blob: a96a99781f4e07a05bdf158fe8f6dd9bfc2a05f3 [file] [log] [blame]
//===-- LegalizeDAG.cpp - Implement SelectionDAG::Legalize ----------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file implements the SelectionDAG::Legalize method.
//
//===----------------------------------------------------------------------===//
#include "llvm/Analysis/DebugInfo.h"
#include "llvm/CodeGen/Analysis.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/Target/TargetFrameLowering.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/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"
using namespace llvm;
//===----------------------------------------------------------------------===//
/// SelectionDAGLegalize - This takes an arbitrary SelectionDAG as input and
/// hacks on it until the target machine can handle it. This involves
/// eliminating value sizes the machine cannot handle (promoting small sizes to
/// large sizes or splitting up large values into small values) as well as
/// eliminating operations the machine cannot handle.
///
/// This code also does a small amount of optimization and recognition of idioms
/// as part of its processing. For example, if a target does not support a
/// 'setcc' instruction efficiently, but does support 'brcc' instruction, this
/// will attempt merge setcc and brc instructions into brcc's.
///
namespace {
class SelectionDAGLegalize : public SelectionDAG::DAGUpdateListener {
const TargetMachine &TM;
const TargetLowering &TLI;
SelectionDAG &DAG;
/// LegalizePosition - The iterator for walking through the node list.
SelectionDAG::allnodes_iterator LegalizePosition;
/// LegalizedNodes - The set of nodes which have already been legalized.
SmallPtrSet<SDNode *, 16> LegalizedNodes;
// Libcall insertion helpers.
public:
explicit SelectionDAGLegalize(SelectionDAG &DAG);
void LegalizeDAG();
private:
/// LegalizeOp - Legalizes the given operation.
void LegalizeOp(SDNode *Node);
SDValue OptimizeFloatStore(StoreSDNode *ST);
/// 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
/// the insert there, and then read the result back.
SDValue PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val,
SDValue Idx, DebugLoc dl);
SDValue ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val,
SDValue Idx, DebugLoc dl);
/// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
/// performs the same shuffe in terms of order or result bytes, but on a type
/// whose vector element type is narrower than the original shuffle type.
/// 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,
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);
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 ExpandSCALAR_TO_VECTOR(SDNode *Node);
void ExpandDYNAMIC_STACKALLOC(SDNode *Node,
SmallVectorImpl<SDValue> &Results);
SDValue ExpandFCOPYSIGN(SDNode *Node);
SDValue ExpandLegalINT_TO_FP(bool isSigned, SDValue LegalOp, EVT DestVT,
DebugLoc dl);
SDValue PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, bool isSigned,
DebugLoc dl);
SDValue PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, bool isSigned,
DebugLoc dl);
SDValue ExpandBSWAP(SDValue Op, DebugLoc dl);
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);
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, this);
} else {
ForgetNode(N);
}
}
void ReplaceNode(SDNode *Old, SDNode *New) {
DAG.ReplaceAllUsesWith(Old, New, this);
ReplacedNode(Old);
}
void ReplaceNode(SDValue Old, SDValue New) {
DAG.ReplaceAllUsesWith(Old, New, this);
ReplacedNode(Old.getNode());
}
void ReplaceNode(SDNode *Old, const SDValue *New) {
DAG.ReplaceAllUsesWith(Old, New, this);
ReplacedNode(Old);
}
};
}
/// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
/// performs the same shuffe in terms of order or result bytes, but on a type
/// whose vector element type is narrower than the original shuffle type.
/// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
SDValue
SelectionDAGLegalize::ShuffleWithNarrowerEltType(EVT NVT, EVT VT, DebugLoc dl,
SDValue N1, SDValue N2,
ArrayRef<int> Mask) const {
unsigned NumMaskElts = VT.getVectorNumElements();
unsigned NumDestElts = NVT.getVectorNumElements();
unsigned NumEltsGrowth = NumDestElts / NumMaskElts;
assert(NumEltsGrowth && "Cannot promote to vector type with fewer elts!");
if (NumEltsGrowth == 1)
return DAG.getVectorShuffle(NVT, dl, N1, N2, &Mask[0]);
SmallVector<int, 8> NewMask;
for (unsigned i = 0; i != NumMaskElts; ++i) {
int Idx = Mask[i];
for (unsigned j = 0; j != NumEltsGrowth; ++j) {
if (Idx < 0)
NewMask.push_back(-1);
else
NewMask.push_back(Idx * NumEltsGrowth + j);
}
}
assert(NewMask.size() == NumDestElts && "Non-integer NumEltsGrowth?");
assert(TLI.isShuffleMaskLegal(NewMask, NVT) && "Shuffle not legal?");
return DAG.getVectorShuffle(NVT, dl, N1, N2, &NewMask[0]);
}
SelectionDAGLegalize::SelectionDAGLegalize(SelectionDAG &dag)
: TM(dag.getTarget()), TLI(dag.getTargetLoweringInfo()),
DAG(dag) {
}
void SelectionDAGLegalize::LegalizeDAG() {
DAG.AssignTopologicalOrder();
// 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);
}
}
if (!AnyLegalized)
break;
}
// Remove dead nodes now.
DAG.RemoveDeadNodes();
}
/// ExpandConstantFP - Expands the ConstantFP node to an integer constant or
/// a load from the constant pool.
SDValue
SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) {
bool Extend = false;
DebugLoc dl = CFP->getDebugLoc();
// If a FP immediate is precise when represented as a float and if the
// target can do an extending load from float to double, we put it into
// the constant pool as a float, even if it's is statically typed as a
// double. This shrinks FP constants and canonicalizes them for targets where
// an FP extending load is the same cost as a normal load (such as on the x87
// fp stack or PPC FP unit).
