SDValue V2);
static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
- switch (TM.getSubtarget<X86Subtarget>().TargetType) {
- default: llvm_unreachable("unknown subtarget type");
- case X86Subtarget::isDarwin:
- if (TM.getSubtarget<X86Subtarget>().is64Bit())
- return new X8664_MachoTargetObjectFile();
+
+ bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
+
+ if (TM.getSubtarget<X86Subtarget>().isTargetDarwin()) {
+ if (is64Bit) return new X8664_MachoTargetObjectFile();
return new TargetLoweringObjectFileMachO();
- case X86Subtarget::isELF:
- if (TM.getSubtarget<X86Subtarget>().is64Bit())
- return new X8664_ELFTargetObjectFile(TM);
+ } else if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
+ if (is64Bit) return new X8664_ELFTargetObjectFile(TM);
return new X8632_ELFTargetObjectFile(TM);
- case X86Subtarget::isMingw:
- case X86Subtarget::isCygwin:
- case X86Subtarget::isWindows:
+ } else if (TM.getSubtarget<X86Subtarget>().isTargetCOFF()) {
return new TargetLoweringObjectFileCOFF();
- }
+ }
+ llvm_unreachable("unknown subtarget type");
}
X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
if (!Subtarget->hasSSE2())
setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
+ // On X86 and X86-64, atomic operations are lowered to locked instructions.
+ // Locked instructions, in turn, have implicit fence semantics (all memory
+ // operations are flushed before issuing the locked instruction, and they
+ // are not buffered), so we can fold away the common pattern of
+ // fence-atomic-fence.
+ setShouldFoldAtomicFences(true);
// Expand certain atomics
setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
- addRegisterClass(MVT::v2f32, X86::VR64RegisterClass, false);
+
addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
setOperationAction(ISD::ADD, MVT::v8i8, Legal);
AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
- setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
- AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
- setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
- setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Custom);
setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Custom);
setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Custom);
- setOperationAction(ISD::BIT_CONVERT, MVT::v2f32, Custom);
setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Custom);
}
}
EVT VT = SVT;
// Do not attempt to promote non-128-bit vectors
- if (!VT.is128BitVector()) {
+ if (!VT.is128BitVector())
continue;
- }
setOperationAction(ISD::AND, SVT, Promote);
AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
setTargetDAGCombine(ISD::SRL);
setTargetDAGCombine(ISD::OR);
setTargetDAGCombine(ISD::STORE);
- setTargetDAGCombine(ISD::MEMBARRIER);
setTargetDAGCombine(ISD::ZERO_EXTEND);
if (Subtarget->is64Bit())
setTargetDAGCombine(ISD::MUL);
return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
}
+bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
+ unsigned &Offset) const {
+ if (!Subtarget->isTargetLinux())
+ return false;
+
+ if (Subtarget->is64Bit()) {
+ // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
+ Offset = 0x28;
+ if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
+ AddressSpace = 256;
+ else
+ AddressSpace = 257;
+ } else {
+ // %gs:0x14 on i386
+ Offset = 0x14;
+ AddressSpace = 256;
+ }
+ return true;
+}
+
+
//===----------------------------------------------------------------------===//
// Return Value Calling Convention Implementation
//===----------------------------------------------------------------------===//
X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
const SmallVectorImpl<EVT> &OutTys,
const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
- SelectionDAG &DAG) const {
+ LLVMContext &Context) const {
SmallVector<CCValAssign, 16> RVLocs;
CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
- RVLocs, *DAG.getContext());
+ RVLocs, Context);
return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
}
// could be overwritten by lowering of arguments in case of a tail call.
