#include "llvm/Analysis/DebugInfo.h"
#include "llvm/CodeGen/Analysis.h"
#include "llvm/CodeGen/MachineFunction.h"
-#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
-#include "llvm/CodeGen/MachineModuleInfo.h"
-#include "llvm/CodeGen/PseudoSourceValue.h"
#include "llvm/CodeGen/SelectionDAG.h"
-#include "llvm/Target/TargetFrameInfo.h"
+#include "llvm/Target/TargetFrameLowering.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetMachine.h"
-#include "llvm/Target/TargetOptions.h"
#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
-#include "llvm/Function.h"
-#include "llvm/GlobalVariable.h"
#include "llvm/LLVMContext.h"
-#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
const TargetMachine &TM;
const TargetLowering &TLI;
SelectionDAG &DAG;
- CodeGenOpt::Level OptLevel;
// Libcall insertion helpers.
- /// LastCALLSEQ_END - This keeps track of the CALLSEQ_END node that has been
+ /// LastCALLSEQ - 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;
+ SmallVector<SDValue, 8> LastCALLSEQ;
enum LegalizeAction {
Legal, // The target natively supports this operation.
// If someone requests legalization of the new node, return itself.
if (From != To)
LegalizedNodes.insert(std::make_pair(To, To));
+
+ // Transfer SDDbgValues.
+ DAG.TransferDbgValues(From, To);
}
public:
- SelectionDAGLegalize(SelectionDAG &DAG, CodeGenOpt::Level ol);
+ explicit SelectionDAGLegalize(SelectionDAG &DAG);
/// 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);
+ return (LegalizeAction)TLI.getTypeAction(*DAG.getContext(), VT);
}
/// isTypeLegal - Return true if this type is legal on this target.
DebugLoc dl);
SDValue ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned);
+ SDValue ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, const SDValue *Ops,
+ unsigned NumOps, bool isSigned, DebugLoc dl);
+
std::pair<SDValue, SDValue> ExpandChainLibCall(RTLIB::Libcall LC,
SDNode *Node, bool isSigned);
SDValue ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32,
RTLIB::Libcall Call_I32,
RTLIB::Libcall Call_I64,
RTLIB::Libcall Call_I128);
+ void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT, DebugLoc dl);
SDValue ExpandBUILD_VECTOR(SDNode *Node);
SDValue ExpandBitCount(unsigned Opc, SDValue Op, DebugLoc dl);
SDValue ExpandExtractFromVectorThroughStack(SDValue Op);
+ SDValue ExpandInsertToVectorThroughStack(SDValue Op);
SDValue ExpandVectorBuildThroughStack(SDNode* Node);
std::pair<SDValue, SDValue> ExpandAtomic(SDNode *Node);
void ExpandNode(SDNode *Node, SmallVectorImpl<SDValue> &Results);
void PromoteNode(SDNode *Node, SmallVectorImpl<SDValue> &Results);
+
+ SDValue getLastCALLSEQ() { return LastCALLSEQ.back(); }
+ void setLastCALLSEQ(const SDValue s) { LastCALLSEQ.back() = s; }
+ void pushLastCALLSEQ(SDValue s) {
+ LastCALLSEQ.push_back(s);
+ }
+ void popLastCALLSEQ() {
+ LastCALLSEQ.pop_back();
+ }
};
}
return DAG.getVectorShuffle(NVT, dl, N1, N2, &NewMask[0]);
}
-SelectionDAGLegalize::SelectionDAGLegalize(SelectionDAG &dag,
- CodeGenOpt::Level ol)
+SelectionDAGLegalize::SelectionDAGLegalize(SelectionDAG &dag)
: TM(dag.getTarget()), TLI(dag.getTargetLoweringInfo()),
- DAG(dag), OptLevel(ol),
+ DAG(dag),
ValueTypeActions(TLI.getValueTypeActions()) {
assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE &&
"Too many value types for ValueTypeActions to hold!");
}
void SelectionDAGLegalize::LegalizeDAG() {
- LastCALLSEQ_END = DAG.getEntryNode();
- IsLegalizingCall = false;
+ pushLastCALLSEQ(DAG.getEntryNode());
// The legalize process is inherently a bottom-up recursive process (users
// legalize their uses before themselves). Given infinite stack space, we
/// 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) {
- if (Node->getOpcode() == ISD::CALLSEQ_END)
- return Node;
+static SDNode *FindCallEndFromCallStart(SDNode *Node, int depth = 0) {
+ int next_depth = depth;
+ if (Node->getOpcode() == ISD::CALLSEQ_START)
+ next_depth = depth + 1;
+ if (Node->getOpcode() == ISD::CALLSEQ_END) {
+ assert(depth > 0 && "negative depth!");
+ if (depth == 1)
+ return Node;
+ else
+ next_depth = depth - 1;
+ }
if (Node->use_empty())
return 0; // No CallSeqEnd
SDNode *User = *UI;
for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i)
if (User->getOperand(i) == TheChain)
- if (SDNode *Result = FindCallEndFromCallStart(User))
+ if (SDNode *Result = FindCallEndFromCallStart(User, next_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??");
- if (Node->getOpcode() == ISD::CALLSEQ_START) return Node;
-
- assert(Node->getOperand(0).getValueType() == MVT::Other &&
- "Node doesn't have a token chain argument!");
- return FindCallStartFromCallEnd(Node->getOperand(0).getNode());
+ 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:
+ break;
+ case ISD::CALLSEQ_START:
+ if (!nested)
+ return Node;
+ Node = Node->getOperand(0).getNode();
+ nested--;
+ break;
+ case ISD::CALLSEQ_END:
+ nested++;
+ break;
+ }
+ }
+ return (Node->getOpcode() == ISD::CALLSEQ_START) ? Node : 0;
}
/// LegalizeAllNodesNotLeadingTo - Recursively walk the uses of N, looking to
SDValue CPIdx = DAG.getConstantPool(LLVMC, TLI.getPointerTy());
unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
if (Extend)
- return DAG.getExtLoad(ISD::EXTLOAD, OrigVT, dl,
+ return DAG.getExtLoad(ISD::EXTLOAD, dl, OrigVT,
DAG.getEntryNode(),
CPIdx, MachinePointerInfo::getConstantPool(),
VT, false, false, Alignment);
SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
return DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
ST->isVolatile(), ST->isNonTemporal(), Alignment);
- } else {
- // Do a (aligned) store to a stack slot, then copy from the stack slot
- // to the final destination using (unaligned) integer loads and stores.
