//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "x86-isel"
#include "X86.h"
#include "X86InstrBuilder.h"
#include "X86MachineFunctionInfo.h"
#include "llvm/Target/TargetOptions.h"
using namespace llvm;
+#define DEBUG_TYPE "x86-isel"
+
STATISTIC(NumLoadMoved, "Number of loads moved below TokenFactor");
//===----------------------------------------------------------------------===//
X86ISelAddressMode()
: BaseType(RegBase), Base_FrameIndex(0), Scale(1), IndexReg(), Disp(0),
- Segment(), GV(0), CP(0), BlockAddr(0), ES(0), JT(-1), Align(0),
- SymbolFlags(X86II::MO_NO_FLAG) {
+ Segment(), GV(nullptr), CP(nullptr), BlockAddr(nullptr), ES(nullptr),
+ JT(-1), Align(0), SymbolFlags(X86II::MO_NO_FLAG) {
}
bool hasSymbolicDisplacement() const {
- return GV != 0 || CP != 0 || ES != 0 || JT != -1 || BlockAddr != 0;
+ return GV != nullptr || CP != nullptr || ES != nullptr ||
+ JT != -1 || BlockAddr != nullptr;
}
bool hasBaseOrIndexReg() const {
- return IndexReg.getNode() != 0 || Base_Reg.getNode() != 0;
+ return BaseType == FrameIndexBase ||
+ IndexReg.getNode() != nullptr || Base_Reg.getNode() != nullptr;
}
/// isRIPRelative - Return true if this addressing mode is already RIP
void dump() {
dbgs() << "X86ISelAddressMode " << this << '\n';
dbgs() << "Base_Reg ";
- if (Base_Reg.getNode() != 0)
+ if (Base_Reg.getNode())
Base_Reg.getNode()->dump();
else
dbgs() << "nul";
dbgs() << " Base.FrameIndex " << Base_FrameIndex << '\n'
<< " Scale" << Scale << '\n'
<< "IndexReg ";
- if (IndexReg.getNode() != 0)
+ if (IndexReg.getNode())
IndexReg.getNode()->dump();
else
dbgs() << "nul";
/// ISel - X86 specific code to select X86 machine instructions for
/// SelectionDAG operations.
///
- class X86DAGToDAGISel : public SelectionDAGISel {
- /// X86Lowering - This object fully describes how to lower LLVM code to an
- /// X86-specific SelectionDAG.
- const X86TargetLowering &X86Lowering;
-
+ class X86DAGToDAGISel final : public SelectionDAGISel {
/// Subtarget - Keep a pointer to the X86Subtarget around so that we can
/// make the right decision when generating code for different targets.
const X86Subtarget *Subtarget;
public:
explicit X86DAGToDAGISel(X86TargetMachine &tm, CodeGenOpt::Level OptLevel)
: SelectionDAGISel(tm, OptLevel),
- X86Lowering(*tm.getTargetLowering()),
Subtarget(&tm.getSubtarget<X86Subtarget>()),
OptForSize(false) {}
- virtual const char *getPassName() const {
+ const char *getPassName() const override {
return "X86 DAG->DAG Instruction Selection";
}
- virtual void EmitFunctionEntryCode();
+ bool runOnMachineFunction(MachineFunction &MF) override {
+ // Reset the subtarget each time through.
+ Subtarget = &TM.getSubtarget<X86Subtarget>();
+ SelectionDAGISel::runOnMachineFunction(MF);
+ return true;
+ }
- virtual bool IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const;
+ void EmitFunctionEntryCode() override;
- virtual void PreprocessISelDAG();
+ bool IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const override;
+
+ void PreprocessISelDAG() override;
inline bool immSext8(SDNode *N) const {
return isInt<8>(cast<ConstantSDNode>(N)->getSExtValue());
#include "X86GenDAGISel.inc"
private:
- SDNode *Select(SDNode *N);
+ SDNode *Select(SDNode *N) override;
SDNode *SelectGather(SDNode *N, unsigned Opc);
SDNode *SelectAtomic64(SDNode *Node, unsigned Opc);
- SDNode *SelectAtomicLoadArith(SDNode *Node, EVT NVT);
+ SDNode *SelectAtomicLoadArith(SDNode *Node, MVT NVT);
bool FoldOffsetIntoAddress(uint64_t Offset, X86ISelAddressMode &AM);
bool MatchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM);
bool SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index, SDValue &Disp,
SDValue &Segment);
+ bool SelectMOV64Imm32(SDValue N, SDValue &Imm);
bool SelectLEAAddr(SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index, SDValue &Disp,
SDValue &Segment);
+ bool SelectLEA64_32Addr(SDValue N, SDValue &Base,
+ SDValue &Scale, SDValue &Index, SDValue &Disp,
+ SDValue &Segment);
bool SelectTLSADDRAddr(SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index, SDValue &Disp,
SDValue &Segment);
/// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
/// inline asm expressions.
- virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
- char ConstraintCode,
- std::vector<SDValue> &OutOps);
+ bool SelectInlineAsmMemoryOperand(const SDValue &Op,
+ char ConstraintCode,
+ std::vector<SDValue> &OutOps) override;
void EmitSpecialCodeForMain(MachineBasicBlock *BB, MachineFrameInfo *MFI);
SDValue &Scale, SDValue &Index,
SDValue &Disp, SDValue &Segment) {
Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
- CurDAG->getTargetFrameIndex(AM.Base_FrameIndex, TLI.getPointerTy()) :
+ CurDAG->getTargetFrameIndex(AM.Base_FrameIndex,
+ getTargetLowering()->getPointerTy()) :
AM.Base_Reg;
Scale = getI8Imm(AM.Scale);
Index = AM.IndexReg;
// These are 32-bit even in 64-bit mode since RIP relative offset
// is 32-bit.
if (AM.GV)
- Disp = CurDAG->getTargetGlobalAddress(AM.GV, DebugLoc(),
+ Disp = CurDAG->getTargetGlobalAddress(AM.GV, SDLoc(),
MVT::i32, AM.Disp,
AM.SymbolFlags);
else if (AM.CP)
/// getTargetMachine - Return a reference to the TargetMachine, casted
/// to the target-specific type.
- const X86TargetMachine &getTargetMachine() {
+ const X86TargetMachine &getTargetMachine() const {
return static_cast<const X86TargetMachine &>(TM);
}
/// getInstrInfo - Return a reference to the TargetInstrInfo, casted
/// to the target-specific type.
- const X86InstrInfo *getInstrInfo() {
+ const X86InstrInfo *getInstrInfo() const {
return getTargetMachine().getInstrInfo();
}
};
// addl %gs:0, %eax
// if the block also has an access to a second TLS address this will save
// a load.
- // FIXME: This is probably also true for non TLS addresses.
+ // FIXME: This is probably also true for non-TLS addresses.
if (Op1.getOpcode() == X86ISD::Wrapper) {
SDValue Val = Op1.getOperand(0);
if (Val.getOpcode() == ISD::TargetGlobalTLSAddress)
else
Ops.push_back(Chain.getOperand(i));
SDValue NewChain =
- CurDAG->getNode(ISD::TokenFactor, Load.getDebugLoc(),
- MVT::Other, &Ops[0], Ops.size());
+ CurDAG->getNode(ISD::TokenFactor, SDLoc(Load), MVT::Other, Ops);
Ops.clear();
Ops.push_back(NewChain);
}
for (unsigned i = 1, e = OrigChain.getNumOperands(); i != e; ++i)
Ops.push_back(OrigChain.getOperand(i));
- CurDAG->UpdateNodeOperands(OrigChain.getNode(), &Ops[0], Ops.size());
+ CurDAG->UpdateNodeOperands(OrigChain.getNode(), Ops);
CurDAG->UpdateNodeOperands(Load.getNode(), Call.getOperand(0),
Load.getOperand(1), Load.getOperand(2));
Ops.push_back(SDValue(Load.getNode(), 1));
for (unsigned i = 1, e = NumOps; i != e; ++i)
Ops.push_back(Call.getOperand(i));
- CurDAG->UpdateNodeOperands(Call.getNode(), &Ops[0], NumOps);
+ CurDAG->UpdateNodeOperands(Call.getNode(), Ops);
}
/// isCalleeLoad - Return true if call address is a load and it can be
SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
if (OptLevel != CodeGenOpt::None &&
- (N->getOpcode() == X86ISD::CALL ||
+ // Only does this when target favors doesn't favor register indirect
+ // call.
+ ((N->getOpcode() == X86ISD::CALL && !Subtarget->callRegIndirect()) ||
(N->getOpcode() == X86ISD::TC_RETURN &&
// Only does this if load can be folded into TC_RETURN.
(Subtarget->is64Bit() ||
if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
continue;
- EVT SrcVT = N->getOperand(0).getValueType();
- EVT DstVT = N->getValueType(0);
+ MVT SrcVT = N->getOperand(0).getSimpleValueType();
+ MVT DstVT = N->getSimpleValueType(0);
// If any of the sources are vectors, no fp stack involved.
if (SrcVT.isVector() || DstVT.isVector())
// If the source and destination are SSE registers, then this is a legal
// conversion that should not be lowered.
- bool SrcIsSSE = X86Lowering.isScalarFPTypeInSSEReg(SrcVT);
- bool DstIsSSE = X86Lowering.isScalarFPTypeInSSEReg(DstVT);
+ const X86TargetLowering *X86Lowering =
+ static_cast<const X86TargetLowering *>(getTargetLowering());
+ bool SrcIsSSE = X86Lowering->isScalarFPTypeInSSEReg(SrcVT);
+ bool DstIsSSE = X86Lowering->isScalarFPTypeInSSEReg(DstVT);
if (SrcIsSSE && DstIsSSE)
continue;
// Here we could have an FP stack truncation or an FPStack <-> SSE convert.
