, Alignment);
- } else if (StWidth & (StWidth - 1)) {
- // If not storing a power-of-2 number of bits, expand as two stores.
- assert(!StVT.isVector() && "Unsupported truncstore!");
- unsigned RoundWidth = 1 << Log2_32(StWidth);
- assert(RoundWidth < StWidth);
- unsigned ExtraWidth = StWidth - RoundWidth;
- assert(ExtraWidth < RoundWidth);
- assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
- "Store size not an integral number of bytes!");
- EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
- EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
- SDValue Lo, Hi;
- unsigned IncrementSize;
-
- if (TLI.isLittleEndian()) {
- // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16)
- // Store the bottom RoundWidth bits.
- Lo = DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2, ST->getPointerInfo(),
- RoundVT,
- isVolatile, isNonTemporal, Alignment);
-
- // Store the remaining ExtraWidth bits.
- IncrementSize = RoundWidth / 8;
- 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()));
- Hi = DAG.getTruncStore(Tmp1, dl, Hi, Tmp2,
- ST->getPointerInfo().getWithOffset(IncrementSize),
- ExtraVT, isVolatile, isNonTemporal,
- MinAlign(Alignment, IncrementSize));
- } else {
- // Big endian - avoid unaligned stores.
- // 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()));
- Hi = DAG.getTruncStore(Tmp1, dl, Hi, Tmp2, ST->getPointerInfo(),
- RoundVT, isVolatile, isNonTemporal, Alignment);
+ }
+ break;
+ }
- // Store the remaining ExtraWidth bits.
- IncrementSize = RoundWidth / 8;
- Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
- DAG.getIntPtrConstant(IncrementSize));
- Lo = DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2,
- ST->getPointerInfo().getWithOffset(IncrementSize),
- ExtraVT, isVolatile, isNonTemporal,
- MinAlign(Alignment, IncrementSize));
- }
+ if (SimpleFinishLegalizing) {
+ SDNode *NewNode = Node;
+ switch (Node->getOpcode()) {
+ default: break;
+ case ISD::SHL:
+ case ISD::SRL:
+ case ISD::SRA:
+ case ISD::ROTL:
+ case ISD::ROTR:
+ // Legalizing shifts/rotates requires adjusting the shift amount
+ // to the appropriate width.
+ if (!Node->getOperand(1).getValueType().isVector()) {
+ SDValue SAO =
+ DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
+ Node->getOperand(1));
+ HandleSDNode Handle(SAO);
+ LegalizeOp(SAO.getNode());
+ NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
+ Handle.getValue());
+ }
+ break;
+ case ISD::SRL_PARTS:
+ case ISD::SRA_PARTS:
+ case ISD::SHL_PARTS:
+ // Legalizing shifts/rotates requires adjusting the shift amount
+ // to the appropriate width.
+ if (!Node->getOperand(2).getValueType().isVector()) {
+ SDValue SAO =
+ DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
+ Node->getOperand(2));
+ HandleSDNode Handle(SAO);
+ LegalizeOp(SAO.getNode());
+ NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
+ Node->getOperand(1),
+ Handle.getValue());
+ }
+ break;
+ }
- // The order of the stores doesn't matter.
- Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
- } else {
- if (Tmp1 != ST->getChain() || Tmp3 != ST->getValue() ||
- Tmp2 != ST->getBasePtr())
- Result = SDValue(DAG.UpdateNodeOperands(Result.getNode(),
- Tmp1, Tmp3, Tmp2,
- ST->getOffset()),
- Result.getResNo());
-
- switch (TLI.getTruncStoreAction(ST->getValue().getValueType(), StVT)) {
- default: assert(0 && "This action is not supported yet!");
- case TargetLowering::Legal:
- // 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());
- unsigned ABIAlignment= TLI.getTargetData()->getABITypeAlignment(Ty);
- if (ST->getAlignment() < ABIAlignment)
- Result = ExpandUnalignedStore(cast<StoreSDNode>(Result.getNode()),
- DAG, TLI);
- }
- break;
- case TargetLowering::Custom:
- Result = TLI.LowerOperation(Result, DAG);
- break;
- case Expand:
- // TRUNCSTORE:i16 i32 -> STORE i16
- assert(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);
- break;
+ if (NewNode != Node) {
+ DAG.ReplaceAllUsesWith(Node, NewNode);
+ for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
+ DAG.TransferDbgValues(SDValue(Node, i), SDValue(NewNode, i));
+ ReplacedNode(Node);
+ Node = NewNode;
+ }
+ switch (Action) {
+ case TargetLowering::Legal:
+ return;
+ case TargetLowering::Custom: {
+ // FIXME: The handling for custom lowering with multiple results is
+ // a complete mess.
+ SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
+ if (Res.getNode()) {
+ SmallVector<SDValue, 8> ResultVals;
+ for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i) {
+ if (e == 1)
+ ResultVals.push_back(Res);
+ else
+ ResultVals.push_back(Res.getValue(i));
+ }
+ if (Res.getNode() != Node || Res.getResNo() != 0) {
+ DAG.ReplaceAllUsesWith(Node, ResultVals.data());
+ for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
+ DAG.TransferDbgValues(SDValue(Node, i), ResultVals[i]);
+ ReplacedNode(Node);
}
+ return;
}
}
- break;
- }
+ // FALL THROUGH
+ case TargetLowering::Expand:
+ ExpandNode(Node);
+ return;
+ case TargetLowering::Promote:
+ PromoteNode(Node);
+ return;
+ }
}
- assert(Result.getValueType() == Op.getValueType() &&
- "Bad legalization!");
- // Make sure that the generated code is itself legal.