if (Flags.isByVal()) {
int FI = MFI->CreateFixedObject(Flags.getByValSize(),
- VA.getLocMemOffset(), isImmutable, false);
+ VA.getLocMemOffset(), isImmutable);
return DAG.getFrameIndex(FI, getPointerTy());
} else {
int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
- VA.getLocMemOffset(), isImmutable, false);
+ VA.getLocMemOffset(), isImmutable);
SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
return DAG.getLoad(ValVT, dl, Chain, FIN,
PseudoSourceValue::getFixedStack(FI), 0,
if (isVarArg) {
if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
CallConv != CallingConv::X86_ThisCall)) {
- FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,
- true, false));
+ FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
}
if (Is64Bit) {
unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
// Calculate the new stack slot for the return address.
int SlotSize = Is64Bit ? 8 : 4;
int NewReturnAddrFI =
- MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false, false);
+ MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
// Create frame index.
int32_t Offset = VA.getLocMemOffset()+FPDiff;
uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
- FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
+ FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
FIN = DAG.getFrameIndex(FI, getPointerTy());
if (Flags.isByVal()) {
FPDiff, dl);
}
- bool WasGlobalOrExternal = false;
if (getTargetMachine().getCodeModel() == CodeModel::Large) {
assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
// In the 64-bit large code model, we have to make all calls
// pc-relative offset may not be large enough to hold the whole
// address.
} else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
- WasGlobalOrExternal = true;
// If the callee is a GlobalAddress node (quite common, every direct call
// is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
// it.
OpFlags = X86II::MO_DARWIN_STUB;
}
- Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
+ Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
G->getOffset(), OpFlags);
}
} else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
- WasGlobalOrExternal = true;
unsigned char OpFlags = 0;
// On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
((X86TargetMachine&)getTargetMachine()).getInstrInfo();
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
CCValAssign &VA = ArgLocs[i];
- EVT RegVT = VA.getLocVT();
SDValue Arg = Outs[i].Val;
ISD::ArgFlagsTy Flags = Outs[i].Flags;
if (VA.getLocInfo() == CCValAssign::Indirect)
// Set up a frame object for the return address.
uint64_t SlotSize = TD->getPointerSize();
ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
- false, false);
+ false);
FuncInfo->setRAIndex(ReturnAddrIndex);
}
}
/// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
-/// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
+/// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
/// done when every pair / quad of shuffle mask elements point to elements in
/// the right sequence. e.g.
/// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
unsigned NumElems = VT.getVectorNumElements();
unsigned NewWidth = (NumElems == 4) ? 2 : 4;
EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
- EVT MaskEltVT = MaskVT.getVectorElementType();
EVT NewVT = MaskVT;
switch (VT.getSimpleVT().SimpleTy) {
default: assert(false && "Unexpected!");
SDValue
X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
DebugLoc dl = Op.getDebugLoc();
- if (Op.getValueType() == MVT::v2f32)
- return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
- DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
- DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
- Op.getOperand(0))));
-
- if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
+
+ if (Op.getValueType() == MVT::v1i64 &&
+ Op.getOperand(0).getValueType() == MVT::i64)
return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
if (OpFlags == X86II::MO_NO_FLAG &&
X86::isOffsetSuitableForCodeModel(Offset, M)) {
// A direct static reference to a global.
- Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
+ Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
Offset = 0;
} else {
- Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
+ Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
}
if (Subtarget->isPICStyleRIPRel() &&
MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
DebugLoc dl = GA->getDebugLoc();
- SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
+ SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
GA->getValueType(0),
GA->getOffset(),
OperandFlags);
// emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
// exec)
- SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
+ SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
+ GA->getValueType(0),
GA->getOffset(), OperandFlags);
SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
OpFlag = X86II::MO_TLVP_PIC_BASE;
else
OpFlag = X86II::MO_TLVP;
-
- SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(),
+ DebugLoc DL = Op.getDebugLoc();
+ SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
getPointerTy(),
GA->getOffset(), OpFlag);
-
- DebugLoc DL = Op.getDebugLoc();
SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
// With PIC32, the address is actually $g + Offset.
bool NeedCF = false;
bool NeedOF = false;
switch (X86CC) {
+ default: break;
case X86::COND_A: case X86::COND_AE:
case X86::COND_B: case X86::COND_BE:
NeedCF = true;
case X86::COND_O: case X86::COND_NO:
NeedOF = true;
break;
- default: break;
}
// See if we can use the EFLAGS value from the operand instead of
// doing a separate TEST. TEST always sets OF and CF to 0, so unless
// we prove that the arithmetic won't overflow, we can't use OF or CF.