// FPStack has extload and truncstore. SSE can fold direct loads into other
// operations. Based on this, decide what we want to do.
- EVT MemVT;
+ MVT MemVT;
if (N->getOpcode() == ISD::FP_ROUND)
MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
else
MemVT = SrcIsSSE ? SrcVT : DstVT;
SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
- DebugLoc dl = N->getDebugLoc();
+ SDLoc dl(N);
// FIXME: optimize the case where the src/dest is a load or store?
SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
// gs:0 (or fs:0 on X86-64) contains its own address.
// For more information see http://people.redhat.com/drepper/tls.pdf
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Address))
- if (C->getSExtValue() == 0 && AM.Segment.getNode() == 0 &&
+ if (C->getSExtValue() == 0 && AM.Segment.getNode() == nullptr &&
Subtarget->isTargetLinux())
switch (N->getPointerInfo().getAddrSpace()) {
case 256:
// a smaller encoding and avoids a scaled-index.
if (AM.Scale == 2 &&
AM.BaseType == X86ISelAddressMode::RegBase &&
- AM.Base_Reg.getNode() == 0) {
+ AM.Base_Reg.getNode() == nullptr) {
AM.Base_Reg = AM.IndexReg;
AM.Scale = 1;
}
Subtarget->is64Bit() &&
AM.Scale == 1 &&
AM.BaseType == X86ISelAddressMode::RegBase &&
- AM.Base_Reg.getNode() == 0 &&
- AM.IndexReg.getNode() == 0 &&
+ AM.Base_Reg.getNode() == nullptr &&
+ AM.IndexReg.getNode() == nullptr &&
AM.SymbolFlags == X86II::MO_NO_FLAG &&
AM.hasSymbolicDisplacement())
AM.Base_Reg = CurDAG->getRegister(X86::RIP, MVT::i64);
Mask != (0xffu << ScaleLog))
return true;
- EVT VT = N.getValueType();
- DebugLoc DL = N.getDebugLoc();
+ MVT VT = N.getSimpleValueType();
+ SDLoc DL(N);
SDValue Eight = DAG.getConstant(8, MVT::i8);
SDValue NewMask = DAG.getConstant(0xff, VT);
SDValue Srl = DAG.getNode(ISD::SRL, DL, VT, X, Eight);
if (ShiftAmt != 1 && ShiftAmt != 2 && ShiftAmt != 3)
return true;
- EVT VT = N.getValueType();
- DebugLoc DL = N.getDebugLoc();
+ MVT VT = N.getSimpleValueType();
+ SDLoc DL(N);
SDValue NewMask = DAG.getConstant(Mask >> ShiftAmt, VT);
SDValue NewAnd = DAG.getNode(ISD::AND, DL, VT, X, NewMask);
SDValue NewShift = DAG.getNode(ISD::SHL, DL, VT, NewAnd, Shift.getOperand(1));
return true;
unsigned ShiftAmt = Shift.getConstantOperandVal(1);
- unsigned MaskLZ = CountLeadingZeros_64(Mask);
- unsigned MaskTZ = CountTrailingZeros_64(Mask);
+ unsigned MaskLZ = countLeadingZeros(Mask);
+ unsigned MaskTZ = countTrailingZeros(Mask);
// The amount of shift we're trying to fit into the addressing mode is taken
// from the trailing zeros of the mask.
// Scale the leading zero count down based on the actual size of the value.
// Also scale it down based on the size of the shift.
- MaskLZ -= (64 - X.getValueSizeInBits()) + ShiftAmt;
+ MaskLZ -= (64 - X.getSimpleValueType().getSizeInBits()) + ShiftAmt;
// The final check is to ensure that any masked out high bits of X are
// already known to be zero. Otherwise, the mask has a semantic impact
// replace them with zero extensions cheaply if necessary.
bool ReplacingAnyExtend = false;
if (X.getOpcode() == ISD::ANY_EXTEND) {
- unsigned ExtendBits =
- X.getValueSizeInBits() - X.getOperand(0).getValueSizeInBits();
+ unsigned ExtendBits = X.getSimpleValueType().getSizeInBits() -
+ X.getOperand(0).getSimpleValueType().getSizeInBits();
// Assume that we'll replace the any-extend with a zero-extend, and
// narrow the search to the extended value.