- if (Result != Op)
- Result = LegalizeOp(Result);
+ switch (Node->getOpcode()) {
+ default:
+#ifndef NDEBUG
+ dbgs() << "NODE: ";
+ Node->dump( &DAG);
+ dbgs() << "\n";
+#endif
+ llvm_unreachable("Do not know how to legalize this operator!");
- // Note that LegalizeOp may be reentered even from single-use nodes, which
- // means that we always must cache transformed nodes.
- AddLegalizedOperand(Op, Result);
- return Result;
+ case ISD::CALLSEQ_START:
+ case ISD::CALLSEQ_END:
+ break;
+ case ISD::LOAD: {
+ return LegalizeLoadOps(Node);
+ }
+ case ISD::STORE: {
+ return LegalizeStoreOps(Node);
+ }
+ }
}
SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
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,
+ false, false, false, 0);
+ 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, 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
StoreChain = DAG.getEntryNode();
// Result is a load from the stack slot.
- return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo, false, false, 0);
+ return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo,
+ false, false, false, 0);
}
SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) {
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(FloatVT.isByteSized() && "Unsupported floating point type!");
// Load out a legal integer with the same sign bit as the float.
SignBit = DAG.getLoad(LoadTy, dl, Ch, StackPtr, MachinePointerInfo(),
- false, false, 0);
+ false, false, false, 0);
} else { // Little endian
SDValue LoadPtr = StackPtr;
// The float may be wider than the integer we are going to load. Advance
LoadPtr, DAG.getIntPtrConstant(ByteOffset));
// Load a legal integer containing the sign bit.
SignBit = DAG.getLoad(LoadTy, dl, Ch, LoadPtr, MachinePointerInfo(),
- false, false, 0);
+ false, false, false, 0);
// Move the sign bit to the top bit of the loaded integer.
unsigned BitShift = LoadTy.getSizeInBits() -
(FloatVT.getSizeInBits() - 8 * ByteOffset);
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.
SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
Chain = SP.getValue(1);
unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
- unsigned StackAlign = TM.getFrameInfo()->getStackAlignment();
+ unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
if (Align > StackAlign)
SP = DAG.getNode(ISD::AND, dl, VT, SP,
DAG.getConstant(-(uint64_t)Align, VT));
SDValue &LHS, SDValue &RHS,
SDValue &CC,
DebugLoc dl) {
- EVT OpVT = LHS.getValueType();
+ MVT OpVT = LHS.getSimpleValueType();
ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
switch (TLI.getCondCodeAction(CCCode, OpVT)) {
- default: assert(0 && "Unknown condition code action!");
+ default: llvm_unreachable("Unknown condition code action!");
case TargetLowering::Legal:
// Nothing to do.
break;
case TargetLowering::Expand: {
ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
+ ISD::CondCode InvCC = ISD::SETCC_INVALID;
unsigned Opc = 0;
switch (CCCode) {
- default: assert(0 && "Don't know how to expand this condition!");
- case ISD::SETOEQ: CC1 = ISD::SETEQ; CC2 = ISD::SETO; Opc = ISD::AND; break;
- case ISD::SETOGT: CC1 = ISD::SETGT; CC2 = ISD::SETO; Opc = ISD::AND; break;
- case ISD::SETOGE: CC1 = ISD::SETGE; CC2 = ISD::SETO; Opc = ISD::AND; break;
- case ISD::SETOLT: CC1 = ISD::SETLT; CC2 = ISD::SETO; Opc = ISD::AND; break;
- case ISD::SETOLE: CC1 = ISD::SETLE; CC2 = ISD::SETO; Opc = ISD::AND; break;
- case ISD::SETONE: CC1 = ISD::SETNE; CC2 = ISD::SETO; Opc = ISD::AND; break;
- case ISD::SETUEQ: CC1 = ISD::SETEQ; CC2 = ISD::SETUO; Opc = ISD::OR; break;
- case ISD::SETUGT: CC1 = ISD::SETGT; CC2 = ISD::SETUO; Opc = ISD::OR; break;
- case ISD::SETUGE: CC1 = ISD::SETGE; CC2 = ISD::SETUO; Opc = ISD::OR; break;
- case ISD::SETULT: CC1 = ISD::SETLT; CC2 = ISD::SETUO; Opc = ISD::OR; break;
- case ISD::SETULE: CC1 = ISD::SETLE; CC2 = ISD::SETUO; Opc = ISD::OR; break;
- case ISD::SETUNE: CC1 = ISD::SETNE; CC2 = ISD::SETUO; Opc = ISD::OR; break;
- // FIXME: Implement more expansions.
- }
-
- SDValue SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
- SDValue SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
+ default: llvm_unreachable("Don't know how to expand this condition!");
+ case ISD::SETO:
+ assert(TLI.getCondCodeAction(ISD::SETOEQ, OpVT)
+ == TargetLowering::Legal
+ && "If SETO is expanded, SETOEQ must be legal!");
+ CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break;
+ case ISD::SETUO:
+ assert(TLI.getCondCodeAction(ISD::SETUNE, OpVT)
+ == TargetLowering::Legal
+ && "If SETUO is expanded, SETUNE must be legal!");
+ CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break;
+ case ISD::SETOEQ:
+ case ISD::SETOGT:
+ case ISD::SETOGE:
+ case ISD::SETOLT:
+ case ISD::SETOLE:
+ case ISD::SETONE:
+ case ISD::SETUEQ:
+ case ISD::SETUNE:
+ case ISD::SETUGT:
+ case ISD::SETUGE:
+ case ISD::SETULT:
+ case ISD::SETULE:
+ // If we are floating point, assign and break, otherwise fall through.
+ if (!OpVT.isInteger()) {
+ // We can use the 4th bit to tell if we are the unordered
+ // or ordered version of the opcode.
+ CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
+ Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND;
+ CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10);
+ break;
+ }
+ // Fallthrough if we are unsigned integer.