- EVT StoredVT = ST->getMemoryVT();
- EVT RegVT =
- TLI.getRegisterType(*DAG.getContext(),
- EVT::getIntegerVT(*DAG.getContext(),
- StoredVT.getSizeInBits()));
- unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
- unsigned RegBytes = RegVT.getSizeInBits() / 8;
- unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
-
- // Make sure the stack slot is also aligned for the register type.
- SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
-
- // Perform the original store, only redirected to the stack slot.
- SDValue Store = DAG.getTruncStore(Chain, dl,
- Val, StackPtr, MachinePointerInfo(),
- StoredVT, false, false, 0);
- SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
- SmallVector<SDValue, 8> Stores;
- unsigned Offset = 0;
-
- // Do all but one copies using the full register width.
- for (unsigned i = 1; i < NumRegs; i++) {
- // Load one integer register's worth from the stack slot.
- SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr,
- MachinePointerInfo(),
- false, false, 0);
- // Store it to the final location. Remember the store.
- Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
- ST->getPointerInfo().getWithOffset(Offset),
- ST->isVolatile(), ST->isNonTemporal(),
- MinAlign(ST->getAlignment(), Offset)));
- // Increment the pointers.
- Offset += RegBytes;
- StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
- Increment);
- Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
- }
+ }
+ // Do a (aligned) store to a stack slot, then copy from the stack slot
+ // to the final destination using (unaligned) integer loads and stores.
+ EVT StoredVT = ST->getMemoryVT();
+ EVT RegVT =
+ TLI.getRegisterType(*DAG.getContext(),
+ EVT::getIntegerVT(*DAG.getContext(),
+ StoredVT.getSizeInBits()));
+ unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
+ unsigned RegBytes = RegVT.getSizeInBits() / 8;
+ unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
- // The last store may be partial. Do a truncating store. On big-endian
- // machines this requires an extending load from the stack slot to ensure
- // that the bits are in the right place.
- EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
- 8 * (StoredBytes - Offset));
+ // Make sure the stack slot is also aligned for the register type.
+ SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
- // Load from the stack slot.
- SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, RegVT, dl, Store, StackPtr,
- MachinePointerInfo(),
- MemVT, false, false, 0);
+ // Perform the original store, only redirected to the stack slot.
+ SDValue Store = DAG.getTruncStore(Chain, dl,
+ Val, StackPtr, MachinePointerInfo(),
+ StoredVT, false, false, 0);
+ SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
+ SmallVector<SDValue, 8> Stores;
+ unsigned Offset = 0;
- Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
- ST->getPointerInfo()
- .getWithOffset(Offset),
- MemVT, ST->isVolatile(),
- ST->isNonTemporal(),
- MinAlign(ST->getAlignment(), Offset)));
- // The order of the stores doesn't matter - say it with a TokenFactor.
- return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
- Stores.size());
+ // Do all but one copies using the full register width.
+ for (unsigned i = 1; i < NumRegs; i++) {
+ // Load one integer register's worth from the stack slot.
+ SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr,
+ MachinePointerInfo(),
+ false, false, 0);
+ // Store it to the final location. Remember the store.
+ Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
+ ST->getPointerInfo().getWithOffset(Offset),
+ ST->isVolatile(), ST->isNonTemporal(),
+ MinAlign(ST->getAlignment(), Offset)));
+ // Increment the pointers.
+ Offset += RegBytes;
+ StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
+ Increment);
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
}
+
+ // The last store may be partial. Do a truncating store. On big-endian
+ // machines this requires an extending load from the stack slot to ensure
+ // that the bits are in the right place.
+ EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
+ 8 * (StoredBytes - Offset));
+
+ // Load from the stack slot.
+ SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Store, StackPtr,
+ MachinePointerInfo(),
+ MemVT, false, false, 0);
+
+ Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
+ ST->getPointerInfo()
+ .getWithOffset(Offset),
+ MemVT, ST->isVolatile(),
+ ST->isNonTemporal(),
+ MinAlign(ST->getAlignment(), Offset)));
+ // The order of the stores doesn't matter - say it with a TokenFactor.
+ return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0],
+ Stores.size());
}
assert(ST->getMemoryVT().isInteger() &&
!ST->getMemoryVT().isVector() &&
int IncrementSize = NumBits / 8;
// Divide the stored value in two parts.
- SDValue ShiftAmount = DAG.getConstant(NumBits, TLI.getShiftAmountTy());
+ SDValue ShiftAmount = DAG.getConstant(NumBits,
+ TLI.getShiftAmountTy(Val.getValueType()));
SDValue Lo = Val;
SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
// The last copy may be partial. Do an extending load.
EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
8 * (LoadedBytes - Offset));
- SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, RegVT, dl, Chain, Ptr,
+ SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr,
LD->getPointerInfo().getWithOffset(Offset),
MemVT, LD->isVolatile(),
LD->isNonTemporal(),
Stores.size());
// Finally, perform the original load only redirected to the stack slot.
- Load = DAG.getExtLoad(LD->getExtensionType(), VT, dl, TF, StackBase,
+ Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
MachinePointerInfo(), LoadedVT, false, false, 0);
// Callers expect a MERGE_VALUES node.
// Load the value in two parts
SDValue Lo, Hi;
if (TLI.isLittleEndian()) {
- Lo = DAG.getExtLoad(ISD::ZEXTLOAD, VT, dl, Chain, Ptr, LD->getPointerInfo(),
+ Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(),
NewLoadedVT, LD->isVolatile(),
LD->isNonTemporal(), Alignment);
Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
DAG.getConstant(IncrementSize, TLI.getPointerTy()));
- Hi = DAG.getExtLoad(HiExtType, VT, dl, Chain, Ptr,
+ Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr,
LD->getPointerInfo().getWithOffset(IncrementSize),
NewLoadedVT, LD->isVolatile(),
LD->isNonTemporal(), MinAlign(Alignment,IncrementSize));
} else {
- Hi = DAG.getExtLoad(HiExtType, VT, dl, Chain, Ptr, LD->getPointerInfo(),
+ Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(),
NewLoadedVT, LD->isVolatile(),
LD->isNonTemporal(), Alignment);
Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
DAG.getConstant(IncrementSize, TLI.getPointerTy()));
- Lo = DAG.getExtLoad(ISD::ZEXTLOAD, VT, dl, Chain, Ptr,
+ Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr,
LD->getPointerInfo().getWithOffset(IncrementSize),
NewLoadedVT, LD->isVolatile(),
LD->isNonTemporal(), MinAlign(Alignment,IncrementSize));
}
// aggregate the two parts
- SDValue ShiftAmount = DAG.getConstant(NumBits, TLI.getShiftAmountTy());
+ SDValue ShiftAmount = DAG.getConstant(NumBits,
+ TLI.getShiftAmountTy(Hi.getValueType()));
SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount);
Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo);
}
}
}
- return SDValue();
+ return SDValue(0, 0);
}
/// LegalizeOp - We know that the specified value has a legal type, and
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::BR_JT:
case ISD::BR_CC:
case ISD::BRCOND:
- // Branches tweak the chain to include LastCALLSEQ_END
+ assert(LastCALLSEQ.size() == 1 && "branch inside CALLSEQ_BEGIN/END?");
+ // Branches tweak the chain to include LastCALLSEQ
Ops[0] = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ops[0],
- LastCALLSEQ_END);
+ getLastCALLSEQ());
Ops[0] = LegalizeOp(Ops[0]);
- LastCALLSEQ_END = DAG.getEntryNode();
+ setLastCALLSEQ(DAG.getEntryNode());
break;
case ISD::SHL:
case ISD::SRL:
// Legalizing shifts/rotates requires adjusting the shift amount
// to the appropriate width.
if (!Ops[1].getValueType().isVector())
- Ops[1] = LegalizeOp(DAG.getShiftAmountOperand(Ops[1]));
+ Ops[1] = LegalizeOp(DAG.getShiftAmountOperand(Ops[0].getValueType(),
+ Ops[1]));
break;
case ISD::SRL_PARTS:
case ISD::SRA_PARTS:
// Legalizing shifts/rotates requires adjusting the shift amount
// to the appropriate width.
if (!Ops[2].getValueType().isVector())
- Ops[2] = LegalizeOp(DAG.getShiftAmountOperand(Ops[2]));
+ Ops[2] = LegalizeOp(DAG.getShiftAmountOperand(Ops[0].getValueType(),
+ Ops[2]));
break;
}
break;
case ISD::CALLSEQ_START: {
SDNode *CallEnd = FindCallEndFromCallStart(Node);
+ assert(CallEnd && "didn't find CALLSEQ_END!");
// 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
// Merge in the last call to ensure that this call starts after the last
// call ended.
- if (LastCALLSEQ_END.getOpcode() != ISD::EntryToken) {
+ if (getLastCALLSEQ().getOpcode() != ISD::EntryToken) {
Tmp1 = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
- Tmp1, LastCALLSEQ_END);
+ Tmp1, getLastCALLSEQ());
Tmp1 = LegalizeOp(Tmp1);
}
// sequence have been legalized, legalize the call itself. During this
// process, no libcalls can/will be inserted, guaranteeing that no calls
// can overlap.
- assert(!IsLegalizingCall && "Inconsistent sequentialization of calls!");
// Note that we are selecting this call!
- LastCALLSEQ_END = SDValue(CallEnd, 0);
- IsLegalizingCall = true;
+ setLastCALLSEQ(SDValue(CallEnd, 0));
// Legalize the call, starting from the CALLSEQ_END.
- LegalizeOp(LastCALLSEQ_END);
- assert(!IsLegalizingCall && "CALLSEQ_END should have cleared this!");
+ LegalizeOp(getLastCALLSEQ());
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;
+ {
+ SDNode *myCALLSEQ_BEGIN = FindCallStartFromCallEnd(Node);
+
+ // 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 (getLastCALLSEQ().getNode() != Node) {
+ LegalizeOp(SDValue(myCALLSEQ_BEGIN, 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;
+ }
+
+ pushLastCALLSEQ(SDValue(myCALLSEQ_BEGIN, 0));
}
// Otherwise, the call start has been legalized and everything is going
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::Flag){
+ 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.getResNo());
}
}
- assert(IsLegalizingCall && "Call sequence imbalance between start/end?");
// This finishes up call legalization.