X = X.getOperand(0);
MaskLZ = ExtendBits > MaskLZ ? 0 : MaskLZ - ExtendBits;
ReplacingAnyExtend = true;
}
- APInt MaskedHighBits = APInt::getHighBitsSet(X.getValueSizeInBits(),
- MaskLZ);
+ APInt MaskedHighBits =
+ APInt::getHighBitsSet(X.getSimpleValueType().getSizeInBits(), MaskLZ);
APInt KnownZero, KnownOne;
- DAG.ComputeMaskedBits(X, KnownZero, KnownOne);
+ DAG.computeKnownBits(X, KnownZero, KnownOne);
if (MaskedHighBits != KnownZero) return true;
// We've identified a pattern that can be transformed into a single shift
// and an addressing mode. Make it so.
- EVT VT = N.getValueType();
+ MVT VT = N.getSimpleValueType();
if (ReplacingAnyExtend) {
assert(X.getValueType() != VT);
// We looked through an ANY_EXTEND node, insert a ZERO_EXTEND.
- SDValue NewX = DAG.getNode(ISD::ZERO_EXTEND, X.getDebugLoc(), VT, X);
+ SDValue NewX = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(X), VT, X);
InsertDAGNode(DAG, N, NewX);
X = NewX;
}
- DebugLoc DL = N.getDebugLoc();
+ SDLoc DL(N);
SDValue NewSRLAmt = DAG.getConstant(ShiftAmt + AMShiftAmt, MVT::i8);
SDValue NewSRL = DAG.getNode(ISD::SRL, DL, VT, X, NewSRLAmt);
SDValue NewSHLAmt = DAG.getConstant(AMShiftAmt, MVT::i8);
bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
unsigned Depth) {
- DebugLoc dl = N.getDebugLoc();
+ SDLoc dl(N);
DEBUG({
dbgs() << "MatchAddress: ";
AM.dump();
case ISD::FrameIndex:
if (AM.BaseType == X86ISelAddressMode::RegBase &&
- AM.Base_Reg.getNode() == 0 &&
+ AM.Base_Reg.getNode() == nullptr &&
(!Subtarget->is64Bit() || isDispSafeForFrameIndex(AM.Disp))) {
AM.BaseType = X86ISelAddressMode::FrameIndexBase;
AM.Base_FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
break;
case ISD::SHL:
- if (AM.IndexReg.getNode() != 0 || AM.Scale != 1)
+ if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1)
break;
if (ConstantSDNode
case ISD::SRL: {
// Scale must not be used already.
- if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
+ if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1) break;
SDValue And = N.getOperand(0);
if (And.getOpcode() != ISD::AND) break;
// We only handle up to 64-bit values here as those are what matter for
// addressing mode optimizations.
- if (X.getValueSizeInBits() > 64) break;
+ if (X.getSimpleValueType().getSizeInBits() > 64) break;
// The mask used for the transform is expected to be post-shift, but we
// found the shift first so just apply the shift to the mask before passing
case X86ISD::MUL_IMM:
// X*[3,5,9] -> X+X*[2,4,8]
if (AM.BaseType == X86ISelAddressMode::RegBase &&
- AM.Base_Reg.getNode() == 0 &&
- AM.IndexReg.getNode() == 0) {
+ AM.Base_Reg.getNode() == nullptr &&
+ AM.IndexReg.getNode() == nullptr) {
if (ConstantSDNode
*CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
// with a constant to enable use of the scaled offset field.
// Scale must not be used already.
- if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
+ if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1) break;
SDValue Shift = N.getOperand(0);
if (Shift.getOpcode() != ISD::SRL && Shift.getOpcode() != ISD::SHL) break;
// We only handle up to 64-bit values here as those are what matter for
// addressing mode optimizations.
- if (X.getValueSizeInBits() > 64) break;
+ if (X.getSimpleValueType().getSizeInBits() > 64) break;
if (!isa<ConstantSDNode>(N.getOperand(1)))
break;
// Is the base register already occupied?
if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base_Reg.getNode()) {
// If so, check to see if the scale index register is set.
- if (AM.IndexReg.getNode() == 0) {
+ if (!AM.IndexReg.getNode()) {
AM.IndexReg = N;
AM.Scale = 1;
return false;
if (MatchAddress(N, AM))
return false;
- EVT VT = N.getValueType();
+ MVT VT = N.getSimpleValueType();
if (AM.BaseType == X86ISelAddressMode::RegBase) {
if (!AM.Base_Reg.getNode())
AM.Base_Reg = CurDAG->getRegister(0, VT);
}
+bool X86DAGToDAGISel::SelectMOV64Imm32(SDValue N, SDValue &Imm) {
+ if (const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
+ uint64_t ImmVal = CN->getZExtValue();
+ if ((uint32_t)ImmVal != (uint64_t)ImmVal)
+ return false;
+
+ Imm = CurDAG->getTargetConstant(ImmVal, MVT::i64);
+ return true;
+ }
+
+ // In static codegen with small code model, we can get the address of a label
+ // into a register with 'movl'. TableGen has already made sure we're looking
+ // at a label of some kind.