+ case ISD::SETLE:
+ case ISD::SETGT:
+ case ISD::SETGE:
+ case ISD::SETLT:
+ case ISD::SETNE:
+ case ISD::SETEQ:
+ InvCC = ISD::getSetCCSwappedOperands(CCCode);
+ if (TLI.getCondCodeAction(InvCC, OpVT) == TargetLowering::Expand) {
+ // We only support using the inverted operation and not a
+ // different manner of supporting expanding these cases.
+ llvm_unreachable("Don't know how to expand this condition!");
+ }
+ LHS = DAG.getSetCC(dl, VT, RHS, LHS, InvCC);
+ RHS = SDValue();
+ CC = SDValue();
+ return;
+ }
+
+ SDValue SetCC1, SetCC2;
+ if (CCCode != ISD::SETO && CCCode != ISD::SETUO) {
+ // If we aren't the ordered or unorder operation,
+ // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS).
+ SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
+ SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
+ } else {
+ // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS)
+ SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1);
+ SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2);
+ }
LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
RHS = SDValue();
CC = SDValue();
DebugLoc dl) {
// Create the stack frame object.
unsigned SrcAlign =
- TLI.getTargetData()->getPrefTypeAlignment(SrcOp.getValueType().
+ TLI.getDataLayout()->getPrefTypeAlignment(SrcOp.getValueType().
getTypeForEVT(*DAG.getContext()));
SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
unsigned SrcSize = SrcOp.getValueType().getSizeInBits();
unsigned SlotSize = SlotVT.getSizeInBits();
unsigned DestSize = DestVT.getSizeInBits();
- const Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
- unsigned DestAlign = TLI.getTargetData()->getPrefTypeAlignment(DestType);
+ Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
+ unsigned DestAlign = TLI.getDataLayout()->getPrefTypeAlignment(DestType);
// Emit a store to the stack slot. Use a truncstore if the input value is
// later than DestVT.
// Result is a load from the stack slot.
if (SlotSize == DestSize)
return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo,
- false, false, DestAlign);
+ false, 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);
}
false, false, 0);
return DAG.getLoad(Node->getValueType(0), dl, Ch, StackPtr,
MachinePointerInfo::getFixedStack(SPFI),
- false, false, 0);
+ false, false, false, 0);
}
// If all elements are constants, create a load from the constant pool.
if (isConstant) {
- std::vector<Constant*> CV;
+ SmallVector<Constant*, 16> CV;
for (unsigned i = 0, e = NumElems; i != e; ++i) {
if (ConstantFPSDNode *V =
dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
}
} 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));
}
}
unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
return DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
MachinePointerInfo::getConstantPool(),
- false, false, Alignment);
+ false, false, false, Alignment);
}
if (!MoreThanTwoValues) {
return ExpandVectorBuildThroughStack(Node);
}
+static bool isInTailCallPosition(SelectionDAG &DAG, SDNode *Node,
+ SDValue &Chain, const TargetLowering &TLI) {
+ const Function *F = DAG.getMachineFunction().getFunction();
+
+ // Conservatively require the attributes of the call to match those of
+ // the return. Ignore noalias because it doesn't affect the call sequence.
+ Attribute CallerRetAttr = F->getAttributes().getRetAttributes();
+ if (AttrBuilder(CallerRetAttr)
+ .removeAttribute(Attribute::NoAlias).hasAttributes())
+ return false;
+
+ // It's not safe to eliminate the sign / zero extension of the return value.
+ if (CallerRetAttr.hasAttribute(Attribute::ZExt) ||
+ CallerRetAttr.hasAttribute(Attribute::SExt))
+ return false;
+
+ // Check if the only use is a function return node.
+ return TLI.isUsedByReturnOnly(Node, Chain);
+}
+
// ExpandLibCall - Expand a node into a call to a libcall. If the result value
// does not fit into a register, return the lo part and set the hi part to the
// by-reg argument. If it does fit into a single register, return the result
// 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.
- SDValue InChain = DAG.getEntryNode();
-
TargetLowering::ArgListTy Args;
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;
SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
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());
+
+ // By default, the input chain to this libcall is the entry node of the
+ // function. If the libcall is going to be emitted as a tail call then
+ // TLI.isUsedByReturnOnly will change it to the right chain if the return
+ // node which is being folded has a non-entry input chain.
+ SDValue InChain = DAG.getEntryNode();
// isTailCall may be true since the callee does not reference caller stack
// frame. Check if it's in the right position.
- bool isTailCall = isInTailCallPosition(DAG, Node, TLI);
- std::pair<SDValue, SDValue> CallInfo =
- TLI.LowerCallTo(InChain, RetTy, isSigned, !isSigned, false, false,
+ SDValue TCChain = InChain;
+ bool isTailCall = isInTailCallPosition(DAG, Node, TCChain, TLI);
+ if (isTailCall)
+ InChain = TCChain;
+
+ TargetLowering::
+ CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, false,
0, TLI.getLibcallCallingConv(LC), isTailCall,
- /*isReturnValueUsed=*/true,
+ /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
Callee, Args, DAG, Node->getDebugLoc());
+ std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
+
if (!CallInfo.second.getNode())
// It's a tailcall, return the chain (which is the DAG root).
return DAG.getRoot();
- // 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;
+}
+
+/// 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());
+ TargetLowering::
+ CallLoweringInfo CLI(DAG.getEntryNode(), RetTy, isSigned, !isSigned, false,
+ false, 0, TLI.getLibcallCallingConv(LC),
+ /*isTailCall=*/false,
+ /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
+ Callee, Args, DAG, dl);
+ std::pair<SDValue,SDValue> CallInfo = TLI.LowerCallTo(CLI);
+
return CallInfo.first;
}
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;
SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
TLI.getPointerTy());
- // Splice the libcall in wherever FindInputOutputChains tells us to.