- if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
- unsigned Opcode = 0;
- unsigned NumOperands = 0;
- switch (Op.getNode()->getOpcode()) {
- case ISD::ADD:
- // Due to an isel shortcoming, be conservative if this add is
- // likely to be selected as part of a load-modify-store
- // instruction. When the root node in a match is a store, isel
- // doesn't know how to remap non-chain non-flag uses of other
- // nodes in the match, such as the ADD in this case. This leads
- // to the ADD being left around and reselected, with the result
- // being two adds in the output. Alas, even if none our users
- // are stores, that doesn't prove we're O.K. Ergo, if we have
- // any parents that aren't CopyToReg or SETCC, eschew INC/DEC.
- // A better fix seems to require climbing the DAG back to the
- // root, and it doesn't seem to be worth the effort.
- for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
- UE = Op.getNode()->use_end(); UI != UE; ++UI)
- if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
- goto default_case;
- if (ConstantSDNode *C =
- dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
- // An add of one will be selected as an INC.
- if (C->getAPIntValue() == 1) {
- Opcode = X86ISD::INC;
- NumOperands = 1;
- break;
- }
- // An add of negative one (subtract of one) will be selected as a DEC.
- if (C->getAPIntValue().isAllOnesValue()) {
- Opcode = X86ISD::DEC;
- NumOperands = 1;
- break;
- }
+ if (Op.getResNo() != 0 || NeedOF || NeedCF)
+ // Emit a CMP with 0, which is the TEST pattern.
+ return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
+ DAG.getConstant(0, Op.getValueType()));
+
+ unsigned Opcode = 0;
+ unsigned NumOperands = 0;
+ switch (Op.getNode()->getOpcode()) {
+ case ISD::ADD:
+ // Due to an isel shortcoming, be conservative if this add is likely to be
+ // selected as part of a load-modify-store instruction. When the root node
+ // in a match is a store, isel doesn't know how to remap non-chain non-flag
+ // uses of other nodes in the match, such as the ADD in this case. This
+ // leads to the ADD being left around and reselected, with the result being
+ // two adds in the output. Alas, even if none our users are stores, that
+ // doesn't prove we're O.K. Ergo, if we have any parents that aren't
+ // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
+ // climbing the DAG back to the root, and it doesn't seem to be worth the
+ // effort.
+ for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
+ UE = Op.getNode()->use_end(); UI != UE; ++UI)
+ if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
+ goto default_case;
+
+ if (ConstantSDNode *C =
+ dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
+ // An add of one will be selected as an INC.
+ if (C->getAPIntValue() == 1) {
+ Opcode = X86ISD::INC;
+ NumOperands = 1;
+ break;
}
- // Otherwise use a regular EFLAGS-setting add.
- Opcode = X86ISD::ADD;
- NumOperands = 2;
- break;
- case ISD::AND: {
- // If the primary and result isn't used, don't bother using X86ISD::AND,
- // because a TEST instruction will be better.
- bool NonFlagUse = false;
- for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
- UE = Op.getNode()->use_end(); UI != UE; ++UI) {
- SDNode *User = *UI;
- unsigned UOpNo = UI.getOperandNo();
- if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
- // Look pass truncate.
- UOpNo = User->use_begin().getOperandNo();
- User = *User->use_begin();
- }
- if (User->getOpcode() != ISD::BRCOND &&
- User->getOpcode() != ISD::SETCC &&
- (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
- NonFlagUse = true;
- break;
- }
+
+ // An add of negative one (subtract of one) will be selected as a DEC.