- IsLegalizingCall = false;
+ popLastCALLSEQ();
// If the CALLSEQ_END node has a flag, remember that we legalized it.
AddLegalizedOperand(SDValue(Node, 0), Result.getValue(0));
ISD::LoadExtType NewExtType =
ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
- Result = DAG.getExtLoad(NewExtType, Node->getValueType(0), dl,
+ Result = DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
Tmp1, Tmp2, LD->getPointerInfo(),
NVT, isVolatile, isNonTemporal, Alignment);
if (TLI.isLittleEndian()) {
// EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
// Load the bottom RoundWidth bits.
- Lo = DAG.getExtLoad(ISD::ZEXTLOAD, Node->getValueType(0), dl,
+ Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
Tmp1, Tmp2,
LD->getPointerInfo(), RoundVT, isVolatile,
isNonTemporal, Alignment);
IncrementSize = RoundWidth / 8;
Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
DAG.getIntPtrConstant(IncrementSize));
- Hi = DAG.getExtLoad(ExtType, Node->getValueType(0), dl, Tmp1, Tmp2,
+ Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Tmp1, Tmp2,
LD->getPointerInfo().getWithOffset(IncrementSize),
ExtraVT, isVolatile, isNonTemporal,
MinAlign(Alignment, IncrementSize));
// Move the top bits to the right place.
Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
- DAG.getConstant(RoundWidth, TLI.getShiftAmountTy()));
+ DAG.getConstant(RoundWidth,
+ TLI.getShiftAmountTy(Hi.getValueType())));
// Join the hi and lo parts.
Result = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
// Big endian - avoid unaligned loads.
// EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
// Load the top RoundWidth bits.
- Hi = DAG.getExtLoad(ExtType, Node->getValueType(0), dl, Tmp1, Tmp2,
+ Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Tmp1, Tmp2,
LD->getPointerInfo(), RoundVT, isVolatile,
isNonTemporal, Alignment);
Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
DAG.getIntPtrConstant(IncrementSize));
Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
- Node->getValueType(0), dl, Tmp1, Tmp2,
+ dl, Node->getValueType(0), Tmp1, Tmp2,
LD->getPointerInfo().getWithOffset(IncrementSize),
ExtraVT, isVolatile, isNonTemporal,
MinAlign(Alignment, IncrementSize));
// Move the top bits to the right place.
Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
- DAG.getConstant(ExtraWidth, TLI.getShiftAmountTy()));
+ DAG.getConstant(ExtraWidth,
+ TLI.getShiftAmountTy(Hi.getValueType())));
// Join the hi and lo parts.
Result = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
Tmp2 = LegalizeOp(Load.getValue(1));
break;
}
+
+ // If this is a promoted vector load, and the vector element types are
+ // legal, then scalarize it.
+ if (ExtType == ISD::EXTLOAD && SrcVT.isVector() &&
+ isTypeLegal(Node->getValueType(0).getScalarType())) {
+ SmallVector<SDValue, 8> LoadVals;
+ SmallVector<SDValue, 8> LoadChains;
+ unsigned NumElem = SrcVT.getVectorNumElements();
+ unsigned Stride = SrcVT.getScalarType().getSizeInBits()/8;
+
+ for (unsigned Idx=0; Idx<NumElem; Idx++) {
+ Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
+ DAG.getIntPtrConstant(Stride));
+ SDValue ScalarLoad = DAG.getExtLoad(ISD::EXTLOAD, dl,
+ Node->getValueType(0).getScalarType(),
+ Tmp1, Tmp2, LD->getPointerInfo().getWithOffset(Idx * Stride),
+ SrcVT.getScalarType(),
+ LD->isVolatile(), LD->isNonTemporal(),
+ LD->getAlignment());
+
+ LoadVals.push_back(ScalarLoad.getValue(0));
+ LoadChains.push_back(ScalarLoad.getValue(1));
+ }
+ Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+ &LoadChains[0], LoadChains.size());
+ SDValue ValRes = DAG.getNode(ISD::BUILD_VECTOR, dl,
+ Node->getValueType(0), &LoadVals[0], LoadVals.size());
+
+ Tmp1 = LegalizeOp(ValRes); // Relegalize new nodes.
+ Tmp2 = LegalizeOp(Result.getValue(0)); // Relegalize new nodes.
+ break;
+ }
+
+ // If this is a promoted vector load, and the vector element types are
+ // illegal, create the promoted vector from bitcasted segments.