+ assert(N->getOpcode() == X86ISD::Wrapper &&
+ "Unexpected node type for MOV32ri64");
+ N = N.getOperand(0);
+
+ if (N->getOpcode() != ISD::TargetConstantPool &&
+ N->getOpcode() != ISD::TargetJumpTable &&
+ N->getOpcode() != ISD::TargetGlobalAddress &&
+ N->getOpcode() != ISD::TargetExternalSymbol &&
+ N->getOpcode() != ISD::TargetBlockAddress)
+ return false;
+
+ Imm = N;
+ return TM.getCodeModel() == CodeModel::Small;
+}
+
+bool X86DAGToDAGISel::SelectLEA64_32Addr(SDValue N, SDValue &Base,
+ SDValue &Scale, SDValue &Index,
+ SDValue &Disp, SDValue &Segment) {
+ if (!SelectLEAAddr(N, Base, Scale, Index, Disp, Segment))
+ return false;
+
+ SDLoc DL(N);
+ RegisterSDNode *RN = dyn_cast<RegisterSDNode>(Base);
+ if (RN && RN->getReg() == 0)
+ Base = CurDAG->getRegister(0, MVT::i64);
+ else if (Base.getValueType() == MVT::i32 && !dyn_cast<FrameIndexSDNode>(N)) {
+ // Base could already be %rip, particularly in the x32 ABI.
+ Base = SDValue(CurDAG->getMachineNode(
+ TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
+ CurDAG->getTargetConstant(0, MVT::i64),
+ Base,
+ CurDAG->getTargetConstant(X86::sub_32bit, MVT::i32)),
+ 0);
+ }
+
+ RN = dyn_cast<RegisterSDNode>(Index);
+ if (RN && RN->getReg() == 0)
+ Index = CurDAG->getRegister(0, MVT::i64);
+ else {
+ assert(Index.getValueType() == MVT::i32 &&
+ "Expect to be extending 32-bit registers for use in LEA");
+ Index = SDValue(CurDAG->getMachineNode(
+ TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
+ CurDAG->getTargetConstant(0, MVT::i64),
+ Index,
+ CurDAG->getTargetConstant(X86::sub_32bit, MVT::i32)),
+ 0);
+ }
+
+ return true;
+}
+
/// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
/// mode it matches can be cost effectively emitted as an LEA instruction.
bool X86DAGToDAGISel::SelectLEAAddr(SDValue N,
assert (T == AM.Segment);
AM.Segment = Copy;
- EVT VT = N.getValueType();
+ MVT VT = N.getSimpleValueType();
unsigned Complexity = 0;
if (AM.BaseType == X86ISelAddressMode::RegBase)
if (AM.Base_Reg.getNode())
///
SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
- return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).getNode();
+ return CurDAG->getRegister(GlobalBaseReg,
+ getTargetLowering()->getPointerTy()).getNode();
}
SDNode *X86DAGToDAGISel::SelectAtomic64(SDNode *Node, unsigned Opc) {
SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
if (!SelectAddr(Node, In1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
- return NULL;
+ return nullptr;
MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
const SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, In2L, In2H, Chain};
- SDNode *ResNode = CurDAG->getMachineNode(Opc, Node->getDebugLoc(),
- MVT::i32, MVT::i32, MVT::Other, Ops,
- array_lengthof(Ops));
+ SDNode *ResNode = CurDAG->getMachineNode(Opc, SDLoc(Node),
+ MVT::i32, MVT::i32, MVT::Other, Ops);
cast<MachineSDNode>(ResNode)->setMemRefs(MemOp, MemOp + 1);
return ResNode;
}
// + empty, the operand is not needed any more with the new op selected.
// + non-empty, otherwise.
static SDValue getAtomicLoadArithTargetConstant(SelectionDAG *CurDAG,
- DebugLoc dl,
- enum AtomicOpc &Op, EVT NVT,
+ SDLoc dl,
+ enum AtomicOpc &Op, MVT NVT,
SDValue Val) {
if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val)) {
int64_t CNVal = CN->getSExtValue();
return Val;
}
-SDNode *X86DAGToDAGISel::SelectAtomicLoadArith(SDNode *Node, EVT NVT) {
+SDNode *X86DAGToDAGISel::SelectAtomicLoadArith(SDNode *Node, MVT NVT) {
if (Node->hasAnyUseOfValue(0))
- return 0;
+ return nullptr;
- DebugLoc dl = Node->getDebugLoc();
+ SDLoc dl(Node);
// Optimize common patterns for __sync_or_and_fetch and similar arith
// operations where the result is not used. This allows us to use the "lock"
SDValue Val = Node->getOperand(2);
SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
if (!SelectAddr(Node, Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
- return 0;
+ return nullptr;
// Which index into the table.