- const Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
- std::pair<SDValue, SDValue> CallInfo =
- TLI.LowerCallTo(InChain, RetTy, isSigned, !isSigned, false, false,
+ Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
+ TargetLowering::
+ CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, false,
0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
- /*isReturnValueUsed=*/true,
+ /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
Callee, Args, DAG, Node->getDebugLoc());
+ std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
- // 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;
}
RTLIB::Libcall Call_F32,
RTLIB::Libcall Call_F64,
RTLIB::Libcall Call_F80,
+ RTLIB::Libcall Call_F128,
RTLIB::Libcall Call_PPCF128) {
RTLIB::Libcall LC;
switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
- default: assert(0 && "Unexpected request for libcall!");
+ default: llvm_unreachable("Unexpected request for libcall!");
case MVT::f32: LC = Call_F32; break;
case MVT::f64: LC = Call_F64; break;
case MVT::f80: LC = Call_F80; break;
+ case MVT::f128: LC = Call_F128; break;
case MVT::ppcf128: LC = Call_PPCF128; break;
}
return ExpandLibCall(LC, Node, false);
RTLIB::Libcall Call_I128) {
RTLIB::Libcall LC;
switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
- default: assert(0 && "Unexpected request for libcall!");
+ default: llvm_unreachable("Unexpected request for libcall!");
case MVT::i8: LC = Call_I8; break;
case MVT::i16: LC = Call_I16; break;
case MVT::i32: LC = Call_I32; break;
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: llvm_unreachable("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) {
+ // The other use might have been replaced with a divrem already.
+ unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
+ 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->getOpcode() == DivRemOpc) &&
+ 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: llvm_unreachable("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());
+
+ DebugLoc dl = Node->getDebugLoc();
+ TargetLowering::
+ CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, false,
+ 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
+ /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
+ Callee, Args, DAG, dl);
+ std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
+
+ // Remainder is loaded back from the stack frame.
+ SDValue Rem = DAG.getLoad(RetVT, dl, CallInfo.second, FIPtr,
+ MachinePointerInfo(), false, 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
SDValue Op0,
EVT DestVT,
DebugLoc dl) {
- if (Op0.getValueType() == MVT::i32) {
+ if (Op0.getValueType() == MVT::i32 && TLI.isTypeLegal(MVT::f64)) {
// simple 32-bit [signed|unsigned] integer to float/double expansion
// Get the stack frame index of a 8 byte buffer.
false, false, 0);
// load the constructed double
SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot,
- MachinePointerInfo(), false, false, 0);
+ MachinePointerInfo(), false, false, false, 0);
// FP constant to bias correct the final result
SDValue Bias = DAG.getConstantFP(isSigned ?
BitsToDouble(0x4330000080000000ULL) :
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));
// offset depending on the data type.
uint64_t FF;
switch (Op0.getValueType().getSimpleVT().SimpleTy) {
- default: assert(0 && "Unsupported integer type!");
+ default: llvm_unreachable("Unsupported integer type!");
case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float)
case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float)
case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float)
if (DestVT == MVT::f32)
FudgeInReg = DAG.getLoad(MVT::f32, dl, DAG.getEntryNode(), CPIdx,
MachinePointerInfo::getConstantPool(),
- false, false, Alignment);
+ false, false, false, Alignment);
else {
- FudgeInReg =
- LegalizeOp(DAG.getExtLoad(ISD::EXTLOAD, DestVT, dl,
- DAG.getEntryNode(), CPIdx,
- MachinePointerInfo::getConstantPool(),
- MVT::f32, false, false, Alignment));
+ SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
+ DAG.getEntryNode(), CPIdx,
+ MachinePointerInfo::getConstantPool(),
+ MVT::f32, false, false, Alignment);
+ HandleSDNode Handle(Load);
+ LegalizeOp(Load.getNode());
+ FudgeInReg = Handle.getValue();
}
return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg);
///
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: llvm_unreachable("Unhandled Expand type in BSWAP!");
case MVT::i16:
Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
}
}
+/// SplatByte - Distribute ByteVal over NumBits bits.
+// FIXME: Move this helper to a common place.
+static APInt SplatByte(unsigned NumBits, uint8_t ByteVal) {
+ APInt Val = APInt(NumBits, ByteVal);
+ unsigned Shift = 8;
+ for (unsigned i = NumBits; i > 8; i >>= 1) {
+ Val = (Val << Shift) | Val;
+ Shift <<= 1;
+ }
+ return Val;
+}
+
/// ExpandBitCount - Expand the specified bitcount instruction into operations.
///
SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
DebugLoc dl) {
switch (Opc) {
- default: assert(0 && "Cannot expand this yet!");
+ default: llvm_unreachable("Cannot expand this yet!");
case ISD::CTPOP: {
- static const uint64_t mask[6] = {
- 0x5555555555555555ULL, 0x3333333333333333ULL,
- 0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL,
- 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL
- };
EVT VT = Op.getValueType();
- EVT ShVT = TLI.getShiftAmountTy();
- unsigned len = VT.getSizeInBits();
- for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
- //x = (x & mask[i][len/8]) + (x >> (1 << i) & mask[i][len/8])
- unsigned EltSize = VT.isVector() ?
- VT.getVectorElementType().getSizeInBits() : len;
- SDValue Tmp2 = DAG.getConstant(APInt(EltSize, mask[i]), VT);
- SDValue Tmp3 = DAG.getConstant(1ULL << i, ShVT);
- Op = DAG.getNode(ISD::ADD, dl, VT,
- DAG.getNode(ISD::AND, dl, VT, Op, Tmp2),
- DAG.getNode(ISD::AND, dl, VT,
- DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3),
- Tmp2));
- }
+ EVT ShVT = TLI.getShiftAmountTy(VT);
+ unsigned Len = VT.getSizeInBits();
+
+ assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
+ "CTPOP not implemented for this type.");
+
+ // This is the "best" algorithm from
+ // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
+
+ SDValue Mask55 = DAG.getConstant(SplatByte(Len, 0x55), VT);
+ SDValue Mask33 = DAG.getConstant(SplatByte(Len, 0x33), VT);
+ SDValue Mask0F = DAG.getConstant(SplatByte(Len, 0x0F), VT);
+ SDValue Mask01 = DAG.getConstant(SplatByte(Len, 0x01), VT);
+
+ // v = v - ((v >> 1) & 0x55555555...)