+ if (ExtType == ISD::EXTLOAD && SrcVT.isVector()) {
+ EVT MemElemTy = Node->getValueType(0).getScalarType();
+ EVT SrcSclrTy = SrcVT.getScalarType();
+ unsigned SizeRatio =
+ (MemElemTy.getSizeInBits() / SrcSclrTy.getSizeInBits());
+
+ SmallVector<SDValue, 8> LoadVals;
+ SmallVector<SDValue, 8> LoadChains;
+ unsigned NumElem = SrcVT.getVectorNumElements();
+ unsigned Stride = SrcVT.getScalarType().getSizeInBits()/8;
+
+ for (unsigned Idx=0; Idx<NumElem; Idx++) {
+ Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
+ DAG.getIntPtrConstant(Stride));
+ SDValue ScalarLoad = DAG.getExtLoad(ISD::EXTLOAD, dl,
+ SrcVT.getScalarType(),
+ Tmp1, Tmp2, LD->getPointerInfo().getWithOffset(Idx * Stride),
+ SrcVT.getScalarType(),
+ LD->isVolatile(), LD->isNonTemporal(),
+ LD->getAlignment());
+ if (TLI.isBigEndian()) {
+ // MSB (which is garbage, comes first)
+ LoadVals.push_back(ScalarLoad.getValue(0));
+ for (unsigned i = 0; i<SizeRatio-1; ++i)
+ LoadVals.push_back(DAG.getUNDEF(SrcVT.getScalarType()));
+ } else {
+ // LSB (which is data, comes first)
+ for (unsigned i = 0; i<SizeRatio-1; ++i)
+ LoadVals.push_back(DAG.getUNDEF(SrcVT.getScalarType()));
+ LoadVals.push_back(ScalarLoad.getValue(0));
+ }
+ LoadChains.push_back(ScalarLoad.getValue(1));
+ }
+
+ Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+ &LoadChains[0], LoadChains.size());
+ EVT TempWideVector = EVT::getVectorVT(*DAG.getContext(),
+ SrcVT.getScalarType(), NumElem*SizeRatio);
+ SDValue ValRes = DAG.getNode(ISD::BUILD_VECTOR, dl,
+ TempWideVector, &LoadVals[0], LoadVals.size());
+
+ // Cast to the correct type
+ ValRes = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), ValRes);
+
+ Tmp1 = LegalizeOp(ValRes); // Relegalize new nodes.
+ Tmp2 = LegalizeOp(Result.getValue(0)); // Relegalize new nodes.
+ break;
+
+ }
+
// FIXME: This does not work for vectors on most targets. Sign- and
// zero-extend operations are currently folded into extending loads,
// whether they are legal or not, and then we end up here without any
"EXTLOAD should always be supported!");
// Turn the unsupported load into an EXTLOAD followed by an explicit
// zero/sign extend inreg.
- Result = DAG.getExtLoad(ISD::EXTLOAD, Node->getValueType(0), dl,
+ Result = DAG.getExtLoad(ISD::EXTLOAD, dl, Node->getValueType(0),
Tmp1, Tmp2, LD->getPointerInfo(), SrcVT,
LD->isVolatile(), LD->isNonTemporal(),
LD->getAlignment());
Result.getValueType(),
Result, DAG.getValueType(SrcVT));
else
- ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT);
+ ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT.getScalarType());
Tmp1 = LegalizeOp(ValRes); // Relegalize new nodes.
Tmp2 = LegalizeOp(Result.getValue(1)); // Relegalize new nodes.
break;
Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
DAG.getIntPtrConstant(IncrementSize));
Hi = DAG.getNode(ISD::SRL, dl, Tmp3.getValueType(), Tmp3,
- DAG.getConstant(RoundWidth, TLI.getShiftAmountTy()));
+ DAG.getConstant(RoundWidth,
+ TLI.getShiftAmountTy(Tmp3.getValueType())));
Hi = DAG.getTruncStore(Tmp1, dl, Hi, Tmp2,
ST->getPointerInfo().getWithOffset(IncrementSize),
ExtraVT, isVolatile, isNonTemporal,
// TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
// Store the top RoundWidth bits.
Hi = DAG.getNode(ISD::SRL, dl, Tmp3.getValueType(), Tmp3,
- DAG.getConstant(ExtraWidth, TLI.getShiftAmountTy()));
+ DAG.getConstant(ExtraWidth,
+ TLI.getShiftAmountTy(Tmp3.getValueType())));
Hi = DAG.getTruncStore(Tmp1, dl, Hi, Tmp2, ST->getPointerInfo(),
RoundVT, isVolatile, isNonTemporal, Alignment);
Result = TLI.LowerOperation(Result, DAG);
break;
case Expand:
+
+ EVT WideScalarVT = Tmp3.getValueType().getScalarType();
+ EVT NarrowScalarVT = StVT.getScalarType();
+
+ // The Store type is illegal, must scalarize the vector store.
+ SmallVector<SDValue, 8> Stores;
+ bool ScalarLegal = isTypeLegal(WideScalarVT);
+ if (!isTypeLegal(StVT) && StVT.isVector() && ScalarLegal) {
+ unsigned NumElem = StVT.getVectorNumElements();
+
+ unsigned ScalarSize = StVT.getScalarType().getSizeInBits();
+ // Round odd types to the next pow of two.
+ if (!isPowerOf2_32(ScalarSize))
+ ScalarSize = NextPowerOf2(ScalarSize);
+ // Types smaller than 8 bits are promoted to 8 bits.
+ ScalarSize = std::max<unsigned>(ScalarSize, 8);
+ // Store stride
+ unsigned Stride = ScalarSize/8;
+ assert(isPowerOf2_32(Stride) && "Stride must be a power of two");
+
+ for (unsigned Idx=0; Idx<NumElem; Idx++) {
+ SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
+ WideScalarVT, Tmp3, DAG.getIntPtrConstant(Idx));
+
+
+ EVT NVT = EVT::getIntegerVT(*DAG.getContext(), ScalarSize);
+
+ Ex = DAG.getNode(ISD::TRUNCATE, dl, NVT, Ex);
+ Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
+ DAG.getIntPtrConstant(Stride));
+ SDValue Store = DAG.getStore(Tmp1, dl, Ex, Tmp2,
+ ST->getPointerInfo().getWithOffset(Idx*Stride),
+ isVolatile, isNonTemporal, Alignment);
+ Stores.push_back(Store);
+ }
+ Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+ &Stores[0], Stores.size());
+ break;
+ }
+
+ // The Store type is illegal, must scalarize the vector store.
+ // However, the scalar type is illegal. Must bitcast the result
+ // and store it in smaller parts.