EVT VT = CFP->getValueType(0);
ConstantFP *LLVMC = const_cast<ConstantFP*>(CFP->getConstantFPValue());
if (!UseCP) {
assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion");
return DAG.getConstant(LLVMC->getValueAPF().bitcastToAPInt(),
(VT == MVT::f64) ? MVT::i64 : MVT::i32);
}
EVT OrigVT = VT;
EVT SVT = VT;
while (SVT != MVT::f32) {
SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1);
if (ConstantFPSDNode::isValueValidForType(SVT, CFP->getValueAPF()) &&
// Only do this if the target has a native EXTLOAD instruction from
// smaller type.
TLI.isLoadExtLegal(ISD::EXTLOAD, SVT) &&
TLI.ShouldShrinkFPConstant(OrigVT)) {
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) {
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 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();
EVT VT = Val.getValueType();
int Alignment = ST->getAlignment();
DebugLoc dl = ST->getDebugLoc();
if (ST->getMemoryVT().isFloatingPoint() ||
ST->getMemoryVT().isVector()) {
EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
if (TLI.isTypeLegal(intVT)) {
// Expand to a bitconvert of the value to the integer type of the
// same size, then a (misaligned) int store.
// FIXME: Does not handle truncating floating point stores!
SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
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();
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, 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() &&
"Unaligned store of unknown type.");
// Get the half-size VT
EVT NewStoredVT = ST->getMemoryVT().getHalfSizedIntegerVT(*DAG.getContext());
int NumBits = NewStoredVT.getSizeInBits();
int IncrementSize = NumBits / 8;
// Divide the stored value in two parts.
SDValue ShiftAmount = DAG.getConstant(NumBits,
TLI.getShiftAmountTy(Val.getValueType()));
SDValue Lo = Val;
SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
// Store the two parts
SDValue Store1, Store2;
Store1 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Lo:Hi, Ptr,
ST->getPointerInfo(), NewStoredVT,
ST->isVolatile(), ST->isNonTemporal(), Alignment);
Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
DAG.getConstant(IncrementSize, TLI.getPointerTy()));
Alignment = MinAlign(Alignment, IncrementSize);
Store2 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Hi:Lo, Ptr,
ST->getPointerInfo().getWithOffset(IncrementSize),
NewStoredVT, ST->isVolatile(), ST->isNonTemporal(),
Alignment);
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 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);
EVT LoadedVT = LD->getMemoryVT();
DebugLoc dl = LD->getDebugLoc();
if (VT.isFloatingPoint() || VT.isVector()) {
EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
if (TLI.isTypeLegal(intVT)) {
// 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->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);
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);
unsigned LoadedBytes = LoadedVT.getSizeInBits() / 8;
unsigned RegBytes = RegVT.getSizeInBits() / 8;
unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes;
// Make sure the stack slot is also aligned for the register type.
SDValue StackBase = DAG.CreateStackTemporary(LoadedVT, RegVT);
SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
SmallVector<SDValue, 8> Stores;
SDValue StackPtr = StackBase;
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 original location.
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,
MachinePointerInfo(), false, false, 0));
// Increment the pointers.
Offset += RegBytes;
Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
Increment);
}
// 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, dl, RegVT, Chain, Ptr,
LD->getPointerInfo().getWithOffset(Offset),
MemVT, LD->isVolatile(),
LD->isNonTemporal(),
MinAlign(LD->getAlignment(), Offset));
// Follow the load with a store to the stack slot. Remember the store.
// On big-endian machines this requires a truncating store to ensure
// that the bits end up in the right place.
Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, StackPtr,
MachinePointerInfo(), MemVT,
false, false, 0));
// The order of the stores doesn't matter - say it with a TokenFactor.
SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
Stores.size());
// Finally, perform the original load only redirected to the stack slot.
Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
MachinePointerInfo(), LoadedVT, false, false, 0);
// Callers expect a MERGE_VALUES node.
ValResult = Load;
ChainResult = TF;
return;
}
assert(LoadedVT.isInteger() && !LoadedVT.isVector() &&
"Unaligned load of unsupported type.");
// Compute the new VT that is half the size of the old one. This is an
// integer MVT.
unsigned NumBits = LoadedVT.getSizeInBits();
EVT NewLoadedVT;
NewLoadedVT = EVT::getIntegerVT(*DAG.getContext(), NumBits/2);
NumBits >>= 1;
unsigned Alignment = LD->getAlignment();
unsigned IncrementSize = NumBits / 8;
ISD::LoadExtType HiExtType = LD->getExtensionType();
// If the original load is NON_EXTLOAD, the hi part load must be ZEXTLOAD.
if (HiExtType == ISD::NON_EXTLOAD)
HiExtType = ISD::ZEXTLOAD;
// Load the value in two parts
SDValue Lo, Hi;
if (TLI.isLittleEndian()) {
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, dl, VT, Chain, Ptr,
LD->getPointerInfo().getWithOffset(IncrementSize),
NewLoadedVT, LD->isVolatile(),
LD->isNonTemporal(), MinAlign(Alignment,IncrementSize));
} else {
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, 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(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));
ValResult = Result;
ChainResult = TF;
}
/// 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
/// the insert there, and then read the result back.
SDValue SelectionDAGLegalize::
PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx,
DebugLoc dl) {
SDValue Tmp1 = Vec;
SDValue Tmp2 = Val;
SDValue Tmp3 = Idx;
// If the target doesn't support this, we have to spill the input vector
// to a temporary stack slot, update the element, then reload it. This is
// badness. We could also load the value into a vector register (either
// with a "move to register" or "extload into register" instruction, then
// permute it into place, if the idx is a constant and if the idx is
// supported by the target.
EVT VT = Tmp1.getValueType();
EVT EltVT = VT.getVectorElementType();
EVT IdxVT = Tmp3.getValueType();
EVT PtrVT = TLI.getPointerTy();
SDValue StackPtr = DAG.CreateStackTemporary(VT);
int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
// Store the vector.
SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Tmp1, StackPtr,
MachinePointerInfo::getFixedStack(SPFI),
false, false, 0);
// Truncate or zero extend offset to target pointer type.
unsigned CastOpc = IdxVT.bitsGT(PtrVT) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
Tmp3 = DAG.getNode(CastOpc, dl, PtrVT, Tmp3);
// Add the offset to the index.
unsigned EltSize = EltVT.getSizeInBits()/8;
Tmp3 = DAG.getNode(ISD::MUL, dl, IdxVT, Tmp3,DAG.getConstant(EltSize, IdxVT));
SDValue StackPtr2 = DAG.getNode(ISD::ADD, dl, IdxVT, Tmp3, StackPtr);
// Store the scalar value.
Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, MachinePointerInfo(), EltVT,
false, false, 0);
// Load the updated vector.
return DAG.getLoad(VT, dl, Ch, StackPtr,
MachinePointerInfo::getFixedStack(SPFI), false, false,
false, 0);
}
SDValue SelectionDAGLegalize::
ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, SDValue Idx, DebugLoc dl) {
if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Idx)) {
// SCALAR_TO_VECTOR requires that the type of the value being inserted
// match the element type of the vector being created, except for
// integers in which case the inserted value can be over width.
EVT EltVT = Vec.getValueType().getVectorElementType();
if (Val.getValueType() == EltVT ||
(EltVT.isInteger() && Val.getValueType().bitsGE(EltVT))) {
SDValue ScVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
Vec.getValueType(), Val);
unsigned NumElts = Vec.getValueType().getVectorNumElements();
// We generate a shuffle of InVec and ScVec, so the shuffle mask
// should be 0,1,2,3,4,5... with the appropriate element replaced with
// elt 0 of the RHS.
SmallVector<int, 8> ShufOps;
for (unsigned i = 0; i != NumElts; ++i)
ShufOps.push_back(i != InsertPos->getZExtValue() ? i : NumElts);
return DAG.getVectorShuffle(Vec.getValueType(), dl, Vec, ScVec,
&ShufOps[0]);
}
}
return PerformInsertVectorEltInMemory(Vec, Val, Idx, dl);
}
SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) {
// Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
// FIXME: We shouldn't do this for TargetConstantFP's.
// FIXME: move this to the DAG Combiner! Note that we can't regress due
// to phase ordering between legalized code and the dag combiner. This
// 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;
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 &&
TLI.isTypeLegal(MVT::i32)) {
Tmp3 = DAG.getConstant(CFP->getValueAPF().
bitcastToAPInt().zextOrTrunc(32),
MVT::i32);
return DAG.getStore(Tmp1, dl, Tmp3, Tmp2, 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 (TLI.isTypeLegal(MVT::i64)) {
Tmp3 = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
zextOrTrunc(64), MVT::i64);
return DAG.getStore(Tmp1, dl, Tmp3, Tmp2, ST->getPointerInfo(),
isVolatile, isNonTemporal, Alignment);
}
if (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.
const APInt &IntVal =CFP->getValueAPF().bitcastToAPInt();
SDValue Lo = DAG.getConstant(IntVal.trunc(32), MVT::i32);
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,
isNonTemporal, Alignment);
Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
DAG.getIntPtrConstant(4));
Hi = DAG.getStore(Tmp1, dl, Hi, Tmp2,
ST->getPointerInfo().getWithOffset(4),
isVolatile, isNonTemporal, MinAlign(Alignment, 4U));
return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
}
}
}
return SDValue(0, 0);
}
/// 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;
DebugLoc dl = Node->getDebugLoc();
for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
assert(TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
TargetLowering::TypeLegal &&
"Unexpected illegal type!");
for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
assert((TLI.getTypeAction(*DAG.getContext(),
Node->getOperand(i).getValueType()) ==
TargetLowering::TypeLegal ||
Node->getOperand(i).getOpcode() == ISD::TargetConstant) &&
"Unexpected illegal type!");
SDValue Tmp1, Tmp2, Tmp3, Tmp4;
bool isCustom = false;
// Figure out the correct action; the way to query this varies by opcode
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::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(),
Node->getOperand(0).getValueType());
break;
case ISD::FP_ROUND_INREG:
case ISD::SIGN_EXTEND_INREG: {
EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT();
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();
ISD::CondCode CCCode =
cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
Action = TLI.getCondCodeAction(CCCode, OpVT);
if (Action == TargetLowering::Legal) {
if (Node->getOpcode() == ISD::SELECT_CC)
Action = TLI.getOperationAction(Node->getOpcode(),
Node->getValueType(0));
else
Action = TLI.getOperationAction(Node->getOpcode(), OpVT);
}
break;
}
case ISD::LOAD:
case ISD::STORE:
// FIXME: Model these properly. LOAD and STORE are complicated, and
// STORE expects the unlegalized operand in some cases.
SimpleFinishLegalizing = false;
break;
case ISD::CALLSEQ_START:
case ISD::CALLSEQ_END:
// FIXME: This shouldn't be necessary. These nodes have special properties
// dealing with the recursive nature of legalization. Removing this
// special case should be done as part of making LegalizeDAG non-recursive.
SimpleFinishLegalizing = false;
break;
case ISD::EXTRACT_ELEMENT:
case ISD::FLT_ROUNDS_:
case ISD::SADDO:
case ISD::SSUBO:
case ISD::UADDO:
case ISD::USUBO:
case ISD::SMULO:
case ISD::UMULO:
case ISD::FPOWI:
case ISD::MERGE_VALUES:
case ISD::EH_RETURN:
case ISD::FRAME_TO_ARGS_OFFSET:
case ISD::EH_SJLJ_SETJMP:
case ISD::EH_SJLJ_LONGJMP:
// 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::INIT_TRAMPOLINE:
case ISD::ADJUST_TRAMPOLINE:
case ISD::FRAMEADDR:
case ISD::RETURNADDR:
// These operations lie about being legal: when they claim to be legal,
// they should actually be custom-lowered.
Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
if (Action == TargetLowering::Legal)
Action = TargetLowering::Custom;
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;
for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
Ops.push_back(Node->getOperand(i));
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 (!Ops[1].getValueType().isVector()) {
SDValue SAO = DAG.getShiftAmountOperand(Ops[0].getValueType(), Ops[1]);
HandleSDNode Handle(SAO);
LegalizeOp(SAO.getNode());
Ops[1] = 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 (!Ops[2].getValueType().isVector()) {
SDValue SAO = DAG.getShiftAmountOperand(Ops[0].getValueType(), Ops[2]);
HandleSDNode Handle(SAO);
LegalizeOp(SAO.getNode());
Ops[2] = Handle.getValue();
}
break;
}
SDNode *NewNode = DAG.UpdateNodeOperands(Node, Ops.data(), Ops.size());
if (NewNode != Node) {
DAG.ReplaceAllUsesWith(Node, NewNode, this);
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.
Tmp1 = TLI.LowerOperation(SDValue(Node, 0), DAG);
if (Tmp1.getNode()) {
SmallVector<SDValue, 8> ResultVals;
for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i) {
if (e == 1)
ResultVals.push_back(Tmp1);
else
ResultVals.push_back(Tmp1.getValue(i));
}
if (Tmp1.getNode() != Node || Tmp1.getResNo() != 0) {
DAG.ReplaceAllUsesWith(Node, ResultVals.data(), this);
for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
DAG.TransferDbgValues(SDValue(Node, i), ResultVals[i]);
ReplacedNode(Node);
}
return;
}
// FALL THROUGH
case TargetLowering::Expand:
ExpandNode(Node);
return;
case TargetLowering::Promote:
PromoteNode(Node);
return;
}
}
switch (Node->getOpcode()) {
default:
#ifndef NDEBUG
dbgs() << "NODE: ";
Node->dump( &DAG);
dbgs() << "\n";
#endif
llvm_unreachable("Do not know how to legalize this operator!");
case ISD::CALLSEQ_START:
case ISD::CALLSEQ_END:
break;
case ISD::LOAD: {
LoadSDNode *LD = cast<LoadSDNode>(Node);
Tmp1 = LD->getChain(); // Legalize the chain.
Tmp2 = LD->getBasePtr(); // Legalize the base pointer.
ISD::LoadExtType ExtType = LD->getExtensionType();
if (ExtType == ISD::NON_EXTLOAD) {
EVT VT = Node->getValueType(0);
Tmp3 = SDValue(Node, 0);
Tmp4 = 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.getTargetData()->getABITypeAlignment(Ty);
if (LD->getAlignment() < ABIAlignment){
ExpandUnalignedLoad(cast<LoadSDNode>(Node),
DAG, TLI, Tmp3, Tmp4);
}
}
break;
case TargetLowering::Custom:
Tmp1 = TLI.LowerOperation(Tmp3, DAG);
if (Tmp1.getNode()) {
Tmp3 = Tmp1;
Tmp4 = 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->isInvariant(), LD->getAlignment());
Tmp3 = DAG.getNode(ISD::BITCAST, dl, VT, Tmp1);
Tmp4 = Tmp1.getValue(1);
break;
}
}
if (Tmp4.getNode() != Node) {
assert(Tmp3.getNode() != Node && "Load must be completely replaced");
DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Tmp3);
DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Tmp4);
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),
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 = Result;
Tmp2 = 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),
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, dl, Node->getValueType(0), 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(Hi.getValueType())));
// Join the hi and lo parts.
Tmp1 = 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), 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,
dl, Node->getValueType(0), 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(Hi.getValueType())));
// Join the hi and lo parts.
Tmp1 = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
}
Tmp2 = Ch;
} else {
switch (TLI.getLoadExtAction(ExtType, SrcVT)) {
default: llvm_unreachable("This action is not supported yet!");
case TargetLowering::Custom:
isCustom = true;
// FALLTHROUGH
case TargetLowering::Legal:
Tmp1 = SDValue(Node, 0);
Tmp2 = SDValue(Node, 1);
if (isCustom) {
Tmp3 = TLI.LowerOperation(SDValue(Node, 0), DAG);
if (Tmp3.getNode()) {
Tmp1 = Tmp3;
Tmp2 = Tmp3.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.getTargetData()->getABITypeAlignment(Ty);
if (LD->getAlignment() < ABIAlignment){
ExpandUnalignedLoad(cast<LoadSDNode>(Node),
DAG, TLI, Tmp1, Tmp2);
}
}
}
break;
case TargetLowering::Expand:
if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) && TLI.isTypeLegal(SrcVT)) {
SDValue Load = DAG.getLoad(SrcVT, dl, Tmp1, Tmp2,
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!");
}
Tmp1 = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
Tmp2 = 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),
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.getScalarType());
Tmp1 = ValRes;
Tmp2 = Result.getValue(1);
break;
}
}
// Since loads produce two values, make sure to remember that we legalized
// both of them.