+ Op = DAG.getNode(ISD::SUB, dl, VT, Op,
+ DAG.getNode(ISD::AND, dl, VT,
+ DAG.getNode(ISD::SRL, dl, VT, Op,
+ DAG.getConstant(1, ShVT)),
+ Mask55));
+ // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
+ Op = DAG.getNode(ISD::ADD, dl, VT,
+ DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
+ DAG.getNode(ISD::AND, dl, VT,
+ DAG.getNode(ISD::SRL, dl, VT, Op,
+ DAG.getConstant(2, ShVT)),
+ Mask33));
+ // v = (v + (v >> 4)) & 0x0F0F0F0F...
+ Op = DAG.getNode(ISD::AND, dl, VT,
+ DAG.getNode(ISD::ADD, dl, VT, Op,
+ DAG.getNode(ISD::SRL, dl, VT, Op,
+ DAG.getConstant(4, ShVT))),
+ Mask0F);
+ // v = (v * 0x01010101...) >> (Len - 8)
+ 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_ZERO_UNDEF:
+ // This trivially expands to CTLZ.
+ return DAG.getNode(ISD::CTLZ, dl, Op.getValueType(), Op);
case ISD::CTLZ: {
// for now, we do this:
// x = x | (x >> 1);
//
// 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);
Op = DAG.getNOT(dl, Op, VT);
return DAG.getNode(ISD::CTPOP, dl, VT, Op);
}
+ case ISD::CTTZ_ZERO_UNDEF:
+ // This trivially expands to CTTZ.
+ return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op);
case ISD::CTTZ: {
// for now, we use: { return popcount(~x & (x - 1)); }
// unless the target has ctlz but not ctpop, in which case we use:
switch (Opc) {
default:
llvm_unreachable("Unhandled atomic intrinsic Expand!");
- break;
case ISD::ATOMIC_SWAP:
switch (VT.SimpleTy) {
default: llvm_unreachable("Unexpected value type for atomic!");
return ExpandChainLibCall(LC, Node, false);
}
-void SelectionDAGLegalize::ExpandNode(SDNode *Node,
- SmallVectorImpl<SDValue> &Results) {
+void SelectionDAGLegalize::ExpandNode(SDNode *Node) {
+ SmallVector<SDValue, 8> Results;
DebugLoc dl = Node->getDebugLoc();
SDValue Tmp1, Tmp2, Tmp3, Tmp4;
switch (Node->getOpcode()) {
case ISD::CTPOP:
case ISD::CTLZ:
+ case ISD::CTLZ_ZERO_UNDEF:
case ISD::CTTZ:
+ case ISD::CTTZ_ZERO_UNDEF:
Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
Results.push_back(Tmp1);
break;
case ISD::PREFETCH:
case ISD::VAEND:
case ISD::EH_SJLJ_LONGJMP:
- case ISD::EH_SJLJ_DISPATCHSETUP:
// If the target didn't expand these, there's nothing to do, so just
// preserve the chain and be done.
Results.push_back(Node->getOperand(0));
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()),
+ TargetLowering::
+ CallLoweringInfo CLI(Node->getOperand(0),
+ Type::getVoidTy(*DAG.getContext()),
false, false, false, false, 0, CallingConv::C,
/*isTailCall=*/false,
- /*isReturnValueUsed=*/true,
+ /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
DAG.getExternalSymbol("__sync_synchronize",
TLI.getPointerTy()),
Args, DAG, dl);
+ std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
+
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, Node->getValueType(0));
+ 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.
- // FIXME: Unimplemented for now. Add libcalls.
case ISD::ATOMIC_SWAP:
case ISD::ATOMIC_LOAD_ADD:
case ISD::ATOMIC_LOAD_SUB:
case ISD::TRAP: {
// If this operation is not supported, lower it to 'abort()' call
TargetLowering::ArgListTy Args;
- std::pair<SDValue, SDValue> CallResult =
- TLI.LowerCallTo(Node->getOperand(0), Type::getVoidTy(*DAG.getContext()),
+ TargetLowering::
+ CallLoweringInfo CLI(Node->getOperand(0),
+ Type::getVoidTy(*DAG.getContext()),
false, false, false, false, 0, CallingConv::C,
/*isTailCall=*/false,
- /*isReturnValueUsed=*/true,
+ /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
DAG.getExternalSymbol("abort", TLI.getPointerTy()),
Args, DAG, dl);
+ std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
+
Results.push_back(CallResult.second);
break;
}
// 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
unsigned Align = Node->getConstantOperandVal(3);
SDValue VAListLoad = DAG.getLoad(TLI.getPointerTy(), dl, Tmp1, Tmp2,
- MachinePointerInfo(V), false, false, 0);
+ MachinePointerInfo(V),
+ false, false, false, 0);
SDValue VAList = VAListLoad;
if (Align > TLI.getMinStackArgumentAlignment()) {
// Increment the pointer, VAList, to the next vaarg
Tmp3 = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), VAList,
- DAG.getConstant(TLI.getTargetData()->
+ DAG.getConstant(TLI.getDataLayout()->
getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())),
TLI.getPointerTy()));
// Store the incremented VAList to the legalized pointer
MachinePointerInfo(V), false, false, 0);
// Load the actual argument out of the pointer VAList
Results.push_back(DAG.getLoad(VT, dl, Tmp3, VAList, MachinePointerInfo(),
- false, false, 0));
+ false, false, false, 0));
Results.push_back(Results[0].getValue(1));
break;
}
const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue();
Tmp1 = DAG.getLoad(TLI.getPointerTy(), dl, Node->getOperand(0),
Node->getOperand(2), MachinePointerInfo(VS),
- false, false, 0);
+ false, false, false, 0);
Tmp1 = DAG.getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1),
MachinePointerInfo(VD), false, false, 0);
Results.push_back(Tmp1);
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;
Node->getOperand(2), dl));
break;
case ISD::VECTOR_SHUFFLE: {
- SmallVector<int, 8> Mask;
- cast<ShuffleVectorSDNode>(Node)->getMask(Mask);
+ SmallVector<int, 32> NewMask;
+ ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
EVT VT = Node->getValueType(0);
EVT EltVT = VT.getVectorElementType();
- if (getTypeAction(EltVT) == Promote)
- EltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
+ SDValue Op0 = Node->getOperand(0);
+ SDValue Op1 = Node->getOperand(1);
+ if (!TLI.isTypeLegal(EltVT)) {
+
+ EVT NewEltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
+
+ // BUILD_VECTOR operands are allowed to be wider than the element type.
+ // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept it
+ if (NewEltVT.bitsLT(EltVT)) {
+
+ // Convert shuffle node.
+ // If original node was v4i64 and the new EltVT is i32,
+ // cast operands to v8i32 and re-build the mask.
+
+ // Calculate new VT, the size of the new VT should be equal to original.
+ EVT NewVT = EVT::getVectorVT(*DAG.getContext(), NewEltVT,
+ VT.getSizeInBits()/NewEltVT.getSizeInBits());
+ assert(NewVT.bitsEq(VT));
+
+ // cast operands to new VT
+ Op0 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op0);
+ Op1 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op1);
+
+ // Convert the shuffle mask
+ unsigned int factor = NewVT.getVectorNumElements()/VT.getVectorNumElements();
+
+ // EltVT gets smaller
+ assert(factor > 0);
+
+ for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
+ if (Mask[i] < 0) {
+ for (unsigned fi = 0; fi < factor; ++fi)
+ NewMask.push_back(Mask[i]);
+ }
+ else {
+ for (unsigned fi = 0; fi < factor; ++fi)
+ NewMask.push_back(Mask[i]*factor+fi);
+ }
+ }
+ Mask = NewMask;
+ VT = NewVT;
+ }
+ EltVT = NewEltVT;
+ }
unsigned NumElems = VT.getVectorNumElements();
- SmallVector<SDValue, 8> Ops;
+ SmallVector<SDValue, 16> Ops;
for (unsigned i = 0; i != NumElems; ++i) {
if (Mask[i] < 0) {
Ops.push_back(DAG.getUNDEF(EltVT));
unsigned Idx = Mask[i];
if (Idx < NumElems)
Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
- Node->getOperand(0),
+ Op0,
DAG.getIntPtrConstant(Idx)));
else
Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
- Node->getOperand(1),
+ Op1,
DAG.getIntPtrConstant(Idx - NumElems)));
}
+
Tmp1 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Ops[0], Ops.size());
+ // We may have changed the BUILD_VECTOR type. Cast it back to the Node type.
+ Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), Tmp1);
Results.push_back(Tmp1);
break;
}
// 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
}
case ISD::FSQRT:
Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
- RTLIB::SQRT_F80, RTLIB::SQRT_PPCF128));
+ RTLIB::SQRT_F80, RTLIB::SQRT_F128,
+ RTLIB::SQRT_PPCF128));
break;
case ISD::FSIN:
Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
- RTLIB::SIN_F80, RTLIB::SIN_PPCF128));
+ RTLIB::SIN_F80, RTLIB::SIN_F128,
+ RTLIB::SIN_PPCF128));
break;
case ISD::FCOS:
Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
- RTLIB::COS_F80, RTLIB::COS_PPCF128));
+ RTLIB::COS_F80, RTLIB::COS_F128,
+ RTLIB::COS_PPCF128));
break;
case ISD::FLOG:
Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
- RTLIB::LOG_F80, RTLIB::LOG_PPCF128));
+ RTLIB::LOG_F80, RTLIB::LOG_F128,
+ RTLIB::LOG_PPCF128));
break;
case ISD::FLOG2:
Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
- RTLIB::LOG2_F80, RTLIB::LOG2_PPCF128));
+ RTLIB::LOG2_F80, RTLIB::LOG2_F128,
+ RTLIB::LOG2_PPCF128));
break;
case ISD::FLOG10:
Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
- RTLIB::LOG10_F80, RTLIB::LOG10_PPCF128));
+ RTLIB::LOG10_F80, RTLIB::LOG10_F128,
+ RTLIB::LOG10_PPCF128));
break;
case ISD::FEXP:
Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
- RTLIB::EXP_F80, RTLIB::EXP_PPCF128));
+ RTLIB::EXP_F80, RTLIB::EXP_F128,
+ RTLIB::EXP_PPCF128));
break;
case ISD::FEXP2:
Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
- RTLIB::EXP2_F80, RTLIB::EXP2_PPCF128));
+ RTLIB::EXP2_F80, RTLIB::EXP2_F128,
+ RTLIB::EXP2_PPCF128));
break;
case ISD::FTRUNC:
Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
- RTLIB::TRUNC_F80, RTLIB::TRUNC_PPCF128));
+ RTLIB::TRUNC_F80, RTLIB::TRUNC_F128,
+ RTLIB::TRUNC_PPCF128));
break;
case ISD::FFLOOR:
Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
- RTLIB::FLOOR_F80, RTLIB::FLOOR_PPCF128));
+ RTLIB::FLOOR_F80, RTLIB::FLOOR_F128,
+ RTLIB::FLOOR_PPCF128));
break;
case ISD::FCEIL:
Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
- RTLIB::CEIL_F80, RTLIB::CEIL_PPCF128));
+ RTLIB::CEIL_F80, RTLIB::CEIL_F128,
+ RTLIB::CEIL_PPCF128));
break;
case ISD::FRINT:
Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
- RTLIB::RINT_F80, RTLIB::RINT_PPCF128));
+ RTLIB::RINT_F80, RTLIB::RINT_F128,
+ RTLIB::RINT_PPCF128));
break;
case ISD::FNEARBYINT:
Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
RTLIB::NEARBYINT_F64,
RTLIB::NEARBYINT_F80,
+ RTLIB::NEARBYINT_F128,
RTLIB::NEARBYINT_PPCF128));
break;
case ISD::FPOWI:
Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
- RTLIB::POWI_F80, RTLIB::POWI_PPCF128));
+ RTLIB::POWI_F80, RTLIB::POWI_F128,
+ RTLIB::POWI_PPCF128));
break;
case ISD::FPOW:
Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
- RTLIB::POW_F80, RTLIB::POW_PPCF128));
+ RTLIB::POW_F80, RTLIB::POW_F128,
+ RTLIB::POW_PPCF128));
break;
case ISD::FDIV:
Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