+ if (!isTypeLegal(StVT) && StVT.isVector()) {
+ unsigned WideNumElem = StVT.getVectorNumElements();
+ unsigned Stride = NarrowScalarVT.getSizeInBits()/8;
+
+ unsigned SizeRatio =
+ (WideScalarVT.getSizeInBits() / NarrowScalarVT.getSizeInBits());
+
+ EVT CastValueVT = EVT::getVectorVT(*DAG.getContext(), NarrowScalarVT,
+ SizeRatio*WideNumElem);
+
+ // Cast the wide elem vector to wider vec with smaller elem type.
+ // Example <2 x i64> -> <4 x i32>
+ Tmp3 = DAG.getNode(ISD::BITCAST, dl, CastValueVT, Tmp3);
+
+ for (unsigned Idx=0; Idx<WideNumElem*SizeRatio; Idx++) {
+ // Extract elment i
+ SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
+ NarrowScalarVT, Tmp3, DAG.getIntPtrConstant(Idx));
+ // bump pointer.
+ Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2,
+ DAG.getIntPtrConstant(Stride));
+
+ // Store if, this element is:
+ // - First element on big endian, or
+ // - Last element on little endian
+ if (( TLI.isBigEndian() && (Idx%SizeRatio == 0)) ||
+ ((!TLI.isBigEndian() && (Idx%SizeRatio == SizeRatio-1)))) {
+ SDValue Store = DAG.getStore(Tmp1, dl, Ex, Tmp2,
+ ST->getPointerInfo().getWithOffset(Idx*Stride),
+ isVolatile, isNonTemporal, Alignment);
+ Stores.push_back(Store);
+ }
+ }
+ Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+ &Stores[0], Stores.size());
+ break;
+ }
+
+
// TRUNCSTORE:i16 i32 -> STORE i16
assert(isTypeLegal(StVT) && "Do not know how to expand this store!");
Tmp3 = DAG.getNode(ISD::TRUNCATE, dl, StVT, Tmp3);
if (Op.getValueType().isVector())
return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr,MachinePointerInfo(),
false, false, 0);
- return DAG.getExtLoad(ISD::EXTLOAD, Op.getValueType(), dl, Ch, StackPtr,
+ return DAG.getExtLoad(ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr,
MachinePointerInfo(),
Vec.getValueType().getVectorElementType(),
false, false, 0);
}
+SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
+ assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
+
+ SDValue Vec = Op.getOperand(0);
+ SDValue Part = Op.getOperand(1);
+ SDValue Idx = Op.getOperand(2);
+ DebugLoc dl = Op.getDebugLoc();
+
+ // Store the value to a temporary stack slot, then LOAD the returned part.
+
+ SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
+ int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
+ MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
+
+ // First store the whole vector.
+ SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo,
+ false, false, 0);
+
+ // Then store the inserted part.
+
+ // Add the offset to the index.
+ unsigned EltSize =
+ Vec.getValueType().getVectorElementType().getSizeInBits()/8;
+
+ Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
+ DAG.getConstant(EltSize, Idx.getValueType()));
+
+ if (Idx.getValueType().bitsGT(TLI.getPointerTy()))
+ Idx = DAG.getNode(ISD::TRUNCATE, dl, TLI.getPointerTy(), Idx);
+ else
+ Idx = DAG.getNode(ISD::ZERO_EXTEND, dl, TLI.getPointerTy(), Idx);
+
+ SDValue SubStackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx,
+ StackPtr);
+
+ // Store the subvector.
+ Ch = DAG.getStore(DAG.getEntryNode(), dl, Part, SubStackPtr,
+ MachinePointerInfo(), false, false, 0);
+
+ // Finally, load the updated vector.
+ return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo,
+ false, false, 0);
+}
+
SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
// We can't handle this case efficiently. Allocate a sufficiently
// aligned object on the stack, store each element into it, then load
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));
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);
}
// 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.
// 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 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());
+
+ const Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
+ std::pair<SDValue,SDValue> CallInfo =
+ TLI.LowerCallTo(DAG.getEntryNode(), RetTy, isSigned, !isSigned, false,
+ false, 0, TLI.getLibcallCallingConv(LC), false,
+ /*isReturnValueUsed=*/true,
+ Callee, Args, DAG, dl);
+
// Legalize the call sequence, starting with the chain. This will advance
// the LastCALLSEQ_END to the legalized version of the CALLSEQ_END node that
// was added by LowerCallTo (guaranteeing proper serialization of calls).
LegalizeOp(CallInfo.second);
+
return CallInfo.first;
}
SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
SDNode *Node,
bool isSigned) {
- assert(!IsLegalizingCall && "Cannot overlap legalization of calls!");
SDValue InChain = Node->getOperand(0);
TargetLowering::ArgListTy Args;
Callee, Args, DAG, Node->getDebugLoc());
// Legalize the call sequence, starting with the chain. This will advance
- // the LastCALLSEQ_END to the legalized version of the CALLSEQ_END node that
+ // the LastCALLSEQ to the legalized version of the CALLSEQ_END node that
// was added by LowerCallTo (guaranteeing proper serialization of calls).
LegalizeOp(CallInfo.second);
return CallInfo;
return ExpandLibCall(LC, Node, isSigned);
}
+/// isDivRemLibcallAvailable - Return true if divmod libcall is available.
+static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
+ const TargetLowering &TLI) {
+ RTLIB::Libcall LC;
+ switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
+ default: assert(0 && "Unexpected request for libcall!");
+ case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
+ case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
+ case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
+ case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
+ case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
+ }
+
+ return TLI.getLibcallName(LC) != 0;
+}
+
+/// UseDivRem - Only issue divrem libcall if both quotient and remainder are
+/// needed.