+ if (C->getAPIntValue().isAllOnesValue()) {
+ Opcode = X86ISD::DEC;
+ NumOperands = 1;
+ break;
}
- if (!NonFlagUse)
+ }
+
+ // Otherwise use a regular EFLAGS-setting add.
+ Opcode = X86ISD::ADD;
+ NumOperands = 2;
+ break;
+ case ISD::AND: {
+ // If the primary and result isn't used, don't bother using X86ISD::AND,
+ // because a TEST instruction will be better.
+ bool NonFlagUse = false;
+ for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
+ UE = Op.getNode()->use_end(); UI != UE; ++UI) {
+ SDNode *User = *UI;
+ unsigned UOpNo = UI.getOperandNo();
+ if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
+ // Look pass truncate.
+ UOpNo = User->use_begin().getOperandNo();
+ User = *User->use_begin();
+ }
+
+ if (User->getOpcode() != ISD::BRCOND &&
+ User->getOpcode() != ISD::SETCC &&
+ (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
+ NonFlagUse = true;
break;
+ }
}
+
+ if (!NonFlagUse)
+ break;
+ }
// FALL THROUGH
- case ISD::SUB:
- case ISD::OR:
- case ISD::XOR:
- // Due to the ISEL shortcoming noted above, be conservative if this op is
- // likely to be selected as part of a load-modify-store instruction.
- for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
+ case ISD::SUB:
+ case ISD::OR:
+ case ISD::XOR:
+ // Due to the ISEL shortcoming noted above, be conservative if this op is
+ // likely to be selected as part of a load-modify-store instruction.
+ for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
UE = Op.getNode()->use_end(); UI != UE; ++UI)
- if (UI->getOpcode() == ISD::STORE)
- goto default_case;
- // Otherwise use a regular EFLAGS-setting instruction.
- switch (Op.getNode()->getOpcode()) {
- case ISD::SUB: Opcode = X86ISD::SUB; break;
- case ISD::OR: Opcode = X86ISD::OR; break;
- case ISD::XOR: Opcode = X86ISD::XOR; break;
- case ISD::AND: Opcode = X86ISD::AND; break;
- default: llvm_unreachable("unexpected operator!");
- }
- NumOperands = 2;
- break;
- case X86ISD::ADD:
- case X86ISD::SUB:
- case X86ISD::INC:
- case X86ISD::DEC:
- case X86ISD::OR:
- case X86ISD::XOR:
- case X86ISD::AND:
- return SDValue(Op.getNode(), 1);
- default:
- default_case:
- break;
- }
- if (Opcode != 0) {
- SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
- SmallVector<SDValue, 4> Ops;
- for (unsigned i = 0; i != NumOperands; ++i)
- Ops.push_back(Op.getOperand(i));
- SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
- DAG.ReplaceAllUsesWith(Op, New);
- return SDValue(New.getNode(), 1);
+ if (UI->getOpcode() == ISD::STORE)
+ goto default_case;
+
+ // Otherwise use a regular EFLAGS-setting instruction.
+ switch (Op.getNode()->getOpcode()) {
+ default: llvm_unreachable("unexpected operator!");
+ case ISD::SUB: Opcode = X86ISD::SUB; break;
+ case ISD::OR: Opcode = X86ISD::OR; break;
+ case ISD::XOR: Opcode = X86ISD::XOR; break;
+ case ISD::AND: Opcode = X86ISD::AND; break;
}
+
+ NumOperands = 2;
+ break;
+ case X86ISD::ADD:
+ case X86ISD::SUB:
+ case X86ISD::INC:
+ case X86ISD::DEC:
+ case X86ISD::OR:
+ case X86ISD::XOR:
+ case X86ISD::AND:
+ return SDValue(Op.getNode(), 1);
+ default:
+ default_case:
+ break;
}
- // Otherwise just emit a CMP with 0, which is the TEST pattern.