- RTLIB::DIV_F80, RTLIB::DIV_PPCF128));
+ RTLIB::DIV_F80, RTLIB::DIV_F128,
+ RTLIB::DIV_PPCF128));
break;
case ISD::FREM:
Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
- RTLIB::REM_F80, RTLIB::REM_PPCF128));
+ RTLIB::REM_F80, RTLIB::REM_F128,
+ RTLIB::REM_PPCF128));
+ break;
+ case ISD::FMA:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
+ RTLIB::FMA_F80, RTLIB::FMA_F128,
+ RTLIB::FMA_PPCF128));
break;
case ISD::FP16_TO_FP32:
Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
// Check to see if this FP immediate is already legal.
// If this is a legal constant, turn it into a TargetConstantFP node.
- if (TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
- Results.push_back(SDValue(Node, 0));
- else
- Results.push_back(ExpandConstantFP(CFP, true, DAG, TLI));
+ if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
+ Results.push_back(ExpandConstantFP(CFP, true));
break;
}
case ISD::EHSELECTION: {
break;
}
case ISD::EXCEPTIONADDR: {
- unsigned Reg = TLI.getExceptionAddressRegister();
+ unsigned Reg = TLI.getExceptionPointerRegister();
assert(Reg && "Can't expand to unknown register!");
Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, Reg,
Node->getValueType(0)));
Results.push_back(Results[0].getValue(1));
break;
}
+ case ISD::FSUB: {
+ EVT VT = Node->getValueType(0);
+ assert(TLI.isOperationLegalOrCustom(ISD::FADD, VT) &&
+ TLI.isOperationLegalOrCustom(ISD::FNEG, VT) &&
+ "Don't know how to expand this FP subtraction!");
+ Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
+ Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1);
+ Results.push_back(Tmp1);
+ break;
+ }
case ISD::SUB: {
EVT VT = Node->getValueType(0);
assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) &&
"Don't know how to expand this subtraction!");
Tmp1 = DAG.getNode(ISD::XOR, dl, VT, Node->getOperand(1),
DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT));
- Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp2, DAG.getConstant(1, VT));
+ Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp1, DAG.getConstant(1, VT));
Results.push_back(DAG.getNode(ISD::ADD, dl, VT, Node->getOperand(0), Tmp1));
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) &&
+ // If div is legal, it's better to do the normal expansion
+ !TLI.isOperationLegalOrCustom(DivOpc, Node->getValueType(0)) &&
+ 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;
BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
RHS);
TopHalf = BottomHalf.getValue(1);
- } else {
- // FIXME: We should be able to fall back to a libcall with an illegal
- // type in some cases.
- // Also, we can fall back to a division in some cases, but that's a big
- // performance hit in the general case.
- assert(TLI.isTypeLegal(EVT::getIntegerVT(*DAG.getContext(),
- VT.getSizeInBits() * 2)) &&
- "Don't know how to expand this operation yet!");
- EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
+ } else if (TLI.isTypeLegal(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);
DAG.getIntPtrConstant(0));
TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
DAG.getIntPtrConstant(1));
+ } else {
+ // We can fall back to a libcall with an illegal type for the MUL if we
+ // 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.
+ RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
+ if (WideVT == MVT::i16)
+ LC = RTLIB::MUL_I16;
+ else if (WideVT == MVT::i32)
+ LC = RTLIB::MUL_I32;
+ else if (WideVT == MVT::i64)
+ LC = RTLIB::MUL_I64;
+ else if (WideVT == MVT::i128)
+ LC = RTLIB::MUL_I128;
+ assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!");
+
+ // 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));
+ // Ret is a node with an illegal type. Because such things are not
+ // generally permitted during this phase of legalization, delete the
+ // node. The above EXTRACT_ELEMENT nodes should have been folded.
+ DAG.DeleteNode(Ret.getNode());
}
+
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;
}
EVT PTy = TLI.getPointerTy();
- const TargetData &TD = *TLI.getTargetData();
+ const DataLayout &TD = *TLI.getDataLayout();
unsigned EntrySize =
DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD);
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);
LegalizeSetCCCondCode(TLI.getSetCCResultType(Tmp2.getValueType()),
Tmp2, Tmp3, Tmp4, dl);
- LastCALLSEQ_END = DAG.getEntryNode();
assert(!Tmp3.getNode() && "Can't legalize BR_CC with legal condition!");
Tmp3 = DAG.getConstant(0, Tmp2.getValueType());
Results.push_back(Tmp1);
break;
}
+ case ISD::BUILD_VECTOR:
+ Results.push_back(ExpandBUILD_VECTOR(Node));
+ break;
+ case ISD::SRA:
+ case ISD::SRL:
+ case ISD::SHL: {
+ // Scalarize vector SRA/SRL/SHL.