if (Tmp2.getNode() != Node) {
assert(Tmp1.getNode() != Node && "Load must be completely replaced");
DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Tmp1);
DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Tmp2);
ReplacedNode(Node);
}
break;
}
case ISD::STORE: {
StoreSDNode *ST = cast<StoreSDNode>(Node);
Tmp1 = ST->getChain();
Tmp2 = ST->getBasePtr();
unsigned Alignment = ST->getAlignment();
bool isVolatile = ST->isVolatile();
bool isNonTemporal = ST->isNonTemporal();
if (!ST->isTruncatingStore()) {
if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
ReplaceNode(ST, OptStore);
break;
}
{
Tmp3 = ST->getValue();
EVT VT = Tmp3.getValueType();
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.getTargetData()->getABITypeAlignment(Ty);
if (ST->getAlignment() < ABIAlignment)
ExpandUnalignedStore(cast<StoreSDNode>(Node),
DAG, TLI, this);
}
break;
case TargetLowering::Custom:
Tmp1 = TLI.LowerOperation(SDValue(Node, 0), DAG);
if (Tmp1.getNode())
ReplaceNode(SDValue(Node, 0), Tmp1);
break;
case TargetLowering::Promote: {
assert(VT.isVector() && "Unknown legal promote case!");
Tmp3 = DAG.getNode(ISD::BITCAST, dl,
TLI.getTypeToPromoteTo(ISD::STORE, VT), Tmp3);
SDValue Result =
DAG.getStore(Tmp1, dl, Tmp3, Tmp2,
ST->getPointerInfo(), isVolatile,
isNonTemporal, Alignment);
ReplaceNode(SDValue(Node, 0), Result);
break;
}
}
break;
}
} else {
Tmp3 = 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);
SDValue Result =
DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2, 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(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(Tmp3.getValueType())));
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(Tmp3.getValueType())));
Hi = DAG.getTruncStore(Tmp1, dl, Hi, 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));
Lo = DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2,
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().getValueType(), StVT)) {
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.getTargetData()->getABITypeAlignment(Ty);
if (ST->getAlignment() < ABIAlignment)
ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
}
break;
case TargetLowering::Custom:
ReplaceNode(SDValue(Node, 0),
TLI.LowerOperation(SDValue(Node, 0), DAG));
break;
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!");
Tmp3 = DAG.getNode(ISD::TRUNCATE, dl, StVT, Tmp3);
SDValue Result =
DAG.getStore(Tmp1, dl, Tmp3, Tmp2, ST->getPointerInfo(),
isVolatile, isNonTemporal, Alignment);
ReplaceNode(SDValue(Node, 0), Result);
break;
}
}
}
break;
}
}
}
SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
SDValue Vec = Op.getOperand(0);
SDValue Idx = Op.getOperand(1);
DebugLoc dl = Op.getDebugLoc();
// Store the value to a temporary stack slot, then LOAD the returned part.
SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
MachinePointerInfo(), false, false, 0);
// 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);
StackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, StackPtr);
if (Op.getValueType().isVector())
return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr,MachinePointerInfo(),
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
// the result as a vector.
// Create the stack frame object.
EVT VT = Node->getValueType(0);
EVT EltVT = VT.getVectorElementType();
DebugLoc dl = Node->getDebugLoc();
SDValue FIPtr = DAG.CreateStackTemporary(VT);
int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex();
MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
// Emit a store of each element to the stack slot.
SmallVector<SDValue, 8> Stores;
unsigned TypeByteSize = EltVT.getSizeInBits() / 8;
// Store (in the right endianness) the elements to memory.
for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
// Ignore undef elements.
if (Node->getOperand(i).getOpcode() == ISD::UNDEF) continue;
unsigned Offset = TypeByteSize*i;
SDValue Idx = DAG.getConstant(Offset, FIPtr.getValueType());
Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx);
// If the destination vector element type is narrower than the source
// element type, only store the bits necessary.
if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) {
Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl,
Node->getOperand(i), Idx,
PtrInfo.getWithOffset(Offset),
EltVT, false, false, 0));
} else
Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl,
Node->getOperand(i), Idx,
PtrInfo.getWithOffset(Offset),
false, false, 0));
}
SDValue StoreChain;
if (!Stores.empty()) // Not all undef elements?
StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
&Stores[0], Stores.size());
else
StoreChain = DAG.getEntryNode();
// Result is a load from the stack slot.
return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo,
false, false, false, 0);
}
SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) {
DebugLoc dl = Node->getDebugLoc();
SDValue Tmp1 = Node->getOperand(0);
SDValue Tmp2 = Node->getOperand(1);
// Get the sign bit of the RHS. First obtain a value that has the same
// sign as the sign bit, i.e. negative if and only if the sign bit is 1.
SDValue SignBit;
EVT FloatVT = Tmp2.getValueType();
EVT IVT = EVT::getIntegerVT(*DAG.getContext(), FloatVT.getSizeInBits());
if (TLI.isTypeLegal(IVT)) {
// Convert to an integer with the same sign bit.
SignBit = DAG.getNode(ISD::BITCAST, dl, IVT, Tmp2);
} else {
// Store the float to memory, then load the sign part out as an integer.
MVT LoadTy = TLI.getPointerTy();
// First create a temporary that is aligned for both the load and store.
SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
// Then store the float to it.
SDValue Ch =
DAG.getStore(DAG.getEntryNode(), dl, Tmp2, StackPtr, MachinePointerInfo(),
false, false, 0);
if (TLI.isBigEndian()) {
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, false, 0);
} else { // Little endian
SDValue LoadPtr = StackPtr;
// The float may be wider than the integer we are going to load. Advance
// the pointer so that the loaded integer will contain the sign bit.
unsigned Strides = (FloatVT.getSizeInBits()-1)/LoadTy.getSizeInBits();
unsigned ByteOffset = (Strides * LoadTy.getSizeInBits()) / 8;
LoadPtr = DAG.getNode(ISD::ADD, dl, LoadPtr.getValueType(),
LoadPtr, DAG.getIntPtrConstant(ByteOffset));
// Load a legal integer containing the sign bit.
SignBit = DAG.getLoad(LoadTy, dl, Ch, LoadPtr, MachinePointerInfo(),
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(SignBit.getValueType())));
}
}
// Now get the sign bit proper, by seeing whether the value is negative.