- return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
- DAG.getConstant(0, Op.getValueType()));
+ if (Opcode == 0)
+ // Emit a CMP with 0, which is the TEST pattern.
+ return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
+ DAG.getConstant(0, Op.getValueType()));
+
+ SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
+ SmallVector<SDValue, 4> Ops;
+ for (unsigned i = 0; i != NumOperands; ++i)
+ Ops.push_back(Op.getOperand(i));
+
+ SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
+ DAG.ReplaceAllUsesWith(Op, New);
+ return SDValue(New.getNode(), 1);
}
/// Emit nodes that will be selected as "cmp Op0,Op1", or something
Op1 = Op1.getOperand(0);
SDValue LHS, RHS;
- if (Op1.getOpcode() == ISD::SHL) {
- if (ConstantSDNode *And10C = dyn_cast<ConstantSDNode>(Op1.getOperand(0)))
- if (And10C->getZExtValue() == 1) {
- LHS = Op0;
- RHS = Op1.getOperand(1);
- }
- } else if (Op0.getOpcode() == ISD::SHL) {
+ if (Op1.getOpcode() == ISD::SHL)
+ std::swap(Op0, Op1);
+ if (Op0.getOpcode() == ISD::SHL) {
if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
if (And00C->getZExtValue() == 1) {
+ // If we looked past a truncate, check that it's only truncating away
+ // known zeros.
+ unsigned BitWidth = Op0.getValueSizeInBits();
+ unsigned AndBitWidth = And.getValueSizeInBits();
+ if (BitWidth > AndBitWidth) {
+ APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
+ DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
+ if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
+ return SDValue();
+ }
LHS = Op1;
RHS = Op0.getOperand(1);
}
SDValue Flag;
- EVT IntPtr = getPointerTy();
EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
// X86-64 va_list is a struct { i32, i32, i8*, i8* }.
assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
- SDValue Chain = Op.getOperand(0);
- SDValue SrcPtr = Op.getOperand(1);
- SDValue SrcSV = Op.getOperand(2);
report_fatal_error("VAArgInst is not yet implemented for x86-64!");
return SDValue();
F->insert(MBBIter, newMBB);
F->insert(MBBIter, nextMBB);
- // Move all successors to thisMBB to nextMBB
- nextMBB->transferSuccessors(thisMBB);
+ // Transfer the remainder of thisMBB and its successor edges to nextMBB.
+ nextMBB->splice(nextMBB->begin(), thisMBB,
+ llvm::next(MachineBasicBlock::iterator(bInstr)),
+ thisMBB->end());
+ nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
// Update thisMBB to fall through to newMBB
thisMBB->addSuccessor(newMBB);
// insert branch
BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
- F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
+ bInstr->eraseFromParent(); // The pseudo instruction is gone now.
return nextMBB;
}
F->insert(MBBIter, newMBB);
F->insert(MBBIter, nextMBB);
- // Move all successors to thisMBB to nextMBB
- nextMBB->transferSuccessors(thisMBB);
+ // Transfer the remainder of thisMBB and its successor edges to nextMBB.
+ nextMBB->splice(nextMBB->begin(), thisMBB,
+ llvm::next(MachineBasicBlock::iterator(bInstr)),
+ thisMBB->end());
+ nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
// Update thisMBB to fall through to newMBB
thisMBB->addSuccessor(newMBB);
// insert branch
BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
- F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
+ bInstr->eraseFromParent(); // The pseudo instruction is gone now.
return nextMBB;
}
F->insert(MBBIter, newMBB);
F->insert(MBBIter, nextMBB);
- // Move all successors of thisMBB to nextMBB
- nextMBB->transferSuccessors(thisMBB);
+ // Transfer the remainder of thisMBB and its successor edges to nextMBB.