+ EVT VT = Node->getValueType(0);
+ assert(VT.isVector() && "Unable to legalize non-vector shift");
+ assert(TLI.isTypeLegal(VT.getScalarType())&& "Element type must be legal");
+ unsigned NumElem = VT.getVectorNumElements();
+
+ SmallVector<SDValue, 8> Scalars;
+ for (unsigned Idx = 0; Idx < NumElem; Idx++) {
+ SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
+ VT.getScalarType(),
+ Node->getOperand(0), DAG.getIntPtrConstant(Idx));
+ SDValue Sh = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
+ VT.getScalarType(),
+ Node->getOperand(1), DAG.getIntPtrConstant(Idx));
+ Scalars.push_back(DAG.getNode(Node->getOpcode(), dl,
+ VT.getScalarType(), Ex, Sh));
+ }
+ SDValue Result =
+ DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0),
+ &Scalars[0], Scalars.size());
+ ReplaceNode(SDValue(Node, 0), Result);
+ break;
+ }
case ISD::GLOBAL_OFFSET_TABLE:
case ISD::GlobalAddress:
case ISD::GlobalTLSAddress:
case ISD::INTRINSIC_WO_CHAIN:
case ISD::INTRINSIC_VOID:
// FIXME: Custom lowering for these operations shouldn't return null!
- for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
- Results.push_back(SDValue(Node, i));
break;
}
+
+ // Replace the original node with the legalized result.
+ if (!Results.empty())
+ ReplaceNode(Node, Results.data());
}
-void SelectionDAGLegalize::PromoteNode(SDNode *Node,
- SmallVectorImpl<SDValue> &Results) {
- EVT OVT = Node->getValueType(0);
+
+void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
+ SmallVector<SDValue, 8> Results;
+ MVT OVT = Node->getSimpleValueType(0);
if (Node->getOpcode() == ISD::UINT_TO_FP ||
Node->getOpcode() == ISD::SINT_TO_FP ||
Node->getOpcode() == ISD::SETCC) {
- OVT = Node->getOperand(0).getValueType();
+ OVT = Node->getOperand(0).getSimpleValueType();
}
- EVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), OVT);
+ MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), OVT);
DebugLoc dl = Node->getDebugLoc();
SDValue Tmp1, Tmp2, Tmp3;
switch (Node->getOpcode()) {
case ISD::CTTZ:
+ case ISD::CTTZ_ZERO_UNDEF:
case ISD::CTLZ:
+ case ISD::CTLZ_ZERO_UNDEF:
case ISD::CTPOP:
// Zero extend the argument.
Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
- // Perform the larger operation.
+ // Perform the larger operation. For CTPOP and CTTZ_ZERO_UNDEF, this is
+ // already the correct result.
Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
if (Node->getOpcode() == ISD::CTTZ) {
- //if Tmp1 == sizeinbits(NVT) then Tmp1 = sizeinbits(Old VT)
+ // FIXME: This should set a bit in the zero extended value instead.
Tmp2 = DAG.getSetCC(dl, TLI.getSetCCResultType(NVT),
Tmp1, DAG.getConstant(NVT.getSizeInBits(), NVT),
ISD::SETEQ);
Tmp1 = DAG.getNode(ISD::SELECT, dl, NVT, Tmp2,
DAG.getConstant(OVT.getSizeInBits(), NVT), Tmp1);
- } else if (Node->getOpcode() == ISD::CTLZ) {
+ } else if (Node->getOpcode() == ISD::CTLZ ||
+ Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF) {
// Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT))
Tmp1 = DAG.getNode(ISD::SUB, dl, NVT, Tmp1,
DAG.getConstant(NVT.getSizeInBits() -
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;
}
Node->getOpcode() == ISD::SINT_TO_FP, dl);
Results.push_back(Tmp1);
break;
+ case ISD::VAARG: {
+ SDValue Chain = Node->getOperand(0); // Get the chain.
+ SDValue Ptr = Node->getOperand(1); // Get the pointer.
+
+ unsigned TruncOp;
+ if (OVT.isVector()) {
+ TruncOp = ISD::BITCAST;
+ } else {
+ assert(OVT.isInteger()
+ && "VAARG promotion is supported only for vectors or integer types");
+ TruncOp = ISD::TRUNCATE;
+ }
+
+ // Perform the larger operation, then convert back
+ Tmp1 = DAG.getVAArg(NVT, dl, Chain, Ptr, Node->getOperand(2),
+ Node->getConstantOperandVal(3));
+ Chain = Tmp1.getValue(1);
+
+ Tmp2 = DAG.getNode(TruncOp, dl, OVT, Tmp1);
+
+ // Modified the chain result - switch anything that used the old chain to
+ // use the new one.
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Tmp2);
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
+ ReplacedNode(Node);
+ break;
+ }
case ISD::AND:
case ISD::OR:
case ISD::XOR: {
break;
}
case ISD::VECTOR_SHUFFLE: {
- SmallVector<int, 8> Mask;
- cast<ShuffleVectorSDNode>(Node)->getMask(Mask);
+ ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
// Cast the two input vectors.
Tmp1 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(0));
Tmp1, Tmp2, Node->getOperand(2)));
break;
}
+ case ISD::FDIV:
+ case ISD::FREM:
+ case ISD::FPOW: {
+ Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
+ Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
+ Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
+ Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
+ Tmp3, DAG.getIntPtrConstant(0)));
+ break;
+ }
+ case ISD::FLOG2:
+ case ISD::FEXP2:
+ case ISD::FLOG:
+ case ISD::FEXP: {
+ Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
+ Tmp2 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
+ Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
+ Tmp2, DAG.getIntPtrConstant(0)));
+ break;
+ }
}
+
+ // Replace the original node with the legalized result.
+ if (!Results.empty())
+ ReplaceNode(Node, Results.data());
}
// 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();
}
-