enum AtomicOpc Op;
switch (Node->getOpcode()) {
default:
- return 0;
+ return nullptr;
case ISD::ATOMIC_LOAD_OR:
Op = OR;
break;
Op = ADD;
break;
}
-
+
Val = getAtomicLoadArithTargetConstant(CurDAG, dl, Op, NVT, Val);
bool isUnOp = !Val.getNode();
bool isCN = Val.getNode() && (Val.getOpcode() == ISD::TargetConstant);
unsigned Opc = 0;
- switch (NVT.getSimpleVT().SimpleTy) {
- default: return 0;
+ switch (NVT.SimpleTy) {
+ default: return nullptr;
case MVT::i8:
if (isCN)
Opc = AtomicOpcTbl[Op][ConstantI8];
MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
if (isUnOp) {
SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Chain };
- Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops,
- array_lengthof(Ops)), 0);
+ Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
} else {
SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Val, Chain };
- Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops,
- array_lengthof(Ops)), 0);
+ Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
}
cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
SDValue RetVals[] = { Undef, Ret };
- return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
+ return CurDAG->getMergeValues(RetVals, dl).getNode();
}
/// HasNoSignedComparisonUses - Test whether the given X86ISD::CMP node has
if (ChainCheck)
// Make a new TokenFactor with all the other input chains except
// for the load.
- InputChain = CurDAG->getNode(ISD::TokenFactor, Chain.getDebugLoc(),
- MVT::Other, &ChainOps[0], ChainOps.size());
+ InputChain = CurDAG->getNode(ISD::TokenFactor, SDLoc(Chain),
+ MVT::Other, ChainOps);
}
if (!ChainCheck)
return false;
SDValue VMask = Node->getOperand(5);
ConstantSDNode *Scale = dyn_cast<ConstantSDNode>(Node->getOperand(6));
if (!Scale)
- return 0;
+ return nullptr;
SDVTList VTs = CurDAG->getVTList(VSrc.getValueType(), VSrc.getValueType(),
MVT::Other);
SDValue Segment = CurDAG->getRegister(0, MVT::i32);
const SDValue Ops[] = { VSrc, Base, getI8Imm(Scale->getSExtValue()), VIdx,
Disp, Segment, VMask, Chain};
- SDNode *ResNode = CurDAG->getMachineNode(Opc, Node->getDebugLoc(),
- VTs, Ops, array_lengthof(Ops));
+ SDNode *ResNode = CurDAG->getMachineNode(Opc, SDLoc(Node), VTs, Ops);
// Node has 2 outputs: VDst and MVT::Other.
// ResNode has 3 outputs: VDst, VMask_wb, and MVT::Other.
// We replace VDst of Node with VDst of ResNode, and Other of Node with Other
}
SDNode *X86DAGToDAGISel::Select(SDNode *Node) {
- EVT NVT = Node->getValueType(0);
+ MVT NVT = Node->getSimpleValueType(0);
unsigned Opc, MOpc;
unsigned Opcode = Node->getOpcode();
- DebugLoc dl = Node->getDebugLoc();
+ SDLoc dl(Node);
DEBUG(dbgs() << "Selecting: "; Node->dump(CurDAG); dbgs() << '\n');
if (Node->isMachineOpcode()) {
DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << '\n');
- return NULL; // Already selected.
+ Node->setNodeId(-1);
+ return nullptr; // Already selected.
}
switch (Opcode) {
case Intrinsic::x86_avx2_gather_d_d_256:
case Intrinsic::x86_avx2_gather_q_d:
case Intrinsic::x86_avx2_gather_q_d_256: {
+ if (!Subtarget->hasAVX2())
+ break;
unsigned Opc;
switch (IntNo) {
default: llvm_unreachable("Impossible intrinsic");
SDNode *RetVal = SelectGather(Node, Opc);
if (RetVal)
// We already called ReplaceUses inside SelectGather.
- return NULL;
+ return nullptr;
break;
}
}
break;
unsigned ShlOp, Op;
- EVT CstVT = NVT;
+ MVT CstVT = NVT;
// Check the minimum bitwidth for the new constant.
// TODO: AND32ri is the same as AND64ri32 with zext imm.