+ nextMBB->splice(nextMBB->begin(), thisMBB,
+ llvm::next(MachineBasicBlock::iterator(mInstr)),
+ thisMBB->end());
+ nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
// Update thisMBB to fall through to newMBB
thisMBB->addSuccessor(newMBB);
// insert branch
BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
- F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
+ mInstr->eraseFromParent(); // The pseudo instruction is gone now.
return nextMBB;
}
X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
unsigned numArgs, bool memArg) const {
- MachineFunction *F = BB->getParent();
DebugLoc dl = MI->getDebugLoc();
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
.addReg(X86::XMM0);
- F->DeleteMachineInstr(MI);
+ MI->eraseFromParent();
return BB;
}
F->insert(MBBIter, XMMSaveMBB);
F->insert(MBBIter, EndMBB);
- // Set up the CFG.
- // Move any original successors of MBB to the end block.
- EndMBB->transferSuccessors(MBB);
+ // Transfer the remainder of MBB and its successor edges to EndMBB.
+ EndMBB->splice(EndMBB->begin(), MBB,
+ llvm::next(MachineBasicBlock::iterator(MI)),
+ MBB->end());
+ EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
+
// The original block will now fall through to the XMM save block.
MBB->addSuccessor(XMMSaveMBB);
// The XMMSaveMBB will fall through to the end block.
.addMemOperand(MMO);
}
- F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
+ MI->eraseFromParent(); // The pseudo instruction is gone now.
return EndMBB;
}
MachineFunction *F = BB->getParent();
MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
- unsigned Opc =
- X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
-
- BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
F->insert(It, copy0MBB);
F->insert(It, sinkMBB);
- // Update machine-CFG edges by first adding all successors of the current
- // block to the new block which will contain the Phi node for the select.
- for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
- E = BB->succ_end(); I != E; ++I)
- sinkMBB->addSuccessor(*I);
-
- // Next, remove all successors of the current block, and add the true
- // and fallthrough blocks as its successors.
- while (!BB->succ_empty())
- BB->removeSuccessor(BB->succ_begin());
-
- // Add the true and fallthrough blocks as its successors.
- BB->addSuccessor(copy0MBB);
- BB->addSuccessor(sinkMBB);
-
// If the EFLAGS register isn't dead in the terminator, then claim that it's
// live into the sink and copy blocks.
const MachineFunction *MF = BB->getParent();
const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
BitVector ReservedRegs = TRI->getReservedRegs(*MF);
- const MachineInstr *Term = BB->getFirstTerminator();
- for (unsigned I = 0, E = Term->getNumOperands(); I != E; ++I) {
- const MachineOperand &MO = Term->getOperand(I);
- if (!MO.isReg() || MO.isKill() || MO.isDead()) continue;
+ for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
+ const MachineOperand &MO = MI->getOperand(I);
+ if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
unsigned Reg = MO.getReg();
if (Reg != X86::EFLAGS) continue;
copy0MBB->addLiveIn(Reg);
sinkMBB->addLiveIn(Reg);
}
+ // Transfer the remainder of BB and its successor edges to sinkMBB.
+ sinkMBB->splice(sinkMBB->begin(), BB,
+ llvm::next(MachineBasicBlock::iterator(MI)),
+ BB->end());
+ sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
+
+ // Add the true and fallthrough blocks as its successors.
+ BB->addSuccessor(copy0MBB);
+ BB->addSuccessor(sinkMBB);
+
+ // Create the conditional branch instruction.
+ unsigned Opc =
+ X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
+ BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
+
// copy0MBB:
// %FalseValue = ...
// # fallthrough to sinkMBB
// sinkMBB:
// %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
// ...