+static bool UseDivRem(SDNode *Node, bool isSigned, bool isDIV) {
+ unsigned OtherOpcode = 0;
+ if (isSigned)
+ OtherOpcode = isDIV ? ISD::SREM : ISD::SDIV;
+ else
+ OtherOpcode = isDIV ? ISD::UREM : ISD::UDIV;
+
+ SDValue Op0 = Node->getOperand(0);
+ SDValue Op1 = Node->getOperand(1);
+ for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
+ UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
+ SDNode *User = *UI;
+ if (User == Node)
+ continue;
+ if (User->getOpcode() == OtherOpcode &&
+ User->getOperand(0) == Op0 &&
+ User->getOperand(1) == Op1)
+ return true;
+ }
+ return false;
+}
+
+/// ExpandDivRemLibCall - Issue libcalls to __{u}divmod to compute div / rem
+/// pairs.
+void
+SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
+ SmallVectorImpl<SDValue> &Results) {
+ unsigned Opcode = Node->getOpcode();
+ bool isSigned = Opcode == ISD::SDIVREM;
+
+ RTLIB::Libcall LC;
+ switch (Node->getValueType(0).getSimpleVT().SimpleTy) {
+ default: assert(0 && "Unexpected request for libcall!");
+ case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
+ case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
+ case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
+ case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
+ case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
+ }
+
+ // The input chain to this libcall is the entry node of the function.
+ // Legalizing the call will automatically add the previous call to the
+ // dependence.
+ SDValue InChain = DAG.getEntryNode();
+
+ EVT RetVT = Node->getValueType(0);
+ const 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();
+ const Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
+ Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
+ Entry.isSExt = isSigned;
+ Entry.isZExt = !isSigned;
+ Args.push_back(Entry);
+ }
+
+ // Also pass the return address of the remainder.
+ SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
+ Entry.Node = FIPtr;
+ Entry.Ty = RetTy->getPointerTo();
+ Entry.isSExt = isSigned;
+ Entry.isZExt = !isSigned;
+ Args.push_back(Entry);
+
+ SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
+ TLI.getPointerTy());
+
+ // Splice the libcall in wherever FindInputOutputChains tells us to.
+ DebugLoc dl = Node->getDebugLoc();
+ std::pair<SDValue, SDValue> CallInfo =
+ TLI.LowerCallTo(InChain, RetTy, isSigned, !isSigned, false, false,
+ 0, TLI.getLibcallCallingConv(LC), /*isTailCall=*/false,
+ /*isReturnValueUsed=*/true, Callee, Args, DAG, dl);
+
+ // Legalize the call sequence, starting with the chain. This will advance
+ // the LastCALLSEQ to the legalized version of the CALLSEQ_END node that
+ // was added by LowerCallTo (guaranteeing proper serialization of calls).
+ LegalizeOp(CallInfo.second);
+
+ // Remainder is loaded back from the stack frame.
+ SDValue Rem = DAG.getLoad(RetVT, dl, getLastCALLSEQ(), FIPtr,
+ MachinePointerInfo(), false, false, 0);
+ Results.push_back(CallInfo.first);
+ Results.push_back(Rem);
+}
+
/// ExpandLegalINT_TO_FP - This function is responsible for legalizing a
/// INT_TO_FP operation of the specified operand when the target requests that
/// we expand it. At this point, we know that the result and operand types are
if (!isSigned) {
SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
- SDValue ShiftConst = DAG.getConstant(1, TLI.getShiftAmountTy());
+ SDValue ShiftConst =
+ DAG.getConstant(1, TLI.getShiftAmountTy(Op0.getValueType()));
SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
SDValue AndConst = DAG.getConstant(1, MVT::i64);
SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
}
// Otherwise, implement the fully general conversion.
- EVT SHVT = TLI.getShiftAmountTy();
SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
DAG.getConstant(UINT64_C(0xfffffffffffff800), MVT::i64));
Op0, DAG.getConstant(UINT64_C(0x0020000000000000), MVT::i64),
ISD::SETUGE);
SDValue Sel2 = DAG.getNode(ISD::SELECT, dl, MVT::i64, Ge, Sel, Op0);
+ EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType());
SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
DAG.getConstant(32, SHVT));
false, false, Alignment);
else {
FudgeInReg =
- LegalizeOp(DAG.getExtLoad(ISD::EXTLOAD, DestVT, dl,
+ LegalizeOp(DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
DAG.getEntryNode(), CPIdx,
MachinePointerInfo::getConstantPool(),
MVT::f32, false, false, Alignment));
///
SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, DebugLoc dl) {
EVT VT = Op.getValueType();
- EVT SHVT = TLI.getShiftAmountTy();
+ EVT SHVT = TLI.getShiftAmountTy(VT);
SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
switch (VT.getSimpleVT().SimpleTy) {
default: assert(0 && "Unhandled Expand type in BSWAP!");
}
}
+/// 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,
switch (Opc) {
default: assert(0 && "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: {
//
// 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);
// 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:
// SAR. However, it is doubtful that any exist.
EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
EVT VT = Node->getValueType(0);
- EVT ShiftAmountTy = TLI.getShiftAmountTy();
+ EVT ShiftAmountTy = TLI.getShiftAmountTy(VT);
if (VT.isVector())
ShiftAmountTy = VT;
unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
}
case ISD::FP_ROUND_INREG: {
// The only way we can lower this is to turn it into a TRUNCSTORE,
- // EXTLOAD pair, targetting a temporary location (a stack slot).
+ // EXTLOAD pair, targeting a temporary location (a stack slot).