if (NVT == CstVT)
break;
- switch (NVT.getSimpleVT().SimpleTy) {
+ switch (NVT.SimpleTy) {
default: llvm_unreachable("Unsupported VT!");
case MVT::i32:
assert(CstVT == MVT::i8);
SDValue N1 = Node->getOperand(1);
unsigned LoReg;
- switch (NVT.getSimpleVT().SimpleTy) {
+ switch (NVT.SimpleTy) {
default: llvm_unreachable("Unsupported VT!");
case MVT::i8: LoReg = X86::AL; Opc = X86::MUL8r; break;
case MVT::i16: LoReg = X86::AX; Opc = X86::MUL16r; break;
SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::i32);
SDValue Ops[] = {N1, InFlag};
- SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops, 2);
+ SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
ReplaceUses(SDValue(Node, 0), SDValue(CNode, 0));
ReplaceUses(SDValue(Node, 1), SDValue(CNode, 1));
ReplaceUses(SDValue(Node, 2), SDValue(CNode, 2));
- return NULL;
+ return nullptr;
}
case ISD::SMUL_LOHI:
bool isSigned = Opcode == ISD::SMUL_LOHI;
bool hasBMI2 = Subtarget->hasBMI2();
if (!isSigned) {
- switch (NVT.getSimpleVT().SimpleTy) {
+ switch (NVT.SimpleTy) {
default: llvm_unreachable("Unsupported VT!");
case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
MOpc = hasBMI2 ? X86::MULX64rm : X86::MUL64m; break;
}
} else {
- switch (NVT.getSimpleVT().SimpleTy) {
+ switch (NVT.SimpleTy) {
default: llvm_unreachable("Unsupported VT!");
case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
InFlag };
if (MOpc == X86::MULX32rm || MOpc == X86::MULX64rm) {
SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::Other, MVT::Glue);
- SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops,
- array_lengthof(Ops));
+ SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
ResHi = SDValue(CNode, 0);
ResLo = SDValue(CNode, 1);
Chain = SDValue(CNode, 2);
InFlag = SDValue(CNode, 3);
} else {
SDVTList VTs = CurDAG->getVTList(MVT::Other, MVT::Glue);
- SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops,
- array_lengthof(Ops));
+ SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
Chain = SDValue(CNode, 0);
InFlag = SDValue(CNode, 1);
}
SDValue Ops[] = { N1, InFlag };
if (Opc == X86::MULX32rr || Opc == X86::MULX64rr) {
SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::Glue);
- SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops,
- array_lengthof(Ops));
+ SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
ResHi = SDValue(CNode, 0);
ResLo = SDValue(CNode, 1);
InFlag = SDValue(CNode, 2);
} else {
SDVTList VTs = CurDAG->getVTList(MVT::Glue);
- SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops,
- array_lengthof(Ops));
+ SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
InFlag = SDValue(CNode, 0);
}
}
}
// Copy the low half of the result, if it is needed.
if (!SDValue(Node, 0).use_empty()) {
- if (ResLo.getNode() == 0) {
+ if (!ResLo.getNode()) {
assert(LoReg && "Register for low half is not defined!");
ResLo = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, LoReg, NVT,
InFlag);
}
// Copy the high half of the result, if it is needed.
if (!SDValue(Node, 1).use_empty()) {
- if (ResHi.getNode() == 0) {
+ if (!ResHi.getNode()) {
assert(HiReg && "Register for high half is not defined!");
ResHi = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, HiReg, NVT,
InFlag);
DEBUG(dbgs() << "=> "; ResHi.getNode()->dump(CurDAG); dbgs() << '\n');
}
- return NULL;
+ return nullptr;
}
case ISD::SDIVREM:
bool isSigned = Opcode == ISD::SDIVREM;
if (!isSigned) {
- switch (NVT.getSimpleVT().SimpleTy) {
+ switch (NVT.SimpleTy) {
default: llvm_unreachable("Unsupported VT!");
case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
}
} else {
- switch (NVT.getSimpleVT().SimpleTy) {
+ switch (NVT.SimpleTy) {
default: llvm_unreachable("Unsupported VT!");
case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
}
unsigned LoReg, HiReg, ClrReg;
- unsigned ClrOpcode, SExtOpcode;
- switch (NVT.getSimpleVT().SimpleTy) {
+ unsigned SExtOpcode;
+ switch (NVT.SimpleTy) {
default: llvm_unreachable("Unsupported VT!");
case MVT::i8:
LoReg = X86::AL; ClrReg = HiReg = X86::AH;
- ClrOpcode = 0;
SExtOpcode = X86::CBW;
break;
case MVT::i16:
LoReg = X86::AX; HiReg = X86::DX;
- ClrOpcode = X86::MOV16r0; ClrReg = X86::DX;
+ ClrReg = X86::DX;
SExtOpcode = X86::CWD;
break;
case MVT::i32:
LoReg = X86::EAX; ClrReg = HiReg = X86::EDX;
- ClrOpcode = X86::MOV32r0;
SExtOpcode = X86::CDQ;
break;
case MVT::i64:
LoReg = X86::RAX; ClrReg = HiReg = X86::RDX;
- ClrOpcode = X86::MOV64r0;
SExtOpcode = X86::CQO;
break;
}
SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
Move =
SDValue(CurDAG->getMachineNode(X86::MOVZX32rm8, dl, MVT::i32,
- MVT::Other, Ops,
- array_lengthof(Ops)), 0);
+ MVT::Other, Ops), 0);
Chain = Move.getValue(1);
ReplaceUses(N0.getValue(1), Chain);
} else {
SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Glue, InFlag),0);
} else {
// Zero out the high part, effectively zero extending the input.