- BuildMI(sinkMBB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
+ BuildMI(*sinkMBB, sinkMBB->begin(), DL,
+ TII->get(X86::PHI), MI->getOperand(0).getReg())
.addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
.addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
- F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
+ MI->eraseFromParent(); // The pseudo instruction is gone now.
return sinkMBB;
}
MachineBasicBlock *BB) const {
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
- MachineFunction *F = BB->getParent();
// The lowering is pretty easy: we're just emitting the call to _alloca. The
// non-trivial part is impdef of ESP.
// FIXME: The code should be tweaked as soon as we'll try to do codegen for
// mingw-w64.
- BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
+ BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
.addExternalSymbol("_alloca")
.addReg(X86::EAX, RegState::Implicit)
.addReg(X86::ESP, RegState::Implicit)
.addReg(X86::EAX, RegState::Define | RegState::Implicit)
.addReg(X86::ESP, RegState::Define | RegState::Implicit);
- F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
+ MI->eraseFromParent(); // The pseudo instruction is gone now.
return BB;
}
assert(MI->getOperand(3).isGlobal() && "This should be a global");
if (Subtarget->is64Bit()) {
- MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV64rm), X86::RDI)
+ MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
+ TII->get(X86::MOV64rm), X86::RDI)
.addReg(X86::RIP)
.addImm(0).addReg(0)
.addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
MI->getOperand(3).getTargetFlags())
.addReg(0);
- MIB = BuildMI(BB, DL, TII->get(X86::CALL64m));
+ MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
addDirectMem(MIB, X86::RDI).addReg(0);
} else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
- MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV32rm), X86::EAX)
+ MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
+ TII->get(X86::MOV32rm), X86::EAX)
.addReg(0)
.addImm(0).addReg(0)
.addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
MI->getOperand(3).getTargetFlags())
.addReg(0);
- MIB = BuildMI(BB, DL, TII->get(X86::CALL32m));
+ MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
addDirectMem(MIB, X86::EAX).addReg(0);
} else {
- MachineInstrBuilder MIB = BuildMI(BB, DL, TII->get(X86::MOV32rm), X86::EAX)
+ MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
+ TII->get(X86::MOV32rm), X86::EAX)
.addReg(TII->getGlobalBaseReg(F))
.addImm(0).addReg(0)
.addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
MI->getOperand(3).getTargetFlags())
.addReg(0);
- MIB = BuildMI(BB, DL, TII->get(X86::CALL32m));
+ MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
addDirectMem(MIB, X86::EAX).addReg(0);
}
- F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
+ MI->eraseFromParent(); // The pseudo instruction is gone now.
return BB;
}
// mode when truncating to an integer value.
MachineFunction *F = BB->getParent();
int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
- addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
+ addFrameReference(BuildMI(*BB, MI, DL,
+ TII->get(X86::FNSTCW16m)), CWFrameIdx);
// Load the old value of the high byte of the control word...
unsigned OldCW =
F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
- addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
+ addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
CWFrameIdx);
// Set the high part to be round to zero...
- addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
+ addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
.addImm(0xC7F);
// Reload the modified control word now...
- addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
+ addFrameReference(BuildMI(*BB, MI, DL,
+ TII->get(X86::FLDCW16m)), CWFrameIdx);
// Restore the memory image of control word to original value
- addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
+ addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
.addReg(OldCW);
// Get the X86 opcode to use.
} else {
AM.Disp = Op.getImm();
}
- addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
+ addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
.addReg(MI->getOperand(X86AddrNumOperands).getReg());
// Reload the original control word now.
- addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
+ addFrameReference(BuildMI(*BB, MI, DL,
+ TII->get(X86::FLDCW16m)), CWFrameIdx);
- F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
+ MI->eraseFromParent(); // The pseudo instruction is gone now.
return BB;
}
// String/text processing lowering.
return SDValue();
}
-// On X86 and X86-64, atomic operations are lowered to locked instructions.
-// Locked instructions, in turn, have implicit fence semantics (all memory
-// operations are flushed before issuing the locked instruction, and the
-// are not buffered), so we can fold away the common pattern of
-// fence-atomic-fence.