// NOTE: there is a choice here between constantly creating new stack
// slots and always reusing the same one. We currently always create
case ISD::EXTRACT_SUBVECTOR:
Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
break;
+ case ISD::INSERT_SUBVECTOR:
+ Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
+ break;
case ISD::CONCAT_VECTORS: {
Results.push_back(ExpandVectorBuildThroughStack(Node));
break;
// 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
unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
Tmp2 = Node->getOperand(0);
Tmp3 = Node->getOperand(1);
- if (TLI.isOperationLegalOrCustom(DivRemOpc, VT)) {
+ if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
+ (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
+ UseDivRem(Node, isSigned, false))) {
Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
} else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
// X % Y -> X-X/Y*Y
Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
- } else if (isSigned) {
+ } else if (isSigned)
Tmp1 = ExpandIntLibCall(Node, true,
RTLIB::SREM_I8,
RTLIB::SREM_I16, RTLIB::SREM_I32,
RTLIB::SREM_I64, RTLIB::SREM_I128);
- } else {
+ else
Tmp1 = ExpandIntLibCall(Node, false,
RTLIB::UREM_I8,
RTLIB::UREM_I16, RTLIB::UREM_I32,
RTLIB::UREM_I64, RTLIB::UREM_I128);
- }
Results.push_back(Tmp1);
break;
}
unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
EVT VT = Node->getValueType(0);
SDVTList VTs = DAG.getVTList(VT, VT);
- if (TLI.isOperationLegalOrCustom(DivRemOpc, VT))
+ if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
+ (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
+ UseDivRem(Node, isSigned, true)))
Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
Node->getOperand(1));
else if (isSigned)
Results.push_back(Tmp1.getValue(1));
break;
}
+ case ISD::SDIVREM:
+ case ISD::UDIVREM:
+ // Expand into divrem libcall
+ ExpandDivRemLibCall(Node, Results);
+ break;
case ISD::MUL: {
EVT VT = Node->getValueType(0);
SDVTList VTs = DAG.getVTList(VT, VT);
case ISD::UMULO:
case ISD::SMULO: {
EVT VT = Node->getValueType(0);
+ EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
SDValue LHS = Node->getOperand(0);
SDValue RHS = Node->getOperand(1);
SDValue BottomHalf;
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));
}
+
if (isSigned) {
- Tmp1 = DAG.getConstant(VT.getSizeInBits() - 1, TLI.getShiftAmountTy());
+ Tmp1 = DAG.getConstant(VT.getSizeInBits() - 1,
+ TLI.getShiftAmountTy(BottomHalf.getValueType()));
Tmp1 = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, Tmp1);
TopHalf = DAG.getSetCC(dl, TLI.getSetCCResultType(VT), TopHalf, Tmp1,
ISD::SETNE);
Tmp2 = DAG.getNode(ISD::ANY_EXTEND, dl, PairTy, Node->getOperand(1));
Tmp2 = DAG.getNode(ISD::SHL, dl, PairTy, Tmp2,
DAG.getConstant(PairTy.getSizeInBits()/2,
- TLI.getShiftAmountTy()));
+ TLI.getShiftAmountTy(PairTy)));
Results.push_back(DAG.getNode(ISD::OR, dl, PairTy, Tmp1, Tmp2));
break;
}
SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
- SDValue LD = DAG.getExtLoad(ISD::SEXTLOAD, PTy, dl, Chain, Addr,
+ SDValue LD = DAG.getExtLoad(ISD::SEXTLOAD, dl, PTy, Chain, Addr,
MachinePointerInfo::getJumpTable(), MemVT,
false, false, 0);
Addr = LD;
Tmp2.getOperand(0), Tmp2.getOperand(1),
Node->getOperand(2));
} else {
+ // We test only the i1 bit. Skip the AND if UNDEF.
+ Tmp3 = (Tmp2.getOpcode() == ISD::UNDEF) ? Tmp2 :
+ DAG.getNode(ISD::AND, dl, Tmp2.getValueType(), Tmp2,
+ DAG.getConstant(1, Tmp2.getValueType()));
Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, Tmp1,
- DAG.getCondCode(ISD::SETNE), Tmp2,
- DAG.getConstant(0, Tmp2.getValueType()),
+ DAG.getCondCode(ISD::SETNE), Tmp3,
+ DAG.getConstant(0, Tmp3.getValueType()),
Node->getOperand(2));
}
Results.push_back(Tmp1);
LegalizeSetCCCondCode(TLI.getSetCCResultType(Tmp2.getValueType()),
Tmp2, Tmp3, Tmp4, dl);
- LastCALLSEQ_END = DAG.getEntryNode();
+ assert(LastCALLSEQ.size() == 1 && "branch inside CALLSEQ_BEGIN/END?");
+ setLastCALLSEQ(DAG.getEntryNode());
assert(!Tmp3.getNode() && "Can't legalize BR_CC with legal condition!");
Tmp3 = DAG.getConstant(0, Tmp2.getValueType());
Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
Tmp1 = DAG.getNode(ISD::BSWAP, dl, NVT, Tmp1);
Tmp1 = DAG.getNode(ISD::SRL, dl, NVT, Tmp1,
- DAG.getConstant(DiffBits, TLI.getShiftAmountTy()));
+ DAG.getConstant(DiffBits, TLI.getShiftAmountTy(NVT)));
Results.push_back(Tmp1);
break;
}
// SelectionDAG::Legalize - This is the entry point for the file.
//
-void SelectionDAG::Legalize(CodeGenOpt::Level OptLevel) {
+void SelectionDAG::Legalize() {
/// run - This is the main entry point to this class.
///
- SelectionDAGLegalize(*this, OptLevel).LegalizeDAG();
+ SelectionDAGLegalize(*this).LegalizeDAG();
}
-