SignBit = DAG.getSetCC(dl, TLI.getSetCCResultType(SignBit.getValueType()),
SignBit, DAG.getConstant(0, SignBit.getValueType()),
ISD::SETLT);
// Get the absolute value of the result.
SDValue AbsVal = DAG.getNode(ISD::FABS, dl, Tmp1.getValueType(), Tmp1);
// Select between the nabs and abs value based on the sign bit of
// the input.
return DAG.getNode(ISD::SELECT, dl, AbsVal.getValueType(), SignBit,
DAG.getNode(ISD::FNEG, dl, AbsVal.getValueType(), AbsVal),
AbsVal);
}
void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
SmallVectorImpl<SDValue> &Results) {
unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
" not tell us which reg is the stack pointer!");
DebugLoc dl = Node->getDebugLoc();
EVT VT = Node->getValueType(0);
SDValue Tmp1 = SDValue(Node, 0);
SDValue Tmp2 = SDValue(Node, 1);
SDValue Tmp3 = Node->getOperand(2);
SDValue Chain = Tmp1.getOperand(0);
// Chain the dynamic stack allocation so that it doesn't modify the stack
// pointer when other instructions are using the stack.
Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
SDValue Size = Tmp2.getOperand(1);
SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
Chain = SP.getValue(1);
unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
if (Align > StackAlign)
SP = DAG.getNode(ISD::AND, dl, VT, SP,
DAG.getConstant(-(uint64_t)Align, VT));
Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
DAG.getIntPtrConstant(0, true), SDValue());
Results.push_back(Tmp1);
Results.push_back(Tmp2);
}
/// LegalizeSetCCCondCode - Legalize a SETCC with given LHS and RHS and
/// condition code CC on the current target. This routine expands SETCC with
/// illegal condition code into AND / OR of multiple SETCC values.
void SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT,
SDValue &LHS, SDValue &RHS,
SDValue &CC,
DebugLoc dl) {
EVT OpVT = LHS.getValueType();
ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
switch (TLI.getCondCodeAction(CCCode, OpVT)) {
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;
unsigned Opc = 0;
switch (CCCode) {
default: llvm_unreachable("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);
LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
RHS = SDValue();
CC = SDValue();
break;
}
}
}
/// EmitStackConvert - Emit a store/load combination to the stack. This stores
/// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does
/// a load from the stack slot to DestVT, extending it if needed.
/// The resultant code need not be legal.
SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp,
EVT SlotVT,
EVT DestVT,
DebugLoc dl) {
// Create the stack frame object.
unsigned SrcAlign =
TLI.getTargetData()->getPrefTypeAlignment(SrcOp.getValueType().
getTypeForEVT(*DAG.getContext()));
SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
int SPFI = StackPtrFI->getIndex();
MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
unsigned SrcSize = SrcOp.getValueType().getSizeInBits();
unsigned SlotSize = SlotVT.getSizeInBits();
unsigned DestSize = DestVT.getSizeInBits();
Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
unsigned DestAlign = TLI.getTargetData()->getPrefTypeAlignment(DestType);
// Emit a store to the stack slot. Use a truncstore if the input value is
// later than DestVT.
SDValue Store;
if (SrcSize > SlotSize)
Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
PtrInfo, SlotVT, false, false, SrcAlign);
else {
assert(SrcSize == SlotSize && "Invalid store");
Store = DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
PtrInfo, false, false, SrcAlign);
}
// Result is a load from the stack slot.
if (SlotSize == DestSize)
return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo,
false, false, false, DestAlign);
assert(SlotSize < DestSize && "Unknown extension!");
return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr,
PtrInfo, SlotVT, false, false, DestAlign);
}
SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
DebugLoc dl = Node->getDebugLoc();
// Create a vector sized/aligned stack slot, store the value to element #0,
// then load the whole vector back out.
SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0));
FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(StackPtr);
int SPFI = StackPtrFI->getIndex();
SDValue Ch = DAG.getTruncStore(DAG.getEntryNode(), dl, Node->getOperand(0),
StackPtr,
MachinePointerInfo::getFixedStack(SPFI),
Node->getValueType(0).getVectorElementType(),
false, false, 0);
return DAG.getLoad(Node->getValueType(0), dl, Ch, StackPtr,
MachinePointerInfo::getFixedStack(SPFI),
false, false, false, 0);
}
/// ExpandBUILD_VECTOR - Expand a BUILD_VECTOR node on targets that don't
/// support the operation, but do support the resultant vector type.
SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) {
unsigned NumElems = Node->getNumOperands();
SDValue Value1, Value2;
DebugLoc dl = Node->getDebugLoc();
EVT VT = Node->getValueType(0);
EVT OpVT = Node->getOperand(0).getValueType();
EVT EltVT = VT.getVectorElementType();
// If the only non-undef value is the low element, turn this into a
// SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X.
bool isOnlyLowElement = true;
bool MoreThanTwoValues = false;
bool isConstant = true;
for (unsigned i = 0; i < NumElems; ++i) {
SDValue V = Node->getOperand(i);
if (V.getOpcode() == ISD::UNDEF)
continue;
if (i > 0)
isOnlyLowElement = false;
if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
isConstant = false;
if (!Value1.getNode()) {
Value1 = V;
} else if (!Value2.getNode()) {
if (V != Value1)
Value2 = V;
} else if (V != Value1 && V != Value2) {
MoreThanTwoValues = true;
}
}
if (!Value1.getNode())
return DAG.getUNDEF(VT);
if (isOnlyLowElement)
return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0));
// If all elements are constants, create a load from the constant pool.