-static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
- SDValue atomic = N->getOperand(0);
- switch (atomic.getOpcode()) {
- case ISD::ATOMIC_CMP_SWAP:
- case ISD::ATOMIC_SWAP:
- case ISD::ATOMIC_LOAD_ADD:
- case ISD::ATOMIC_LOAD_SUB:
- case ISD::ATOMIC_LOAD_AND:
- case ISD::ATOMIC_LOAD_OR:
- case ISD::ATOMIC_LOAD_XOR:
- case ISD::ATOMIC_LOAD_NAND:
- case ISD::ATOMIC_LOAD_MIN:
- case ISD::ATOMIC_LOAD_MAX:
- case ISD::ATOMIC_LOAD_UMIN:
- case ISD::ATOMIC_LOAD_UMAX:
- break;
- default:
- return SDValue();
- }
-
- SDValue fence = atomic.getOperand(0);
- if (fence.getOpcode() != ISD::MEMBARRIER)
- return SDValue();
-
- switch (atomic.getOpcode()) {
- case ISD::ATOMIC_CMP_SWAP:
- return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
- fence.getOperand(0),
- atomic.getOperand(1), atomic.getOperand(2),
- atomic.getOperand(3)), atomic.getResNo());
- case ISD::ATOMIC_SWAP:
- case ISD::ATOMIC_LOAD_ADD:
- case ISD::ATOMIC_LOAD_SUB:
- case ISD::ATOMIC_LOAD_AND:
- case ISD::ATOMIC_LOAD_OR:
- case ISD::ATOMIC_LOAD_XOR:
- case ISD::ATOMIC_LOAD_NAND:
- case ISD::ATOMIC_LOAD_MIN:
- case ISD::ATOMIC_LOAD_MAX:
- case ISD::ATOMIC_LOAD_UMIN:
- case ISD::ATOMIC_LOAD_UMAX:
- return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
- fence.getOperand(0),
- atomic.getOperand(1), atomic.getOperand(2)),
- atomic.getResNo());
- default:
- return SDValue();
- }
-}
-
static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
// (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
// (and (i32 x86isd::setcc_carry), 1)
case X86ISD::FAND: return PerformFANDCombine(N, DAG);
case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
- case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
}
// so don't worry about this.
// Verify this is a simple bswap.
- if (CI->getNumOperands() != 2 ||
- CI->getType() != CI->getOperand(1)->getType() ||
+ if (CI->getNumArgOperands() != 1 ||
+ CI->getType() != CI->getArgOperand(0)->getType() ||
!CI->getType()->isIntegerTy())
return false;
Module *M = CI->getParent()->getParent()->getParent();
Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
- Value *Op = CI->getOperand(1);
+ Value *Op = CI->getArgOperand(0);
Op = CallInst::Create(Int, Op, CI->getName(), CI);
CI->replaceAllUsesWith(Op);
/// vector. If it is invalid, don't add anything to Ops.
void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
char Constraint,
- bool hasMemory,
std::vector<SDValue>&Ops,
SelectionDAG &DAG) const {
SDValue Result(0, 0);
break;
}
+ // In any sort of PIC mode addresses need to be computed at runtime by
+ // adding in a register or some sort of table lookup. These can't
+ // be used as immediates.
+ if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC() ||
+ Subtarget->isPICStyleRIPRel())
+ return;
+
// If we are in non-pic codegen mode, we allow the address of a global (with
// an optional displacement) to be used with 'i'.
GlobalAddressSDNode *GA = 0;
getTargetMachine())))
return;
- if (hasMemory)
- Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
- else
- Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
- Result = Op;
+ Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
+ GA->getValueType(0), Offset);
break;
}
}
Ops.push_back(Result);
return;
}
- return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
- Ops, DAG);
+ return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
}
std::vector<unsigned> X86TargetLowering::