- SDValue ClrNode =
- SDValue(CurDAG->getMachineNode(ClrOpcode, dl, NVT), 0);
+ SDValue ClrNode = SDValue(CurDAG->getMachineNode(X86::MOV32r0, dl, NVT), 0);
+ switch (NVT.SimpleTy) {
+ case MVT::i16:
+ ClrNode =
+ SDValue(CurDAG->getMachineNode(
+ TargetOpcode::EXTRACT_SUBREG, dl, MVT::i16, ClrNode,
+ CurDAG->getTargetConstant(X86::sub_16bit, MVT::i32)),
+ 0);
+ break;
+ case MVT::i32:
+ break;
+ case MVT::i64:
+ ClrNode =
+ SDValue(CurDAG->getMachineNode(
+ TargetOpcode::SUBREG_TO_REG, dl, MVT::i64,
+ CurDAG->getTargetConstant(0, MVT::i64), ClrNode,
+ CurDAG->getTargetConstant(X86::sub_32bit, MVT::i32)),
+ 0);
+ break;
+ default:
+ llvm_unreachable("Unexpected division source");
+ }
+
InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ClrReg,
ClrNode, InFlag).getValue(1);
}
SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
InFlag };
SDNode *CNode =
- CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Glue, Ops,
- array_lengthof(Ops));
+ CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Glue, Ops);
InFlag = SDValue(CNode, 1);
// Update the chain.
ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
// Prevent use of AH in a REX instruction by referencing AX instead.
// Shift it down 8 bits.
+ //
+ // The current assumption of the register allocator is that isel
+ // won't generate explicit references to the GPR8_NOREX registers. If
+ // the allocator and/or the backend get enhanced to be more robust in
+ // that regard, this can be, and should be, removed.
if (HiReg == X86::AH && Subtarget->is64Bit() &&
!SDValue(Node, 1).use_empty()) {
SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
ReplaceUses(SDValue(Node, 1), Result);
DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
}
- return NULL;
+ return nullptr;
}
case X86ISD::CMP:
// On x86-32, only the ABCD registers have 8-bit subregisters.
if (!Subtarget->is64Bit()) {
const TargetRegisterClass *TRC;
- switch (N0.getValueType().getSimpleVT().SimpleTy) {
+ switch (N0.getSimpleValueType().SimpleTy) {
case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
default: llvm_unreachable("Unsupported TEST operand type!");
// one, do not call ReplaceAllUsesWith.
ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
SDValue(NewNode, 0));
- return NULL;
+ return nullptr;
}
// For example, "testl %eax, $2048" to "testb %ah, $8".
// Put the value in an ABCD register.
const TargetRegisterClass *TRC;
- switch (N0.getValueType().getSimpleVT().SimpleTy) {
+ switch (N0.getSimpleValueType().SimpleTy) {
case MVT::i64: TRC = &X86::GR64_ABCDRegClass; break;
case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
// one, do not call ReplaceAllUsesWith.
ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
SDValue(NewNode, 0));
- return NULL;
+ return nullptr;
}
// For example, "testl %eax, $32776" to "testw %ax, $32776".
// one, do not call ReplaceAllUsesWith.
ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
SDValue(NewNode, 0));
- return NULL;
+ return nullptr;
}
// For example, "testq %rax, $268468232" to "testl %eax, $268468232".
// one, do not call ReplaceAllUsesWith.
ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
SDValue(NewNode, 0));
- return NULL;
+ return nullptr;
}
}
break;
SDValue StoredVal = StoreNode->getOperand(1);
unsigned Opc = StoredVal->getOpcode();
- LoadSDNode *LoadNode = 0;
+ LoadSDNode *LoadNode = nullptr;
SDValue InputChain;
if (!isLoadIncOrDecStore(StoreNode, Opc, StoredVal, CurDAG,
LoadNode, InputChain))
EVT LdVT = LoadNode->getMemoryVT();
unsigned newOpc = getFusedLdStOpcode(LdVT, Opc);
MachineSDNode *Result = CurDAG->getMachineNode(newOpc,
- Node->getDebugLoc(),
- MVT::i32, MVT::Other, Ops,
- array_lengthof(Ops));
+ SDLoc(Node),
+ MVT::i32, MVT::Other, Ops);
Result->setMemRefs(MemOp, MemOp + 2);
ReplaceUses(SDValue(StoreNode, 0), SDValue(Result, 1));
SDNode *ResNode = SelectCode(Node);
DEBUG(dbgs() << "=> ";
- if (ResNode == NULL || ResNode == Node)
+ if (ResNode == nullptr || ResNode == Node)
Node->dump(CurDAG);
else
ResNode->dump(CurDAG);
case 'v': // not offsetable ??
default: return true;
case 'm': // memory
- if (!SelectAddr(0, Op, Op0, Op1, Op2, Op3, Op4))
+ if (!SelectAddr(nullptr, Op, Op0, Op1, Op2, Op3, Op4))
return true;
break;
}