if (isConstant) {
SmallVector<Constant*, 16> CV;
for (unsigned i = 0, e = NumElems; i != e; ++i) {
if (ConstantFPSDNode *V =
dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue()));
} else if (ConstantSDNode *V =
dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
if (OpVT==EltVT)
CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
else {
// If OpVT and EltVT don't match, EltVT is not legal and the
// element values have been promoted/truncated earlier. Undo this;
// we don't want a v16i8 to become a v16i32 for example.
const ConstantInt *CI = V->getConstantIntValue();
CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()),
CI->getZExtValue()));
}
} else {
assert(Node->getOperand(i).getOpcode() == ISD::UNDEF);
Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
CV.push_back(UndefValue::get(OpNTy));
}
}
Constant *CP = ConstantVector::get(CV);
SDValue CPIdx = DAG.getConstantPool(CP, TLI.getPointerTy());
unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
return DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
MachinePointerInfo::getConstantPool(),
false, false, false, Alignment);
}
if (!MoreThanTwoValues) {
SmallVector<int, 8> ShuffleVec(NumElems, -1);
for (unsigned i = 0; i < NumElems; ++i) {
SDValue V = Node->getOperand(i);
if (V.getOpcode() == ISD::UNDEF)
continue;
ShuffleVec[i] = V == Value1 ? 0 : NumElems;
}
if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) {
// Get the splatted value into the low element of a vector register.
SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1);
SDValue Vec2;
if (Value2.getNode())
Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2);
else
Vec2 = DAG.getUNDEF(VT);
// Return shuffle(LowValVec, undef, <0,0,0,0>)
return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
}
}
// Otherwise, we can't handle this case efficiently.
return ExpandVectorBuildThroughStack(Node);
}
// 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) {
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);
}
SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
TLI.getPointerTy());
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.
SDValue TCChain = InChain;
bool isTailCall = isInTailCallPosition(DAG, Node, TCChain, TLI);
if (isTailCall)
InChain = TCChain;
std::pair<SDValue, SDValue> CallInfo =
TLI.LowerCallTo(InChain, RetTy, isSigned, !isSigned, false, false,
0, TLI.getLibcallCallingConv(LC), isTailCall,
/*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
Callee, Args, DAG, Node->getDebugLoc());
if (!CallInfo.second.getNode())
// It's a tailcall, return the chain (which is the DAG root).
return DAG.getRoot();
return CallInfo.first;
}
/// ExpandLibCall - Generate a libcall taking the given operands as arguments
/// and returning a result of type RetVT.
SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
const SDValue *Ops, unsigned NumOps,
bool isSigned, DebugLoc dl) {
TargetLowering::ArgListTy Args;
Args.reserve(NumOps);
TargetLowering::ArgListEntry Entry;
for (unsigned i = 0; i != NumOps; ++i) {
Entry.Node = Ops[i];
Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
Entry.isSExt = isSigned;
Entry.isZExt = !isSigned;
Args.push_back(Entry);
}
SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
TLI.getPointerTy());
Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
std::pair<SDValue,SDValue> CallInfo =
TLI.LowerCallTo(DAG.getEntryNode(), RetTy, isSigned, !isSigned, false,
false, 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
/*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
Callee, Args, DAG, dl);
return CallInfo.first;
}
// ExpandChainLibCall - Expand a node into a call to a libcall. Similar to
// ExpandLibCall except that the first operand is the in-chain.
std::pair<SDValue, SDValue>
SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
SDNode *Node,
bool isSigned) {
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();
Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
Entry.Node = Node->getOperand(i);
Entry.Ty = ArgTy;
Entry.isSExt = isSigned;
Entry.isZExt = !isSigned;
Args.push_back(Entry);
}
SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
TLI.getPointerTy());
Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
std::pair<SDValue, SDValue> CallInfo =
TLI.LowerCallTo(InChain, RetTy, isSigned, !isSigned, false, false,
0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
/*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
Callee, Args, DAG, Node->getDebugLoc());
return CallInfo;
}
SDValue SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node,
RTLIB::Libcall Call_F32,
RTLIB::Libcall Call_F64,
RTLIB::Libcall Call_F80,
RTLIB::Libcall Call_PPCF128) {
RTLIB::Libcall LC;
switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
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::ppcf128: LC = Call_PPCF128; break;
}
return ExpandLibCall(LC, Node, false);
}
SDValue SelectionDAGLegalize::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) {
RTLIB::Libcall LC;
switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
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;
case MVT::i64: LC = Call_I64; break;
case MVT::i128: LC = Call_I128; 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) {
unsigned OtherOpcode = 0;
if (isSigned)
OtherOpcode = isDIV ? ISD::SREM : ISD::SDIV;
else
OtherOpcode = isDIV ? ISD::UREM : ISD::UDIV;
SDValue Op0 = Node->getOperand(0);
SDValue Op1 = Node->getOperand(1);
for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
SDNode *User = *UI;
if (User == Node)
continue;
if (User->getOpcode() == OtherOpcode &&
User->getOperand(0) == Op0 &&
User->getOperand(1) == Op1)
return true;
}
return false;
}
/// ExpandDivRemLibCall - Issue libcalls to __{u}divmod to compute div / rem
/// pairs.
void
SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
SmallVectorImpl<SDValue> &Results) {
unsigned Opcode = Node->getOpcode();
bool isSigned = Opcode == ISD::SDIVREM;
RTLIB::Libcall LC;
switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
default: 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();
std::pair<SDValue, SDValue> CallInfo =
TLI.LowerCallTo(InChain, RetTy, isSigned, !isSigned, false, false,
0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
/*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
Callee, Args, DAG, dl);
// 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
/// legal for the target.
SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned,
SDValue Op0,
EVT DestVT,
DebugLoc dl) {
if (Op0.getValueType() == MVT::i32) {
// simple 32-bit [signed|unsigned] integer to float/double expansion
// Get the stack frame index of a 8 byte buffer.
SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64);
// word offset constant for Hi/Lo address computation
SDValue WordOff = DAG.getConstant(sizeof(int), TLI.getPointerTy());
// set up Hi and Lo (into buffer) address based on endian
SDValue Hi = StackSlot;
SDValue Lo = DAG.getNode(ISD::ADD, dl,
TLI.getPointerTy(), StackSlot, WordOff);
if (TLI.isLittleEndian())
std::swap(Hi, Lo);
// if signed map to unsigned space
SDValue Op0Mapped;
if (isSigned) {
// constant used to invert sign bit (signed to unsigned mapping)
SDValue SignBit = DAG.getConstant(0x80000000u, MVT::i32);
Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit);
} else {
Op0Mapped = Op0;
}
// store the lo of the constructed double - based on integer input
SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl,
Op0Mapped, Lo, MachinePointerInfo(),
false, false, 0);
// initial hi portion of constructed double
SDValue InitialHi = DAG.getConstant(0x43300000u, MVT::i32);
// store the hi of the constructed double - biased exponent
SDValue Store2 = DAG.getStore(Store1, dl, InitialHi, Hi,
MachinePointerInfo(),
false, false, 0);
// load the constructed double
SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot,
MachinePointerInfo(), false, false, false, 0);
// FP constant to bias correct the final result
SDValue Bias = DAG.getConstantFP(isSigned ?
BitsToDouble(0x4330000080000000ULL) :
BitsToDouble(0x4330000000000000ULL),
MVT::f64);
// subtract the bias
SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias);
// final result
SDValue Result;
// handle final rounding
if (DestVT == MVT::f64) {
// do nothing
Result = Sub;
} else if (DestVT.bitsLT(MVT::f64)) {
Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
DAG.getIntPtrConstant(0));
} else if (DestVT.bitsGT(MVT::f64)) {
Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
}
return Result;
}
assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet");
// Code below here assumes !isSigned without checking again.
// Implementation of unsigned i64 to f64 following the algorithm in
// __floatundidf in compiler_rt. This implementation has the advantage
// of performing rounding correctly, both in the default rounding mode
// and in all alternate rounding modes.
// TODO: Generalize this for use with other types.
if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) {
SDValue TwoP52 =
DAG.getConstant(UINT64_C(0x4330000000000000), MVT::i64);
SDValue TwoP84PlusTwoP52 =
DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), MVT::f64);
SDValue TwoP84 =
DAG.getConstant(UINT64_C(0x4530000000000000), MVT::i64);
SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32);
SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0,
DAG.getConstant(32, MVT::i64));
SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52);
SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84);
SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr);
SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr);
SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt,
TwoP84PlusTwoP52);
return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub);
}
// Implementation of unsigned i64 to f32.
// TODO: Generalize this for use with other types.
if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) {
// For unsigned conversions, convert them to signed conversions using the
// algorithm from the x86_64 __floatundidf in compiler_rt.
if (!isSigned) {
SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
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);
SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr);
SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or);
SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt);
// TODO: This really should be implemented using a branch rather than a
// select. We happen to get lucky and machinesink does the right
// thing most of the time. This would be a good candidate for a
//pseudo-op, or, even better, for whole-function isel.
SDValue SignBitTest = DAG.getSetCC(dl, TLI.getSetCCResultType(MVT::i64),
Op0, DAG.getConstant(0, MVT::i64), ISD::SETLT);
return DAG.getNode(ISD::SELECT, dl, MVT::f32, SignBitTest, Slow, Fast);
}
// Otherwise, implement the fully general conversion.
SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
DAG.getConstant(UINT64_C(0xfffffffffffff800), MVT::i64));
SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And,
DAG.getConstant(UINT64_C(0x800), MVT::i64));
SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
DAG.getConstant(UINT64_C(0x7ff), MVT::i64));
SDValue Ne = DAG.getSetCC(dl, TLI.getSetCCResultType(MVT::i64),
And2, DAG.getConstant(UINT64_C(0), MVT::i64), ISD::SETNE);
SDValue Sel = DAG.getNode(ISD::SELECT, dl, MVT::i64, Ne, Or, Op0);
SDValue Ge = DAG.getSetCC(dl, TLI.getSetCCResultType(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));
SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh);
SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc);
SDValue TwoP32 =
DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), MVT::f64);
SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt);
SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2);
SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo);
SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2);
return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd,
DAG.getIntPtrConstant(0));
}
SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0);
SDValue SignSet = DAG.getSetCC(dl, TLI.getSetCCResultType(Op0.getValueType()),
Op0, DAG.getConstant(0, Op0.getValueType()),
ISD::SETLT);
SDValue Zero = DAG.getIntPtrConstant(0), Four = DAG.getIntPtrConstant(4);
SDValue CstOffset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(),
SignSet, Four, Zero);
// If the sign bit of the integer is set, the large number will be treated
// as a negative number. To counteract this, the dynamic code adds an
// offset depending on the data type.
uint64_t FF;
switch (Op0.getValueType().getSimpleVT().SimpleTy) {
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)
case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float)
}
if (TLI.isLittleEndian()) FF <<= 32;
Constant *FudgeFactor = ConstantInt::get(
Type::getInt64Ty(*DAG.getContext()), FF);
SDValue CPIdx = DAG.getConstantPool(FudgeFactor, TLI.getPointerTy());
unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
CPIdx = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), CPIdx, CstOffset);
Alignment = std::min(Alignment, 4u);
SDValue FudgeInReg;
if (DestVT == MVT::f32)
FudgeInReg = DAG.getLoad(MVT::f32, dl, DAG.getEntryNode(), CPIdx,
MachinePointerInfo::getConstantPool(),
false, false, false, Alignment);
else {
SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
DAG.getEntryNode(), CPIdx,